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/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/DerivedTypes.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 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
58 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
60 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
62 // PowerPC has pre-inc load and store's.
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
67 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
72 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
74 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
75 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
76 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
77 // This is used in the ppcf128->int sequence. Note it has different semantics
78 // from FP_ROUND: that rounds to nearest, this rounds to zero.
79 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
81 // PowerPC has no SREM/UREM instructions
82 setOperationAction(ISD::SREM, MVT::i32, Expand);
83 setOperationAction(ISD::UREM, MVT::i32, Expand);
84 setOperationAction(ISD::SREM, MVT::i64, Expand);
85 setOperationAction(ISD::UREM, MVT::i64, Expand);
87 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
88 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
89 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
90 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
91 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
92 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
93 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
94 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
95 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
97 // We don't support sin/cos/sqrt/fmod/pow
98 setOperationAction(ISD::FSIN , MVT::f64, Expand);
99 setOperationAction(ISD::FCOS , MVT::f64, Expand);
100 setOperationAction(ISD::FREM , MVT::f64, Expand);
101 setOperationAction(ISD::FPOW , MVT::f64, Expand);
102 setOperationAction(ISD::FSIN , MVT::f32, Expand);
103 setOperationAction(ISD::FCOS , MVT::f32, Expand);
104 setOperationAction(ISD::FREM , MVT::f32, Expand);
105 setOperationAction(ISD::FPOW , MVT::f32, Expand);
107 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
109 // If we're enabling GP optimizations, use hardware square root
110 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
111 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
112 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
115 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
116 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
118 // PowerPC does not have BSWAP, CTPOP or CTTZ
119 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
120 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
121 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
122 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
123 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
124 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
126 // PowerPC does not have ROTR
127 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
128 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
130 // PowerPC does not have Select
131 setOperationAction(ISD::SELECT, MVT::i32, Expand);
132 setOperationAction(ISD::SELECT, MVT::i64, Expand);
133 setOperationAction(ISD::SELECT, MVT::f32, Expand);
134 setOperationAction(ISD::SELECT, MVT::f64, Expand);
136 // PowerPC wants to turn select_cc of FP into fsel when possible.
137 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
138 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
140 // PowerPC wants to optimize integer setcc a bit
141 setOperationAction(ISD::SETCC, MVT::i32, Custom);
143 // PowerPC does not have BRCOND which requires SetCC
144 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
146 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
148 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
149 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
151 // PowerPC does not have [U|S]INT_TO_FP
152 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
153 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
155 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
156 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
157 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
158 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
160 // We cannot sextinreg(i1). Expand to shifts.
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
163 // Support label based line numbers.
164 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
165 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
167 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
168 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
169 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
170 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
173 // We want to legalize GlobalAddress and ConstantPool nodes into the
174 // appropriate instructions to materialize the address.
175 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
176 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
177 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
178 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
179 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
180 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
181 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
182 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
184 // RET must be custom lowered, to meet ABI requirements.
185 setOperationAction(ISD::RET , MVT::Other, Custom);
188 setOperationAction(ISD::TRAP, MVT::Other, Legal);
190 // TRAMPOLINE is custom lowered.
191 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
193 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
194 setOperationAction(ISD::VASTART , MVT::Other, Custom);
196 // VAARG is custom lowered with ELF 32 ABI
197 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
198 setOperationAction(ISD::VAARG, MVT::Other, Custom);
200 setOperationAction(ISD::VAARG, MVT::Other, Expand);
202 // Use the default implementation.
203 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
204 setOperationAction(ISD::VAEND , MVT::Other, Expand);
205 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
206 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
208 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
210 // We want to custom lower some of our intrinsics.
211 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
213 // Comparisons that require checking two conditions.
214 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
215 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
216 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
217 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
218 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
219 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
220 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
221 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
222 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
223 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
224 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
225 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
227 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
228 // They also have instructions for converting between i64 and fp.
229 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
230 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
231 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
232 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
233 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
235 // FIXME: disable this lowered code. This generates 64-bit register values,
236 // and we don't model the fact that the top part is clobbered by calls. We
237 // need to flag these together so that the value isn't live across a call.
238 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
240 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
241 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
243 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
244 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
247 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
248 // 64-bit PowerPC implementations can support i64 types directly
249 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
250 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
251 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
252 // 64-bit PowerPC wants to expand i128 shifts itself.
253 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
254 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
255 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
257 // 32-bit PowerPC wants to expand i64 shifts itself.
258 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
259 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
260 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
263 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
264 // First set operation action for all vector types to expand. Then we
265 // will selectively turn on ones that can be effectively codegen'd.
266 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
267 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
268 MVT VT = (MVT::SimpleValueType)i;
270 // add/sub are legal for all supported vector VT's.
271 setOperationAction(ISD::ADD , VT, Legal);
272 setOperationAction(ISD::SUB , VT, Legal);
274 // We promote all shuffles to v16i8.
275 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
276 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
278 // We promote all non-typed operations to v4i32.
279 setOperationAction(ISD::AND , VT, Promote);
280 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
281 setOperationAction(ISD::OR , VT, Promote);
282 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
283 setOperationAction(ISD::XOR , VT, Promote);
284 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
285 setOperationAction(ISD::LOAD , VT, Promote);
286 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
287 setOperationAction(ISD::SELECT, VT, Promote);
288 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
289 setOperationAction(ISD::STORE, VT, Promote);
290 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
292 // No other operations are legal.
293 setOperationAction(ISD::MUL , VT, Expand);
294 setOperationAction(ISD::SDIV, VT, Expand);
295 setOperationAction(ISD::SREM, VT, Expand);
296 setOperationAction(ISD::UDIV, VT, Expand);
297 setOperationAction(ISD::UREM, VT, Expand);
298 setOperationAction(ISD::FDIV, VT, Expand);
299 setOperationAction(ISD::FNEG, VT, Expand);
300 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
301 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
302 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
303 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
304 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
305 setOperationAction(ISD::UDIVREM, VT, Expand);
306 setOperationAction(ISD::SDIVREM, VT, Expand);
307 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
308 setOperationAction(ISD::FPOW, VT, Expand);
309 setOperationAction(ISD::CTPOP, VT, Expand);
310 setOperationAction(ISD::CTLZ, VT, Expand);
311 setOperationAction(ISD::CTTZ, VT, Expand);
314 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
315 // with merges, splats, etc.
316 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
318 setOperationAction(ISD::AND , MVT::v4i32, Legal);
319 setOperationAction(ISD::OR , MVT::v4i32, Legal);
320 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
321 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
322 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
323 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
325 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
326 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
327 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
328 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
330 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
331 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
332 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
333 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
335 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
336 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
338 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
339 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
340 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
341 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
344 setShiftAmountType(MVT::i32);
345 setBooleanContents(ZeroOrOneBooleanContent);
347 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
348 setStackPointerRegisterToSaveRestore(PPC::X1);
349 setExceptionPointerRegister(PPC::X3);
350 setExceptionSelectorRegister(PPC::X4);
352 setStackPointerRegisterToSaveRestore(PPC::R1);
353 setExceptionPointerRegister(PPC::R3);
354 setExceptionSelectorRegister(PPC::R4);
357 // We have target-specific dag combine patterns for the following nodes:
358 setTargetDAGCombine(ISD::SINT_TO_FP);
359 setTargetDAGCombine(ISD::STORE);
360 setTargetDAGCombine(ISD::BR_CC);
361 setTargetDAGCombine(ISD::BSWAP);
363 // Darwin long double math library functions have $LDBL128 appended.
364 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
365 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
366 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
367 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
368 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
369 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
370 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
371 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
372 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
373 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
374 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
377 computeRegisterProperties();
380 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
381 /// function arguments in the caller parameter area.
382 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
383 TargetMachine &TM = getTargetMachine();
384 // Darwin passes everything on 4 byte boundary.
385 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
391 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
394 case PPCISD::FSEL: return "PPCISD::FSEL";
395 case PPCISD::FCFID: return "PPCISD::FCFID";
396 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
397 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
398 case PPCISD::STFIWX: return "PPCISD::STFIWX";
399 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
400 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
401 case PPCISD::VPERM: return "PPCISD::VPERM";
402 case PPCISD::Hi: return "PPCISD::Hi";
403 case PPCISD::Lo: return "PPCISD::Lo";
404 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
405 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
406 case PPCISD::SRL: return "PPCISD::SRL";
407 case PPCISD::SRA: return "PPCISD::SRA";
408 case PPCISD::SHL: return "PPCISD::SHL";
409 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
410 case PPCISD::STD_32: return "PPCISD::STD_32";
411 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
412 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
413 case PPCISD::MTCTR: return "PPCISD::MTCTR";
414 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
415 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
416 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
417 case PPCISD::MFCR: return "PPCISD::MFCR";
418 case PPCISD::VCMP: return "PPCISD::VCMP";
419 case PPCISD::VCMPo: return "PPCISD::VCMPo";
420 case PPCISD::LBRX: return "PPCISD::LBRX";
421 case PPCISD::STBRX: return "PPCISD::STBRX";
422 case PPCISD::LARX: return "PPCISD::LARX";
423 case PPCISD::STCX: return "PPCISD::STCX";
424 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
425 case PPCISD::MFFS: return "PPCISD::MFFS";
426 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
427 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
428 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
429 case PPCISD::MTFSF: return "PPCISD::MTFSF";
430 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
431 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
436 MVT PPCTargetLowering::getSetCCResultType(MVT VT) const {
441 //===----------------------------------------------------------------------===//
442 // Node matching predicates, for use by the tblgen matching code.
443 //===----------------------------------------------------------------------===//
445 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
446 static bool isFloatingPointZero(SDValue Op) {
447 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
448 return CFP->getValueAPF().isZero();
449 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
450 // Maybe this has already been legalized into the constant pool?
451 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
452 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
453 return CFP->getValueAPF().isZero();
458 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
459 /// true if Op is undef or if it matches the specified value.
460 static bool isConstantOrUndef(int Op, int Val) {
461 return Op < 0 || Op == Val;
464 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
465 /// VPKUHUM instruction.
466 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
468 for (unsigned i = 0; i != 16; ++i)
469 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
472 for (unsigned i = 0; i != 8; ++i)
473 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
474 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
480 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
481 /// VPKUWUM instruction.
482 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
484 for (unsigned i = 0; i != 16; i += 2)
485 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
486 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
489 for (unsigned i = 0; i != 8; i += 2)
490 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
491 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
492 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
493 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
499 /// isVMerge - Common function, used to match vmrg* shuffles.
501 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
502 unsigned LHSStart, unsigned RHSStart) {
503 assert(N->getValueType(0) == MVT::v16i8 &&
504 "PPC only supports shuffles by bytes!");
505 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
506 "Unsupported merge size!");
508 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
509 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
510 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
511 LHSStart+j+i*UnitSize) ||
512 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
513 RHSStart+j+i*UnitSize))
519 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
520 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
521 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
524 return isVMerge(N, UnitSize, 8, 24);
525 return isVMerge(N, UnitSize, 8, 8);
528 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
529 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
530 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
533 return isVMerge(N, UnitSize, 0, 16);
534 return isVMerge(N, UnitSize, 0, 0);
538 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
539 /// amount, otherwise return -1.
540 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
541 assert(N->getValueType(0) == MVT::v16i8 &&
542 "PPC only supports shuffles by bytes!");
544 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
546 // Find the first non-undef value in the shuffle mask.
548 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
551 if (i == 16) return -1; // all undef.
553 // Otherwise, check to see if the rest of the elements are consecutively
554 // numbered from this value.
555 unsigned ShiftAmt = SVOp->getMaskElt(i);
556 if (ShiftAmt < i) return -1;
560 // Check the rest of the elements to see if they are consecutive.
561 for (++i; i != 16; ++i)
562 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
565 // Check the rest of the elements to see if they are consecutive.
566 for (++i; i != 16; ++i)
567 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
573 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
574 /// specifies a splat of a single element that is suitable for input to
575 /// VSPLTB/VSPLTH/VSPLTW.
576 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
577 assert(N->getValueType(0) == MVT::v16i8 &&
578 (EltSize == 1 || EltSize == 2 || EltSize == 4));
580 // This is a splat operation if each element of the permute is the same, and
581 // if the value doesn't reference the second vector.
582 unsigned ElementBase = N->getMaskElt(0);
584 // FIXME: Handle UNDEF elements too!
585 if (ElementBase >= 16)
588 // Check that the indices are consecutive, in the case of a multi-byte element
589 // splatted with a v16i8 mask.
590 for (unsigned i = 1; i != EltSize; ++i)
591 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
594 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
595 if (N->getMaskElt(i) < 0) continue;
596 for (unsigned j = 0; j != EltSize; ++j)
597 if (N->getMaskElt(i+j) != N->getMaskElt(j))
603 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
605 bool PPC::isAllNegativeZeroVector(SDNode *N) {
606 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
608 APInt APVal, APUndef;
612 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32))
613 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
614 return CFP->getValueAPF().isNegZero();
619 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
620 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
621 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
622 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
623 assert(isSplatShuffleMask(SVOp, EltSize));
624 return SVOp->getMaskElt(0) / EltSize;
627 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
628 /// by using a vspltis[bhw] instruction of the specified element size, return
629 /// the constant being splatted. The ByteSize field indicates the number of
630 /// bytes of each element [124] -> [bhw].
631 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
634 // If ByteSize of the splat is bigger than the element size of the
635 // build_vector, then we have a case where we are checking for a splat where
636 // multiple elements of the buildvector are folded together into a single
637 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
638 unsigned EltSize = 16/N->getNumOperands();
639 if (EltSize < ByteSize) {
640 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
641 SDValue UniquedVals[4];
642 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
644 // See if all of the elements in the buildvector agree across.
645 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
646 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
647 // If the element isn't a constant, bail fully out.
648 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
651 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
652 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
653 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
654 return SDValue(); // no match.
657 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
658 // either constant or undef values that are identical for each chunk. See
659 // if these chunks can form into a larger vspltis*.
661 // Check to see if all of the leading entries are either 0 or -1. If
662 // neither, then this won't fit into the immediate field.
663 bool LeadingZero = true;
664 bool LeadingOnes = true;
665 for (unsigned i = 0; i != Multiple-1; ++i) {
666 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
668 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
669 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
671 // Finally, check the least significant entry.
673 if (UniquedVals[Multiple-1].getNode() == 0)
674 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
675 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
677 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
680 if (UniquedVals[Multiple-1].getNode() == 0)
681 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
682 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
683 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
684 return DAG.getTargetConstant(Val, MVT::i32);
690 // Check to see if this buildvec has a single non-undef value in its elements.
