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"
37 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
38 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
41 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
42 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
46 // Use _setjmp/_longjmp instead of setjmp/longjmp.
47 setUseUnderscoreSetJmp(true);
48 setUseUnderscoreLongJmp(true);
50 // Set up the register classes.
51 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
52 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
53 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
55 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
56 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
57 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
59 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
61 // PowerPC has pre-inc load and store's.
62 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
67 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
73 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
74 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
75 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
76 // This is used in the ppcf128->int sequence. Note it has different semantics
77 // from FP_ROUND: that rounds to nearest, this rounds to zero.
78 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
80 // PowerPC has no SREM/UREM instructions
81 setOperationAction(ISD::SREM, MVT::i32, Expand);
82 setOperationAction(ISD::UREM, MVT::i32, Expand);
83 setOperationAction(ISD::SREM, MVT::i64, Expand);
84 setOperationAction(ISD::UREM, MVT::i64, Expand);
86 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
87 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
88 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
89 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
90 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
91 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
92 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
93 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
94 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
96 // We don't support sin/cos/sqrt/fmod/pow
97 setOperationAction(ISD::FSIN , MVT::f64, Expand);
98 setOperationAction(ISD::FCOS , MVT::f64, Expand);
99 setOperationAction(ISD::FREM , MVT::f64, Expand);
100 setOperationAction(ISD::FPOW , MVT::f64, Expand);
101 setOperationAction(ISD::FSIN , MVT::f32, Expand);
102 setOperationAction(ISD::FCOS , MVT::f32, Expand);
103 setOperationAction(ISD::FREM , MVT::f32, Expand);
104 setOperationAction(ISD::FPOW , MVT::f32, Expand);
106 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
108 // If we're enabling GP optimizations, use hardware square root
109 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
110 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
111 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
114 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
115 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
117 // PowerPC does not have BSWAP, CTPOP or CTTZ
118 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
119 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
120 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
121 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
122 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
123 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
125 // PowerPC does not have ROTR
126 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
127 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
129 // PowerPC does not have Select
130 setOperationAction(ISD::SELECT, MVT::i32, Expand);
131 setOperationAction(ISD::SELECT, MVT::i64, Expand);
132 setOperationAction(ISD::SELECT, MVT::f32, Expand);
133 setOperationAction(ISD::SELECT, MVT::f64, Expand);
135 // PowerPC wants to turn select_cc of FP into fsel when possible.
136 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
137 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
139 // PowerPC wants to optimize integer setcc a bit
140 setOperationAction(ISD::SETCC, MVT::i32, Custom);
142 // PowerPC does not have BRCOND which requires SetCC
143 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
145 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
147 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
148 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
150 // PowerPC does not have [U|S]INT_TO_FP
151 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
152 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
154 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
155 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
156 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
157 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
159 // We cannot sextinreg(i1). Expand to shifts.
160 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
162 // Support label based line numbers.
163 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
164 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
166 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
167 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
168 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
169 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
172 // We want to legalize GlobalAddress and ConstantPool nodes into the
173 // appropriate instructions to materialize the address.
174 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
175 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
176 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
177 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
178 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
179 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
180 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
181 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
183 // RET must be custom lowered, to meet ABI requirements.
184 setOperationAction(ISD::RET , MVT::Other, Custom);
187 setOperationAction(ISD::TRAP, MVT::Other, Legal);
189 // TRAMPOLINE is custom lowered.
190 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
192 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
193 setOperationAction(ISD::VASTART , MVT::Other, Custom);
195 // VAARG is custom lowered with ELF 32 ABI
196 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
197 setOperationAction(ISD::VAARG, MVT::Other, Custom);
199 setOperationAction(ISD::VAARG, MVT::Other, Expand);
201 // Use the default implementation.
202 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
203 setOperationAction(ISD::VAEND , MVT::Other, Expand);
204 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
205 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
206 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
209 // We want to custom lower some of our intrinsics.
210 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
212 // Comparisons that require checking two conditions.
213 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
214 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
215 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
216 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
217 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
218 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
219 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
220 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
221 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
222 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
223 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
224 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
226 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
227 // They also have instructions for converting between i64 and fp.
228 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
229 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
230 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
231 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
232 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
234 // FIXME: disable this lowered code. This generates 64-bit register values,
235 // and we don't model the fact that the top part is clobbered by calls. We
236 // need to flag these together so that the value isn't live across a call.
237 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
239 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
240 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
242 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
243 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
246 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
247 // 64-bit PowerPC implementations can support i64 types directly
248 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
249 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
250 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
251 // 64-bit PowerPC wants to expand i128 shifts itself.
252 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
253 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
254 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
256 // 32-bit PowerPC wants to expand i64 shifts itself.
257 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
258 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
259 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
262 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
263 // First set operation action for all vector types to expand. Then we
264 // will selectively turn on ones that can be effectively codegen'd.
265 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
266 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
267 MVT VT = (MVT::SimpleValueType)i;
269 // add/sub are legal for all supported vector VT's.
270 setOperationAction(ISD::ADD , VT, Legal);
271 setOperationAction(ISD::SUB , VT, Legal);
273 // We promote all shuffles to v16i8.
274 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
275 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
277 // We promote all non-typed operations to v4i32.
278 setOperationAction(ISD::AND , VT, Promote);
279 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
280 setOperationAction(ISD::OR , VT, Promote);
281 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
282 setOperationAction(ISD::XOR , VT, Promote);
283 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
284 setOperationAction(ISD::LOAD , VT, Promote);
285 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
286 setOperationAction(ISD::SELECT, VT, Promote);
287 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
288 setOperationAction(ISD::STORE, VT, Promote);
289 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
291 // No other operations are legal.
292 setOperationAction(ISD::MUL , VT, Expand);
293 setOperationAction(ISD::SDIV, VT, Expand);
294 setOperationAction(ISD::SREM, VT, Expand);
295 setOperationAction(ISD::UDIV, VT, Expand);
296 setOperationAction(ISD::UREM, VT, Expand);
297 setOperationAction(ISD::FDIV, VT, Expand);
298 setOperationAction(ISD::FNEG, VT, Expand);
299 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
300 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
301 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
302 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
303 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
304 setOperationAction(ISD::UDIVREM, VT, Expand);
305 setOperationAction(ISD::SDIVREM, VT, Expand);
306 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
307 setOperationAction(ISD::FPOW, VT, Expand);
308 setOperationAction(ISD::CTPOP, VT, Expand);
309 setOperationAction(ISD::CTLZ, VT, Expand);
310 setOperationAction(ISD::CTTZ, VT, Expand);
313 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
314 // with merges, splats, etc.
315 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
317 setOperationAction(ISD::AND , MVT::v4i32, Legal);
318 setOperationAction(ISD::OR , MVT::v4i32, Legal);
319 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
320 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
321 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
322 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
324 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
325 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
326 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
327 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
329 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
330 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
331 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
332 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
334 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
335 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
337 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
338 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
339 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
340 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
343 setShiftAmountType(MVT::i32);
344 setBooleanContents(ZeroOrOneBooleanContent);
346 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
347 setStackPointerRegisterToSaveRestore(PPC::X1);
348 setExceptionPointerRegister(PPC::X3);
349 setExceptionSelectorRegister(PPC::X4);
351 setStackPointerRegisterToSaveRestore(PPC::R1);
352 setExceptionPointerRegister(PPC::R3);
353 setExceptionSelectorRegister(PPC::R4);
356 // We have target-specific dag combine patterns for the following nodes:
357 setTargetDAGCombine(ISD::SINT_TO_FP);
358 setTargetDAGCombine(ISD::STORE);
359 setTargetDAGCombine(ISD::BR_CC);
360 setTargetDAGCombine(ISD::BSWAP);
362 // Darwin long double math library functions have $LDBL128 appended.
363 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
364 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
365 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
366 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
367 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
368 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
369 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
370 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
371 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
372 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
373 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
376 computeRegisterProperties();
379 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
380 /// function arguments in the caller parameter area.
381 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
382 TargetMachine &TM = getTargetMachine();
383 // Darwin passes everything on 4 byte boundary.
384 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
390 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
393 case PPCISD::FSEL: return "PPCISD::FSEL";
394 case PPCISD::FCFID: return "PPCISD::FCFID";
395 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
396 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
397 case PPCISD::STFIWX: return "PPCISD::STFIWX";
398 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
399 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
400 case PPCISD::VPERM: return "PPCISD::VPERM";
401 case PPCISD::Hi: return "PPCISD::Hi";
402 case PPCISD::Lo: return "PPCISD::Lo";
403 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
404 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
405 case PPCISD::SRL: return "PPCISD::SRL";
406 case PPCISD::SRA: return "PPCISD::SRA";
407 case PPCISD::SHL: return "PPCISD::SHL";
408 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
409 case PPCISD::STD_32: return "PPCISD::STD_32";
410 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
411 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
412 case PPCISD::MTCTR: return "PPCISD::MTCTR";
413 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
414 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
415 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
416 case PPCISD::MFCR: return "PPCISD::MFCR";
417 case PPCISD::VCMP: return "PPCISD::VCMP";
418 case PPCISD::VCMPo: return "PPCISD::VCMPo";
419 case PPCISD::LBRX: return "PPCISD::LBRX";
420 case PPCISD::STBRX: return "PPCISD::STBRX";
421 case PPCISD::LARX: return "PPCISD::LARX";
422 case PPCISD::STCX: return "PPCISD::STCX";
423 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
424 case PPCISD::MFFS: return "PPCISD::MFFS";
425 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
426 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
427 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
428 case PPCISD::MTFSF: return "PPCISD::MTFSF";
429 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
430 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
435 MVT PPCTargetLowering::getSetCCResultType(MVT VT) const {
440 //===----------------------------------------------------------------------===//
441 // Node matching predicates, for use by the tblgen matching code.
442 //===----------------------------------------------------------------------===//
444 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
445 static bool isFloatingPointZero(SDValue Op) {
446 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
447 return CFP->getValueAPF().isZero();
448 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
449 // Maybe this has already been legalized into the constant pool?
450 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
451 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
452 return CFP->getValueAPF().isZero();
457 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
458 /// true if Op is undef or if it matches the specified value.
459 static bool isConstantOrUndef(int Op, int Val) {
460 return Op < 0 || Op == Val;
463 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
464 /// VPKUHUM instruction.
465 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
466 const int *Mask = N->getMask();
468 for (unsigned i = 0; i != 16; ++i)
469 if (!isConstantOrUndef(Mask[i], i*2+1))
472 for (unsigned i = 0; i != 8; ++i)
473 if (!isConstantOrUndef(Mask[i], i*2+1) ||
474 !isConstantOrUndef(Mask[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) {
483 const int *Mask = N->getMask();
485 for (unsigned i = 0; i != 16; i += 2)
486 if (!isConstantOrUndef(Mask[i ], i*2+2) ||
487 !isConstantOrUndef(Mask[i+1], i*2+3))
490 for (unsigned i = 0; i != 8; i += 2)
491 if (!isConstantOrUndef(Mask[i ], i*2+2) ||
492 !isConstantOrUndef(Mask[i+1], i*2+3) ||
493 !isConstantOrUndef(Mask[i+8], i*2+2) ||
494 !isConstantOrUndef(Mask[i+9], i*2+3))
500 /// isVMerge - Common function, used to match vmrg* shuffles.
502 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
503 unsigned LHSStart, unsigned RHSStart) {
504 assert(N->getValueType(0) == MVT::v16i8 &&
505 "PPC only supports shuffles by bytes!");
506 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
507 "Unsupported merge size!");
509 const int *Mask = N->getMask();
510 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
511 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
512 if (!isConstantOrUndef(Mask[i*UnitSize*2+j],
513 LHSStart+j+i*UnitSize) ||
514 !isConstantOrUndef(Mask[i*UnitSize*2+UnitSize+j],
515 RHSStart+j+i*UnitSize))
521 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
522 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
523 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
526 return isVMerge(N, UnitSize, 8, 24);
527 return isVMerge(N, UnitSize, 8, 8);
530 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
531 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
532 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
535 return isVMerge(N, UnitSize, 0, 16);
536 return isVMerge(N, UnitSize, 0, 0);
540 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
541 /// amount, otherwise return -1.
542 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
543 assert(N->getValueType(0) == MVT::v16i8 &&
544 "PPC only supports shuffles by bytes!");
546 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
548 // Find the first non-undef value in the shuffle mask.
549 const int *Mask = SVOp->getMask();
551 for (i = 0; i != 16 && Mask[i] < 0; ++i)
554 if (i == 16) return -1; // all undef.
556 // Otherwise, check to see if the rest of the elements are consecutively
557 // numbered from this value.
558 unsigned ShiftAmt = Mask[i];
559 if (ShiftAmt < i) return -1;
563 // Check the rest of the elements to see if they are consecutive.
564 for (++i; i != 16; ++i)
565 if (!isConstantOrUndef(Mask[i], ShiftAmt+i))
568 // Check the rest of the elements to see if they are consecutive.
569 for (++i; i != 16; ++i)
570 if (!isConstantOrUndef(Mask[i], (ShiftAmt+i) & 15))
576 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
577 /// specifies a splat of a single element that is suitable for input to
578 /// VSPLTB/VSPLTH/VSPLTW.
579 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
580 assert(N->getValueType(0) == MVT::v16i8 &&
581 (EltSize == 1 || EltSize == 2 || EltSize == 4));
583 // This is a splat operation if each element of the permute is the same, and
584 // if the value doesn't reference the second vector.
585 const int *Mask = N->getMask();
586 unsigned ElementBase = Mask[0];
588 // FIXME: Handle UNDEF elements too!
589 if (ElementBase >= 16)
592 // Check that the indices are consecutive, in the case of a multi-byte element
593 // splatted with a v16i8 mask.
594 for (unsigned i = 1; i != EltSize; ++i)
595 if (Mask[i] < 0 || Mask[i] != (int)(i+ElementBase))
598 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
599 if (Mask[i] < 0) continue;
600 for (unsigned j = 0; j != EltSize; ++j)
601 if (Mask[i+j] != Mask[j])
607 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
609 bool PPC::isAllNegativeZeroVector(SDNode *N) {
610 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
612 APInt APVal, APUndef;
616 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32))
617 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
618 return CFP->getValueAPF().isNegZero();
623 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
624 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
625 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
626 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
627 assert(isSplatShuffleMask(SVOp, EltSize));
628 return SVOp->getMask()[0] / EltSize;
631 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
632 /// by using a vspltis[bhw] instruction of the specified element size, return
633 /// the constant being splatted. The ByteSize field indicates the number of
634 /// bytes of each element [124] -> [bhw].
635 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
638 // If ByteSize of the splat is bigger than the element size of the
639 // build_vector, then we have a case where we are checking for a splat where
640 // multiple elements of the buildvector are folded together into a single
641 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
642 unsigned EltSize = 16/N->getNumOperands();
643 if (EltSize < ByteSize) {
644 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
645 SDValue UniquedVals[4];
646 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
648 // See if all of the elements in the buildvector agree across.