691 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
692 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
693 if (OpVal.getNode() == 0)
694 OpVal = N->getOperand(i);
695 else if (OpVal != N->getOperand(i))
699 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
701 unsigned ValSizeInBytes = EltSize;
703 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
704 Value = CN->getZExtValue();
705 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
706 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
707 Value = FloatToBits(CN->getValueAPF().convertToFloat());
710 // If the splat value is larger than the element value, then we can never do
711 // this splat. The only case that we could fit the replicated bits into our
712 // immediate field for would be zero, and we prefer to use vxor for it.
713 if (ValSizeInBytes < ByteSize) return SDValue();
715 // If the element value is larger than the splat value, cut it in half and
716 // check to see if the two halves are equal. Continue doing this until we
717 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
718 while (ValSizeInBytes > ByteSize) {
719 ValSizeInBytes >>= 1;
721 // If the top half equals the bottom half, we're still ok.
722 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
723 (Value & ((1 << (8*ValSizeInBytes))-1)))
727 // Properly sign extend the value.
728 int ShAmt = (4-ByteSize)*8;
729 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
731 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
732 if (MaskVal == 0) return SDValue();
734 // Finally, if this value fits in a 5 bit sext field, return it
735 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
736 return DAG.getTargetConstant(MaskVal, MVT::i32);
740 //===----------------------------------------------------------------------===//
741 // Addressing Mode Selection
742 //===----------------------------------------------------------------------===//
744 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
745 /// or 64-bit immediate, and if the value can be accurately represented as a
746 /// sign extension from a 16-bit value. If so, this returns true and the
748 static bool isIntS16Immediate(SDNode *N, short &Imm) {
749 if (N->getOpcode() != ISD::Constant)
752 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
753 if (N->getValueType(0) == MVT::i32)
754 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
756 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
758 static bool isIntS16Immediate(SDValue Op, short &Imm) {
759 return isIntS16Immediate(Op.getNode(), Imm);
763 /// SelectAddressRegReg - Given the specified addressed, check to see if it
764 /// can be represented as an indexed [r+r] operation. Returns false if it
765 /// can be more efficiently represented with [r+imm].
766 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
768 SelectionDAG &DAG) const {
770 if (N.getOpcode() == ISD::ADD) {
771 if (isIntS16Immediate(N.getOperand(1), imm))
773 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
776 Base = N.getOperand(0);
777 Index = N.getOperand(1);
779 } else if (N.getOpcode() == ISD::OR) {
780 if (isIntS16Immediate(N.getOperand(1), imm))
781 return false; // r+i can fold it if we can.
783 // If this is an or of disjoint bitfields, we can codegen this as an add
784 // (for better address arithmetic) if the LHS and RHS of the OR are provably
786 APInt LHSKnownZero, LHSKnownOne;
787 APInt RHSKnownZero, RHSKnownOne;
788 DAG.ComputeMaskedBits(N.getOperand(0),
789 APInt::getAllOnesValue(N.getOperand(0)
790 .getValueSizeInBits()),
791 LHSKnownZero, LHSKnownOne);
793 if (LHSKnownZero.getBoolValue()) {
794 DAG.ComputeMaskedBits(N.getOperand(1),
795 APInt::getAllOnesValue(N.getOperand(1)
796 .getValueSizeInBits()),
797 RHSKnownZero, RHSKnownOne);
798 // If all of the bits are known zero on the LHS or RHS, the add won't
800 if (~(LHSKnownZero | RHSKnownZero) == 0) {
801 Base = N.getOperand(0);
802 Index = N.getOperand(1);
811 /// Returns true if the address N can be represented by a base register plus
812 /// a signed 16-bit displacement [r+imm], and if it is not better
813 /// represented as reg+reg.
814 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
816 SelectionDAG &DAG) const {
817 // FIXME dl should come from parent load or store, not from address
818 DebugLoc dl = N.getDebugLoc();
819 // If this can be more profitably realized as r+r, fail.
820 if (SelectAddressRegReg(N, Disp, Base, DAG))
823 if (N.getOpcode() == ISD::ADD) {
825 if (isIntS16Immediate(N.getOperand(1), imm)) {
826 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
827 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
828 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
830 Base = N.getOperand(0);
832 return true; // [r+i]
833 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
834 // Match LOAD (ADD (X, Lo(G))).
835 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
836 && "Cannot handle constant offsets yet!");
837 Disp = N.getOperand(1).getOperand(0); // The global address.
838 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
839 Disp.getOpcode() == ISD::TargetConstantPool ||
840 Disp.getOpcode() == ISD::TargetJumpTable);
841 Base = N.getOperand(0);
842 return true; // [&g+r]
844 } else if (N.getOpcode() == ISD::OR) {
846 if (isIntS16Immediate(N.getOperand(1), imm)) {
847 // If this is an or of disjoint bitfields, we can codegen this as an add
848 // (for better address arithmetic) if the LHS and RHS of the OR are
849 // provably disjoint.
850 APInt LHSKnownZero, LHSKnownOne;
851 DAG.ComputeMaskedBits(N.getOperand(0),
852 APInt::getAllOnesValue(N.getOperand(0)
853 .getValueSizeInBits()),
854 LHSKnownZero, LHSKnownOne);
856 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
857 // If all of the bits are known zero on the LHS or RHS, the add won't
859 Base = N.getOperand(0);
860 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
864 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
865 // Loading from a constant address.
867 // If this address fits entirely in a 16-bit sext immediate field, codegen
870 if (isIntS16Immediate(CN, Imm)) {
871 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
872 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
876 // Handle 32-bit sext immediates with LIS + addr mode.
877 if (CN->getValueType(0) == MVT::i32 ||
878 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
879 int Addr = (int)CN->getZExtValue();
881 // Otherwise, break this down into an LIS + disp.
882 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
884 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
885 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
886 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base), 0);
891 Disp = DAG.getTargetConstant(0, getPointerTy());
892 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
893 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
896 return true; // [r+0]
899 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
900 /// represented as an indexed [r+r] operation.
901 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
903 SelectionDAG &DAG) const {
904 // Check to see if we can easily represent this as an [r+r] address. This
905 // will fail if it thinks that the address is more profitably represented as
906 // reg+imm, e.g. where imm = 0.
907 if (SelectAddressRegReg(N, Base, Index, DAG))
910 // If the operand is an addition, always emit this as [r+r], since this is
911 // better (for code size, and execution, as the memop does the add for free)
912 // than emitting an explicit add.
913 if (N.getOpcode() == ISD::ADD) {
914 Base = N.getOperand(0);
915 Index = N.getOperand(1);
919 // Otherwise, do it the hard way, using R0 as the base register.
920 Base = DAG.getRegister(PPC::R0, N.getValueType());
925 /// SelectAddressRegImmShift - Returns true if the address N can be
926 /// represented by a base register plus a signed 14-bit displacement
927 /// [r+imm*4]. Suitable for use by STD and friends.
928 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
930 SelectionDAG &DAG) const {
931 // FIXME dl should come from the parent load or store, not the address
932 DebugLoc dl = N.getDebugLoc();
933 // If this can be more profitably realized as r+r, fail.
934 if (SelectAddressRegReg(N, Disp, Base, DAG))
937 if (N.getOpcode() == ISD::ADD) {
939 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
940 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
941 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
942 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
944 Base = N.getOperand(0);
946 return true; // [r+i]
947 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
948 // Match LOAD (ADD (X, Lo(G))).
949 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
950 && "Cannot handle constant offsets yet!");
951 Disp = N.getOperand(1).getOperand(0); // The global address.
952 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
953 Disp.getOpcode() == ISD::TargetConstantPool ||
954 Disp.getOpcode() == ISD::TargetJumpTable);
955 Base = N.getOperand(0);
956 return true; // [&g+r]
958 } else if (N.getOpcode() == ISD::OR) {
960 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
961 // If this is an or of disjoint bitfields, we can codegen this as an add
962 // (for better address arithmetic) if the LHS and RHS of the OR are
963 // provably disjoint.
964 APInt LHSKnownZero, LHSKnownOne;
965 DAG.ComputeMaskedBits(N.getOperand(0),
966 APInt::getAllOnesValue(N.getOperand(0)
967 .getValueSizeInBits()),
968 LHSKnownZero, LHSKnownOne);
969 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
970 // If all of the bits are known zero on the LHS or RHS, the add won't
972 Base = N.getOperand(0);
973 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
977 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
978 // Loading from a constant address. Verify low two bits are clear.
979 if ((CN->getZExtValue() & 3) == 0) {
980 // If this address fits entirely in a 14-bit sext immediate field, codegen
983 if (isIntS16Immediate(CN, Imm)) {
984 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
985 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
989 // Fold the low-part of 32-bit absolute addresses into addr mode.
990 if (CN->getValueType(0) == MVT::i32 ||
991 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
992 int Addr = (int)CN->getZExtValue();
994 // Otherwise, break this down into an LIS + disp.
995 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
996 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
997 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
998 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base),0);
1004 Disp = DAG.getTargetConstant(0, getPointerTy());
1005 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1006 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1009 return true; // [r+0]
1013 /// getPreIndexedAddressParts - returns true by value, base pointer and
1014 /// offset pointer and addressing mode by reference if the node's address
1015 /// can be legally represented as pre-indexed load / store address.
1016 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1018 ISD::MemIndexedMode &AM,
1019 SelectionDAG &DAG) const {
1020 // Disabled by default for now.
1021 if (!EnablePPCPreinc) return false;
1025 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1026 Ptr = LD->getBasePtr();
1027 VT = LD->getMemoryVT();
1029 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1031 Ptr = ST->getBasePtr();
1032 VT = ST->getMemoryVT();
1036 // PowerPC doesn't have preinc load/store instructions for vectors.
1040 // TODO: Check reg+reg first.
1042 // LDU/STU use reg+imm*4, others use reg+imm.
1043 if (VT != MVT::i64) {
1045 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1049 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1053 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1054 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1055 // sext i32 to i64 when addr mode is r+i.
1056 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1057 LD->getExtensionType() == ISD::SEXTLOAD &&
1058 isa<ConstantSDNode>(Offset))
1066 //===----------------------------------------------------------------------===//
1067 // LowerOperation implementation
1068 //===----------------------------------------------------------------------===//
1070 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1071 SelectionDAG &DAG) {
1072 MVT PtrVT = Op.getValueType();
1073 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1074 Constant *C = CP->getConstVal();
1075 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1076 SDValue Zero = DAG.getConstant(0, PtrVT);
1077 // FIXME there isn't really any debug info here
1078 DebugLoc dl = Op.getDebugLoc();
1080 const TargetMachine &TM = DAG.getTarget();
1082 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, CPI, Zero);
1083 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, CPI, Zero);
1085 // If this is a non-darwin platform, we don't support non-static relo models
1087 if (TM.getRelocationModel() == Reloc::Static ||
1088 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1089 // Generate non-pic code that has direct accesses to the constant pool.
1090 // The address of the global is just (hi(&g)+lo(&g)).
1091 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1094 if (TM.getRelocationModel() == Reloc::PIC_) {
1095 // With PIC, the first instruction is actually "GR+hi(&G)".
1096 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1097 DAG.getNode(PPCISD::GlobalBaseReg,
1098 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1101 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1105 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1106 MVT PtrVT = Op.getValueType();
1107 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1108 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1109 SDValue Zero = DAG.getConstant(0, PtrVT);
1110 // FIXME there isn't really any debug loc here
1111 DebugLoc dl = Op.getDebugLoc();
1113 const TargetMachine &TM = DAG.getTarget();
1115 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, JTI, Zero);
1116 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, JTI, Zero);
1118 // If this is a non-darwin platform, we don't support non-static relo models
1120 if (TM.getRelocationModel() == Reloc::Static ||
1121 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1122 // Generate non-pic code that has direct accesses to the constant pool.
1123 // The address of the global is just (hi(&g)+lo(&g)).
1124 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1127 if (TM.getRelocationModel() == Reloc::PIC_) {
1128 // With PIC, the first instruction is actually "GR+hi(&G)".
1129 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1130 DAG.getNode(PPCISD::GlobalBaseReg,
1131 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1134 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1138 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1139 SelectionDAG &DAG) {
1140 assert(0 && "TLS not implemented for PPC.");
1141 return SDValue(); // Not reached
1144 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1145 SelectionDAG &DAG) {
1146 MVT PtrVT = Op.getValueType();
1147 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1148 GlobalValue *GV = GSDN->getGlobal();
1149 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1150 SDValue Zero = DAG.getConstant(0, PtrVT);
1151 // FIXME there isn't really any debug info here
1152 DebugLoc dl = GSDN->getDebugLoc();
1154 const TargetMachine &TM = DAG.getTarget();
1156 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, GA, Zero);
1157 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, GA, Zero);
1159 // If this is a non-darwin platform, we don't support non-static relo models
1161 if (TM.getRelocationModel() == Reloc::Static ||
1162 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1163 // Generate non-pic code that has direct accesses to globals.
1164 // The address of the global is just (hi(&g)+lo(&g)).
1165 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1168 if (TM.getRelocationModel() == Reloc::PIC_) {
1169 // With PIC, the first instruction is actually "GR+hi(&G)".
1170 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1171 DAG.getNode(PPCISD::GlobalBaseReg,
1172 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1175 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1177 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1180 // If the global is weak or external, we have to go through the lazy
1182 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Lo, NULL, 0);
1185 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1186 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1187 DebugLoc dl = Op.getDebugLoc();
1189 // If we're comparing for equality to zero, expose the fact that this is
1190 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1191 // fold the new nodes.
1192 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1193 if (C->isNullValue() && CC == ISD::SETEQ) {
1194 MVT VT = Op.getOperand(0).getValueType();
1195 SDValue Zext = Op.getOperand(0);
1196 if (VT.bitsLT(MVT::i32)) {
1198 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1200 unsigned Log2b = Log2_32(VT.getSizeInBits());
1201 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1202 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1203 DAG.getConstant(Log2b, MVT::i32));
1204 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1206 // Leave comparisons against 0 and -1 alone for now, since they're usually
1207 // optimized. FIXME: revisit this when we can custom lower all setcc
1209 if (C->isAllOnesValue() || C->isNullValue())
1213 // If we have an integer seteq/setne, turn it into a compare against zero
1214 // by xor'ing the rhs with the lhs, which is faster than setting a
1215 // condition register, reading it back out, and masking the correct bit. The
1216 // normal approach here uses sub to do this instead of xor. Using xor exposes
1217 // the result to other bit-twiddling opportunities.
1218 MVT LHSVT = Op.getOperand(0).getValueType();
1219 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1220 MVT VT = Op.getValueType();
1221 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1223 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1228 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1229 int VarArgsFrameIndex,
1230 int VarArgsStackOffset,
1231 unsigned VarArgsNumGPR,
1232 unsigned VarArgsNumFPR,
1233 const PPCSubtarget &Subtarget) {
1235 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1236 return SDValue(); // Not reached
1239 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1240 SDValue Chain = Op.getOperand(0);
1241 SDValue Trmp = Op.getOperand(1); // trampoline
1242 SDValue FPtr = Op.getOperand(2); // nested function
1243 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1244 DebugLoc dl = Op.getDebugLoc();
1246 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1247 bool isPPC64 = (PtrVT == MVT::i64);
1248 const Type *IntPtrTy =
1249 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType();
1251 TargetLowering::ArgListTy Args;
1252 TargetLowering::ArgListEntry Entry;
1254 Entry.Ty = IntPtrTy;
1255 Entry.Node = Trmp; Args.push_back(Entry);
1257 // TrampSize == (isPPC64 ? 48 : 40);
1258 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1259 isPPC64 ? MVT::i64 : MVT::i32);
1260 Args.push_back(Entry);
1262 Entry.Node = FPtr; Args.push_back(Entry);
1263 Entry.Node = Nest; Args.push_back(Entry);
1265 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1266 std::pair<SDValue, SDValue> CallResult =
1267 LowerCallTo(Chain, Op.getValueType().getTypeForMVT(), false, false,
1268 false, false, CallingConv::C, false,
1269 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1273 { CallResult.first, CallResult.second };
1275 return DAG.getMergeValues(Ops, 2, dl);
1278 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1279 int VarArgsFrameIndex,
1280 int VarArgsStackOffset,
1281 unsigned VarArgsNumGPR,
1282 unsigned VarArgsNumFPR,
1283 const PPCSubtarget &Subtarget) {
1284 DebugLoc dl = Op.getDebugLoc();
1286 if (Subtarget.isMachoABI()) {
1287 // vastart just stores the address of the VarArgsFrameIndex slot into the
1288 // memory location argument.
1289 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1290 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1291 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1292 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1295 // For ELF 32 ABI we follow the layout of the va_list struct.
1296 // We suppose the given va_list is already allocated.
1299 // char gpr; /* index into the array of 8 GPRs
1300 // * stored in the register save area
1301 // * gpr=0 corresponds to r3,
1302 // * gpr=1 to r4, etc.
1304 // char fpr; /* index into the array of 8 FPRs
1305 // * stored in the register save area
1306 // * fpr=0 corresponds to f1,
1307 // * fpr=1 to f2, etc.
1309 // char *overflow_arg_area;
1310 // /* location on stack that holds
1311 // * the next overflow argument
1313 // char *reg_save_area;
1314 // /* where r3:r10 and f1:f8 (if saved)
1320 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1321 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1324 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1326 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1327 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1329 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1330 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1332 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1333 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1335 uint64_t FPROffset = 1;
1336 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1338 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1340 // Store first byte : number of int regs
1341 SDValue firstStore = DAG.getStore(Op.getOperand(0), dl, ArgGPR,
1342 Op.getOperand(1), SV, 0);
1343 uint64_t nextOffset = FPROffset;
1344 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1347 // Store second byte : number of float regs
1348 SDValue secondStore =
1349 DAG.getStore(firstStore, dl, ArgFPR, nextPtr, SV, nextOffset);
1350 nextOffset += StackOffset;
1351 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1353 // Store second word : arguments given on stack
1354 SDValue thirdStore =
1355 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, SV, nextOffset);
1356 nextOffset += FrameOffset;
1357 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1359 // Store third word : arguments given in registers
1360 return DAG.getStore(thirdStore, dl, FR, nextPtr, SV, nextOffset);
1364 #include "PPCGenCallingConv.inc"
1366 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1367 /// depending on which subtarget is selected.
1368 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1369 if (Subtarget.isMachoABI()) {
1370 static const unsigned FPR[] = {
1371 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1372 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1378 static const unsigned FPR[] = {
1379 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1385 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1387 static unsigned CalculateStackSlotSize(SDValue Arg, ISD::ArgFlagsTy Flags,
1388 bool isVarArg, unsigned PtrByteSize) {
1389 MVT ArgVT = Arg.getValueType();
1390 unsigned ArgSize =ArgVT.getSizeInBits()/8;
1391 if (Flags.isByVal())
1392 ArgSize = Flags.getByValSize();
1393 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1399 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1401 int &VarArgsFrameIndex,
1402 int &VarArgsStackOffset,
1403 unsigned &VarArgsNumGPR,
1404 unsigned &VarArgsNumFPR,
1405 const PPCSubtarget &Subtarget) {
1406 // TODO: add description of PPC stack frame format, or at least some docs.
1408 MachineFunction &MF = DAG.getMachineFunction();
1409 MachineFrameInfo *MFI = MF.getFrameInfo();
1410 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1411 SmallVector<SDValue, 8> ArgValues;
1412 SDValue Root = Op.getOperand(0);
1413 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1414 DebugLoc dl = Op.getDebugLoc();
1416 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1417 bool isPPC64 = PtrVT == MVT::i64;
1418 bool isMachoABI = Subtarget.isMachoABI();
1419 bool isELF32_ABI = Subtarget.isELF32_ABI();
1420 // Potential tail calls could cause overwriting of argument stack slots.
1421 unsigned CC = MF.getFunction()->getCallingConv();
1422 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1423 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1425 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1426 // Area that is at least reserved in caller of this function.
1427 unsigned MinReservedArea = ArgOffset;
1429 static const unsigned GPR_32[] = { // 32-bit registers.
1430 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1431 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1433 static const unsigned GPR_64[] = { // 64-bit registers.
1434 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1435 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1438 static const unsigned *FPR = GetFPR(Subtarget);
1440 static const unsigned VR[] = {
1441 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1442 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1445 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1446 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1447 const unsigned Num_VR_Regs = array_lengthof( VR);
1449 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1451 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1453 // In 32-bit non-varargs functions, the stack space for vectors is after the
1454 // stack space for non-vectors. We do not use this space unless we have
1455 // too many vectors to fit in registers, something that only occurs in
1456 // constructed examples:), but we have to walk the arglist to figure
1457 // that out...for the pathological case, compute VecArgOffset as the
1458 // start of the vector parameter area. Computing VecArgOffset is the
1459 // entire point of the following loop.
1460 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1461 // to handle Elf here.
1462 unsigned VecArgOffset = ArgOffset;
1463 if (!isVarArg && !isPPC64) {
1464 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1466 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1467 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1468 ISD::ArgFlagsTy Flags =
1469 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1471 if (Flags.isByVal()) {
1472 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1473 ObjSize = Flags.getByValSize();
1475 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1476 VecArgOffset += ArgSize;
1480 switch(ObjectVT.getSimpleVT()) {
1481 default: assert(0 && "Unhandled argument type!");
1484 VecArgOffset += isPPC64 ? 8 : 4;
1486 case MVT::i64: // PPC64
1494 // Nothing to do, we're only looking at Nonvector args here.
1499 // We've found where the vector parameter area in memory is. Skip the
1500 // first 12 parameters; these don't use that memory.
1501 VecArgOffset = ((VecArgOffset+15)/16)*16;
1502 VecArgOffset += 12*16;
1504 // Add DAG nodes to load the arguments or copy them out of registers. On
1505 // entry to a function on PPC, the arguments start after the linkage area,
1506 // although the first ones are often in registers.
1508 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1509 // represented with two words (long long or double) must be copied to an
1510 // even GPR_idx value or to an even ArgOffset value.
1512 SmallVector<SDValue, 8> MemOps;
1513 unsigned nAltivecParamsAtEnd = 0;
1514 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1515 ArgNo != e; ++ArgNo) {
1517 bool needsLoad = false;
1518 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1519 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1520 unsigned ArgSize = ObjSize;
1521 ISD::ArgFlagsTy Flags =
1522 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1523 // See if next argument requires stack alignment in ELF
1524 bool Align = Flags.isSplit();
1526 unsigned CurArgOffset = ArgOffset;
1528 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1529 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1530 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1531 if (isVarArg || isPPC64) {
1532 MinReservedArea = ((MinReservedArea+15)/16)*16;
1533 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1537 } else nAltivecParamsAtEnd++;
1539 // Calculate min reserved area.
1540 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1545 // FIXME alignment for ELF may not be right
1546 // FIXME the codegen can be much improved in some cases.
1547 // We do not have to keep everything in memory.
1548 if (Flags.isByVal()) {
1549 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1550 ObjSize = Flags.getByValSize();
1551 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1552 // Double word align in ELF
1553 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1554 // Objects of size 1 and 2 are right justified, everything else is
1555 // left justified. This means the memory address is adjusted forwards.
1556 if (ObjSize==1 || ObjSize==2) {
1557 CurArgOffset = CurArgOffset + (4 - ObjSize);
1559 // The value of the object is its address.
1560 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1561 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1562 ArgValues.push_back(FIN);
1563 if (ObjSize==1 || ObjSize==2) {
1564 if (GPR_idx != Num_GPR_Regs) {
1565 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1566 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1567 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1568 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1569 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1570 MemOps.push_back(Store);
1572 if (isMachoABI) ArgOffset += PtrByteSize;
1574 ArgOffset += PtrByteSize;
1578 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1579 // Store whatever pieces of the object are in registers
1580 // to memory. ArgVal will be address of the beginning of
1582 if (GPR_idx != Num_GPR_Regs) {
1583 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1584 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1585 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1586 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1587 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1588 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1589 MemOps.push_back(Store);
1591 if (isMachoABI) ArgOffset += PtrByteSize;
1593 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1600 switch (ObjectVT.getSimpleVT()) {
1601 default: assert(0 && "Unhandled argument type!");
1604 // Double word align in ELF
1605 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1607 if (GPR_idx != Num_GPR_Regs) {
1608 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1609 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1610 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1614 ArgSize = PtrByteSize;
1616 // Stack align in ELF
1617 if (needsLoad && Align && isELF32_ABI)
1618 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1619 // All int arguments reserve stack space in Macho ABI.
1620 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1624 case MVT::i64: // PPC64
1625 if (GPR_idx != Num_GPR_Regs) {
1626 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1627 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1628 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1630 if (ObjectVT == MVT::i32) {
1631 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1632 // value to MVT::i64 and then truncate to the correct register size.
1634 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1635 DAG.getValueType(ObjectVT));
1636 else if (Flags.isZExt())
1637 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1638 DAG.getValueType(ObjectVT));
1640 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1646 ArgSize = PtrByteSize;
1648 // All int arguments reserve stack space in Macho ABI.
1649 if (isMachoABI || needsLoad) ArgOffset += 8;
1654 // Every 4 bytes of argument space consumes one of the GPRs available for
1655 // argument passing.
1656 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1658 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1661 if (FPR_idx != Num_FPR_Regs) {
1663 if (ObjectVT == MVT::f32)
1664 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1666 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1667 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1668 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1674 // Stack align in ELF
1675 if (needsLoad && Align && isELF32_ABI)
1676 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1677 // All FP arguments reserve stack space in Macho ABI.
1678 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1684 // Note that vector arguments in registers don't reserve stack space,
1685 // except in varargs functions.
1686 if (VR_idx != Num_VR_Regs) {
1687 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1688 RegInfo.addLiveIn(VR[VR_idx], VReg);
1689 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1691 while ((ArgOffset % 16) != 0) {
1692 ArgOffset += PtrByteSize;
1693 if (GPR_idx != Num_GPR_Regs)
1697 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1701 if (!isVarArg && !isPPC64) {
1702 // Vectors go after all the nonvectors.
1703 CurArgOffset = VecArgOffset;
1706 // Vectors are aligned.
1707 ArgOffset = ((ArgOffset+15)/16)*16;
1708 CurArgOffset = ArgOffset;
1716 // We need to load the argument to a virtual register if we determined above
1717 // that we ran out of physical registers of the appropriate type.
1719 int FI = MFI->CreateFixedObject(ObjSize,
1720 CurArgOffset + (ArgSize - ObjSize),
1722 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1723 ArgVal = DAG.getLoad(ObjectVT, dl, Root, FIN, NULL, 0);
1726 ArgValues.push_back(ArgVal);
1729 // Set the size that is at least reserved in caller of this function. Tail
1730 // call optimized function's reserved stack space needs to be aligned so that
1731 // taking the difference between two stack areas will result in an aligned
1733 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1734 // Add the Altivec parameters at the end, if needed.
1735 if (nAltivecParamsAtEnd) {
1736 MinReservedArea = ((MinReservedArea+15)/16)*16;
1737 MinReservedArea += 16*nAltivecParamsAtEnd;
1740 std::max(MinReservedArea,
1741 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1742 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1743 getStackAlignment();
1744 unsigned AlignMask = TargetAlign-1;
1745 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1746 FI->setMinReservedArea(MinReservedArea);
1748 // If the function takes variable number of arguments, make a frame index for
1749 // the start of the first vararg value... for expansion of llvm.va_start.
1754 VarArgsNumGPR = GPR_idx;
1755 VarArgsNumFPR = FPR_idx;
1757 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1759 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
1760 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
1761 PtrVT.getSizeInBits()/8);
1763 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1770 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1772 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1774 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1775 // stored to the VarArgsFrameIndex on the stack.
1777 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1778 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1779 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1780 MemOps.push_back(Store);
1781 // Increment the address by four for the next argument to store
1782 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1783 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1787 // If this function is vararg, store any remaining integer argument regs
1788 // to their spots on the stack so that they may be loaded by deferencing the
1789 // result of va_next.
1790 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1793 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1795 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1797 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1798 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1799 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1800 MemOps.push_back(Store);
1801 // Increment the address by four for the next argument to store
1802 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1803 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1806 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1809 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1810 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1811 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1812 MemOps.push_back(Store);
1813 // Increment the address by eight for the next argument to store
1814 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1816 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1819 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1821 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1823 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1824 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::f64);
1825 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1826 MemOps.push_back(Store);
1827 // Increment the address by eight for the next argument to store
1828 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1830 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1835 if (!MemOps.empty())
1836 Root = DAG.getNode(ISD::TokenFactor, dl,
1837 MVT::Other, &MemOps[0], MemOps.size());
1839 ArgValues.push_back(Root);
1841 // Return the new list of results.
1842 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1843 &ArgValues[0], ArgValues.size());
1846 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
1849 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
1854 CallSDNode *TheCall,
1855 unsigned &nAltivecParamsAtEnd) {
1856 // Count how many bytes are to be pushed on the stack, including the linkage
1857 // area, and parameter passing area. We start with 24/48 bytes, which is
1858 // prereserved space for [SP][CR][LR][3 x unused].
1859 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1860 unsigned NumOps = TheCall->getNumArgs();
1861 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1863 // Add up all the space actually used.