649 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
650 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
651 // If the element isn't a constant, bail fully out.
652 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
655 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
656 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
657 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
658 return SDValue(); // no match.
661 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
662 // either constant or undef values that are identical for each chunk. See
663 // if these chunks can form into a larger vspltis*.
665 // Check to see if all of the leading entries are either 0 or -1. If
666 // neither, then this won't fit into the immediate field.
667 bool LeadingZero = true;
668 bool LeadingOnes = true;
669 for (unsigned i = 0; i != Multiple-1; ++i) {
670 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
672 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
673 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
675 // Finally, check the least significant entry.
677 if (UniquedVals[Multiple-1].getNode() == 0)
678 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
679 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
681 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
684 if (UniquedVals[Multiple-1].getNode() == 0)
685 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
686 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
687 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
688 return DAG.getTargetConstant(Val, MVT::i32);
694 // Check to see if this buildvec has a single non-undef value in its elements.
695 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
696 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
697 if (OpVal.getNode() == 0)
698 OpVal = N->getOperand(i);
699 else if (OpVal != N->getOperand(i))
703 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
705 unsigned ValSizeInBytes = 0;
707 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
708 Value = CN->getZExtValue();
709 ValSizeInBytes = CN->getValueType(0).getSizeInBits()/8;
710 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
711 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
712 Value = FloatToBits(CN->getValueAPF().convertToFloat());
716 // If the splat value is larger than the element value, then we can never do
717 // this splat. The only case that we could fit the replicated bits into our
718 // immediate field for would be zero, and we prefer to use vxor for it.
719 if (ValSizeInBytes < ByteSize) return SDValue();
721 // If the element value is larger than the splat value, cut it in half and
722 // check to see if the two halves are equal. Continue doing this until we
723 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
724 while (ValSizeInBytes > ByteSize) {
725 ValSizeInBytes >>= 1;
727 // If the top half equals the bottom half, we're still ok.
728 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
729 (Value & ((1 << (8*ValSizeInBytes))-1)))
733 // Properly sign extend the value.
734 int ShAmt = (4-ByteSize)*8;
735 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
737 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
738 if (MaskVal == 0) return SDValue();
740 // Finally, if this value fits in a 5 bit sext field, return it
741 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
742 return DAG.getTargetConstant(MaskVal, MVT::i32);
746 //===----------------------------------------------------------------------===//
747 // Addressing Mode Selection
748 //===----------------------------------------------------------------------===//
750 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
751 /// or 64-bit immediate, and if the value can be accurately represented as a
752 /// sign extension from a 16-bit value. If so, this returns true and the
754 static bool isIntS16Immediate(SDNode *N, short &Imm) {
755 if (N->getOpcode() != ISD::Constant)
758 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
759 if (N->getValueType(0) == MVT::i32)
760 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
762 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
764 static bool isIntS16Immediate(SDValue Op, short &Imm) {
765 return isIntS16Immediate(Op.getNode(), Imm);
769 /// SelectAddressRegReg - Given the specified addressed, check to see if it
770 /// can be represented as an indexed [r+r] operation. Returns false if it
771 /// can be more efficiently represented with [r+imm].
772 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
774 SelectionDAG &DAG) const {
776 if (N.getOpcode() == ISD::ADD) {
777 if (isIntS16Immediate(N.getOperand(1), imm))
779 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
782 Base = N.getOperand(0);
783 Index = N.getOperand(1);
785 } else if (N.getOpcode() == ISD::OR) {
786 if (isIntS16Immediate(N.getOperand(1), imm))
787 return false; // r+i can fold it if we can.
789 // If this is an or of disjoint bitfields, we can codegen this as an add
790 // (for better address arithmetic) if the LHS and RHS of the OR are provably
792 APInt LHSKnownZero, LHSKnownOne;
793 APInt RHSKnownZero, RHSKnownOne;
794 DAG.ComputeMaskedBits(N.getOperand(0),
795 APInt::getAllOnesValue(N.getOperand(0)
796 .getValueSizeInBits()),
797 LHSKnownZero, LHSKnownOne);
799 if (LHSKnownZero.getBoolValue()) {
800 DAG.ComputeMaskedBits(N.getOperand(1),
801 APInt::getAllOnesValue(N.getOperand(1)
802 .getValueSizeInBits()),
803 RHSKnownZero, RHSKnownOne);
804 // If all of the bits are known zero on the LHS or RHS, the add won't
806 if (~(LHSKnownZero | RHSKnownZero) == 0) {
807 Base = N.getOperand(0);
808 Index = N.getOperand(1);
817 /// Returns true if the address N can be represented by a base register plus
818 /// a signed 16-bit displacement [r+imm], and if it is not better
819 /// represented as reg+reg.
820 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
822 SelectionDAG &DAG) const {
823 // FIXME dl should come from parent load or store, not from address
824 DebugLoc dl = N.getDebugLoc();
825 // If this can be more profitably realized as r+r, fail.
826 if (SelectAddressRegReg(N, Disp, Base, DAG))
829 if (N.getOpcode() == ISD::ADD) {
831 if (isIntS16Immediate(N.getOperand(1), imm)) {
832 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
833 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
834 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
836 Base = N.getOperand(0);
838 return true; // [r+i]
839 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
840 // Match LOAD (ADD (X, Lo(G))).
841 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
842 && "Cannot handle constant offsets yet!");
843 Disp = N.getOperand(1).getOperand(0); // The global address.
844 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
845 Disp.getOpcode() == ISD::TargetConstantPool ||
846 Disp.getOpcode() == ISD::TargetJumpTable);
847 Base = N.getOperand(0);
848 return true; // [&g+r]
850 } else if (N.getOpcode() == ISD::OR) {
852 if (isIntS16Immediate(N.getOperand(1), imm)) {
853 // If this is an or of disjoint bitfields, we can codegen this as an add
854 // (for better address arithmetic) if the LHS and RHS of the OR are
855 // provably disjoint.
856 APInt LHSKnownZero, LHSKnownOne;
857 DAG.ComputeMaskedBits(N.getOperand(0),
858 APInt::getAllOnesValue(N.getOperand(0)
859 .getValueSizeInBits()),
860 LHSKnownZero, LHSKnownOne);
862 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
863 // If all of the bits are known zero on the LHS or RHS, the add won't
865 Base = N.getOperand(0);
866 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
870 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
871 // Loading from a constant address.
873 // If this address fits entirely in a 16-bit sext immediate field, codegen
876 if (isIntS16Immediate(CN, Imm)) {
877 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
878 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
882 // Handle 32-bit sext immediates with LIS + addr mode.
883 if (CN->getValueType(0) == MVT::i32 ||
884 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
885 int Addr = (int)CN->getZExtValue();
887 // Otherwise, break this down into an LIS + disp.
888 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
890 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
891 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
892 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base), 0);
897 Disp = DAG.getTargetConstant(0, getPointerTy());
898 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
899 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
902 return true; // [r+0]
905 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
906 /// represented as an indexed [r+r] operation.
907 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
909 SelectionDAG &DAG) const {
910 // Check to see if we can easily represent this as an [r+r] address. This
911 // will fail if it thinks that the address is more profitably represented as
912 // reg+imm, e.g. where imm = 0.
913 if (SelectAddressRegReg(N, Base, Index, DAG))
916 // If the operand is an addition, always emit this as [r+r], since this is
917 // better (for code size, and execution, as the memop does the add for free)
918 // than emitting an explicit add.
919 if (N.getOpcode() == ISD::ADD) {
920 Base = N.getOperand(0);
921 Index = N.getOperand(1);
925 // Otherwise, do it the hard way, using R0 as the base register.
926 Base = DAG.getRegister(PPC::R0, N.getValueType());
931 /// SelectAddressRegImmShift - Returns true if the address N can be
932 /// represented by a base register plus a signed 14-bit displacement
933 /// [r+imm*4]. Suitable for use by STD and friends.
934 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
936 SelectionDAG &DAG) const {
937 // FIXME dl should come from the parent load or store, not the address
938 DebugLoc dl = N.getDebugLoc();
939 // If this can be more profitably realized as r+r, fail.
940 if (SelectAddressRegReg(N, Disp, Base, DAG))
943 if (N.getOpcode() == ISD::ADD) {
945 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
946 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
947 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
948 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
950 Base = N.getOperand(0);
952 return true; // [r+i]
953 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
954 // Match LOAD (ADD (X, Lo(G))).
955 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
956 && "Cannot handle constant offsets yet!");
957 Disp = N.getOperand(1).getOperand(0); // The global address.
958 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
959 Disp.getOpcode() == ISD::TargetConstantPool ||
960 Disp.getOpcode() == ISD::TargetJumpTable);
961 Base = N.getOperand(0);
962 return true; // [&g+r]
964 } else if (N.getOpcode() == ISD::OR) {
966 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
967 // If this is an or of disjoint bitfields, we can codegen this as an add
968 // (for better address arithmetic) if the LHS and RHS of the OR are
969 // provably disjoint.
970 APInt LHSKnownZero, LHSKnownOne;
971 DAG.ComputeMaskedBits(N.getOperand(0),
972 APInt::getAllOnesValue(N.getOperand(0)
973 .getValueSizeInBits()),
974 LHSKnownZero, LHSKnownOne);
975 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
976 // If all of the bits are known zero on the LHS or RHS, the add won't
978 Base = N.getOperand(0);
979 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
983 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
984 // Loading from a constant address. Verify low two bits are clear.
985 if ((CN->getZExtValue() & 3) == 0) {
986 // If this address fits entirely in a 14-bit sext immediate field, codegen
989 if (isIntS16Immediate(CN, Imm)) {
990 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
991 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
995 // Fold the low-part of 32-bit absolute addresses into addr mode.
996 if (CN->getValueType(0) == MVT::i32 ||
997 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
998 int Addr = (int)CN->getZExtValue();
1000 // Otherwise, break this down into an LIS + disp.
1001 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1002 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1003 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1004 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base),0);
1010 Disp = DAG.getTargetConstant(0, getPointerTy());
1011 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1012 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1015 return true; // [r+0]
1019 /// getPreIndexedAddressParts - returns true by value, base pointer and
1020 /// offset pointer and addressing mode by reference if the node's address
1021 /// can be legally represented as pre-indexed load / store address.
1022 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1024 ISD::MemIndexedMode &AM,
1025 SelectionDAG &DAG) const {
1026 // Disabled by default for now.
1027 if (!EnablePPCPreinc) return false;
1031 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1032 Ptr = LD->getBasePtr();
1033 VT = LD->getMemoryVT();
1035 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1037 Ptr = ST->getBasePtr();
1038 VT = ST->getMemoryVT();
1042 // PowerPC doesn't have preinc load/store instructions for vectors.
1046 // TODO: Check reg+reg first.
1048 // LDU/STU use reg+imm*4, others use reg+imm.
1049 if (VT != MVT::i64) {
1051 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1055 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1059 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1060 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1061 // sext i32 to i64 when addr mode is r+i.
1062 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1063 LD->getExtensionType() == ISD::SEXTLOAD &&
1064 isa<ConstantSDNode>(Offset))
1072 //===----------------------------------------------------------------------===//
1073 // LowerOperation implementation
1074 //===----------------------------------------------------------------------===//
1076 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1077 SelectionDAG &DAG) {
1078 MVT PtrVT = Op.getValueType();
1079 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1080 Constant *C = CP->getConstVal();
1081 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1082 SDValue Zero = DAG.getConstant(0, PtrVT);
1083 // FIXME there isn't really any debug info here
1084 DebugLoc dl = Op.getDebugLoc();
1086 const TargetMachine &TM = DAG.getTarget();
1088 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, CPI, Zero);
1089 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, CPI, Zero);
1091 // If this is a non-darwin platform, we don't support non-static relo models
1093 if (TM.getRelocationModel() == Reloc::Static ||
1094 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1095 // Generate non-pic code that has direct accesses to the constant pool.
1096 // The address of the global is just (hi(&g)+lo(&g)).
1097 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1100 if (TM.getRelocationModel() == Reloc::PIC_) {
1101 // With PIC, the first instruction is actually "GR+hi(&G)".
1102 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1103 DAG.getNode(PPCISD::GlobalBaseReg,
1104 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1107 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1111 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1112 MVT PtrVT = Op.getValueType();
1113 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1114 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1115 SDValue Zero = DAG.getConstant(0, PtrVT);
1116 // FIXME there isn't really any debug loc here
1117 DebugLoc dl = Op.getDebugLoc();
1119 const TargetMachine &TM = DAG.getTarget();
1121 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, JTI, Zero);
1122 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, JTI, Zero);
1124 // If this is a non-darwin platform, we don't support non-static relo models
1126 if (TM.getRelocationModel() == Reloc::Static ||
1127 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1128 // Generate non-pic code that has direct accesses to the constant pool.
1129 // The address of the global is just (hi(&g)+lo(&g)).
1130 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1133 if (TM.getRelocationModel() == Reloc::PIC_) {
1134 // With PIC, the first instruction is actually "GR+hi(&G)".
1135 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1136 DAG.getNode(PPCISD::GlobalBaseReg,
1137 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1140 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1144 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1145 SelectionDAG &DAG) {
1146 assert(0 && "TLS not implemented for PPC.");
1147 return SDValue(); // Not reached
1150 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1151 SelectionDAG &DAG) {
1152 MVT PtrVT = Op.getValueType();
1153 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1154 GlobalValue *GV = GSDN->getGlobal();
1155 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1156 SDValue Zero = DAG.getConstant(0, PtrVT);
1157 // FIXME there isn't really any debug info here
1158 DebugLoc dl = GSDN->getDebugLoc();
1160 const TargetMachine &TM = DAG.getTarget();
1162 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, GA, Zero);
1163 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, GA, Zero);
1165 // If this is a non-darwin platform, we don't support non-static relo models
1167 if (TM.getRelocationModel() == Reloc::Static ||
1168 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1169 // Generate non-pic code that has direct accesses to globals.
1170 // The address of the global is just (hi(&g)+lo(&g)).
1171 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1174 if (TM.getRelocationModel() == Reloc::PIC_) {
1175 // With PIC, the first instruction is actually "GR+hi(&G)".
1176 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1177 DAG.getNode(PPCISD::GlobalBaseReg,
1178 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1181 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1183 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1186 // If the global is weak or external, we have to go through the lazy
1188 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Lo, NULL, 0);
1191 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1192 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1193 DebugLoc dl = Op.getDebugLoc();
1195 // If we're comparing for equality to zero, expose the fact that this is
1196 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1197 // fold the new nodes.