1864 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1865 // they all go in registers, but we must reserve stack space for them for
1866 // possible use by the caller. In varargs or 64-bit calls, parameters are
1867 // assigned stack space in order, with padding so Altivec parameters are
1869 nAltivecParamsAtEnd = 0;
1870 for (unsigned i = 0; i != NumOps; ++i) {
1871 SDValue Arg = TheCall->getArg(i);
1872 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1873 MVT ArgVT = Arg.getValueType();
1874 // Varargs Altivec parameters are padded to a 16 byte boundary.
1875 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1876 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1877 if (!isVarArg && !isPPC64) {
1878 // Non-varargs Altivec parameters go after all the non-Altivec
1879 // parameters; handle those later so we know how much padding we need.
1880 nAltivecParamsAtEnd++;
1883 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1884 NumBytes = ((NumBytes+15)/16)*16;
1886 NumBytes += CalculateStackSlotSize(Arg, Flags, isVarArg, PtrByteSize);
1889 // Allow for Altivec parameters at the end, if needed.
1890 if (nAltivecParamsAtEnd) {
1891 NumBytes = ((NumBytes+15)/16)*16;
1892 NumBytes += 16*nAltivecParamsAtEnd;
1895 // The prolog code of the callee may store up to 8 GPR argument registers to
1896 // the stack, allowing va_start to index over them in memory if its varargs.
1897 // Because we cannot tell if this is needed on the caller side, we have to
1898 // conservatively assume that it is needed. As such, make sure we have at
1899 // least enough stack space for the caller to store the 8 GPRs.
1900 NumBytes = std::max(NumBytes,
1901 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1903 // Tail call needs the stack to be aligned.
1904 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1905 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1906 getStackAlignment();
1907 unsigned AlignMask = TargetAlign-1;
1908 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
1914 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
1915 /// adjusted to accomodate the arguments for the tailcall.
1916 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
1917 unsigned ParamSize) {
1919 if (!IsTailCall) return 0;
1921 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
1922 unsigned CallerMinReservedArea = FI->getMinReservedArea();
1923 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
1924 // Remember only if the new adjustement is bigger.
1925 if (SPDiff < FI->getTailCallSPDelta())
1926 FI->setTailCallSPDelta(SPDiff);
1931 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1932 /// following the call is a return. A function is eligible if caller/callee
1933 /// calling conventions match, currently only fastcc supports tail calls, and
1934 /// the function CALL is immediatly followed by a RET.
1936 PPCTargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1938 SelectionDAG& DAG) const {
1939 // Variable argument functions are not supported.
1940 if (!PerformTailCallOpt || TheCall->isVarArg())
1943 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1944 MachineFunction &MF = DAG.getMachineFunction();
1945 unsigned CallerCC = MF.getFunction()->getCallingConv();
1946 unsigned CalleeCC = TheCall->getCallingConv();
1947 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1948 // Functions containing by val parameters are not supported.
1949 for (unsigned i = 0; i != TheCall->getNumArgs(); i++) {
1950 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1951 if (Flags.isByVal()) return false;
1954 SDValue Callee = TheCall->getCallee();
1955 // Non PIC/GOT tail calls are supported.
1956 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1959 // At the moment we can only do local tail calls (in same module, hidden
1960 // or protected) if we are generating PIC.
1961 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1962 return G->getGlobal()->hasHiddenVisibility()
1963 || G->getGlobal()->hasProtectedVisibility();
1970 /// isCallCompatibleAddress - Return the immediate to use if the specified
1971 /// 32-bit value is representable in the immediate field of a BxA instruction.
1972 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
1973 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1976 int Addr = C->getZExtValue();
1977 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1978 (Addr << 6 >> 6) != Addr)
1979 return 0; // Top 6 bits have to be sext of immediate.
1981 return DAG.getConstant((int)C->getZExtValue() >> 2,
1982 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
1987 struct TailCallArgumentInfo {
1992 TailCallArgumentInfo() : FrameIdx(0) {}
1997 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
1999 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2001 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2002 SmallVector<SDValue, 8> &MemOpChains,
2004 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2005 SDValue Arg = TailCallArgs[i].Arg;
2006 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2007 int FI = TailCallArgs[i].FrameIdx;
2008 // Store relative to framepointer.
2009 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2010 PseudoSourceValue::getFixedStack(FI),
2015 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2016 /// the appropriate stack slot for the tail call optimized function call.
2017 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2018 MachineFunction &MF,
2027 // Calculate the new stack slot for the return address.
2028 int SlotSize = isPPC64 ? 8 : 4;
2029 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2031 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2033 int NewFPLoc = SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64,
2035 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
2037 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2038 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2039 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2040 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2041 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2042 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2043 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2048 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2049 /// the position of the argument.
2051 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2052 SDValue Arg, int SPDiff, unsigned ArgOffset,
2053 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2054 int Offset = ArgOffset + SPDiff;
2055 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2056 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2057 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2058 SDValue FIN = DAG.getFrameIndex(FI, VT);
2059 TailCallArgumentInfo Info;
2061 Info.FrameIdxOp = FIN;
2063 TailCallArguments.push_back(Info);
2066 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2067 /// stack slot. Returns the chain as result and the loaded frame pointers in
2068 /// LROpOut/FPOpout. Used when tail calling.
2069 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2076 // Load the LR and FP stack slot for later adjusting.
2077 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2078 LROpOut = getReturnAddrFrameIndex(DAG);
2079 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, NULL, 0);
2080 Chain = SDValue(LROpOut.getNode(), 1);
2081 FPOpOut = getFramePointerFrameIndex(DAG);
2082 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, NULL, 0);
2083 Chain = SDValue(FPOpOut.getNode(), 1);
2088 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2089 /// by "Src" to address "Dst" of size "Size". Alignment information is
2090 /// specified by the specific parameter attribute. The copy will be passed as
2091 /// a byval function parameter.
2092 /// Sometimes what we are copying is the end of a larger object, the part that
2093 /// does not fit in registers.
2095 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2096 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2097 unsigned Size, DebugLoc dl) {
2098 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
2099 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2100 false, NULL, 0, NULL, 0);
2103 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2106 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2107 SDValue Arg, SDValue PtrOff, int SPDiff,
2108 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2109 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2110 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments,
2112 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2117 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2119 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2120 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2121 DAG.getConstant(ArgOffset, PtrVT));
2123 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0));
2124 // Calculate and remember argument location.
2125 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2129 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2130 const PPCSubtarget &Subtarget,
2131 TargetMachine &TM) {
2132 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2133 SDValue Chain = TheCall->getChain();
2134 bool isVarArg = TheCall->isVarArg();
2135 unsigned CC = TheCall->getCallingConv();
2136 bool isTailCall = TheCall->isTailCall()
2137 && CC == CallingConv::Fast && PerformTailCallOpt;
2138 SDValue Callee = TheCall->getCallee();
2139 unsigned NumOps = TheCall->getNumArgs();
2140 DebugLoc dl = TheCall->getDebugLoc();
2142 bool isMachoABI = Subtarget.isMachoABI();
2143 bool isELF32_ABI = Subtarget.isELF32_ABI();
2145 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2146 bool isPPC64 = PtrVT == MVT::i64;
2147 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2149 MachineFunction &MF = DAG.getMachineFunction();
2151 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
2152 // SelectExpr to use to put the arguments in the appropriate registers.
2153 std::vector<SDValue> args_to_use;
2155 // Mark this function as potentially containing a function that contains a
2156 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2157 // and restoring the callers stack pointer in this functions epilog. This is
2158 // done because by tail calling the called function might overwrite the value
2159 // in this function's (MF) stack pointer stack slot 0(SP).
2160 if (PerformTailCallOpt && CC==CallingConv::Fast)
2161 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2163 unsigned nAltivecParamsAtEnd = 0;
2165 // Count how many bytes are to be pushed on the stack, including the linkage
2166 // area, and parameter passing area. We start with 24/48 bytes, which is
2167 // prereserved space for [SP][CR][LR][3 x unused].
2169 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2170 TheCall, nAltivecParamsAtEnd);
2172 // Calculate by how many bytes the stack has to be adjusted in case of tail
2173 // call optimization.
2174 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2176 // Adjust the stack pointer for the new arguments...
2177 // These operations are automatically eliminated by the prolog/epilog pass
2178 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2179 SDValue CallSeqStart = Chain;
2181 // Load the return address and frame pointer so it can be move somewhere else
2184 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
2186 // Set up a copy of the stack pointer for use loading and storing any
2187 // arguments that may not fit in the registers available for argument
2191 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2193 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2195 // Figure out which arguments are going to go in registers, and which in
2196 // memory. Also, if this is a vararg function, floating point operations
2197 // must be stored to our stack, and loaded into integer regs as well, if
2198 // any integer regs are available for argument passing.
2199 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2200 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2202 static const unsigned GPR_32[] = { // 32-bit registers.
2203 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2204 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2206 static const unsigned GPR_64[] = { // 64-bit registers.
2207 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2208 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2210 static const unsigned *FPR = GetFPR(Subtarget);
2212 static const unsigned VR[] = {
2213 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2214 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2216 const unsigned NumGPRs = array_lengthof(GPR_32);
2217 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2218 const unsigned NumVRs = array_lengthof( VR);
2220 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2222 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2223 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2225 SmallVector<SDValue, 8> MemOpChains;
2226 for (unsigned i = 0; i != NumOps; ++i) {
2228 SDValue Arg = TheCall->getArg(i);
2229 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2230 // See if next argument requires stack alignment in ELF
2231 bool Align = Flags.isSplit();
2233 // PtrOff will be used to store the current argument to the stack if a
2234 // register cannot be found for it.
2237 // Stack align in ELF 32
2238 if (isELF32_ABI && Align)
2239 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2240 StackPtr.getValueType());
2242 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2244 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
2246 // On PPC64, promote integers to 64-bit values.
2247 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2248 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2249 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2250 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
2253 // FIXME Elf untested, what are alignment rules?
2254 // FIXME memcpy is used way more than necessary. Correctness first.
2255 if (Flags.isByVal()) {
2256 unsigned Size = Flags.getByValSize();
2257 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2258 if (Size==1 || Size==2) {
2259 // Very small objects are passed right-justified.
2260 // Everything else is passed left-justified.
2261 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2262 if (GPR_idx != NumGPRs) {
2263 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
2265 MemOpChains.push_back(Load.getValue(1));
2266 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2268 ArgOffset += PtrByteSize;
2270 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2271 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
2272 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2273 CallSeqStart.getNode()->getOperand(0),
2274 Flags, DAG, Size, dl);
2275 // This must go outside the CALLSEQ_START..END.
2276 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2277 CallSeqStart.getNode()->getOperand(1));
2278 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2279 NewCallSeqStart.getNode());
2280 Chain = CallSeqStart = NewCallSeqStart;
2281 ArgOffset += PtrByteSize;
2285 // Copy entire object into memory. There are cases where gcc-generated
2286 // code assumes it is there, even if it could be put entirely into
2287 // registers. (This is not what the doc says.)
2288 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2289 CallSeqStart.getNode()->getOperand(0),
2290 Flags, DAG, Size, dl);
2291 // This must go outside the CALLSEQ_START..END.
2292 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2293 CallSeqStart.getNode()->getOperand(1));
2294 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2295 Chain = CallSeqStart = NewCallSeqStart;
2296 // And copy the pieces of it that fit into registers.
2297 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2298 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2299 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
2300 if (GPR_idx != NumGPRs) {
2301 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, NULL, 0);
2302 MemOpChains.push_back(Load.getValue(1));
2303 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2305 ArgOffset += PtrByteSize;
2307 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2314 switch (Arg.getValueType().getSimpleVT()) {
2315 default: assert(0 && "Unexpected ValueType for argument!");
2318 // Double word align in ELF
2319 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2320 if (GPR_idx != NumGPRs) {
2321 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2323 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2324 isPPC64, isTailCall, false, MemOpChains,
2325 TailCallArguments, dl);
2328 if (inMem || isMachoABI) {
2329 // Stack align in ELF
2330 if (isELF32_ABI && Align)
2331 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2333 ArgOffset += PtrByteSize;
2338 if (FPR_idx != NumFPRs) {
2339 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2342 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
2343 MemOpChains.push_back(Store);
2345 // Float varargs are always shadowed in available integer registers
2346 if (GPR_idx != NumGPRs) {
2347 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2348 MemOpChains.push_back(Load.getValue(1));
2349 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2352 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2353 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2354 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
2355 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2356 MemOpChains.push_back(Load.getValue(1));
2357 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2361 // If we have any FPRs remaining, we may also have GPRs remaining.
2362 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2365 if (GPR_idx != NumGPRs)
2367 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
2368 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
2373 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2374 isPPC64, isTailCall, false, MemOpChains,
2375 TailCallArguments, dl);
2378 if (inMem || isMachoABI) {
2379 // Stack align in ELF
2380 if (isELF32_ABI && Align)
2381 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2385 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2393 // These go aligned on the stack, or in the corresponding R registers
2394 // when within range. The Darwin PPC ABI doc claims they also go in
2395 // V registers; in fact gcc does this only for arguments that are
2396 // prototyped, not for those that match the ... We do it for all
2397 // arguments, seems to work.
2398 while (ArgOffset % 16 !=0) {
2399 ArgOffset += PtrByteSize;
2400 if (GPR_idx != NumGPRs)
2403 // We could elide this store in the case where the object fits
2404 // entirely in R registers. Maybe later.
2405 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2406 DAG.getConstant(ArgOffset, PtrVT));
2407 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
2408 MemOpChains.push_back(Store);
2409 if (VR_idx != NumVRs) {
2410 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, NULL, 0);
2411 MemOpChains.push_back(Load.getValue(1));
2412 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2415 for (unsigned i=0; i<16; i+=PtrByteSize) {
2416 if (GPR_idx == NumGPRs)
2418 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
2419 DAG.getConstant(i, PtrVT));
2420 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, NULL, 0);
2421 MemOpChains.push_back(Load.getValue(1));
2422 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2427 // Non-varargs Altivec params generally go in registers, but have
2428 // stack space allocated at the end.
2429 if (VR_idx != NumVRs) {
2430 // Doesn't have GPR space allocated.
2431 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2432 } else if (nAltivecParamsAtEnd==0) {
2433 // We are emitting Altivec params in order.
2434 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2435 isPPC64, isTailCall, true, MemOpChains,
2436 TailCallArguments, dl);
2442 // If all Altivec parameters fit in registers, as they usually do,
2443 // they get stack space following the non-Altivec parameters. We
2444 // don't track this here because nobody below needs it.
2445 // If there are more Altivec parameters than fit in registers emit
2447 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2449 // Offset is aligned; skip 1st 12 params which go in V registers.
2450 ArgOffset = ((ArgOffset+15)/16)*16;
2452 for (unsigned i = 0; i != NumOps; ++i) {
2453 SDValue Arg = TheCall->getArg(i);
2454 MVT ArgType = Arg.getValueType();
2455 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2456 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2459 // We are emitting Altivec params in order.
2460 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2461 isPPC64, isTailCall, true, MemOpChains,
2462 TailCallArguments, dl);
2469 if (!MemOpChains.empty())
2470 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2471 &MemOpChains[0], MemOpChains.size());
2473 // Build a sequence of copy-to-reg nodes chained together with token chain
2474 // and flag operands which copy the outgoing args into the appropriate regs.
2476 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2477 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2478 RegsToPass[i].second, InFlag);
2479 InFlag = Chain.getValue(1);
2482 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2483 if (isVarArg && isELF32_ABI) {
2484 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, dl, MVT::i32), 0);
2485 Chain = DAG.getCopyToReg(Chain, dl, PPC::CR1EQ, SetCR, InFlag);
2486 InFlag = Chain.getValue(1);
2489 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2490 // might overwrite each other in case of tail call optimization.
2492 SmallVector<SDValue, 8> MemOpChains2;
2493 // Do not flag preceeding copytoreg stuff together with the following stuff.
2495 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2497 if (!MemOpChains2.empty())
2498 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2499 &MemOpChains2[0], MemOpChains2.size());
2501 // Store the return address to the appropriate stack slot.
2502 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2503 isPPC64, isMachoABI, dl);
2506 // Emit callseq_end just before tailcall node.
2508 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2509 DAG.getIntPtrConstant(0, true), InFlag);
2510 InFlag = Chain.getValue(1);
2513 std::vector<MVT> NodeTys;
2514 NodeTys.push_back(MVT::Other); // Returns a chain
2515 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2517 SmallVector<SDValue, 8> Ops;
2518 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2520 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2521 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2522 // node so that legalize doesn't hack it.
2523 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2524 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2525 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2526 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2527 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2528 // If this is an absolute destination address, use the munged value.
2529 Callee = SDValue(Dest, 0);
2531 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2532 // to do the call, we can't use PPCISD::CALL.
2533 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2534 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2535 2 + (InFlag.getNode() != 0));
2536 InFlag = Chain.getValue(1);
2538 // Copy the callee address into R12/X12 on darwin.
2540 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2541 Chain = DAG.getCopyToReg(Chain, dl, Reg, Callee, InFlag);
2542 InFlag = Chain.getValue(1);
2546 NodeTys.push_back(MVT::Other);
2547 NodeTys.push_back(MVT::Flag);
2548 Ops.push_back(Chain);
2549 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2551 // Add CTR register as callee so a bctr can be emitted later.
2553 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2556 // If this is a direct call, pass the chain and the callee.
2557 if (Callee.getNode()) {
2558 Ops.push_back(Chain);
2559 Ops.push_back(Callee);
2561 // If this is a tail call add stack pointer delta.
2563 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2565 // Add argument registers to the end of the list so that they are known live
2567 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2568 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2569 RegsToPass[i].second.getValueType()));
2571 // When performing tail call optimization the callee pops its arguments off
2572 // the stack. Account for this here so these bytes can be pushed back on in
2573 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2574 int BytesCalleePops =
2575 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2577 if (InFlag.getNode())
2578 Ops.push_back(InFlag);
2582 assert(InFlag.getNode() &&
2583 "Flag must be set. Depend on flag being set in LowerRET");
2584 Chain = DAG.getNode(PPCISD::TAILCALL, dl,
2585 TheCall->getVTList(), &Ops[0], Ops.size());
2586 return SDValue(Chain.getNode(), Op.getResNo());
2589 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2590 InFlag = Chain.getValue(1);
2592 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2593 DAG.getIntPtrConstant(BytesCalleePops, true),
2595 if (TheCall->getValueType(0) != MVT::Other)
2596 InFlag = Chain.getValue(1);
2598 SmallVector<SDValue, 16> ResultVals;
2599 SmallVector<CCValAssign, 16> RVLocs;
2600 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2601 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
2602 CCInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
2604 // Copy all of the result registers out of their specified physreg.
2605 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2606 CCValAssign &VA = RVLocs[i];
2607 MVT VT = VA.getValVT();
2608 assert(VA.isRegLoc() && "Can only return in registers!");
2609 Chain = DAG.getCopyFromReg(Chain, dl,
2610 VA.getLocReg(), VT, InFlag).getValue(1);
2611 ResultVals.push_back(Chain.getValue(0));
2612 InFlag = Chain.getValue(2);
2615 // If the function returns void, just return the chain.
2619 // Otherwise, merge everything together with a MERGE_VALUES node.
2620 ResultVals.push_back(Chain);
2621 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
2622 &ResultVals[0], ResultVals.size());
2623 return Res.getValue(Op.getResNo());
2626 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
2627 TargetMachine &TM) {
2628 SmallVector<CCValAssign, 16> RVLocs;
2629 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2630 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2631 DebugLoc dl = Op.getDebugLoc();
2632 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2633 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
2635 // If this is the first return lowered for this function, add the regs to the
2636 // liveout set for the function.
2637 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2638 for (unsigned i = 0; i != RVLocs.size(); ++i)
2639 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2642 SDValue Chain = Op.getOperand(0);
2644 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
2645 if (Chain.getOpcode() == PPCISD::TAILCALL) {
2646 SDValue TailCall = Chain;
2647 SDValue TargetAddress = TailCall.getOperand(1);
2648 SDValue StackAdjustment = TailCall.getOperand(2);
2650 assert(((TargetAddress.getOpcode() == ISD::Register &&
2651 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
2652 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
2653 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
2654 isa<ConstantSDNode>(TargetAddress)) &&
2655 "Expecting an global address, external symbol, absolute value or register");
2657 assert(StackAdjustment.getOpcode() == ISD::Constant &&
2658 "Expecting a const value");
2660 SmallVector<SDValue,8> Operands;
2661 Operands.push_back(Chain.getOperand(0));
2662 Operands.push_back(TargetAddress);
2663 Operands.push_back(StackAdjustment);
2664 // Copy registers used by the call. Last operand is a flag so it is not
2666 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
2667 Operands.push_back(Chain.getOperand(i));
2669 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Operands[0],
2675 // Copy the result values into the output registers.
2676 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2677 CCValAssign &VA = RVLocs[i];
2678 assert(VA.isRegLoc() && "Can only return in registers!");
2679 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2680 Op.getOperand(i*2+1), Flag);
2681 Flag = Chain.getValue(1);
2685 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
2687 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
2690 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
2691 const PPCSubtarget &Subtarget) {
2692 // When we pop the dynamic allocation we need to restore the SP link.
2693 DebugLoc dl = Op.getDebugLoc();
2695 // Get the corect type for pointers.
2696 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2698 // Construct the stack pointer operand.
2699 bool IsPPC64 = Subtarget.isPPC64();
2700 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2701 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
2703 // Get the operands for the STACKRESTORE.
2704 SDValue Chain = Op.getOperand(0);
2705 SDValue SaveSP = Op.getOperand(1);
2707 // Load the old link SP.
2708 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr, NULL, 0);
2710 // Restore the stack pointer.
2711 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
2713 // Store the old link SP.
2714 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, NULL, 0);
2720 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
2721 MachineFunction &MF = DAG.getMachineFunction();
2722 bool IsPPC64 = PPCSubTarget.isPPC64();
2723 bool isMachoABI = PPCSubTarget.isMachoABI();
2724 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2726 // Get current frame pointer save index. The users of this index will be
2727 // primarily DYNALLOC instructions.
2728 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2729 int RASI = FI->getReturnAddrSaveIndex();
2731 // If the frame pointer save index hasn't been defined yet.
2733 // Find out what the fix offset of the frame pointer save area.
2734 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
2735 // Allocate the frame index for frame pointer save area.
2736 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
2738 FI->setReturnAddrSaveIndex(RASI);
2740 return DAG.getFrameIndex(RASI, PtrVT);
2744 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
2745 MachineFunction &MF = DAG.getMachineFunction();
2746 bool IsPPC64 = PPCSubTarget.isPPC64();
2747 bool isMachoABI = PPCSubTarget.isMachoABI();
2748 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2750 // Get current frame pointer save index. The users of this index will be
2751 // primarily DYNALLOC instructions.
2752 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2753 int FPSI = FI->getFramePointerSaveIndex();
2755 // If the frame pointer save index hasn't been defined yet.
2757 // Find out what the fix offset of the frame pointer save area.
2758 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2760 // Allocate the frame index for frame pointer save area.
2761 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2763 FI->setFramePointerSaveIndex(FPSI);
2765 return DAG.getFrameIndex(FPSI, PtrVT);
2768 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
2770 const PPCSubtarget &Subtarget) {
2772 SDValue Chain = Op.getOperand(0);
2773 SDValue Size = Op.getOperand(1);
2774 DebugLoc dl = Op.getDebugLoc();
2776 // Get the corect type for pointers.
2777 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2779 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
2780 DAG.getConstant(0, PtrVT), Size);
2781 // Construct a node for the frame pointer save index.
2782 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
2783 // Build a DYNALLOC node.
2784 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
2785 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2786 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
2789 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2791 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2792 // Not FP? Not a fsel.
2793 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
2794 !Op.getOperand(2).getValueType().isFloatingPoint())
2797 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2799 // Cannot handle SETEQ/SETNE.
2800 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDValue();
2802 MVT ResVT = Op.getValueType();
2803 MVT CmpVT = Op.getOperand(0).getValueType();
2804 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2805 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
2806 DebugLoc dl = Op.getDebugLoc();
2808 // If the RHS of the comparison is a 0.0, we don't need to do the
2809 // subtraction at all.
2810 if (isFloatingPointZero(RHS))
2812 default: break; // SETUO etc aren't handled by fsel.
2815 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2818 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2819 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
2820 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
2823 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2826 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2827 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
2828 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
2829 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
2834 default: break; // SETUO etc aren't handled by fsel.
2837 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
2838 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2839 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
2840 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
2843 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
2844 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2845 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
2846 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
2849 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
2850 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2851 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
2852 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
2855 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
2856 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2857 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
2858 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
2863 // FIXME: Split this code up when LegalizeDAGTypes lands.
2864 SDValue PPCTargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG,
2866 assert(Op.getOperand(0).getValueType().isFloatingPoint());
2867 SDValue Src = Op.getOperand(0);
2868 if (Src.getValueType() == MVT::f32)
2869 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
2872 switch (Op.getValueType().getSimpleVT()) {
2873 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2875 Tmp = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Src);
2878 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
2882 // Convert the FP value to an int value through memory.
2883 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
2885 // Emit a store to the stack slot.
2886 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr, NULL, 0);
2888 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2890 if (Op.getValueType() == MVT::i32)
2891 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
2892 DAG.getConstant(4, FIPtr.getValueType()));
2893 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, NULL, 0);
2896 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2897 DebugLoc dl = Op.getDebugLoc();
2898 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2899 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2902 if (Op.getOperand(0).getValueType() == MVT::i64) {
2903 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, dl,
2904 MVT::f64, Op.getOperand(0));
2905 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
2906 if (Op.getValueType() == MVT::f32)
2907 FP = DAG.getNode(ISD::FP_ROUND, dl,
2908 MVT::f32, FP, DAG.getIntPtrConstant(0));
2912 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2913 "Unhandled SINT_TO_FP type in custom expander!");
2914 // Since we only generate this in 64-bit mode, we can take advantage of
2915 // 64-bit registers. In particular, sign extend the input value into the
2916 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2917 // then lfd it and fcfid it.
2918 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2919 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2920 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2921 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2923 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
2926 // STD the extended value into the stack slot.
2927 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
2928 MachineMemOperand::MOStore, 0, 8, 8);
2929 SDValue Store = DAG.getNode(PPCISD::STD_32, dl, MVT::Other,
2930 DAG.getEntryNode(), Ext64, FIdx,
2931 DAG.getMemOperand(MO));
2932 // Load the value as a double.
2933 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, NULL, 0);
2935 // FCFID it and return it.
2936 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
2937 if (Op.getValueType() == MVT::f32)
2938 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
2942 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
2943 DebugLoc dl = Op.getDebugLoc();
2945 The rounding mode is in bits 30:31 of FPSR, and has the following
2952 FLT_ROUNDS, on the other hand, expects the following:
2959 To perform the conversion, we do:
2960 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2963 MachineFunction &MF = DAG.getMachineFunction();
2964 MVT VT = Op.getValueType();
2965 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2966 std::vector<MVT> NodeTys;
2967 SDValue MFFSreg, InFlag;
2969 // Save FP Control Word to register
2970 NodeTys.push_back(MVT::f64); // return register
2971 NodeTys.push_back(MVT::Flag); // unused in this context
2972 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
2974 // Save FP register to stack slot
2975 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2976 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2977 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
2978 StackSlot, NULL, 0);
2980 // Load FP Control Word from low 32 bits of stack slot.
2981 SDValue Four = DAG.getConstant(4, PtrVT);
2982 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
2983 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, NULL, 0);
2985 // Transform as necessary
2987 DAG.getNode(ISD::AND, dl, MVT::i32,
2988 CWD, DAG.getConstant(3, MVT::i32));
2990 DAG.getNode(ISD::SRL, dl, MVT::i32,
2991 DAG.getNode(ISD::AND, dl, MVT::i32,
2992 DAG.getNode(ISD::XOR, dl, MVT::i32,
2993 CWD, DAG.getConstant(3, MVT::i32)),
2994 DAG.getConstant(3, MVT::i32)),
2995 DAG.getConstant(1, MVT::i32));
2998 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3000 return DAG.getNode((VT.getSizeInBits() < 16 ?
3001 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3004 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3005 MVT VT = Op.getValueType();
3006 unsigned BitWidth = VT.getSizeInBits();
3007 DebugLoc dl = Op.getDebugLoc();
3008 assert(Op.getNumOperands() == 3 &&
3009 VT == Op.getOperand(1).getValueType() &&
3012 // Expand into a bunch of logical ops. Note that these ops
3013 // depend on the PPC behavior for oversized shift amounts.