1198 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1199 if (C->isNullValue() && CC == ISD::SETEQ) {
1200 MVT VT = Op.getOperand(0).getValueType();
1201 SDValue Zext = Op.getOperand(0);
1202 if (VT.bitsLT(MVT::i32)) {
1204 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1206 unsigned Log2b = Log2_32(VT.getSizeInBits());
1207 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1208 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1209 DAG.getConstant(Log2b, MVT::i32));
1210 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1212 // Leave comparisons against 0 and -1 alone for now, since they're usually
1213 // optimized. FIXME: revisit this when we can custom lower all setcc
1215 if (C->isAllOnesValue() || C->isNullValue())
1219 // If we have an integer seteq/setne, turn it into a compare against zero
1220 // by xor'ing the rhs with the lhs, which is faster than setting a
1221 // condition register, reading it back out, and masking the correct bit. The
1222 // normal approach here uses sub to do this instead of xor. Using xor exposes
1223 // the result to other bit-twiddling opportunities.
1224 MVT LHSVT = Op.getOperand(0).getValueType();
1225 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1226 MVT VT = Op.getValueType();
1227 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1229 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1234 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1235 int VarArgsFrameIndex,
1236 int VarArgsStackOffset,
1237 unsigned VarArgsNumGPR,
1238 unsigned VarArgsNumFPR,
1239 const PPCSubtarget &Subtarget) {
1241 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1242 return SDValue(); // Not reached
1245 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1246 SDValue Chain = Op.getOperand(0);
1247 SDValue Trmp = Op.getOperand(1); // trampoline
1248 SDValue FPtr = Op.getOperand(2); // nested function
1249 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1250 DebugLoc dl = Op.getDebugLoc();
1252 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1253 bool isPPC64 = (PtrVT == MVT::i64);
1254 const Type *IntPtrTy =
1255 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType();
1257 TargetLowering::ArgListTy Args;
1258 TargetLowering::ArgListEntry Entry;
1260 Entry.Ty = IntPtrTy;
1261 Entry.Node = Trmp; Args.push_back(Entry);
1263 // TrampSize == (isPPC64 ? 48 : 40);
1264 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1265 isPPC64 ? MVT::i64 : MVT::i32);
1266 Args.push_back(Entry);
1268 Entry.Node = FPtr; Args.push_back(Entry);
1269 Entry.Node = Nest; Args.push_back(Entry);
1271 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1272 std::pair<SDValue, SDValue> CallResult =
1273 LowerCallTo(Chain, Op.getValueType().getTypeForMVT(), false, false,
1274 false, false, CallingConv::C, false,
1275 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1279 { CallResult.first, CallResult.second };
1281 return DAG.getMergeValues(Ops, 2, dl);
1284 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1285 int VarArgsFrameIndex,
1286 int VarArgsStackOffset,
1287 unsigned VarArgsNumGPR,
1288 unsigned VarArgsNumFPR,
1289 const PPCSubtarget &Subtarget) {
1290 DebugLoc dl = Op.getDebugLoc();
1292 if (Subtarget.isMachoABI()) {
1293 // vastart just stores the address of the VarArgsFrameIndex slot into the
1294 // memory location argument.
1295 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1296 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1297 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1298 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1301 // For ELF 32 ABI we follow the layout of the va_list struct.
1302 // We suppose the given va_list is already allocated.
1305 // char gpr; /* index into the array of 8 GPRs
1306 // * stored in the register save area
1307 // * gpr=0 corresponds to r3,
1308 // * gpr=1 to r4, etc.
1310 // char fpr; /* index into the array of 8 FPRs
1311 // * stored in the register save area
1312 // * fpr=0 corresponds to f1,
1313 // * fpr=1 to f2, etc.
1315 // char *overflow_arg_area;
1316 // /* location on stack that holds
1317 // * the next overflow argument
1319 // char *reg_save_area;
1320 // /* where r3:r10 and f1:f8 (if saved)
1326 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1327 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1330 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1332 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1333 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1335 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1336 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1338 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1339 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1341 uint64_t FPROffset = 1;
1342 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1344 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1346 // Store first byte : number of int regs
1347 SDValue firstStore = DAG.getStore(Op.getOperand(0), dl, ArgGPR,
1348 Op.getOperand(1), SV, 0);
1349 uint64_t nextOffset = FPROffset;
1350 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1353 // Store second byte : number of float regs
1354 SDValue secondStore =
1355 DAG.getStore(firstStore, dl, ArgFPR, nextPtr, SV, nextOffset);
1356 nextOffset += StackOffset;
1357 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1359 // Store second word : arguments given on stack
1360 SDValue thirdStore =
1361 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, SV, nextOffset);
1362 nextOffset += FrameOffset;
1363 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1365 // Store third word : arguments given in registers
1366 return DAG.getStore(thirdStore, dl, FR, nextPtr, SV, nextOffset);
1370 #include "PPCGenCallingConv.inc"
1372 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1373 /// depending on which subtarget is selected.
1374 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1375 if (Subtarget.isMachoABI()) {
1376 static const unsigned FPR[] = {
1377 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1378 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1384 static const unsigned FPR[] = {
1385 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1391 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1393 static unsigned CalculateStackSlotSize(SDValue Arg, ISD::ArgFlagsTy Flags,
1394 bool isVarArg, unsigned PtrByteSize) {
1395 MVT ArgVT = Arg.getValueType();
1396 unsigned ArgSize =ArgVT.getSizeInBits()/8;
1397 if (Flags.isByVal())
1398 ArgSize = Flags.getByValSize();
1399 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1405 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1407 int &VarArgsFrameIndex,
1408 int &VarArgsStackOffset,
1409 unsigned &VarArgsNumGPR,
1410 unsigned &VarArgsNumFPR,
1411 const PPCSubtarget &Subtarget) {
1412 // TODO: add description of PPC stack frame format, or at least some docs.
1414 MachineFunction &MF = DAG.getMachineFunction();
1415 MachineFrameInfo *MFI = MF.getFrameInfo();
1416 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1417 SmallVector<SDValue, 8> ArgValues;
1418 SDValue Root = Op.getOperand(0);
1419 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1420 DebugLoc dl = Op.getDebugLoc();
1422 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1423 bool isPPC64 = PtrVT == MVT::i64;
1424 bool isMachoABI = Subtarget.isMachoABI();
1425 bool isELF32_ABI = Subtarget.isELF32_ABI();
1426 // Potential tail calls could cause overwriting of argument stack slots.
1427 unsigned CC = MF.getFunction()->getCallingConv();
1428 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1429 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1431 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1432 // Area that is at least reserved in caller of this function.
1433 unsigned MinReservedArea = ArgOffset;
1435 static const unsigned GPR_32[] = { // 32-bit registers.
1436 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1437 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1439 static const unsigned GPR_64[] = { // 64-bit registers.
1440 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1441 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1444 static const unsigned *FPR = GetFPR(Subtarget);
1446 static const unsigned VR[] = {
1447 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1448 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1451 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1452 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1453 const unsigned Num_VR_Regs = array_lengthof( VR);
1455 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1457 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1459 // In 32-bit non-varargs functions, the stack space for vectors is after the
1460 // stack space for non-vectors. We do not use this space unless we have
1461 // too many vectors to fit in registers, something that only occurs in
1462 // constructed examples:), but we have to walk the arglist to figure
1463 // that out...for the pathological case, compute VecArgOffset as the
1464 // start of the vector parameter area. Computing VecArgOffset is the
1465 // entire point of the following loop.
1466 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1467 // to handle Elf here.
1468 unsigned VecArgOffset = ArgOffset;
1469 if (!isVarArg && !isPPC64) {
1470 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1472 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1473 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1474 ISD::ArgFlagsTy Flags =
1475 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1477 if (Flags.isByVal()) {
1478 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1479 ObjSize = Flags.getByValSize();
1481 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1482 VecArgOffset += ArgSize;
1486 switch(ObjectVT.getSimpleVT()) {
1487 default: assert(0 && "Unhandled argument type!");
1490 VecArgOffset += isPPC64 ? 8 : 4;
1492 case MVT::i64: // PPC64
1500 // Nothing to do, we're only looking at Nonvector args here.
1505 // We've found where the vector parameter area in memory is. Skip the
1506 // first 12 parameters; these don't use that memory.
1507 VecArgOffset = ((VecArgOffset+15)/16)*16;
1508 VecArgOffset += 12*16;
1510 // Add DAG nodes to load the arguments or copy them out of registers. On
1511 // entry to a function on PPC, the arguments start after the linkage area,
1512 // although the first ones are often in registers.
1514 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1515 // represented with two words (long long or double) must be copied to an
1516 // even GPR_idx value or to an even ArgOffset value.
1518 SmallVector<SDValue, 8> MemOps;
1519 unsigned nAltivecParamsAtEnd = 0;
1520 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1521 ArgNo != e; ++ArgNo) {
1523 bool needsLoad = false;
1524 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1525 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1526 unsigned ArgSize = ObjSize;
1527 ISD::ArgFlagsTy Flags =
1528 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1529 // See if next argument requires stack alignment in ELF
1530 bool Align = Flags.isSplit();
1532 unsigned CurArgOffset = ArgOffset;
1534 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1535 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1536 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1537 if (isVarArg || isPPC64) {
1538 MinReservedArea = ((MinReservedArea+15)/16)*16;
1539 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1543 } else nAltivecParamsAtEnd++;
1545 // Calculate min reserved area.
1546 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1551 // FIXME alignment for ELF may not be right
1552 // FIXME the codegen can be much improved in some cases.
1553 // We do not have to keep everything in memory.
1554 if (Flags.isByVal()) {
1555 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1556 ObjSize = Flags.getByValSize();
1557 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1558 // Double word align in ELF
1559 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1560 // Objects of size 1 and 2 are right justified, everything else is
1561 // left justified. This means the memory address is adjusted forwards.
1562 if (ObjSize==1 || ObjSize==2) {
1563 CurArgOffset = CurArgOffset + (4 - ObjSize);
1565 // The value of the object is its address.
1566 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1567 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1568 ArgValues.push_back(FIN);
1569 if (ObjSize==1 || ObjSize==2) {
1570 if (GPR_idx != Num_GPR_Regs) {
1571 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1572 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1573 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1574 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1575 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1576 MemOps.push_back(Store);
1578 if (isMachoABI) ArgOffset += PtrByteSize;
1580 ArgOffset += PtrByteSize;
1584 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1585 // Store whatever pieces of the object are in registers
1586 // to memory. ArgVal will be address of the beginning of
1588 if (GPR_idx != Num_GPR_Regs) {
1589 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1590 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1591 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1592 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1593 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1594 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1595 MemOps.push_back(Store);
1597 if (isMachoABI) ArgOffset += PtrByteSize;
1599 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1606 switch (ObjectVT.getSimpleVT()) {
1607 default: assert(0 && "Unhandled argument type!");
1610 // Double word align in ELF
1611 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1613 if (GPR_idx != Num_GPR_Regs) {
1614 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1615 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1616 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1620 ArgSize = PtrByteSize;
1622 // Stack align in ELF
1623 if (needsLoad && Align && isELF32_ABI)
1624 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1625 // All int arguments reserve stack space in Macho ABI.
1626 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1630 case MVT::i64: // PPC64
1631 if (GPR_idx != Num_GPR_Regs) {
1632 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1633 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1634 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1636 if (ObjectVT == MVT::i32) {
1637 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1638 // value to MVT::i64 and then truncate to the correct register size.
1640 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1641 DAG.getValueType(ObjectVT));
1642 else if (Flags.isZExt())
1643 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1644 DAG.getValueType(ObjectVT));
1646 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1652 ArgSize = PtrByteSize;
1654 // All int arguments reserve stack space in Macho ABI.
1655 if (isMachoABI || needsLoad) ArgOffset += 8;
1660 // Every 4 bytes of argument space consumes one of the GPRs available for
1661 // argument passing.
1662 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1664 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1667 if (FPR_idx != Num_FPR_Regs) {
1669 if (ObjectVT == MVT::f32)
1670 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1672 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1673 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1674 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1680 // Stack align in ELF
1681 if (needsLoad && Align && isELF32_ABI)
1682 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1683 // All FP arguments reserve stack space in Macho ABI.
1684 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1690 // Note that vector arguments in registers don't reserve stack space,
1691 // except in varargs functions.
1692 if (VR_idx != Num_VR_Regs) {
1693 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1694 RegInfo.addLiveIn(VR[VR_idx], VReg);
1695 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1697 while ((ArgOffset % 16) != 0) {
1698 ArgOffset += PtrByteSize;
1699 if (GPR_idx != Num_GPR_Regs)
1703 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1707 if (!isVarArg && !isPPC64) {
1708 // Vectors go after all the nonvectors.
1709 CurArgOffset = VecArgOffset;
1712 // Vectors are aligned.
1713 ArgOffset = ((ArgOffset+15)/16)*16;
1714 CurArgOffset = ArgOffset;
1722 // We need to load the argument to a virtual register if we determined above
1723 // that we ran out of physical registers of the appropriate type.
1725 int FI = MFI->CreateFixedObject(ObjSize,
1726 CurArgOffset + (ArgSize - ObjSize),
1728 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1729 ArgVal = DAG.getLoad(ObjectVT, dl, Root, FIN, NULL, 0);
1732 ArgValues.push_back(ArgVal);
1735 // Set the size that is at least reserved in caller of this function. Tail
1736 // call optimized function's reserved stack space needs to be aligned so that
1737 // taking the difference between two stack areas will result in an aligned
1739 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1740 // Add the Altivec parameters at the end, if needed.
1741 if (nAltivecParamsAtEnd) {
1742 MinReservedArea = ((MinReservedArea+15)/16)*16;
1743 MinReservedArea += 16*nAltivecParamsAtEnd;
1746 std::max(MinReservedArea,
1747 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1748 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1749 getStackAlignment();
1750 unsigned AlignMask = TargetAlign-1;
1751 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1752 FI->setMinReservedArea(MinReservedArea);
1754 // If the function takes variable number of arguments, make a frame index for
1755 // the start of the first vararg value... for expansion of llvm.va_start.
1760 VarArgsNumGPR = GPR_idx;
1761 VarArgsNumFPR = FPR_idx;
1763 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1765 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
1766 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
1767 PtrVT.getSizeInBits()/8);
1769 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1776 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1778 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1780 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1781 // stored to the VarArgsFrameIndex on the stack.
1783 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1784 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1785 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1786 MemOps.push_back(Store);
1787 // Increment the address by four for the next argument to store
1788 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1789 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1793 // If this function is vararg, store any remaining integer argument regs
1794 // to their spots on the stack so that they may be loaded by deferencing the
1795 // result of va_next.
1796 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1799 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1801 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1803 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1804 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1805 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1806 MemOps.push_back(Store);
1807 // Increment the address by four for the next argument to store
1808 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1809 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1812 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1815 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1816 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1817 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1818 MemOps.push_back(Store);
1819 // Increment the address by eight for the next argument to store
1820 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1822 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1825 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1827 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1829 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1830 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::f64);
1831 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1832 MemOps.push_back(Store);
1833 // Increment the address by eight for the next argument to store
1834 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1836 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1841 if (!MemOps.empty())
1842 Root = DAG.getNode(ISD::TokenFactor, dl,
1843 MVT::Other, &MemOps[0], MemOps.size());
1845 ArgValues.push_back(Root);
1847 // Return the new list of results.
1848 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1849 &ArgValues[0], ArgValues.size());
1852 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
1855 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
1860 CallSDNode *TheCall,
1861 unsigned &nAltivecParamsAtEnd) {
1862 // Count how many bytes are to be pushed on the stack, including the linkage
1863 // area, and parameter passing area. We start with 24/48 bytes, which is
1864 // prereserved space for [SP][CR][LR][3 x unused].
1865 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1866 unsigned NumOps = TheCall->getNumArgs();
1867 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1869 // Add up all the space actually used.
1870 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1871 // they all go in registers, but we must reserve stack space for them for
1872 // possible use by the caller. In varargs or 64-bit calls, parameters are
1873 // assigned stack space in order, with padding so Altivec parameters are
1875 nAltivecParamsAtEnd = 0;
1876 for (unsigned i = 0; i != NumOps; ++i) {
1877 SDValue Arg = TheCall->getArg(i);
1878 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1879 MVT ArgVT = Arg.getValueType();
1880 // Varargs Altivec parameters are padded to a 16 byte boundary.
1881 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1882 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1883 if (!isVarArg && !isPPC64) {
1884 // Non-varargs Altivec parameters go after all the non-Altivec
1885 // parameters; handle those later so we know how much padding we need.
1886 nAltivecParamsAtEnd++;
1889 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1890 NumBytes = ((NumBytes+15)/16)*16;
1892 NumBytes += CalculateStackSlotSize(Arg, Flags, isVarArg, PtrByteSize);
1895 // Allow for Altivec parameters at the end, if needed.
1896 if (nAltivecParamsAtEnd) {
1897 NumBytes = ((NumBytes+15)/16)*16;
1898 NumBytes += 16*nAltivecParamsAtEnd;
1901 // The prolog code of the callee may store up to 8 GPR argument registers to
1902 // the stack, allowing va_start to index over them in memory if its varargs.
1903 // Because we cannot tell if this is needed on the caller side, we have to
1904 // conservatively assume that it is needed. As such, make sure we have at
1905 // least enough stack space for the caller to store the 8 GPRs.
1906 NumBytes = std::max(NumBytes,
1907 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1909 // Tail call needs the stack to be aligned.
1910 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1911 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1912 getStackAlignment();
1913 unsigned AlignMask = TargetAlign-1;
1914 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
1920 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
1921 /// adjusted to accomodate the arguments for the tailcall.
1922 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
1923 unsigned ParamSize) {
1925 if (!IsTailCall) return 0;
1927 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
1928 unsigned CallerMinReservedArea = FI->getMinReservedArea();
1929 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
1930 // Remember only if the new adjustement is bigger.
1931 if (SPDiff < FI->getTailCallSPDelta())
1932 FI->setTailCallSPDelta(SPDiff);
1937 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1938 /// following the call is a return. A function is eligible if caller/callee
1939 /// calling conventions match, currently only fastcc supports tail calls, and
1940 /// the function CALL is immediatly followed by a RET.
1942 PPCTargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1944 SelectionDAG& DAG) const {
1945 // Variable argument functions are not supported.
1946 if (!PerformTailCallOpt || TheCall->isVarArg())
1949 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1950 MachineFunction &MF = DAG.getMachineFunction();
1951 unsigned CallerCC = MF.getFunction()->getCallingConv();
1952 unsigned CalleeCC = TheCall->getCallingConv();
1953 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1954 // Functions containing by val parameters are not supported.
1955 for (unsigned i = 0; i != TheCall->getNumArgs(); i++) {
1956 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1957 if (Flags.isByVal()) return false;
1960 SDValue Callee = TheCall->getCallee();
1961 // Non PIC/GOT tail calls are supported.
1962 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1965 // At the moment we can only do local tail calls (in same module, hidden
1966 // or protected) if we are generating PIC.
1967 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1968 return G->getGlobal()->hasHiddenVisibility()
1969 || G->getGlobal()->hasProtectedVisibility();
1976 /// isCallCompatibleAddress - Return the immediate to use if the specified
1977 /// 32-bit value is representable in the immediate field of a BxA instruction.
1978 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
1979 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1982 int Addr = C->getZExtValue();
1983 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1984 (Addr << 6 >> 6) != Addr)
1985 return 0; // Top 6 bits have to be sext of immediate.
1987 return DAG.getConstant((int)C->getZExtValue() >> 2,
1988 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
1993 struct TailCallArgumentInfo {
1998 TailCallArgumentInfo() : FrameIdx(0) {}
2003 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2005 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2007 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2008 SmallVector<SDValue, 8> &MemOpChains,
2010 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2011 SDValue Arg = TailCallArgs[i].Arg;
2012 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2013 int FI = TailCallArgs[i].FrameIdx;
2014 // Store relative to framepointer.
2015 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2016 PseudoSourceValue::getFixedStack(FI),
2021 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2022 /// the appropriate stack slot for the tail call optimized function call.
2023 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2024 MachineFunction &MF,
2033 // Calculate the new stack slot for the return address.
2034 int SlotSize = isPPC64 ? 8 : 4;
2035 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2037 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2039 int NewFPLoc = SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64,
2041 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
2043 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2044 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2045 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2046 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2047 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2048 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2049 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2054 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2055 /// the position of the argument.
2057 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2058 SDValue Arg, int SPDiff, unsigned ArgOffset,
2059 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2060 int Offset = ArgOffset + SPDiff;
2061 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2062 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2063 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2064 SDValue FIN = DAG.getFrameIndex(FI, VT);
2065 TailCallArgumentInfo Info;
2067 Info.FrameIdxOp = FIN;
2069 TailCallArguments.push_back(Info);
2072 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2073 /// stack slot. Returns the chain as result and the loaded frame pointers in
2074 /// LROpOut/FPOpout. Used when tail calling.
2075 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2082 // Load the LR and FP stack slot for later adjusting.
2083 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2084 LROpOut = getReturnAddrFrameIndex(DAG);
2085 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, NULL, 0);
2086 Chain = SDValue(LROpOut.getNode(), 1);
2087 FPOpOut = getFramePointerFrameIndex(DAG);
2088 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, NULL, 0);
2089 Chain = SDValue(FPOpOut.getNode(), 1);
2094 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2095 /// by "Src" to address "Dst" of size "Size". Alignment information is
2096 /// specified by the specific parameter attribute. The copy will be passed as
2097 /// a byval function parameter.
2098 /// Sometimes what we are copying is the end of a larger object, the part that
2099 /// does not fit in registers.
2101 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2102 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2103 unsigned Size, DebugLoc dl) {
2104 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
2105 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2106 false, NULL, 0, NULL, 0);
2109 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2112 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2113 SDValue Arg, SDValue PtrOff, int SPDiff,
2114 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2115 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2116 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments,
2118 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2123 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2125 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2126 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2127 DAG.getConstant(ArgOffset, PtrVT));
2129 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0));
2130 // Calculate and remember argument location.
2131 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2135 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2136 const PPCSubtarget &Subtarget,
2137 TargetMachine &TM) {
2138 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2139 SDValue Chain = TheCall->getChain();
2140 bool isVarArg = TheCall->isVarArg();
2141 unsigned CC = TheCall->getCallingConv();
2142 bool isTailCall = TheCall->isTailCall()
2143 && CC == CallingConv::Fast && PerformTailCallOpt;
2144 SDValue Callee = TheCall->getCallee();
2145 unsigned NumOps = TheCall->getNumArgs();
2146 DebugLoc dl = TheCall->getDebugLoc();
2148 bool isMachoABI = Subtarget.isMachoABI();
2149 bool isELF32_ABI = Subtarget.isELF32_ABI();
2151 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2152 bool isPPC64 = PtrVT == MVT::i64;
2153 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2155 MachineFunction &MF = DAG.getMachineFunction();
2157 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
2158 // SelectExpr to use to put the arguments in the appropriate registers.
2159 std::vector<SDValue> args_to_use;
2161 // Mark this function as potentially containing a function that contains a
2162 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2163 // and restoring the callers stack pointer in this functions epilog. This is
2164 // done because by tail calling the called function might overwrite the value
2165 // in this function's (MF) stack pointer stack slot 0(SP).
2166 if (PerformTailCallOpt && CC==CallingConv::Fast)
2167 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2169 unsigned nAltivecParamsAtEnd = 0;
2171 // Count how many bytes are to be pushed on the stack, including the linkage
2172 // area, and parameter passing area. We start with 24/48 bytes, which is
2173 // prereserved space for [SP][CR][LR][3 x unused].
2175 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2176 TheCall, nAltivecParamsAtEnd);
2178 // Calculate by how many bytes the stack has to be adjusted in case of tail
2179 // call optimization.
2180 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2182 // Adjust the stack pointer for the new arguments...
2183 // These operations are automatically eliminated by the prolog/epilog pass
2184 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2185 SDValue CallSeqStart = Chain;
2187 // Load the return address and frame pointer so it can be move somewhere else
2190 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
2192 // Set up a copy of the stack pointer for use loading and storing any
2193 // arguments that may not fit in the registers available for argument
2197 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2199 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2201 // Figure out which arguments are going to go in registers, and which in
2202 // memory. Also, if this is a vararg function, floating point operations
2203 // must be stored to our stack, and loaded into integer regs as well, if
2204 // any integer regs are available for argument passing.
2205 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2206 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2208 static const unsigned GPR_32[] = { // 32-bit registers.
2209 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2210 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2212 static const unsigned GPR_64[] = { // 64-bit registers.
2213 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2214 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2216 static const unsigned *FPR = GetFPR(Subtarget);
2218 static const unsigned VR[] = {
2219 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2220 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2222 const unsigned NumGPRs = array_lengthof(GPR_32);
2223 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2224 const unsigned NumVRs = array_lengthof( VR);
2226 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2228 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2229 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2231 SmallVector<SDValue, 8> MemOpChains;
2232 for (unsigned i = 0; i != NumOps; ++i) {
2234 SDValue Arg = TheCall->getArg(i);
2235 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2236 // See if next argument requires stack alignment in ELF
2237 bool Align = Flags.isSplit();
2239 // PtrOff will be used to store the current argument to the stack if a
2240 // register cannot be found for it.
2243 // Stack align in ELF 32
2244 if (isELF32_ABI && Align)
2245 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2246 StackPtr.getValueType());
2248 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2250 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
2252 // On PPC64, promote integers to 64-bit values.
2253 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2254 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2255 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2256 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
2259 // FIXME Elf untested, what are alignment rules?
2260 // FIXME memcpy is used way more than necessary. Correctness first.
2261 if (Flags.isByVal()) {
2262 unsigned Size = Flags.getByValSize();
2263 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2264 if (Size==1 || Size==2) {
2265 // Very small objects are passed right-justified.
2266 // Everything else is passed left-justified.
2267 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2268 if (GPR_idx != NumGPRs) {
2269 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
2271 MemOpChains.push_back(Load.getValue(1));
2272 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2274 ArgOffset += PtrByteSize;
2276 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2277 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
2278 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2279 CallSeqStart.getNode()->getOperand(0),
2280 Flags, DAG, Size, dl);
2281 // This must go outside the CALLSEQ_START..END.
2282 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2283 CallSeqStart.getNode()->getOperand(1));
2284 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2285 NewCallSeqStart.getNode());
2286 Chain = CallSeqStart = NewCallSeqStart;
2287 ArgOffset += PtrByteSize;
2291 // Copy entire object into memory. There are cases where gcc-generated
2292 // code assumes it is there, even if it could be put entirely into
2293 // registers. (This is not what the doc says.)
2294 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2295 CallSeqStart.getNode()->getOperand(0),
2296 Flags, DAG, Size, dl);
2297 // This must go outside the CALLSEQ_START..END.
2298 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2299 CallSeqStart.getNode()->getOperand(1));
2300 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2301 Chain = CallSeqStart = NewCallSeqStart;
2302 // And copy the pieces of it that fit into registers.
2303 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2304 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2305 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
2306 if (GPR_idx != NumGPRs) {
2307 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, NULL, 0);
2308 MemOpChains.push_back(Load.getValue(1));
2309 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2311 ArgOffset += PtrByteSize;
2313 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2320 switch (Arg.getValueType().getSimpleVT()) {
2321 default: assert(0 && "Unexpected ValueType for argument!");
2324 // Double word align in ELF
2325 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2326 if (GPR_idx != NumGPRs) {
2327 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2329 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2330 isPPC64, isTailCall, false, MemOpChains,
2331 TailCallArguments, dl);
2334 if (inMem || isMachoABI) {
2335 // Stack align in ELF
2336 if (isELF32_ABI && Align)
2337 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2339 ArgOffset += PtrByteSize;
2344 if (FPR_idx != NumFPRs) {
2345 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2348 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
2349 MemOpChains.push_back(Store);
2351 // Float varargs are always shadowed in available integer registers
2352 if (GPR_idx != NumGPRs) {
2353 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2354 MemOpChains.push_back(Load.getValue(1));
2355 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2358 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2359 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2360 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
2361 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2362 MemOpChains.push_back(Load.getValue(1));
2363 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2367 // If we have any FPRs remaining, we may also have GPRs remaining.
2368 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2371 if (GPR_idx != NumGPRs)
2373 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
2374 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
2379 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2380 isPPC64, isTailCall, false, MemOpChains,
2381 TailCallArguments, dl);
2384 if (inMem || isMachoABI) {
2385 // Stack align in ELF
2386 if (isELF32_ABI && Align)
2387 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2391 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2399 // These go aligned on the stack, or in the corresponding R registers
2400 // when within range. The Darwin PPC ABI doc claims they also go in
2401 // V registers; in fact gcc does this only for arguments that are
2402 // prototyped, not for those that match the ... We do it for all
2403 // arguments, seems to work.
2404 while (ArgOffset % 16 !=0) {
2405 ArgOffset += PtrByteSize;
2406 if (GPR_idx != NumGPRs)
2409 // We could elide this store in the case where the object fits
2410 // entirely in R registers. Maybe later.
2411 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2412 DAG.getConstant(ArgOffset, PtrVT));
2413 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
2414 MemOpChains.push_back(Store);
2415 if (VR_idx != NumVRs) {
2416 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, NULL, 0);
2417 MemOpChains.push_back(Load.getValue(1));
2418 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2421 for (unsigned i=0; i<16; i+=PtrByteSize) {
2422 if (GPR_idx == NumGPRs)
2424 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
2425 DAG.getConstant(i, PtrVT));
2426 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, NULL, 0);
2427 MemOpChains.push_back(Load.getValue(1));
2428 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2433 // Non-varargs Altivec params generally go in registers, but have
2434 // stack space allocated at the end.
2435 if (VR_idx != NumVRs) {
2436 // Doesn't have GPR space allocated.