3014 SDValue Lo = Op.getOperand(0);
3015 SDValue Hi = Op.getOperand(1);
3016 SDValue Amt = Op.getOperand(2);
3017 MVT AmtVT = Amt.getValueType();
3019 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3020 DAG.getConstant(BitWidth, AmtVT), Amt);
3021 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3022 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3023 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3024 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3025 DAG.getConstant(-BitWidth, AmtVT));
3026 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3027 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3028 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3029 SDValue OutOps[] = { OutLo, OutHi };
3030 return DAG.getMergeValues(OutOps, 2, dl);
3033 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3034 MVT VT = Op.getValueType();
3035 DebugLoc dl = Op.getDebugLoc();
3036 unsigned BitWidth = VT.getSizeInBits();
3037 assert(Op.getNumOperands() == 3 &&
3038 VT == Op.getOperand(1).getValueType() &&
3041 // Expand into a bunch of logical ops. Note that these ops
3042 // depend on the PPC behavior for oversized shift amounts.
3043 SDValue Lo = Op.getOperand(0);
3044 SDValue Hi = Op.getOperand(1);
3045 SDValue Amt = Op.getOperand(2);
3046 MVT AmtVT = Amt.getValueType();
3048 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3049 DAG.getConstant(BitWidth, AmtVT), Amt);
3050 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3051 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3052 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3053 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3054 DAG.getConstant(-BitWidth, AmtVT));
3055 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3056 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3057 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3058 SDValue OutOps[] = { OutLo, OutHi };
3059 return DAG.getMergeValues(OutOps, 2, dl);
3062 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3063 DebugLoc dl = Op.getDebugLoc();
3064 MVT VT = Op.getValueType();
3065 unsigned BitWidth = VT.getSizeInBits();
3066 assert(Op.getNumOperands() == 3 &&
3067 VT == Op.getOperand(1).getValueType() &&
3070 // Expand into a bunch of logical ops, followed by a select_cc.
3071 SDValue Lo = Op.getOperand(0);
3072 SDValue Hi = Op.getOperand(1);
3073 SDValue Amt = Op.getOperand(2);
3074 MVT AmtVT = Amt.getValueType();
3076 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3077 DAG.getConstant(BitWidth, AmtVT), Amt);
3078 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3079 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3080 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3081 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3082 DAG.getConstant(-BitWidth, AmtVT));
3083 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3084 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3085 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3086 Tmp4, Tmp6, ISD::SETLE);
3087 SDValue OutOps[] = { OutLo, OutHi };
3088 return DAG.getMergeValues(OutOps, 2, dl);
3091 //===----------------------------------------------------------------------===//
3092 // Vector related lowering.
3095 /// BuildSplatI - Build a canonical splati of Val with an element size of
3096 /// SplatSize. Cast the result to VT.
3097 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3098 SelectionDAG &DAG, DebugLoc dl) {
3099 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3101 static const MVT VTys[] = { // canonical VT to use for each size.
3102 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3105 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3107 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3111 MVT CanonicalVT = VTys[SplatSize-1];
3113 // Build a canonical splat for this value.
3114 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3115 SmallVector<SDValue, 8> Ops;
3116 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3117 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3118 &Ops[0], Ops.size());
3119 return DAG.getNode(ISD::BIT_CONVERT, dl, ReqVT, Res);
3122 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3123 /// specified intrinsic ID.
3124 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3125 SelectionDAG &DAG, DebugLoc dl,
3126 MVT DestVT = MVT::Other) {
3127 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3128 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3129 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3132 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3133 /// specified intrinsic ID.
3134 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3135 SDValue Op2, SelectionDAG &DAG,
3136 DebugLoc dl, MVT DestVT = MVT::Other) {
3137 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3138 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3139 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3143 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3144 /// amount. The result has the specified value type.
3145 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3146 MVT VT, SelectionDAG &DAG, DebugLoc dl) {
3147 // Force LHS/RHS to be the right type.
3148 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, LHS);
3149 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, RHS);
3152 for (unsigned i = 0; i != 16; ++i)
3154 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3155 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3158 // If this is a case we can't handle, return null and let the default
3159 // expansion code take care of it. If we CAN select this case, and if it
3160 // selects to a single instruction, return Op. Otherwise, if we can codegen
3161 // this case more efficiently than a constant pool load, lower it to the
3162 // sequence of ops that should be used.
3163 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3164 DebugLoc dl = Op.getDebugLoc();
3165 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3166 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3168 // Check if this is a splat of a constant value.
3169 APInt APSplatBits, APSplatUndef;
3170 unsigned SplatBitSize;
3172 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3173 HasAnyUndefs) || SplatBitSize > 32)
3176 unsigned SplatBits = APSplatBits.getZExtValue();
3177 unsigned SplatUndef = APSplatUndef.getZExtValue();
3178 unsigned SplatSize = SplatBitSize / 8;
3180 // First, handle single instruction cases.
3183 if (SplatBits == 0) {
3184 // Canonicalize all zero vectors to be v4i32.
3185 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3186 SDValue Z = DAG.getConstant(0, MVT::i32);
3187 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3188 Op = DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Z);
3193 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3194 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3196 if (SextVal >= -16 && SextVal <= 15)
3197 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3200 // Two instruction sequences.
3202 // If this value is in the range [-32,30] and is even, use:
3203 // tmp = VSPLTI[bhw], result = add tmp, tmp
3204 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3205 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3206 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3207 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3210 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3211 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3213 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3214 // Make -1 and vspltisw -1:
3215 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3217 // Make the VSLW intrinsic, computing 0x8000_0000.
3218 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3221 // xor by OnesV to invert it.
3222 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3223 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3226 // Check to see if this is a wide variety of vsplti*, binop self cases.
3227 static const signed char SplatCsts[] = {
3228 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3229 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3232 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3233 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3234 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3235 int i = SplatCsts[idx];
3237 // Figure out what shift amount will be used by altivec if shifted by i in
3239 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3241 // vsplti + shl self.
3242 if (SextVal == (i << (int)TypeShiftAmt)) {
3243 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3244 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3245 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3246 Intrinsic::ppc_altivec_vslw
3248 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3249 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3252 // vsplti + srl self.
3253 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3254 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3255 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3256 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3257 Intrinsic::ppc_altivec_vsrw
3259 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3260 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3263 // vsplti + sra self.
3264 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3265 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3266 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3267 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3268 Intrinsic::ppc_altivec_vsraw
3270 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3271 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3274 // vsplti + rol self.
3275 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3276 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3277 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3278 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3279 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3280 Intrinsic::ppc_altivec_vrlw
3282 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3283 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3286 // t = vsplti c, result = vsldoi t, t, 1
3287 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3288 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3289 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3291 // t = vsplti c, result = vsldoi t, t, 2
3292 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3293 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3294 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3296 // t = vsplti c, result = vsldoi t, t, 3
3297 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3298 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3299 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3303 // Three instruction sequences.
3305 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3306 if (SextVal >= 0 && SextVal <= 31) {
3307 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3308 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3309 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3310 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3312 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3313 if (SextVal >= -31 && SextVal <= 0) {
3314 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3315 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3316 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3317 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3323 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3324 /// the specified operations to build the shuffle.
3325 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3326 SDValue RHS, SelectionDAG &DAG,
3328 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3329 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3330 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3333 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3345 if (OpNum == OP_COPY) {
3346 if (LHSID == (1*9+2)*9+3) return LHS;
3347 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3351 SDValue OpLHS, OpRHS;
3352 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3353 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3357 default: assert(0 && "Unknown i32 permute!");
3359 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3360 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3361 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3362 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3365 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3366 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3367 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3368 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3371 for (unsigned i = 0; i != 16; ++i)
3372 ShufIdxs[i] = (i&3)+0;
3375 for (unsigned i = 0; i != 16; ++i)
3376 ShufIdxs[i] = (i&3)+4;
3379 for (unsigned i = 0; i != 16; ++i)
3380 ShufIdxs[i] = (i&3)+8;
3383 for (unsigned i = 0; i != 16; ++i)
3384 ShufIdxs[i] = (i&3)+12;
3387 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
3389 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
3391 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
3393 MVT VT = OpLHS.getValueType();
3394 OpLHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpLHS);
3395 OpRHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpRHS);
3396 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
3397 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3400 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3401 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3402 /// return the code it can be lowered into. Worst case, it can always be
3403 /// lowered into a vperm.
3404 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3405 SelectionDAG &DAG) {
3406 DebugLoc dl = Op.getDebugLoc();
3407 SDValue V1 = Op.getOperand(0);
3408 SDValue V2 = Op.getOperand(1);
3409 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3410 MVT VT = Op.getValueType();
3412 // Cases that are handled by instructions that take permute immediates
3413 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3414 // selected by the instruction selector.
3415 if (V2.getOpcode() == ISD::UNDEF) {
3416 if (PPC::isSplatShuffleMask(SVOp, 1) ||
3417 PPC::isSplatShuffleMask(SVOp, 2) ||
3418 PPC::isSplatShuffleMask(SVOp, 4) ||
3419 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
3420 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
3421 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
3422 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
3423 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
3424 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
3425 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
3426 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
3427 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
3432 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3433 // and produce a fixed permutation. If any of these match, do not lower to
3435 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
3436 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
3437 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
3438 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
3439 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
3440 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
3441 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
3442 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
3443 PPC::isVMRGHShuffleMask(SVOp, 4, false))
3446 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3447 // perfect shuffle table to emit an optimal matching sequence.
3448 SmallVector<int, 16> PermMask;
3449 SVOp->getMask(PermMask);
3451 unsigned PFIndexes[4];
3452 bool isFourElementShuffle = true;
3453 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3454 unsigned EltNo = 8; // Start out undef.
3455 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3456 if (PermMask[i*4+j] < 0)
3457 continue; // Undef, ignore it.
3459 unsigned ByteSource = PermMask[i*4+j];
3460 if ((ByteSource & 3) != j) {
3461 isFourElementShuffle = false;
3466 EltNo = ByteSource/4;
3467 } else if (EltNo != ByteSource/4) {
3468 isFourElementShuffle = false;
3472 PFIndexes[i] = EltNo;
3475 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3476 // perfect shuffle vector to determine if it is cost effective to do this as
3477 // discrete instructions, or whether we should use a vperm.
3478 if (isFourElementShuffle) {
3479 // Compute the index in the perfect shuffle table.
3480 unsigned PFTableIndex =
3481 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3483 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3484 unsigned Cost = (PFEntry >> 30);
3486 // Determining when to avoid vperm is tricky. Many things affect the cost
3487 // of vperm, particularly how many times the perm mask needs to be computed.
3488 // For example, if the perm mask can be hoisted out of a loop or is already
3489 // used (perhaps because there are multiple permutes with the same shuffle
3490 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3491 // the loop requires an extra register.
3493 // As a compromise, we only emit discrete instructions if the shuffle can be
3494 // generated in 3 or fewer operations. When we have loop information
3495 // available, if this block is within a loop, we should avoid using vperm
3496 // for 3-operation perms and use a constant pool load instead.
3498 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3501 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3502 // vector that will get spilled to the constant pool.
3503 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3505 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3506 // that it is in input element units, not in bytes. Convert now.
3507 MVT EltVT = V1.getValueType().getVectorElementType();
3508 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
3510 SmallVector<SDValue, 16> ResultMask;
3511 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
3512 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
3514 for (unsigned j = 0; j != BytesPerElement; ++j)
3515 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3519 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
3520 &ResultMask[0], ResultMask.size());
3521 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
3524 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3525 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3526 /// information about the intrinsic.
3527 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
3529 unsigned IntrinsicID =
3530 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
3533 switch (IntrinsicID) {
3534 default: return false;
3535 // Comparison predicates.
3536 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3537 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3538 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3539 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3540 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3541 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3542 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3543 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3544 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3545 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3546 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3547 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3548 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3550 // Normal Comparisons.
3551 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3552 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3553 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3554 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3555 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3556 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3557 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3558 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3559 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3560 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3561 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3562 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3563 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3568 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3569 /// lower, do it, otherwise return null.
3570 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
3571 SelectionDAG &DAG) {
3572 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3573 // opcode number of the comparison.
3574 DebugLoc dl = Op.getDebugLoc();
3577 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3578 return SDValue(); // Don't custom lower most intrinsics.
3580 // If this is a non-dot comparison, make the VCMP node and we are done.
3582 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
3583 Op.getOperand(1), Op.getOperand(2),
3584 DAG.getConstant(CompareOpc, MVT::i32));
3585 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Tmp);
3588 // Create the PPCISD altivec 'dot' comparison node.
3590 Op.getOperand(2), // LHS
3591 Op.getOperand(3), // RHS
3592 DAG.getConstant(CompareOpc, MVT::i32)
3594 std::vector<MVT> VTs;
3595 VTs.push_back(Op.getOperand(2).getValueType());
3596 VTs.push_back(MVT::Flag);
3597 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
3599 // Now that we have the comparison, emit a copy from the CR to a GPR.
3600 // This is flagged to the above dot comparison.
3601 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
3602 DAG.getRegister(PPC::CR6, MVT::i32),
3603 CompNode.getValue(1));
3605 // Unpack the result based on how the target uses it.
3606 unsigned BitNo; // Bit # of CR6.
3607 bool InvertBit; // Invert result?
3608 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
3609 default: // Can't happen, don't crash on invalid number though.
3610 case 0: // Return the value of the EQ bit of CR6.
3611 BitNo = 0; InvertBit = false;
3613 case 1: // Return the inverted value of the EQ bit of CR6.
3614 BitNo = 0; InvertBit = true;
3616 case 2: // Return the value of the LT bit of CR6.
3617 BitNo = 2; InvertBit = false;
3619 case 3: // Return the inverted value of the LT bit of CR6.
3620 BitNo = 2; InvertBit = true;
3624 // Shift the bit into the low position.
3625 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
3626 DAG.getConstant(8-(3-BitNo), MVT::i32));
3628 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
3629 DAG.getConstant(1, MVT::i32));
3631 // If we are supposed to, toggle the bit.
3633 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
3634 DAG.getConstant(1, MVT::i32));
3638 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
3639 SelectionDAG &DAG) {
3640 DebugLoc dl = Op.getDebugLoc();
3641 // Create a stack slot that is 16-byte aligned.
3642 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3643 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3644 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3645 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3647 // Store the input value into Value#0 of the stack slot.
3648 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
3649 Op.getOperand(0), FIdx, NULL, 0);
3651 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, NULL, 0);
3654 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
3655 DebugLoc dl = Op.getDebugLoc();
3656 if (Op.getValueType() == MVT::v4i32) {
3657 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3659 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
3660 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
3662 SDValue RHSSwap = // = vrlw RHS, 16
3663 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
3665 // Shrinkify inputs to v8i16.
3666 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, LHS);
3667 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHS);
3668 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHSSwap);
3670 // Low parts multiplied together, generating 32-bit results (we ignore the
3672 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3673 LHS, RHS, DAG, dl, MVT::v4i32);
3675 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3676 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
3677 // Shift the high parts up 16 bits.