2437 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2438 } else if (nAltivecParamsAtEnd==0) {
2439 // We are emitting Altivec params in order.
2440 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2441 isPPC64, isTailCall, true, MemOpChains,
2442 TailCallArguments, dl);
2448 // If all Altivec parameters fit in registers, as they usually do,
2449 // they get stack space following the non-Altivec parameters. We
2450 // don't track this here because nobody below needs it.
2451 // If there are more Altivec parameters than fit in registers emit
2453 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2455 // Offset is aligned; skip 1st 12 params which go in V registers.
2456 ArgOffset = ((ArgOffset+15)/16)*16;
2458 for (unsigned i = 0; i != NumOps; ++i) {
2459 SDValue Arg = TheCall->getArg(i);
2460 MVT ArgType = Arg.getValueType();
2461 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2462 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2465 // We are emitting Altivec params in order.
2466 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2467 isPPC64, isTailCall, true, MemOpChains,
2468 TailCallArguments, dl);
2475 if (!MemOpChains.empty())
2476 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2477 &MemOpChains[0], MemOpChains.size());
2479 // Build a sequence of copy-to-reg nodes chained together with token chain
2480 // and flag operands which copy the outgoing args into the appropriate regs.
2482 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2483 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2484 RegsToPass[i].second, InFlag);
2485 InFlag = Chain.getValue(1);
2488 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2489 if (isVarArg && isELF32_ABI) {
2490 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, dl, MVT::i32), 0);
2491 Chain = DAG.getCopyToReg(Chain, dl, PPC::CR1EQ, SetCR, InFlag);
2492 InFlag = Chain.getValue(1);
2495 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2496 // might overwrite each other in case of tail call optimization.
2498 SmallVector<SDValue, 8> MemOpChains2;
2499 // Do not flag preceeding copytoreg stuff together with the following stuff.
2501 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2503 if (!MemOpChains2.empty())
2504 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2505 &MemOpChains2[0], MemOpChains2.size());
2507 // Store the return address to the appropriate stack slot.
2508 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2509 isPPC64, isMachoABI, dl);
2512 // Emit callseq_end just before tailcall node.
2514 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2515 DAG.getIntPtrConstant(0, true), InFlag);
2516 InFlag = Chain.getValue(1);
2519 std::vector<MVT> NodeTys;
2520 NodeTys.push_back(MVT::Other); // Returns a chain
2521 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2523 SmallVector<SDValue, 8> Ops;
2524 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2526 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2527 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2528 // node so that legalize doesn't hack it.
2529 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2530 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2531 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2532 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2533 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2534 // If this is an absolute destination address, use the munged value.
2535 Callee = SDValue(Dest, 0);
2537 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2538 // to do the call, we can't use PPCISD::CALL.
2539 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2540 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2541 2 + (InFlag.getNode() != 0));
2542 InFlag = Chain.getValue(1);
2544 // Copy the callee address into R12/X12 on darwin.
2546 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2547 Chain = DAG.getCopyToReg(Chain, dl, Reg, Callee, InFlag);
2548 InFlag = Chain.getValue(1);
2552 NodeTys.push_back(MVT::Other);
2553 NodeTys.push_back(MVT::Flag);
2554 Ops.push_back(Chain);
2555 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2557 // Add CTR register as callee so a bctr can be emitted later.
2559 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2562 // If this is a direct call, pass the chain and the callee.
2563 if (Callee.getNode()) {
2564 Ops.push_back(Chain);
2565 Ops.push_back(Callee);
2567 // If this is a tail call add stack pointer delta.
2569 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2571 // Add argument registers to the end of the list so that they are known live
2573 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2574 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2575 RegsToPass[i].second.getValueType()));
2577 // When performing tail call optimization the callee pops its arguments off
2578 // the stack. Account for this here so these bytes can be pushed back on in
2579 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2580 int BytesCalleePops =
2581 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2583 if (InFlag.getNode())
2584 Ops.push_back(InFlag);
2588 assert(InFlag.getNode() &&
2589 "Flag must be set. Depend on flag being set in LowerRET");
2590 Chain = DAG.getNode(PPCISD::TAILCALL, dl,
2591 TheCall->getVTList(), &Ops[0], Ops.size());
2592 return SDValue(Chain.getNode(), Op.getResNo());
2595 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2596 InFlag = Chain.getValue(1);
2598 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2599 DAG.getIntPtrConstant(BytesCalleePops, true),
2601 if (TheCall->getValueType(0) != MVT::Other)
2602 InFlag = Chain.getValue(1);
2604 SmallVector<SDValue, 16> ResultVals;
2605 SmallVector<CCValAssign, 16> RVLocs;
2606 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2607 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
2608 CCInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
2610 // Copy all of the result registers out of their specified physreg.
2611 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2612 CCValAssign &VA = RVLocs[i];
2613 MVT VT = VA.getValVT();
2614 assert(VA.isRegLoc() && "Can only return in registers!");
2615 Chain = DAG.getCopyFromReg(Chain, dl,
2616 VA.getLocReg(), VT, InFlag).getValue(1);
2617 ResultVals.push_back(Chain.getValue(0));
2618 InFlag = Chain.getValue(2);
2621 // If the function returns void, just return the chain.
2625 // Otherwise, merge everything together with a MERGE_VALUES node.
2626 ResultVals.push_back(Chain);
2627 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
2628 &ResultVals[0], ResultVals.size());
2629 return Res.getValue(Op.getResNo());
2632 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
2633 TargetMachine &TM) {
2634 SmallVector<CCValAssign, 16> RVLocs;
2635 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2636 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2637 DebugLoc dl = Op.getDebugLoc();
2638 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2639 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
2641 // If this is the first return lowered for this function, add the regs to the
2642 // liveout set for the function.
2643 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2644 for (unsigned i = 0; i != RVLocs.size(); ++i)
2645 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2648 SDValue Chain = Op.getOperand(0);
2650 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
2651 if (Chain.getOpcode() == PPCISD::TAILCALL) {
2652 SDValue TailCall = Chain;
2653 SDValue TargetAddress = TailCall.getOperand(1);
2654 SDValue StackAdjustment = TailCall.getOperand(2);
2656 assert(((TargetAddress.getOpcode() == ISD::Register &&
2657 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
2658 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
2659 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
2660 isa<ConstantSDNode>(TargetAddress)) &&
2661 "Expecting an global address, external symbol, absolute value or register");
2663 assert(StackAdjustment.getOpcode() == ISD::Constant &&
2664 "Expecting a const value");
2666 SmallVector<SDValue,8> Operands;
2667 Operands.push_back(Chain.getOperand(0));
2668 Operands.push_back(TargetAddress);
2669 Operands.push_back(StackAdjustment);
2670 // Copy registers used by the call. Last operand is a flag so it is not
2672 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
2673 Operands.push_back(Chain.getOperand(i));
2675 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Operands[0],
2681 // Copy the result values into the output registers.
2682 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2683 CCValAssign &VA = RVLocs[i];
2684 assert(VA.isRegLoc() && "Can only return in registers!");
2685 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2686 Op.getOperand(i*2+1), Flag);
2687 Flag = Chain.getValue(1);
2691 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
2693 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
2696 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
2697 const PPCSubtarget &Subtarget) {
2698 // When we pop the dynamic allocation we need to restore the SP link.
2699 DebugLoc dl = Op.getDebugLoc();
2701 // Get the corect type for pointers.
2702 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2704 // Construct the stack pointer operand.
2705 bool IsPPC64 = Subtarget.isPPC64();
2706 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2707 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
2709 // Get the operands for the STACKRESTORE.
2710 SDValue Chain = Op.getOperand(0);
2711 SDValue SaveSP = Op.getOperand(1);
2713 // Load the old link SP.
2714 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr, NULL, 0);
2716 // Restore the stack pointer.
2717 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
2719 // Store the old link SP.
2720 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, NULL, 0);
2726 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
2727 MachineFunction &MF = DAG.getMachineFunction();
2728 bool IsPPC64 = PPCSubTarget.isPPC64();
2729 bool isMachoABI = PPCSubTarget.isMachoABI();
2730 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2732 // Get current frame pointer save index. The users of this index will be
2733 // primarily DYNALLOC instructions.
2734 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2735 int RASI = FI->getReturnAddrSaveIndex();
2737 // If the frame pointer save index hasn't been defined yet.
2739 // Find out what the fix offset of the frame pointer save area.
2740 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
2741 // Allocate the frame index for frame pointer save area.
2742 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
2744 FI->setReturnAddrSaveIndex(RASI);
2746 return DAG.getFrameIndex(RASI, PtrVT);
2750 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
2751 MachineFunction &MF = DAG.getMachineFunction();
2752 bool IsPPC64 = PPCSubTarget.isPPC64();
2753 bool isMachoABI = PPCSubTarget.isMachoABI();
2754 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2756 // Get current frame pointer save index. The users of this index will be
2757 // primarily DYNALLOC instructions.
2758 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2759 int FPSI = FI->getFramePointerSaveIndex();
2761 // If the frame pointer save index hasn't been defined yet.
2763 // Find out what the fix offset of the frame pointer save area.
2764 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2766 // Allocate the frame index for frame pointer save area.
2767 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2769 FI->setFramePointerSaveIndex(FPSI);
2771 return DAG.getFrameIndex(FPSI, PtrVT);
2774 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
2776 const PPCSubtarget &Subtarget) {
2778 SDValue Chain = Op.getOperand(0);
2779 SDValue Size = Op.getOperand(1);
2780 DebugLoc dl = Op.getDebugLoc();
2782 // Get the corect type for pointers.
2783 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2785 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
2786 DAG.getConstant(0, PtrVT), Size);
2787 // Construct a node for the frame pointer save index.
2788 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
2789 // Build a DYNALLOC node.
2790 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
2791 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2792 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
2795 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2797 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2798 // Not FP? Not a fsel.
2799 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
2800 !Op.getOperand(2).getValueType().isFloatingPoint())
2803 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2805 // Cannot handle SETEQ/SETNE.
2806 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDValue();
2808 MVT ResVT = Op.getValueType();
2809 MVT CmpVT = Op.getOperand(0).getValueType();
2810 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2811 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
2812 DebugLoc dl = Op.getDebugLoc();
2814 // If the RHS of the comparison is a 0.0, we don't need to do the
2815 // subtraction at all.
2816 if (isFloatingPointZero(RHS))
2818 default: break; // SETUO etc aren't handled by fsel.
2821 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2824 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2825 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
2826 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
2829 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2832 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2833 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
2834 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
2835 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
2840 default: break; // SETUO etc aren't handled by fsel.
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, FV, TV);
2849 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
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, TV, FV);
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, FV, TV);
2861 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
2862 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2863 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
2864 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
2869 // FIXME: Split this code up when LegalizeDAGTypes lands.
2870 SDValue PPCTargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG,
2872 assert(Op.getOperand(0).getValueType().isFloatingPoint());
2873 SDValue Src = Op.getOperand(0);
2874 if (Src.getValueType() == MVT::f32)
2875 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
2878 switch (Op.getValueType().getSimpleVT()) {
2879 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2881 Tmp = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Src);
2884 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
2888 // Convert the FP value to an int value through memory.
2889 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
2891 // Emit a store to the stack slot.
2892 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr, NULL, 0);
2894 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2896 if (Op.getValueType() == MVT::i32)
2897 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
2898 DAG.getConstant(4, FIPtr.getValueType()));
2899 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, NULL, 0);
2902 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2903 DebugLoc dl = Op.getDebugLoc();
2904 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2905 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2908 if (Op.getOperand(0).getValueType() == MVT::i64) {
2909 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, dl,
2910 MVT::f64, Op.getOperand(0));
2911 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
2912 if (Op.getValueType() == MVT::f32)
2913 FP = DAG.getNode(ISD::FP_ROUND, dl,
2914 MVT::f32, FP, DAG.getIntPtrConstant(0));
2918 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2919 "Unhandled SINT_TO_FP type in custom expander!");
2920 // Since we only generate this in 64-bit mode, we can take advantage of
2921 // 64-bit registers. In particular, sign extend the input value into the
2922 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2923 // then lfd it and fcfid it.
2924 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2925 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2926 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2927 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2929 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
2932 // STD the extended value into the stack slot.
2933 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
2934 MachineMemOperand::MOStore, 0, 8, 8);
2935 SDValue Store = DAG.getNode(PPCISD::STD_32, dl, MVT::Other,
2936 DAG.getEntryNode(), Ext64, FIdx,
2937 DAG.getMemOperand(MO));
2938 // Load the value as a double.
2939 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, NULL, 0);
2941 // FCFID it and return it.
2942 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
2943 if (Op.getValueType() == MVT::f32)
2944 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
2948 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
2949 DebugLoc dl = Op.getDebugLoc();
2951 The rounding mode is in bits 30:31 of FPSR, and has the following
2958 FLT_ROUNDS, on the other hand, expects the following:
2965 To perform the conversion, we do:
2966 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2969 MachineFunction &MF = DAG.getMachineFunction();
2970 MVT VT = Op.getValueType();
2971 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2972 std::vector<MVT> NodeTys;
2973 SDValue MFFSreg, InFlag;
2975 // Save FP Control Word to register
2976 NodeTys.push_back(MVT::f64); // return register
2977 NodeTys.push_back(MVT::Flag); // unused in this context
2978 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
2980 // Save FP register to stack slot
2981 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2982 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2983 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
2984 StackSlot, NULL, 0);
2986 // Load FP Control Word from low 32 bits of stack slot.
2987 SDValue Four = DAG.getConstant(4, PtrVT);
2988 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
2989 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, NULL, 0);
2991 // Transform as necessary
2993 DAG.getNode(ISD::AND, dl, MVT::i32,
2994 CWD, DAG.getConstant(3, MVT::i32));
2996 DAG.getNode(ISD::SRL, dl, MVT::i32,
2997 DAG.getNode(ISD::AND, dl, MVT::i32,
2998 DAG.getNode(ISD::XOR, dl, MVT::i32,
2999 CWD, DAG.getConstant(3, MVT::i32)),
3000 DAG.getConstant(3, MVT::i32)),
3001 DAG.getConstant(1, MVT::i32));
3004 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3006 return DAG.getNode((VT.getSizeInBits() < 16 ?
3007 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3010 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3011 MVT VT = Op.getValueType();
3012 unsigned BitWidth = VT.getSizeInBits();
3013 DebugLoc dl = Op.getDebugLoc();
3014 assert(Op.getNumOperands() == 3 &&
3015 VT == Op.getOperand(1).getValueType() &&
3018 // Expand into a bunch of logical ops. Note that these ops
3019 // depend on the PPC behavior for oversized shift amounts.