3678 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
3680 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
3681 } else if (Op.getValueType() == MVT::v8i16) {
3682 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3684 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
3686 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3687 LHS, RHS, Zero, DAG, dl);
3688 } else if (Op.getValueType() == MVT::v16i8) {
3689 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3691 // Multiply the even 8-bit parts, producing 16-bit sums.
3692 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3693 LHS, RHS, DAG, dl, MVT::v8i16);
3694 EvenParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, EvenParts);
3696 // Multiply the odd 8-bit parts, producing 16-bit sums.
3697 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3698 LHS, RHS, DAG, dl, MVT::v8i16);
3699 OddParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OddParts);
3701 // Merge the results together.
3703 for (unsigned i = 0; i != 8; ++i) {
3705 Ops[i*2+1] = 2*i+1+16;
3707 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
3709 assert(0 && "Unknown mul to lower!");
3714 /// LowerOperation - Provide custom lowering hooks for some operations.
3716 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3717 switch (Op.getOpcode()) {
3718 default: assert(0 && "Wasn't expecting to be able to lower this!");
3719 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3720 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3721 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3722 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3723 case ISD::SETCC: return LowerSETCC(Op, DAG);
3724 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
3726 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3727 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3730 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3731 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3733 case ISD::FORMAL_ARGUMENTS:
3734 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3735 VarArgsStackOffset, VarArgsNumGPR,
3736 VarArgsNumFPR, PPCSubTarget);
3738 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3739 getTargetMachine());
3740 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3741 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3742 case ISD::DYNAMIC_STACKALLOC:
3743 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3745 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3746 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG,
3748 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3749 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3751 // Lower 64-bit shifts.
3752 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3753 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3754 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3756 // Vector-related lowering.
3757 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3758 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3759 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3760 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3761 case ISD::MUL: return LowerMUL(Op, DAG);
3763 // Frame & Return address.
3764 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3765 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3770 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
3771 SmallVectorImpl<SDValue>&Results,
3772 SelectionDAG &DAG) {
3773 DebugLoc dl = N->getDebugLoc();
3774 switch (N->getOpcode()) {
3776 assert(false && "Do not know how to custom type legalize this operation!");
3778 case ISD::FP_ROUND_INREG: {
3779 assert(N->getValueType(0) == MVT::ppcf128);
3780 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
3781 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
3782 MVT::f64, N->getOperand(0),
3783 DAG.getIntPtrConstant(0));
3784 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
3785 MVT::f64, N->getOperand(0),
3786 DAG.getIntPtrConstant(1));
3788 // This sequence changes FPSCR to do round-to-zero, adds the two halves
3789 // of the long double, and puts FPSCR back the way it was. We do not
3790 // actually model FPSCR.
3791 std::vector<MVT> NodeTys;
3792 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
3794 NodeTys.push_back(MVT::f64); // Return register
3795 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
3796 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3797 MFFSreg = Result.getValue(0);
3798 InFlag = Result.getValue(1);
3801 NodeTys.push_back(MVT::Flag); // Returns a flag
3802 Ops[0] = DAG.getConstant(31, MVT::i32);
3804 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
3805 InFlag = Result.getValue(0);
3808 NodeTys.push_back(MVT::Flag); // Returns a flag
3809 Ops[0] = DAG.getConstant(30, MVT::i32);
3811 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
3812 InFlag = Result.getValue(0);
3815 NodeTys.push_back(MVT::f64); // result of add
3816 NodeTys.push_back(MVT::Flag); // Returns a flag
3820 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
3821 FPreg = Result.getValue(0);
3822 InFlag = Result.getValue(1);
3825 NodeTys.push_back(MVT::f64);
3826 Ops[0] = DAG.getConstant(1, MVT::i32);
3830 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
3831 FPreg = Result.getValue(0);
3833 // We know the low half is about to be thrown away, so just use something
3835 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
3839 case ISD::FP_TO_SINT:
3840 Results.push_back(LowerFP_TO_SINT(SDValue(N, 0), DAG, dl));
3846 //===----------------------------------------------------------------------===//
3847 // Other Lowering Code
3848 //===----------------------------------------------------------------------===//
3851 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3852 bool is64bit, unsigned BinOpcode) const {
3853 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3854 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3856 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3857 MachineFunction *F = BB->getParent();
3858 MachineFunction::iterator It = BB;
3861 unsigned dest = MI->getOperand(0).getReg();
3862 unsigned ptrA = MI->getOperand(1).getReg();
3863 unsigned ptrB = MI->getOperand(2).getReg();
3864 unsigned incr = MI->getOperand(3).getReg();
3865 DebugLoc dl = MI->getDebugLoc();
3867 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3868 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3869 F->insert(It, loopMBB);
3870 F->insert(It, exitMBB);
3871 exitMBB->transferSuccessors(BB);
3873 MachineRegisterInfo &RegInfo = F->getRegInfo();
3874 unsigned TmpReg = (!BinOpcode) ? incr :
3875 RegInfo.createVirtualRegister(
3876 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3877 (const TargetRegisterClass *) &PPC::GPRCRegClass);
3881 // fallthrough --> loopMBB
3882 BB->addSuccessor(loopMBB);
3885 // l[wd]arx dest, ptr
3886 // add r0, dest, incr
3887 // st[wd]cx. r0, ptr
3889 // fallthrough --> exitMBB
3891 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
3892 .addReg(ptrA).addReg(ptrB);
3894 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
3895 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
3896 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
3897 BuildMI(BB, dl, TII->get(PPC::BCC))
3898 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
3899 BB->addSuccessor(loopMBB);
3900 BB->addSuccessor(exitMBB);
3909 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
3910 MachineBasicBlock *BB,
3911 bool is8bit, // operation
3912 unsigned BinOpcode) const {
3913 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3914 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3915 // In 64 bit mode we have to use 64 bits for addresses, even though the
3916 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
3917 // registers without caring whether they're 32 or 64, but here we're
3918 // doing actual arithmetic on the addresses.
3919 bool is64bit = PPCSubTarget.isPPC64();
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();
3930 DebugLoc dl = MI->getDebugLoc();
3932 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3933 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3934 F->insert(It, loopMBB);
3935 F->insert(It, exitMBB);
3936 exitMBB->transferSuccessors(BB);
3938 MachineRegisterInfo &RegInfo = F->getRegInfo();
3939 const TargetRegisterClass *RC =
3940 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3941 (const TargetRegisterClass *) &PPC::GPRCRegClass;
3942 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
3943 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
3944 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
3945 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
3946 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
3947 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
3948 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
3949 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
3950 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
3951 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
3952 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
3954 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
3958 // fallthrough --> loopMBB
3959 BB->addSuccessor(loopMBB);
3961 // The 4-byte load must be aligned, while a char or short may be
3962 // anywhere in the word. Hence all this nasty bookkeeping code.
3963 // add ptr1, ptrA, ptrB [copy if ptrA==0]
3964 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
3965 // xori shift, shift1, 24 [16]
3966 // rlwinm ptr, ptr1, 0, 0, 29
3967 // slw incr2, incr, shift
3968 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
3969 // slw mask, mask2, shift
3971 // lwarx tmpDest, ptr
3972 // add tmp, tmpDest, incr2
3973 // andc tmp2, tmpDest, mask
3974 // and tmp3, tmp, mask
3975 // or tmp4, tmp3, tmp2
3978 // fallthrough --> exitMBB
3979 // srw dest, tmpDest, shift
3981 if (ptrA!=PPC::R0) {
3982 Ptr1Reg = RegInfo.createVirtualRegister(RC);
3983 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
3984 .addReg(ptrA).addReg(ptrB);
3988 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
3989 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
3990 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
3991 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
3993 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
3994 .addReg(Ptr1Reg).addImm(0).addImm(61);
3996 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
3997 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
3998 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
3999 .addReg(incr).addReg(ShiftReg);
4001 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4003 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4004 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4006 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4007 .addReg(Mask2Reg).addReg(ShiftReg);
4010 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4011 .addReg(PPC::R0).addReg(PtrReg);
4013 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4014 .addReg(Incr2Reg).addReg(TmpDestReg);
4015 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4016 .addReg(TmpDestReg).addReg(MaskReg);
4017 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4018 .addReg(TmpReg).addReg(MaskReg);
4019 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4020 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4021 BuildMI(BB, dl, TII->get(PPC::STWCX))
4022 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4023 BuildMI(BB, dl, TII->get(PPC::BCC))
4024 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4025 BB->addSuccessor(loopMBB);
4026 BB->addSuccessor(exitMBB);
4031 BuildMI(BB, dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4036 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4037 MachineBasicBlock *BB) const {
4038 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4040 // To "insert" these instructions we actually have to insert their
4041 // control-flow patterns.
4042 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4043 MachineFunction::iterator It = BB;
4046 MachineFunction *F = BB->getParent();
4048 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4049 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4050 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4051 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4052 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4054 // The incoming instruction knows the destination vreg to set, the
4055 // condition code register to branch on, the true/false values to
4056 // select between, and a branch opcode to use.
4061 // cmpTY ccX, r1, r2
4063 // fallthrough --> copy0MBB
4064 MachineBasicBlock *thisMBB = BB;
4065 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4066 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4067 unsigned SelectPred = MI->getOperand(4).getImm();
4068 DebugLoc dl = MI->getDebugLoc();
4069 BuildMI(BB, dl, TII->get(PPC::BCC))
4070 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4071 F->insert(It, copy0MBB);
4072 F->insert(It, sinkMBB);
4073 // Update machine-CFG edges by transferring all successors of the current
4074 // block to the new block which will contain the Phi node for the select.
4075 sinkMBB->transferSuccessors(BB);
4076 // Next, add the true and fallthrough blocks as its successors.
4077 BB->addSuccessor(copy0MBB);
4078 BB->addSuccessor(sinkMBB);
4081 // %FalseValue = ...
4082 // # fallthrough to sinkMBB
4085 // Update machine-CFG edges
4086 BB->addSuccessor(sinkMBB);
4089 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4092 BuildMI(BB, dl, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4093 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4094 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4096 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4097 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4098 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4099 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4100 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4101 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4102 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4103 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4105 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4106 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4107 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4108 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4109 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4110 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4111 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4112 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4114 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4115 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4116 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4117 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4118 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4119 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4120 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4121 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4123 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4124 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4125 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4126 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4127 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4128 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4129 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4130 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4132 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4133 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4134 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4135 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4136 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4137 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4138 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4139 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4141 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4142 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4143 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4144 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4145 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4146 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4147 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4148 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4150 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4151 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4152 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4153 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4154 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4155 BB = EmitAtomicBinary(MI, BB, false, 0);
4156 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4157 BB = EmitAtomicBinary(MI, BB, true, 0);
4159 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4160 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4161 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4163 unsigned dest = MI->getOperand(0).getReg();
4164 unsigned ptrA = MI->getOperand(1).getReg();
4165 unsigned ptrB = MI->getOperand(2).getReg();
4166 unsigned oldval = MI->getOperand(3).getReg();
4167 unsigned newval = MI->getOperand(4).getReg();
4168 DebugLoc dl = MI->getDebugLoc();
4170 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4171 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4172 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4173 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4174 F->insert(It, loop1MBB);
4175 F->insert(It, loop2MBB);
4176 F->insert(It, midMBB);
4177 F->insert(It, exitMBB);
4178 exitMBB->transferSuccessors(BB);
4182 // fallthrough --> loopMBB
4183 BB->addSuccessor(loop1MBB);
4186 // l[wd]arx dest, ptr
4187 // cmp[wd] dest, oldval
4190 // st[wd]cx. newval, ptr
4194 // st[wd]cx. dest, ptr
4197 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4198 .addReg(ptrA).addReg(ptrB);
4199 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4200 .addReg(oldval).addReg(dest);
4201 BuildMI(BB, dl, TII->get(PPC::BCC))
4202 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4203 BB->addSuccessor(loop2MBB);
4204 BB->addSuccessor(midMBB);
4207 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4208 .addReg(newval).addReg(ptrA).addReg(ptrB);
4209 BuildMI(BB, dl, TII->get(PPC::BCC))
4210 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4211 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4212 BB->addSuccessor(loop1MBB);
4213 BB->addSuccessor(exitMBB);
4216 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4217 .addReg(dest).addReg(ptrA).addReg(ptrB);
4218 BB->addSuccessor(exitMBB);
4223 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4224 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4225 // We must use 64-bit registers for addresses when targeting 64-bit,
4226 // since we're actually doing arithmetic on them. Other registers
4228 bool is64bit = PPCSubTarget.isPPC64();
4229 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4231 unsigned dest = MI->getOperand(0).getReg();
4232 unsigned ptrA = MI->getOperand(1).getReg();
4233 unsigned ptrB = MI->getOperand(2).getReg();
4234 unsigned oldval = MI->getOperand(3).getReg();
4235 unsigned newval = MI->getOperand(4).getReg();
4236 DebugLoc dl = MI->getDebugLoc();
4238 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4239 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4240 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4241 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4242 F->insert(It, loop1MBB);
4243 F->insert(It, loop2MBB);
4244 F->insert(It, midMBB);
4245 F->insert(It, exitMBB);
4246 exitMBB->transferSuccessors(BB);
4248 MachineRegisterInfo &RegInfo = F->getRegInfo();
4249 const TargetRegisterClass *RC =
4250 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4251 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4252 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4253 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4254 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4255 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4256 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4257 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4258 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4259 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4260 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4261 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4262 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4263 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4264 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4266 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4269 // fallthrough --> loopMBB
4270 BB->addSuccessor(loop1MBB);
4272 // The 4-byte load must be aligned, while a char or short may be
4273 // anywhere in the word. Hence all this nasty bookkeeping code.