3020 SDValue Lo = Op.getOperand(0);
3021 SDValue Hi = Op.getOperand(1);
3022 SDValue Amt = Op.getOperand(2);
3023 MVT AmtVT = Amt.getValueType();
3025 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3026 DAG.getConstant(BitWidth, AmtVT), Amt);
3027 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3028 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3029 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3030 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3031 DAG.getConstant(-BitWidth, AmtVT));
3032 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3033 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3034 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3035 SDValue OutOps[] = { OutLo, OutHi };
3036 return DAG.getMergeValues(OutOps, 2, dl);
3039 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3040 MVT VT = Op.getValueType();
3041 DebugLoc dl = Op.getDebugLoc();
3042 unsigned BitWidth = VT.getSizeInBits();
3043 assert(Op.getNumOperands() == 3 &&
3044 VT == Op.getOperand(1).getValueType() &&
3047 // Expand into a bunch of logical ops. Note that these ops
3048 // depend on the PPC behavior for oversized shift amounts.
3049 SDValue Lo = Op.getOperand(0);
3050 SDValue Hi = Op.getOperand(1);
3051 SDValue Amt = Op.getOperand(2);
3052 MVT AmtVT = Amt.getValueType();
3054 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3055 DAG.getConstant(BitWidth, AmtVT), Amt);
3056 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3057 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3058 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3059 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3060 DAG.getConstant(-BitWidth, AmtVT));
3061 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3062 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3063 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3064 SDValue OutOps[] = { OutLo, OutHi };
3065 return DAG.getMergeValues(OutOps, 2, dl);
3068 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3069 DebugLoc dl = Op.getDebugLoc();
3070 MVT VT = Op.getValueType();
3071 unsigned BitWidth = VT.getSizeInBits();
3072 assert(Op.getNumOperands() == 3 &&
3073 VT == Op.getOperand(1).getValueType() &&
3076 // Expand into a bunch of logical ops, followed by a select_cc.
3077 SDValue Lo = Op.getOperand(0);
3078 SDValue Hi = Op.getOperand(1);
3079 SDValue Amt = Op.getOperand(2);
3080 MVT AmtVT = Amt.getValueType();
3082 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3083 DAG.getConstant(BitWidth, AmtVT), Amt);
3084 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3085 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3086 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3087 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3088 DAG.getConstant(-BitWidth, AmtVT));
3089 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3090 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3091 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3092 Tmp4, Tmp6, ISD::SETLE);
3093 SDValue OutOps[] = { OutLo, OutHi };
3094 return DAG.getMergeValues(OutOps, 2, dl);
3097 //===----------------------------------------------------------------------===//
3098 // Vector related lowering.
3101 /// BuildSplatI - Build a canonical splati of Val with an element size of
3102 /// SplatSize. Cast the result to VT.
3103 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3104 SelectionDAG &DAG, DebugLoc dl) {
3105 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3107 static const MVT VTys[] = { // canonical VT to use for each size.
3108 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3111 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3113 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3117 MVT CanonicalVT = VTys[SplatSize-1];
3119 // Build a canonical splat for this value.
3120 SDValue Elt = DAG.getConstant(Val, CanonicalVT.getVectorElementType());
3121 SmallVector<SDValue, 8> Ops;
3122 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3123 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3124 &Ops[0], Ops.size());
3125 return DAG.getNode(ISD::BIT_CONVERT, dl, ReqVT, Res);
3128 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3129 /// specified intrinsic ID.
3130 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3131 SelectionDAG &DAG, DebugLoc dl,
3132 MVT DestVT = MVT::Other) {
3133 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3134 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3135 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3138 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3139 /// specified intrinsic ID.
3140 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3141 SDValue Op2, SelectionDAG &DAG,
3142 DebugLoc dl, MVT DestVT = MVT::Other) {
3143 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3144 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3145 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3149 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3150 /// amount. The result has the specified value type.
3151 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3152 MVT VT, SelectionDAG &DAG, DebugLoc dl) {
3153 // Force LHS/RHS to be the right type.
3154 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, LHS);
3155 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, RHS);
3158 for (unsigned i = 0; i != 16; ++i)
3160 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3161 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3164 // If this is a case we can't handle, return null and let the default
3165 // expansion code take care of it. If we CAN select this case, and if it
3166 // selects to a single instruction, return Op. Otherwise, if we can codegen
3167 // this case more efficiently than a constant pool load, lower it to the
3168 // sequence of ops that should be used.
3169 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3170 DebugLoc dl = Op.getDebugLoc();
3171 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3172 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3174 // Check if this is a splat of a constant value.
3175 APInt APSplatBits, APSplatUndef;
3176 unsigned SplatBitSize;
3178 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3179 HasAnyUndefs) || SplatBitSize > 32)
3182 unsigned SplatBits = APSplatBits.getZExtValue();
3183 unsigned SplatUndef = APSplatUndef.getZExtValue();
3184 unsigned SplatSize = SplatBitSize / 8;
3186 // First, handle single instruction cases.
3189 if (SplatBits == 0) {
3190 // Canonicalize all zero vectors to be v4i32.
3191 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3192 SDValue Z = DAG.getConstant(0, MVT::i32);
3193 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3194 Op = DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Z);
3199 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3200 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3202 if (SextVal >= -16 && SextVal <= 15)
3203 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3206 // Two instruction sequences.
3208 // If this value is in the range [-32,30] and is even, use:
3209 // tmp = VSPLTI[bhw], result = add tmp, tmp
3210 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3211 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3212 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3213 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3216 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3217 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3219 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3220 // Make -1 and vspltisw -1:
3221 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3223 // Make the VSLW intrinsic, computing 0x8000_0000.
3224 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3227 // xor by OnesV to invert it.
3228 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3229 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3232 // Check to see if this is a wide variety of vsplti*, binop self cases.
3233 static const signed char SplatCsts[] = {
3234 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3235 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3238 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3239 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3240 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3241 int i = SplatCsts[idx];
3243 // Figure out what shift amount will be used by altivec if shifted by i in
3245 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3247 // vsplti + shl self.
3248 if (SextVal == (i << (int)TypeShiftAmt)) {
3249 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3250 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3251 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3252 Intrinsic::ppc_altivec_vslw
3254 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3255 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3258 // vsplti + srl self.
3259 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3260 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3261 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3262 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3263 Intrinsic::ppc_altivec_vsrw
3265 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3266 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3269 // vsplti + sra self.
3270 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3271 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3272 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3273 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3274 Intrinsic::ppc_altivec_vsraw
3276 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3277 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3280 // vsplti + rol self.
3281 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3282 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3283 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3284 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3285 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3286 Intrinsic::ppc_altivec_vrlw
3288 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3289 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3292 // t = vsplti c, result = vsldoi t, t, 1
3293 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3294 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3295 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3297 // t = vsplti c, result = vsldoi t, t, 2
3298 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3299 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3300 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3302 // t = vsplti c, result = vsldoi t, t, 3
3303 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3304 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3305 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3309 // Three instruction sequences.
3311 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3312 if (SextVal >= 0 && SextVal <= 31) {
3313 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3314 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3315 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3316 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3318 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3319 if (SextVal >= -31 && SextVal <= 0) {
3320 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3321 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3322 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3323 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3329 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3330 /// the specified operations to build the shuffle.
3331 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3332 SDValue RHS, SelectionDAG &DAG,
3334 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3335 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3336 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3339 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3351 if (OpNum == OP_COPY) {
3352 if (LHSID == (1*9+2)*9+3) return LHS;
3353 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3357 SDValue OpLHS, OpRHS;
3358 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3359 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3363 default: assert(0 && "Unknown i32 permute!");
3365 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3366 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3367 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3368 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3371 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3372 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3373 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3374 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3377 for (unsigned i = 0; i != 16; ++i)
3378 ShufIdxs[i] = (i&3)+0;
3381 for (unsigned i = 0; i != 16; ++i)
3382 ShufIdxs[i] = (i&3)+4;
3385 for (unsigned i = 0; i != 16; ++i)
3386 ShufIdxs[i] = (i&3)+8;
3389 for (unsigned i = 0; i != 16; ++i)
3390 ShufIdxs[i] = (i&3)+12;
3393 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
3395 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
3397 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
3399 MVT VT = OpLHS.getValueType();
3400 OpLHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpLHS);
3401 OpRHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpRHS);
3402 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
3403 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3406 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3407 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3408 /// return the code it can be lowered into. Worst case, it can always be
3409 /// lowered into a vperm.
3410 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3411 SelectionDAG &DAG) {
3412 DebugLoc dl = Op.getDebugLoc();
3413 SDValue V1 = Op.getOperand(0);
3414 SDValue V2 = Op.getOperand(1);
3415 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3416 const int *PermMask = SVOp->getMask();
3417 MVT VT = Op.getValueType();
3419 // Cases that are handled by instructions that take permute immediates
3420 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3421 // selected by the instruction selector.
3422 if (V2.getOpcode() == ISD::UNDEF) {
3423 if (PPC::isSplatShuffleMask(SVOp, 1) ||
3424 PPC::isSplatShuffleMask(SVOp, 2) ||
3425 PPC::isSplatShuffleMask(SVOp, 4) ||
3426 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
3427 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
3428 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
3429 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
3430 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
3431 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
3432 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
3433 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
3434 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
3439 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3440 // and produce a fixed permutation. If any of these match, do not lower to
3442 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
3443 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
3444 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
3445 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
3446 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
3447 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
3448 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
3449 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
3450 PPC::isVMRGHShuffleMask(SVOp, 4, false))
3453 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3454 // perfect shuffle table to emit an optimal matching sequence.
3455 unsigned PFIndexes[4];
3456 bool isFourElementShuffle = true;
3457 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3458 unsigned EltNo = 8; // Start out undef.
3459 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3460 if (PermMask[i*4+j] < 0)
3461 continue; // Undef, ignore it.
3463 unsigned ByteSource = PermMask[i*4+j];
3464 if ((ByteSource & 3) != j) {
3465 isFourElementShuffle = false;
3470 EltNo = ByteSource/4;
3471 } else if (EltNo != ByteSource/4) {
3472 isFourElementShuffle = false;
3476 PFIndexes[i] = EltNo;
3479 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3480 // perfect shuffle vector to determine if it is cost effective to do this as
3481 // discrete instructions, or whether we should use a vperm.
3482 if (isFourElementShuffle) {
3483 // Compute the index in the perfect shuffle table.
3484 unsigned PFTableIndex =
3485 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3487 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3488 unsigned Cost = (PFEntry >> 30);
3490 // Determining when to avoid vperm is tricky. Many things affect the cost
3491 // of vperm, particularly how many times the perm mask needs to be computed.
3492 // For example, if the perm mask can be hoisted out of a loop or is already
3493 // used (perhaps because there are multiple permutes with the same shuffle
3494 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3495 // the loop requires an extra register.
3497 // As a compromise, we only emit discrete instructions if the shuffle can be
3498 // generated in 3 or fewer operations. When we have loop information
3499 // available, if this block is within a loop, we should avoid using vperm
3500 // for 3-operation perms and use a constant pool load instead.
3502 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3505 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3506 // vector that will get spilled to the constant pool.
3507 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3509 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3510 // that it is in input element units, not in bytes. Convert now.
3511 MVT EltVT = V1.getValueType().getVectorElementType();
3512 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
3514 SmallVector<SDValue, 16> ResultMask;
3515 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
3516 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
3518 for (unsigned j = 0; j != BytesPerElement; ++j)
3519 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3523 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
3524 &ResultMask[0], ResultMask.size());
3525 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
3528 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3529 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3530 /// information about the intrinsic.
3531 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
3533 unsigned IntrinsicID =
3534 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
3537 switch (IntrinsicID) {
3538 default: return false;
3539 // Comparison predicates.
3540 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3541 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3542 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3543 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3544 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3545 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3546 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3547 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3548 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3549 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3550 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3551 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3552 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3554 // Normal Comparisons.
3555 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3556 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3557 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3558 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3559 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3560 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3561 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3562 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3563 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3564 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3565 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3566 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3567 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3572 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3573 /// lower, do it, otherwise return null.
3574 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
3575 SelectionDAG &DAG) {
3576 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3577 // opcode number of the comparison.
3578 DebugLoc dl = Op.getDebugLoc();
3581 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3582 return SDValue(); // Don't custom lower most intrinsics.
3584 // If this is a non-dot comparison, make the VCMP node and we are done.
3586 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
3587 Op.getOperand(1), Op.getOperand(2),
3588 DAG.getConstant(CompareOpc, MVT::i32));
3589 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Tmp);
3592 // Create the PPCISD altivec 'dot' comparison node.
3594 Op.getOperand(2), // LHS
3595 Op.getOperand(3), // RHS
3596 DAG.getConstant(CompareOpc, MVT::i32)
3598 std::vector<MVT> VTs;
3599 VTs.push_back(Op.getOperand(2).getValueType());
3600 VTs.push_back(MVT::Flag);
3601 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
3603 // Now that we have the comparison, emit a copy from the CR to a GPR.
3604 // This is flagged to the above dot comparison.
3605 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
3606 DAG.getRegister(PPC::CR6, MVT::i32),
3607 CompNode.getValue(1));
3609 // Unpack the result based on how the target uses it.
3610 unsigned BitNo; // Bit # of CR6.
3611 bool InvertBit; // Invert result?
3612 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
3613 default: // Can't happen, don't crash on invalid number though.
3614 case 0: // Return the value of the EQ bit of CR6.
3615 BitNo = 0; InvertBit = false;
3617 case 1: // Return the inverted value of the EQ bit of CR6.
3618 BitNo = 0; InvertBit = true;
3620 case 2: // Return the value of the LT bit of CR6.
3621 BitNo = 2; InvertBit = false;
3623 case 3: // Return the inverted value of the LT bit of CR6.
3624 BitNo = 2; InvertBit = true;
3628 // Shift the bit into the low position.
3629 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
3630 DAG.getConstant(8-(3-BitNo), MVT::i32));
3632 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
3633 DAG.getConstant(1, MVT::i32));
3635 // If we are supposed to, toggle the bit.
3637 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
3638 DAG.getConstant(1, MVT::i32));
3642 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
3643 SelectionDAG &DAG) {
3644 DebugLoc dl = Op.getDebugLoc();
3645 // Create a stack slot that is 16-byte aligned.
3646 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3647 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3648 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3649 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3651 // Store the input value into Value#0 of the stack slot.
3652 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
3653 Op.getOperand(0), FIdx, NULL, 0);
3655 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, NULL, 0);
3658 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
3659 DebugLoc dl = Op.getDebugLoc();
3660 if (Op.getValueType() == MVT::v4i32) {
3661 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3663 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
3664 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
3666 SDValue RHSSwap = // = vrlw RHS, 16
3667 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
3669 // Shrinkify inputs to v8i16.
3670 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, LHS);
3671 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHS);
3672 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHSSwap);
3674 // Low parts multiplied together, generating 32-bit results (we ignore the
3676 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3677 LHS, RHS, DAG, dl, MVT::v4i32);
3679 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3680 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
3681 // Shift the high parts up 16 bits.