4274 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4275 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4276 // xori shift, shift1, 24 [16]
4277 // rlwinm ptr, ptr1, 0, 0, 29
4278 // slw newval2, newval, shift
4279 // slw oldval2, oldval,shift
4280 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4281 // slw mask, mask2, shift
4282 // and newval3, newval2, mask
4283 // and oldval3, oldval2, mask
4285 // lwarx tmpDest, ptr
4286 // and tmp, tmpDest, mask
4287 // cmpw tmp, oldval3
4290 // andc tmp2, tmpDest, mask
4291 // or tmp4, tmp2, newval3
4296 // stwcx. tmpDest, ptr
4298 // srw dest, tmpDest, shift
4299 if (ptrA!=PPC::R0) {
4300 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4301 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4302 .addReg(ptrA).addReg(ptrB);
4306 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4307 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4308 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4309 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4311 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4312 .addReg(Ptr1Reg).addImm(0).addImm(61);
4314 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4315 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4316 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4317 .addReg(newval).addReg(ShiftReg);
4318 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4319 .addReg(oldval).addReg(ShiftReg);
4321 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4323 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4324 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
4325 .addReg(Mask3Reg).addImm(65535);
4327 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4328 .addReg(Mask2Reg).addReg(ShiftReg);
4329 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
4330 .addReg(NewVal2Reg).addReg(MaskReg);
4331 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
4332 .addReg(OldVal2Reg).addReg(MaskReg);
4335 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4336 .addReg(PPC::R0).addReg(PtrReg);
4337 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
4338 .addReg(TmpDestReg).addReg(MaskReg);
4339 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
4340 .addReg(TmpReg).addReg(OldVal3Reg);
4341 BuildMI(BB, dl, TII->get(PPC::BCC))
4342 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4343 BB->addSuccessor(loop2MBB);
4344 BB->addSuccessor(midMBB);
4347 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
4348 .addReg(TmpDestReg).addReg(MaskReg);
4349 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
4350 .addReg(Tmp2Reg).addReg(NewVal3Reg);
4351 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4352 .addReg(PPC::R0).addReg(PtrReg);
4353 BuildMI(BB, dl, TII->get(PPC::BCC))
4354 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4355 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4356 BB->addSuccessor(loop1MBB);
4357 BB->addSuccessor(exitMBB);
4360 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4361 .addReg(PPC::R0).addReg(PtrReg);
4362 BB->addSuccessor(exitMBB);
4367 BuildMI(BB, dl, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4369 assert(0 && "Unexpected instr type to insert");
4372 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4376 //===----------------------------------------------------------------------===//
4377 // Target Optimization Hooks
4378 //===----------------------------------------------------------------------===//
4380 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4381 DAGCombinerInfo &DCI) const {
4382 TargetMachine &TM = getTargetMachine();
4383 SelectionDAG &DAG = DCI.DAG;
4384 DebugLoc dl = N->getDebugLoc();
4385 switch (N->getOpcode()) {
4388 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4389 if (C->getZExtValue() == 0) // 0 << V -> 0.
4390 return N->getOperand(0);
4394 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4395 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4396 return N->getOperand(0);
4400 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4401 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4402 C->isAllOnesValue()) // -1 >>s V -> -1.
4403 return N->getOperand(0);
4407 case ISD::SINT_TO_FP:
4408 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4409 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4410 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4411 // We allow the src/dst to be either f32/f64, but the intermediate
4412 // type must be i64.
4413 if (N->getOperand(0).getValueType() == MVT::i64 &&
4414 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4415 SDValue Val = N->getOperand(0).getOperand(0);
4416 if (Val.getValueType() == MVT::f32) {
4417 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4418 DCI.AddToWorklist(Val.getNode());
4421 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
4422 DCI.AddToWorklist(Val.getNode());
4423 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
4424 DCI.AddToWorklist(Val.getNode());
4425 if (N->getValueType(0) == MVT::f32) {
4426 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
4427 DAG.getIntPtrConstant(0));
4428 DCI.AddToWorklist(Val.getNode());
4431 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4432 // If the intermediate type is i32, we can avoid the load/store here
4439 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4440 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4441 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4442 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4443 N->getOperand(1).getValueType() == MVT::i32 &&
4444 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4445 SDValue Val = N->getOperand(1).getOperand(0);
4446 if (Val.getValueType() == MVT::f32) {
4447 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4448 DCI.AddToWorklist(Val.getNode());
4450 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
4451 DCI.AddToWorklist(Val.getNode());
4453 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
4454 N->getOperand(2), N->getOperand(3));
4455 DCI.AddToWorklist(Val.getNode());
4459 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4460 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
4461 N->getOperand(1).getNode()->hasOneUse() &&
4462 (N->getOperand(1).getValueType() == MVT::i32 ||
4463 N->getOperand(1).getValueType() == MVT::i16)) {
4464 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4465 // Do an any-extend to 32-bits if this is a half-word input.
4466 if (BSwapOp.getValueType() == MVT::i16)
4467 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
4469 return DAG.getNode(PPCISD::STBRX, dl, MVT::Other, N->getOperand(0),
4470 BSwapOp, N->getOperand(2), N->getOperand(3),
4471 DAG.getValueType(N->getOperand(1).getValueType()));
4475 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4476 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4477 N->getOperand(0).hasOneUse() &&
4478 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4479 SDValue Load = N->getOperand(0);
4480 LoadSDNode *LD = cast<LoadSDNode>(Load);
4481 // Create the byte-swapping load.
4482 std::vector<MVT> VTs;
4483 VTs.push_back(MVT::i32);
4484 VTs.push_back(MVT::Other);
4485 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
4487 LD->getChain(), // Chain
4488 LD->getBasePtr(), // Ptr
4490 DAG.getValueType(N->getValueType(0)) // VT
4492 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, dl, VTs, Ops, 4);
4494 // If this is an i16 load, insert the truncate.
4495 SDValue ResVal = BSLoad;
4496 if (N->getValueType(0) == MVT::i16)
4497 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
4499 // First, combine the bswap away. This makes the value produced by the
4501 DCI.CombineTo(N, ResVal);
4503 // Next, combine the load away, we give it a bogus result value but a real
4504 // chain result. The result value is dead because the bswap is dead.
4505 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
4507 // Return N so it doesn't get rechecked!
4508 return SDValue(N, 0);
4512 case PPCISD::VCMP: {
4513 // If a VCMPo node already exists with exactly the same operands as this
4514 // node, use its result instead of this node (VCMPo computes both a CR6 and
4515 // a normal output).
4517 if (!N->getOperand(0).hasOneUse() &&
4518 !N->getOperand(1).hasOneUse() &&
4519 !N->getOperand(2).hasOneUse()) {
4521 // Scan all of the users of the LHS, looking for VCMPo's that match.
4522 SDNode *VCMPoNode = 0;
4524 SDNode *LHSN = N->getOperand(0).getNode();
4525 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
4527 if (UI->getOpcode() == PPCISD::VCMPo &&
4528 UI->getOperand(1) == N->getOperand(1) &&
4529 UI->getOperand(2) == N->getOperand(2) &&
4530 UI->getOperand(0) == N->getOperand(0)) {
4535 // If there is no VCMPo node, or if the flag value has a single use, don't
4537 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
4540 // Look at the (necessarily single) use of the flag value. If it has a
4541 // chain, this transformation is more complex. Note that multiple things
4542 // could use the value result, which we should ignore.
4543 SDNode *FlagUser = 0;
4544 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
4545 FlagUser == 0; ++UI) {
4546 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
4548 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
4549 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
4556 // If the user is a MFCR instruction, we know this is safe. Otherwise we
4557 // give up for right now.
4558 if (FlagUser->getOpcode() == PPCISD::MFCR)
4559 return SDValue(VCMPoNode, 0);
4564 // If this is a branch on an altivec predicate comparison, lower this so
4565 // that we don't have to do a MFCR: instead, branch directly on CR6. This
4566 // lowering is done pre-legalize, because the legalizer lowers the predicate
4567 // compare down to code that is difficult to reassemble.
4568 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
4569 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
4573 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
4574 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
4575 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
4576 assert(isDot && "Can't compare against a vector result!");
4578 // If this is a comparison against something other than 0/1, then we know
4579 // that the condition is never/always true.
4580 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
4581 if (Val != 0 && Val != 1) {
4582 if (CC == ISD::SETEQ) // Cond never true, remove branch.
4583 return N->getOperand(0);
4584 // Always !=, turn it into an unconditional branch.
4585 return DAG.getNode(ISD::BR, dl, MVT::Other,
4586 N->getOperand(0), N->getOperand(4));
4589 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
4591 // Create the PPCISD altivec 'dot' comparison node.
4592 std::vector<MVT> VTs;
4594 LHS.getOperand(2), // LHS of compare
4595 LHS.getOperand(3), // RHS of compare
4596 DAG.getConstant(CompareOpc, MVT::i32)
4598 VTs.push_back(LHS.getOperand(2).getValueType());
4599 VTs.push_back(MVT::Flag);
4600 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4602 // Unpack the result based on how the target uses it.
4603 PPC::Predicate CompOpc;
4604 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
4605 default: // Can't happen, don't crash on invalid number though.
4606 case 0: // Branch on the value of the EQ bit of CR6.
4607 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
4609 case 1: // Branch on the inverted value of the EQ bit of CR6.
4610 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
4612 case 2: // Branch on the value of the LT bit of CR6.
4613 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
4615 case 3: // Branch on the inverted value of the LT bit of CR6.
4616 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
4620 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
4621 DAG.getConstant(CompOpc, MVT::i32),
4622 DAG.getRegister(PPC::CR6, MVT::i32),
4623 N->getOperand(4), CompNode.getValue(1));
4632 //===----------------------------------------------------------------------===//
4633 // Inline Assembly Support
4634 //===----------------------------------------------------------------------===//
4636 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4640 const SelectionDAG &DAG,
4641 unsigned Depth) const {
4642 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4643 switch (Op.getOpcode()) {
4645 case PPCISD::LBRX: {
4646 // lhbrx is known to have the top bits cleared out.
4647 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
4648 KnownZero = 0xFFFF0000;
4651 case ISD::INTRINSIC_WO_CHAIN: {
4652 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
4654 case Intrinsic::ppc_altivec_vcmpbfp_p:
4655 case Intrinsic::ppc_altivec_vcmpeqfp_p:
4656 case Intrinsic::ppc_altivec_vcmpequb_p:
4657 case Intrinsic::ppc_altivec_vcmpequh_p:
4658 case Intrinsic::ppc_altivec_vcmpequw_p:
4659 case Intrinsic::ppc_altivec_vcmpgefp_p:
4660 case Intrinsic::ppc_altivec_vcmpgtfp_p:
4661 case Intrinsic::ppc_altivec_vcmpgtsb_p:
4662 case Intrinsic::ppc_altivec_vcmpgtsh_p:
4663 case Intrinsic::ppc_altivec_vcmpgtsw_p:
4664 case Intrinsic::ppc_altivec_vcmpgtub_p:
4665 case Intrinsic::ppc_altivec_vcmpgtuh_p:
4666 case Intrinsic::ppc_altivec_vcmpgtuw_p:
4667 KnownZero = ~1U; // All bits but the low one are known to be zero.
4675 /// getConstraintType - Given a constraint, return the type of
4676 /// constraint it is for this target.
4677 PPCTargetLowering::ConstraintType
4678 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
4679 if (Constraint.size() == 1) {
4680 switch (Constraint[0]) {
4687 return C_RegisterClass;
4690 return TargetLowering::getConstraintType(Constraint);
4693 std::pair<unsigned, const TargetRegisterClass*>
4694 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4696 if (Constraint.size() == 1) {
4697 // GCC RS6000 Constraint Letters
4698 switch (Constraint[0]) {
4701 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
4702 return std::make_pair(0U, PPC::G8RCRegisterClass);
4703 return std::make_pair(0U, PPC::GPRCRegisterClass);
4706 return std::make_pair(0U, PPC::F4RCRegisterClass);
4707 else if (VT == MVT::f64)
4708 return std::make_pair(0U, PPC::F8RCRegisterClass);
4711 return std::make_pair(0U, PPC::VRRCRegisterClass);
4713 return std::make_pair(0U, PPC::CRRCRegisterClass);
4717 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4721 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4722 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is true
4723 /// it means one of the asm constraint of the inline asm instruction being
4724 /// processed is 'm'.
4725 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
4727 std::vector<SDValue>&Ops,
4728 SelectionDAG &DAG) const {
4729 SDValue Result(0,0);
4740 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
4741 if (!CST) return; // Must be an immediate to match.
4742 unsigned Value = CST->getZExtValue();
4744 default: assert(0 && "Unknown constraint letter!");
4745 case 'I': // "I" is a signed 16-bit constant.
4746 if ((short)Value == (int)Value)
4747 Result = DAG.getTargetConstant(Value, Op.getValueType());
4749 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
4750 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
4751 if ((short)Value == 0)
4752 Result = DAG.getTargetConstant(Value, Op.getValueType());
4754 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
4755 if ((Value >> 16) == 0)
4756 Result = DAG.getTargetConstant(Value, Op.getValueType());
4758 case 'M': // "M" is a constant that is greater than 31.
4760 Result = DAG.getTargetConstant(Value, Op.getValueType());
4762 case 'N': // "N" is a positive constant that is an exact power of two.
4763 if ((int)Value > 0 && isPowerOf2_32(Value))
4764 Result = DAG.getTargetConstant(Value, Op.getValueType());
4766 case 'O': // "O" is the constant zero.
4768 Result = DAG.getTargetConstant(Value, Op.getValueType());
4770 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
4771 if ((short)-Value == (int)-Value)
4772 Result = DAG.getTargetConstant(Value, Op.getValueType());
4779 if (Result.getNode()) {
4780 Ops.push_back(Result);
4784 // Handle standard constraint letters.
4785 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, hasMemory, Ops, DAG);
4788 // isLegalAddressingMode - Return true if the addressing mode represented
4789 // by AM is legal for this target, for a load/store of the specified type.
4790 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4791 const Type *Ty) const {
4792 // FIXME: PPC does not allow r+i addressing modes for vectors!
4794 // PPC allows a sign-extended 16-bit immediate field.
4795 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
4798 // No global is ever allowed as a base.
4802 // PPC only support r+r,
4804 case 0: // "r+i" or just "i", depending on HasBaseReg.
4807 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
4809 // Otherwise we have r+r or r+i.
4812 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
4814 // Allow 2*r as r+r.
4817 // No other scales are supported.
4824 /// isLegalAddressImmediate - Return true if the integer value can be used
4825 /// as the offset of the target addressing mode for load / store of the
4827 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
4828 // PPC allows a sign-extended 16-bit immediate field.
4829 return (V > -(1 << 16) && V < (1 << 16)-1);
4832 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
4836 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
4837 DebugLoc dl = Op.getDebugLoc();
4838 // Depths > 0 not supported yet!
4839 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4842 MachineFunction &MF = DAG.getMachineFunction();
4843 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4845 // Just load the return address off the stack.
4846 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
4848 // Make sure the function really does not optimize away the store of the RA
4850 FuncInfo->setLRStoreRequired();
4851 return DAG.getLoad(getPointerTy(), dl,
4852 DAG.getEntryNode(), RetAddrFI, NULL, 0);
4855 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
4856 DebugLoc dl = Op.getDebugLoc();
4857 // Depths > 0 not supported yet!
4858 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4861 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4862 bool isPPC64 = PtrVT == MVT::i64;
4864 MachineFunction &MF = DAG.getMachineFunction();
4865 MachineFrameInfo *MFI = MF.getFrameInfo();
4866 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4867 && MFI->getStackSize();
4870 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::X31 : PPC::X1,
4873 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::R31 : PPC::R1,
4878 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4879 // The PowerPC target isn't yet aware of offsets.