3682 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
3684 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
3685 } else if (Op.getValueType() == MVT::v8i16) {
3686 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3688 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
3690 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3691 LHS, RHS, Zero, DAG, dl);
3692 } else if (Op.getValueType() == MVT::v16i8) {
3693 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3695 // Multiply the even 8-bit parts, producing 16-bit sums.
3696 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3697 LHS, RHS, DAG, dl, MVT::v8i16);
3698 EvenParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, EvenParts);
3700 // Multiply the odd 8-bit parts, producing 16-bit sums.
3701 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3702 LHS, RHS, DAG, dl, MVT::v8i16);
3703 OddParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OddParts);
3705 // Merge the results together.
3707 for (unsigned i = 0; i != 8; ++i) {
3709 Ops[i*2+1] = 2*i+1+16;
3711 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
3713 assert(0 && "Unknown mul to lower!");
3718 /// LowerOperation - Provide custom lowering hooks for some operations.
3720 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3721 switch (Op.getOpcode()) {
3722 default: assert(0 && "Wasn't expecting to be able to lower this!");
3723 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3724 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3725 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3726 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3727 case ISD::SETCC: return LowerSETCC(Op, DAG);
3728 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
3730 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3731 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3734 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3735 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3737 case ISD::FORMAL_ARGUMENTS:
3738 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3739 VarArgsStackOffset, VarArgsNumGPR,
3740 VarArgsNumFPR, PPCSubTarget);
3742 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3743 getTargetMachine());
3744 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3745 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3746 case ISD::DYNAMIC_STACKALLOC:
3747 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3749 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3750 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG,
3752 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3753 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3755 // Lower 64-bit shifts.
3756 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3757 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3758 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3760 // Vector-related lowering.
3761 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3762 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3763 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3764 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3765 case ISD::MUL: return LowerMUL(Op, DAG);
3767 // Frame & Return address.
3768 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3769 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3774 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
3775 SmallVectorImpl<SDValue>&Results,
3776 SelectionDAG &DAG) {
3777 DebugLoc dl = N->getDebugLoc();
3778 switch (N->getOpcode()) {
3780 assert(false && "Do not know how to custom type legalize this operation!");
3782 case ISD::FP_ROUND_INREG: {
3783 assert(N->getValueType(0) == MVT::ppcf128);
3784 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
3785 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
3786 MVT::f64, N->getOperand(0),
3787 DAG.getIntPtrConstant(0));
3788 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
3789 MVT::f64, N->getOperand(0),
3790 DAG.getIntPtrConstant(1));
3792 // This sequence changes FPSCR to do round-to-zero, adds the two halves
3793 // of the long double, and puts FPSCR back the way it was. We do not
3794 // actually model FPSCR.
3795 std::vector<MVT> NodeTys;
3796 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
3798 NodeTys.push_back(MVT::f64); // Return register
3799 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
3800 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3801 MFFSreg = Result.getValue(0);
3802 InFlag = Result.getValue(1);
3805 NodeTys.push_back(MVT::Flag); // Returns a flag
3806 Ops[0] = DAG.getConstant(31, MVT::i32);
3808 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
3809 InFlag = Result.getValue(0);
3812 NodeTys.push_back(MVT::Flag); // Returns a flag
3813 Ops[0] = DAG.getConstant(30, MVT::i32);
3815 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
3816 InFlag = Result.getValue(0);
3819 NodeTys.push_back(MVT::f64); // result of add
3820 NodeTys.push_back(MVT::Flag); // Returns a flag
3824 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
3825 FPreg = Result.getValue(0);
3826 InFlag = Result.getValue(1);
3829 NodeTys.push_back(MVT::f64);
3830 Ops[0] = DAG.getConstant(1, MVT::i32);
3834 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
3835 FPreg = Result.getValue(0);
3837 // We know the low half is about to be thrown away, so just use something
3839 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
3843 case ISD::FP_TO_SINT:
3844 Results.push_back(LowerFP_TO_SINT(SDValue(N, 0), DAG, dl));
3850 //===----------------------------------------------------------------------===//
3851 // Other Lowering Code
3852 //===----------------------------------------------------------------------===//
3855 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3856 bool is64bit, unsigned BinOpcode) const {
3857 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3858 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3860 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3861 MachineFunction *F = BB->getParent();
3862 MachineFunction::iterator It = BB;
3865 unsigned dest = MI->getOperand(0).getReg();
3866 unsigned ptrA = MI->getOperand(1).getReg();
3867 unsigned ptrB = MI->getOperand(2).getReg();
3868 unsigned incr = MI->getOperand(3).getReg();
3869 DebugLoc dl = MI->getDebugLoc();
3871 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3872 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3873 F->insert(It, loopMBB);
3874 F->insert(It, exitMBB);
3875 exitMBB->transferSuccessors(BB);
3877 MachineRegisterInfo &RegInfo = F->getRegInfo();
3878 unsigned TmpReg = (!BinOpcode) ? incr :
3879 RegInfo.createVirtualRegister(
3880 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3881 (const TargetRegisterClass *) &PPC::GPRCRegClass);
3885 // fallthrough --> loopMBB
3886 BB->addSuccessor(loopMBB);
3889 // l[wd]arx dest, ptr
3890 // add r0, dest, incr
3891 // st[wd]cx. r0, ptr
3893 // fallthrough --> exitMBB
3895 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
3896 .addReg(ptrA).addReg(ptrB);
3898 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
3899 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
3900 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
3901 BuildMI(BB, dl, TII->get(PPC::BCC))
3902 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
3903 BB->addSuccessor(loopMBB);
3904 BB->addSuccessor(exitMBB);
3913 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
3914 MachineBasicBlock *BB,
3915 bool is8bit, // operation
3916 unsigned BinOpcode) const {
3917 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3918 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3919 // In 64 bit mode we have to use 64 bits for addresses, even though the
3920 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
3921 // registers without caring whether they're 32 or 64, but here we're
3922 // doing actual arithmetic on the addresses.
3923 bool is64bit = PPCSubTarget.isPPC64();
3925 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3926 MachineFunction *F = BB->getParent();
3927 MachineFunction::iterator It = BB;
3930 unsigned dest = MI->getOperand(0).getReg();
3931 unsigned ptrA = MI->getOperand(1).getReg();
3932 unsigned ptrB = MI->getOperand(2).getReg();
3933 unsigned incr = MI->getOperand(3).getReg();
3934 DebugLoc dl = MI->getDebugLoc();
3936 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3937 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3938 F->insert(It, loopMBB);
3939 F->insert(It, exitMBB);
3940 exitMBB->transferSuccessors(BB);
3942 MachineRegisterInfo &RegInfo = F->getRegInfo();
3943 const TargetRegisterClass *RC =
3944 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3945 (const TargetRegisterClass *) &PPC::GPRCRegClass;
3946 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
3947 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
3948 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
3949 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
3950 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
3951 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
3952 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
3953 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
3954 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
3955 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
3956 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
3958 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
3962 // fallthrough --> loopMBB
3963 BB->addSuccessor(loopMBB);
3965 // The 4-byte load must be aligned, while a char or short may be
3966 // anywhere in the word. Hence all this nasty bookkeeping code.
3967 // add ptr1, ptrA, ptrB [copy if ptrA==0]
3968 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
3969 // xori shift, shift1, 24 [16]
3970 // rlwinm ptr, ptr1, 0, 0, 29
3971 // slw incr2, incr, shift
3972 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
3973 // slw mask, mask2, shift
3975 // lwarx tmpDest, ptr
3976 // add tmp, tmpDest, incr2
3977 // andc tmp2, tmpDest, mask
3978 // and tmp3, tmp, mask
3979 // or tmp4, tmp3, tmp2
3982 // fallthrough --> exitMBB
3983 // srw dest, tmpDest, shift
3985 if (ptrA!=PPC::R0) {
3986 Ptr1Reg = RegInfo.createVirtualRegister(RC);
3987 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
3988 .addReg(ptrA).addReg(ptrB);
3992 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
3993 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
3994 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
3995 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
3997 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
3998 .addReg(Ptr1Reg).addImm(0).addImm(61);
4000 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4001 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4002 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4003 .addReg(incr).addReg(ShiftReg);
4005 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4007 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4008 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4010 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4011 .addReg(Mask2Reg).addReg(ShiftReg);
4014 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4015 .addReg(PPC::R0).addReg(PtrReg);
4017 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4018 .addReg(Incr2Reg).addReg(TmpDestReg);
4019 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4020 .addReg(TmpDestReg).addReg(MaskReg);
4021 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4022 .addReg(TmpReg).addReg(MaskReg);
4023 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4024 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4025 BuildMI(BB, dl, TII->get(PPC::STWCX))
4026 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4027 BuildMI(BB, dl, TII->get(PPC::BCC))
4028 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4029 BB->addSuccessor(loopMBB);
4030 BB->addSuccessor(exitMBB);
4035 BuildMI(BB, dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4040 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4041 MachineBasicBlock *BB) const {
4042 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4044 // To "insert" these instructions we actually have to insert their
4045 // control-flow patterns.
4046 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4047 MachineFunction::iterator It = BB;
4050 MachineFunction *F = BB->getParent();
4052 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4053 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4054 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4055 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4056 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4058 // The incoming instruction knows the destination vreg to set, the
4059 // condition code register to branch on, the true/false values to
4060 // select between, and a branch opcode to use.
4065 // cmpTY ccX, r1, r2
4067 // fallthrough --> copy0MBB
4068 MachineBasicBlock *thisMBB = BB;
4069 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4070 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4071 unsigned SelectPred = MI->getOperand(4).getImm();
4072 DebugLoc dl = MI->getDebugLoc();
4073 BuildMI(BB, dl, TII->get(PPC::BCC))
4074 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4075 F->insert(It, copy0MBB);
4076 F->insert(It, sinkMBB);
4077 // Update machine-CFG edges by transferring all successors of the current
4078 // block to the new block which will contain the Phi node for the select.
4079 sinkMBB->transferSuccessors(BB);
4080 // Next, add the true and fallthrough blocks as its successors.
4081 BB->addSuccessor(copy0MBB);
4082 BB->addSuccessor(sinkMBB);
4085 // %FalseValue = ...
4086 // # fallthrough to sinkMBB
4089 // Update machine-CFG edges
4090 BB->addSuccessor(sinkMBB);
4093 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4096 BuildMI(BB, dl, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4097 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4098 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4100 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4101 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4102 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4103 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4104 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4105 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4106 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4107 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4109 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4110 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4111 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4112 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4113 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4114 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4115 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4116 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4118 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4119 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4120 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4121 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4122 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4123 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4124 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4125 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4127 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4128 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4129 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4130 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4131 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4132 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4133 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4134 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4136 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4137 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4138 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4139 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4140 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4141 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4142 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4143 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4145 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4146 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4147 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4148 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4149 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4150 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4151 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4152 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4154 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4155 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4156 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4157 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4158 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4159 BB = EmitAtomicBinary(MI, BB, false, 0);
4160 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4161 BB = EmitAtomicBinary(MI, BB, true, 0);
4163 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4164 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4165 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4167 unsigned dest = MI->getOperand(0).getReg();
4168 unsigned ptrA = MI->getOperand(1).getReg();
4169 unsigned ptrB = MI->getOperand(2).getReg();
4170 unsigned oldval = MI->getOperand(3).getReg();
4171 unsigned newval = MI->getOperand(4).getReg();
4172 DebugLoc dl = MI->getDebugLoc();
4174 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4175 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4176 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4177 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4178 F->insert(It, loop1MBB);
4179 F->insert(It, loop2MBB);
4180 F->insert(It, midMBB);
4181 F->insert(It, exitMBB);
4182 exitMBB->transferSuccessors(BB);
4186 // fallthrough --> loopMBB
4187 BB->addSuccessor(loop1MBB);
4190 // l[wd]arx dest, ptr
4191 // cmp[wd] dest, oldval
4194 // st[wd]cx. newval, ptr
4198 // st[wd]cx. dest, ptr
4201 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4202 .addReg(ptrA).addReg(ptrB);
4203 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4204 .addReg(oldval).addReg(dest);
4205 BuildMI(BB, dl, TII->get(PPC::BCC))
4206 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4207 BB->addSuccessor(loop2MBB);
4208 BB->addSuccessor(midMBB);
4211 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4212 .addReg(newval).addReg(ptrA).addReg(ptrB);
4213 BuildMI(BB, dl, TII->get(PPC::BCC))
4214 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4215 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4216 BB->addSuccessor(loop1MBB);
4217 BB->addSuccessor(exitMBB);
4220 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4221 .addReg(dest).addReg(ptrA).addReg(ptrB);
4222 BB->addSuccessor(exitMBB);
4227 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4228 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4229 // We must use 64-bit registers for addresses when targeting 64-bit,
4230 // since we're actually doing arithmetic on them. Other registers
4232 bool is64bit = PPCSubTarget.isPPC64();
4233 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4235 unsigned dest = MI->getOperand(0).getReg();
4236 unsigned ptrA = MI->getOperand(1).getReg();
4237 unsigned ptrB = MI->getOperand(2).getReg();
4238 unsigned oldval = MI->getOperand(3).getReg();
4239 unsigned newval = MI->getOperand(4).getReg();
4240 DebugLoc dl = MI->getDebugLoc();
4242 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4243 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4244 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4245 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4246 F->insert(It, loop1MBB);
4247 F->insert(It, loop2MBB);
4248 F->insert(It, midMBB);
4249 F->insert(It, exitMBB);
4250 exitMBB->transferSuccessors(BB);
4252 MachineRegisterInfo &RegInfo = F->getRegInfo();
4253 const TargetRegisterClass *RC =
4254 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4255 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4256 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4257 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4258 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4259 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4260 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4261 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4262 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4263 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4264 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4265 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4266 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4267 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4268 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4270 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4273 // fallthrough --> loopMBB
4274 BB->addSuccessor(loop1MBB);
4276 // The 4-byte load must be aligned, while a char or short may be
4277 // anywhere in the word. Hence all this nasty bookkeeping code.
4278 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4279 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4280 // xori shift, shift1, 24 [16]
4281 // rlwinm ptr, ptr1, 0, 0, 29
4282 // slw newval2, newval, shift
4283 // slw oldval2, oldval,shift
4284 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4285 // slw mask, mask2, shift
4286 // and newval3, newval2, mask
4287 // and oldval3, oldval2, mask
4289 // lwarx tmpDest, ptr
4290 // and tmp, tmpDest, mask
4291 // cmpw tmp, oldval3
4294 // andc tmp2, tmpDest, mask
4295 // or tmp4, tmp2, newval3
4300 // stwcx. tmpDest, ptr
4302 // srw dest, tmpDest, shift
4303 if (ptrA!=PPC::R0) {
4304 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4305 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4306 .addReg(ptrA).addReg(ptrB);
4310 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4311 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4312 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4313 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4315 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4316 .addReg(Ptr1Reg).addImm(0).addImm(61);
4318 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4319 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4320 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4321 .addReg(newval).addReg(ShiftReg);
4322 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4323 .addReg(oldval).addReg(ShiftReg);
4325 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4327 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4328 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
4329 .addReg(Mask3Reg).addImm(65535);
4331 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4332 .addReg(Mask2Reg).addReg(ShiftReg);
4333 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
4334 .addReg(NewVal2Reg).addReg(MaskReg);
4335 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
4336 .addReg(OldVal2Reg).addReg(MaskReg);
4339 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4340 .addReg(PPC::R0).addReg(PtrReg);
4341 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
4342 .addReg(TmpDestReg).addReg(MaskReg);
4343 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
4344 .addReg(TmpReg).addReg(OldVal3Reg);
4345 BuildMI(BB, dl, TII->get(PPC::BCC))
4346 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4347 BB->addSuccessor(loop2MBB);
4348 BB->addSuccessor(midMBB);
4351 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
4352 .addReg(TmpDestReg).addReg(MaskReg);
4353 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
4354 .addReg(Tmp2Reg).addReg(NewVal3Reg);
4355 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4356 .addReg(PPC::R0).addReg(PtrReg);
4357 BuildMI(BB, dl, TII->get(PPC::BCC))
4358 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4359 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4360 BB->addSuccessor(loop1MBB);
4361 BB->addSuccessor(exitMBB);
4364 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4365 .addReg(PPC::R0).addReg(PtrReg);
4366 BB->addSuccessor(exitMBB);
4371 BuildMI(BB, dl, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4373 assert(0 && "Unexpected instr type to insert");
4376 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4380 //===----------------------------------------------------------------------===//
4381 // Target Optimization Hooks
4382 //===----------------------------------------------------------------------===//
4384 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4385 DAGCombinerInfo &DCI) const {
4386 TargetMachine &TM = getTargetMachine();
4387 SelectionDAG &DAG = DCI.DAG;
4388 DebugLoc dl = N->getDebugLoc();
4389 switch (N->getOpcode()) {
4392 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4393 if (C->getZExtValue() == 0) // 0 << V -> 0.
4394 return N->getOperand(0);
4398 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4399 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4400 return N->getOperand(0);
4404 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4405 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4406 C->isAllOnesValue()) // -1 >>s V -> -1.
4407 return N->getOperand(0);
4411 case ISD::SINT_TO_FP:
4412 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4413 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4414 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4415 // We allow the src/dst to be either f32/f64, but the intermediate
4416 // type must be i64.
4417 if (N->getOperand(0).getValueType() == MVT::i64 &&
4418 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4419 SDValue Val = N->getOperand(0).getOperand(0);
4420 if (Val.getValueType() == MVT::f32) {
4421 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4422 DCI.AddToWorklist(Val.getNode());
4425 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
4426 DCI.AddToWorklist(Val.getNode());
4427 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
4428 DCI.AddToWorklist(Val.getNode());
4429 if (N->getValueType(0) == MVT::f32) {
4430 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
4431 DAG.getIntPtrConstant(0));
4432 DCI.AddToWorklist(Val.getNode());
4435 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4436 // If the intermediate type is i32, we can avoid the load/store here
4443 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4444 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4445 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4446 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4447 N->getOperand(1).getValueType() == MVT::i32 &&
4448 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4449 SDValue Val = N->getOperand(1).getOperand(0);
4450 if (Val.getValueType() == MVT::f32) {
4451 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4452 DCI.AddToWorklist(Val.getNode());
4454 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
4455 DCI.AddToWorklist(Val.getNode());
4457 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
4458 N->getOperand(2), N->getOperand(3));
4459 DCI.AddToWorklist(Val.getNode());
4463 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4464 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
4465 N->getOperand(1).getNode()->hasOneUse() &&
4466 (N->getOperand(1).getValueType() == MVT::i32 ||
4467 N->getOperand(1).getValueType() == MVT::i16)) {
4468 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4469 // Do an any-extend to 32-bits if this is a half-word input.
4470 if (BSwapOp.getValueType() == MVT::i16)
4471 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
4473 return DAG.getNode(PPCISD::STBRX, dl, MVT::Other, N->getOperand(0),
4474 BSwapOp, N->getOperand(2), N->getOperand(3),
4475 DAG.getValueType(N->getOperand(1).getValueType()));
4479 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4480 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4481 N->getOperand(0).hasOneUse() &&
4482 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4483 SDValue Load = N->getOperand(0);
4484 LoadSDNode *LD = cast<LoadSDNode>(Load);
4485 // Create the byte-swapping load.
4486 std::vector<MVT> VTs;
4487 VTs.push_back(MVT::i32);
4488 VTs.push_back(MVT::Other);
4489 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
4491 LD->getChain(), // Chain
4492 LD->getBasePtr(), // Ptr
4494 DAG.getValueType(N->getValueType(0)) // VT
4496 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, dl, VTs, Ops, 4);
4498 // If this is an i16 load, insert the truncate.
4499 SDValue ResVal = BSLoad;
4500 if (N->getValueType(0) == MVT::i16)
4501 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
4503 // First, combine the bswap away. This makes the value produced by the
4505 DCI.CombineTo(N, ResVal);
4507 // Next, combine the load away, we give it a bogus result value but a real
4508 // chain result. The result value is dead because the bswap is dead.
4509 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
4511 // Return N so it doesn't get rechecked!
4512 return SDValue(N, 0);
4516 case PPCISD::VCMP: {
4517 // If a VCMPo node already exists with exactly the same operands as this
4518 // node, use its result instead of this node (VCMPo computes both a CR6 and
4519 // a normal output).
4521 if (!N->getOperand(0).hasOneUse() &&
4522 !N->getOperand(1).hasOneUse() &&
4523 !N->getOperand(2).hasOneUse()) {
4525 // Scan all of the users of the LHS, looking for VCMPo's that match.
4526 SDNode *VCMPoNode = 0;
4528 SDNode *LHSN = N->getOperand(0).getNode();
4529 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
4531 if (UI->getOpcode() == PPCISD::VCMPo &&
4532 UI->getOperand(1) == N->getOperand(1) &&
4533 UI->getOperand(2) == N->getOperand(2) &&
4534 UI->getOperand(0) == N->getOperand(0)) {
4539 // If there is no VCMPo node, or if the flag value has a single use, don't
4541 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
4544 // Look at the (necessarily single) use of the flag value. If it has a
4545 // chain, this transformation is more complex. Note that multiple things
4546 // could use the value result, which we should ignore.
4547 SDNode *FlagUser = 0;
4548 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
4549 FlagUser == 0; ++UI) {
4550 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
4552 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
4553 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
4560 // If the user is a MFCR instruction, we know this is safe. Otherwise we
4561 // give up for right now.
4562 if (FlagUser->getOpcode() == PPCISD::MFCR)
4563 return SDValue(VCMPoNode, 0);
4568 // If this is a branch on an altivec predicate comparison, lower this so
4569 // that we don't have to do a MFCR: instead, branch directly on CR6. This
4570 // lowering is done pre-legalize, because the legalizer lowers the predicate
4571 // compare down to code that is difficult to reassemble.
4572 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
4573 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
4577 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
4578 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
4579 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
4580 assert(isDot && "Can't compare against a vector result!");
4582 // If this is a comparison against something other than 0/1, then we know
4583 // that the condition is never/always true.
4584 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
4585 if (Val != 0 && Val != 1) {
4586 if (CC == ISD::SETEQ) // Cond never true, remove branch.
4587 return N->getOperand(0);
4588 // Always !=, turn it into an unconditional branch.
4589 return DAG.getNode(ISD::BR, dl, MVT::Other,
4590 N->getOperand(0), N->getOperand(4));
4593 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
4595 // Create the PPCISD altivec 'dot' comparison node.
4596 std::vector<MVT> VTs;
4598 LHS.getOperand(2), // LHS of compare
4599 LHS.getOperand(3), // RHS of compare
4600 DAG.getConstant(CompareOpc, MVT::i32)
4602 VTs.push_back(LHS.getOperand(2).getValueType());
4603 VTs.push_back(MVT::Flag);
4604 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4606 // Unpack the result based on how the target uses it.
4607 PPC::Predicate CompOpc;
4608 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
4609 default: // Can't happen, don't crash on invalid number though.
4610 case 0: // Branch on the value of the EQ bit of CR6.
4611 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
4613 case 1: // Branch on the inverted value of the EQ bit of CR6.
4614 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
4616 case 2: // Branch on the value of the LT bit of CR6.
4617 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
4619 case 3: // Branch on the inverted value of the LT bit of CR6.
4620 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
4624 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
4625 DAG.getConstant(CompOpc, MVT::i32),
4626 DAG.getRegister(PPC::CR6, MVT::i32),
4627 N->getOperand(4), CompNode.getValue(1));
4636 //===----------------------------------------------------------------------===//
4637 // Inline Assembly Support
4638 //===----------------------------------------------------------------------===//
4640 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4644 const SelectionDAG &DAG,
4645 unsigned Depth) const {
4646 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4647 switch (Op.getOpcode()) {
4649 case PPCISD::LBRX: {
4650 // lhbrx is known to have the top bits cleared out.
4651 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
4652 KnownZero = 0xFFFF0000;
4655 case ISD::INTRINSIC_WO_CHAIN: {
4656 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
4658 case Intrinsic::ppc_altivec_vcmpbfp_p:
4659 case Intrinsic::ppc_altivec_vcmpeqfp_p:
4660 case Intrinsic::ppc_altivec_vcmpequb_p:
4661 case Intrinsic::ppc_altivec_vcmpequh_p:
4662 case Intrinsic::ppc_altivec_vcmpequw_p:
4663 case Intrinsic::ppc_altivec_vcmpgefp_p:
4664 case Intrinsic::ppc_altivec_vcmpgtfp_p:
4665 case Intrinsic::ppc_altivec_vcmpgtsb_p:
4666 case Intrinsic::ppc_altivec_vcmpgtsh_p:
4667 case Intrinsic::ppc_altivec_vcmpgtsw_p:
4668 case Intrinsic::ppc_altivec_vcmpgtub_p:
4669 case Intrinsic::ppc_altivec_vcmpgtuh_p:
4670 case Intrinsic::ppc_altivec_vcmpgtuw_p:
4671 KnownZero = ~1U; // All bits but the low one are known to be zero.
4679 /// getConstraintType - Given a constraint, return the type of
4680 /// constraint it is for this target.
4681 PPCTargetLowering::ConstraintType
4682 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
4683 if (Constraint.size() == 1) {
4684 switch (Constraint[0]) {
4691 return C_RegisterClass;
4694 return TargetLowering::getConstraintType(Constraint);
4697 std::pair<unsigned, const TargetRegisterClass*>
4698 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4700 if (Constraint.size() == 1) {
4701 // GCC RS6000 Constraint Letters
4702 switch (Constraint[0]) {
4705 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
4706 return std::make_pair(0U, PPC::G8RCRegisterClass);
4707 return std::make_pair(0U, PPC::GPRCRegisterClass);
4710 return std::make_pair(0U, PPC::F4RCRegisterClass);
4711 else if (VT == MVT::f64)
4712 return std::make_pair(0U, PPC::F8RCRegisterClass);
4715 return std::make_pair(0U, PPC::VRRCRegisterClass);
4717 return std::make_pair(0U, PPC::CRRCRegisterClass);
4721 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4725 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4726 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is true
4727 /// it means one of the asm constraint of the inline asm instruction being
4728 /// processed is 'm'.
4729 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
4731 std::vector<SDValue>&Ops,
4732 SelectionDAG &DAG) const {
4733 SDValue Result(0,0);
4744 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
4745 if (!CST) return; // Must be an immediate to match.
4746 unsigned Value = CST->getZExtValue();
4748 default: assert(0 && "Unknown constraint letter!");
4749 case 'I': // "I" is a signed 16-bit constant.
4750 if ((short)Value == (int)Value)
4751 Result = DAG.getTargetConstant(Value, Op.getValueType());
4753 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
4754 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
4755 if ((short)Value == 0)
4756 Result = DAG.getTargetConstant(Value, Op.getValueType());
4758 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
4759 if ((Value >> 16) == 0)
4760 Result = DAG.getTargetConstant(Value, Op.getValueType());
4762 case 'M': // "M" is a constant that is greater than 31.
4764 Result = DAG.getTargetConstant(Value, Op.getValueType());
4766 case 'N': // "N" is a positive constant that is an exact power of two.
4767 if ((int)Value > 0 && isPowerOf2_32(Value))
4768 Result = DAG.getTargetConstant(Value, Op.getValueType());
4770 case 'O': // "O" is the constant zero.
4772 Result = DAG.getTargetConstant(Value, Op.getValueType());
4774 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
4775 if ((short)-Value == (int)-Value)
4776 Result = DAG.getTargetConstant(Value, Op.getValueType());
4783 if (Result.getNode()) {
4784 Ops.push_back(Result);
4788 // Handle standard constraint letters.
4789 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, hasMemory, Ops, DAG);
4792 // isLegalAddressingMode - Return true if the addressing mode represented
4793 // by AM is legal for this target, for a load/store of the specified type.
4794 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4795 const Type *Ty) const {
4796 // FIXME: PPC does not allow r+i addressing modes for vectors!
4798 // PPC allows a sign-extended 16-bit immediate field.
4799 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
4802 // No global is ever allowed as a base.
4806 // PPC only support r+r,
4808 case 0: // "r+i" or just "i", depending on HasBaseReg.
4811 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
4813 // Otherwise we have r+r or r+i.
4816 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
4818 // Allow 2*r as r+r.
4821 // No other scales are supported.
4828 /// isLegalAddressImmediate - Return true if the integer value can be used
4829 /// as the offset of the target addressing mode for load / store of the
4831 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
4832 // PPC allows a sign-extended 16-bit immediate field.
4833 return (V > -(1 << 16) && V < (1 << 16)-1);
4836 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
4840 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
4841 DebugLoc dl = Op.getDebugLoc();
4842 // Depths > 0 not supported yet!
4843 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4846 MachineFunction &MF = DAG.getMachineFunction();
4847 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4849 // Just load the return address off the stack.
4850 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
4852 // Make sure the function really does not optimize away the store of the RA
4854 FuncInfo->setLRStoreRequired();
4855 return DAG.getLoad(getPointerTy(), dl,
4856 DAG.getEntryNode(), RetAddrFI, NULL, 0);
4859 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
4860 DebugLoc dl = Op.getDebugLoc();
4861 // Depths > 0 not supported yet!
4862 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4865 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4866 bool isPPC64 = PtrVT == MVT::i64;
4868 MachineFunction &MF = DAG.getMachineFunction();
4869 MachineFrameInfo *MFI = MF.getFrameInfo();
4870 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4871 && MFI->getStackSize();
4874 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::X31 : PPC::X1,
4877 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::R31 : PPC::R1,
4882 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4883 // The PowerPC target isn't yet aware of offsets.