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 "MCTargetDesc/PPCPredicates.h"
16 #include "PPCMachineFunctionInfo.h"
17 #include "PPCPerfectShuffle.h"
18 #include "PPCTargetMachine.h"
19 #include "PPCTargetObjectFile.h"
20 #include "llvm/ADT/STLExtras.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/SelectionDAG.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetOptions.h"
40 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
41 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
43 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
44 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
46 static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned",
47 cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden);
49 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
50 if (TM.getSubtargetImpl()->isDarwin())
51 return new TargetLoweringObjectFileMachO();
53 if (TM.getSubtargetImpl()->isSVR4ABI())
54 return new PPC64LinuxTargetObjectFile();
56 return new TargetLoweringObjectFileELF();
59 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
60 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
61 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
65 // Use _setjmp/_longjmp instead of setjmp/longjmp.
66 setUseUnderscoreSetJmp(true);
67 setUseUnderscoreLongJmp(true);
69 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
70 // arguments are at least 4/8 bytes aligned.
71 bool isPPC64 = Subtarget->isPPC64();
72 setMinStackArgumentAlignment(isPPC64 ? 8:4);
74 // Set up the register classes.
75 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
76 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
77 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
79 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
80 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
81 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
83 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
85 // PowerPC has pre-inc load and store's.
86 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
87 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
88 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
89 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
90 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
91 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
92 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
93 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
94 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
95 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
97 // This is used in the ppcf128->int sequence. Note it has different semantics
98 // from FP_ROUND: that rounds to nearest, this rounds to zero.
99 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
101 // We do not currently implement these libm ops for PowerPC.
102 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
103 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
104 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
105 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
106 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
107 setOperationAction(ISD::FREM, MVT::ppcf128, Expand);
109 // PowerPC has no SREM/UREM instructions
110 setOperationAction(ISD::SREM, MVT::i32, Expand);
111 setOperationAction(ISD::UREM, MVT::i32, Expand);
112 setOperationAction(ISD::SREM, MVT::i64, Expand);
113 setOperationAction(ISD::UREM, MVT::i64, Expand);
115 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
116 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
117 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
118 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
119 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
120 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
121 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
122 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
123 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
125 // We don't support sin/cos/sqrt/fmod/pow
126 setOperationAction(ISD::FSIN , MVT::f64, Expand);
127 setOperationAction(ISD::FCOS , MVT::f64, Expand);
128 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
129 setOperationAction(ISD::FREM , MVT::f64, Expand);
130 setOperationAction(ISD::FPOW , MVT::f64, Expand);
131 setOperationAction(ISD::FMA , MVT::f64, Legal);
132 setOperationAction(ISD::FSIN , MVT::f32, Expand);
133 setOperationAction(ISD::FCOS , MVT::f32, Expand);
134 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
135 setOperationAction(ISD::FREM , MVT::f32, Expand);
136 setOperationAction(ISD::FPOW , MVT::f32, Expand);
137 setOperationAction(ISD::FMA , MVT::f32, Legal);
139 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
141 // If we're enabling GP optimizations, use hardware square root
142 if (!Subtarget->hasFSQRT() &&
143 !(TM.Options.UnsafeFPMath &&
144 Subtarget->hasFRSQRTE() && Subtarget->hasFRE()))
145 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
147 if (!Subtarget->hasFSQRT() &&
148 !(TM.Options.UnsafeFPMath &&
149 Subtarget->hasFRSQRTES() && Subtarget->hasFRES()))
150 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
152 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
153 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
155 if (Subtarget->hasFPRND()) {
156 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
157 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
158 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
160 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
161 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
162 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
164 // frin does not implement "ties to even." Thus, this is safe only in
166 if (TM.Options.UnsafeFPMath) {
167 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
168 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
170 // These need to set FE_INEXACT, and use a custom inserter.
171 setOperationAction(ISD::FRINT, MVT::f64, Legal);
172 setOperationAction(ISD::FRINT, MVT::f32, Legal);
176 // PowerPC does not have BSWAP, CTPOP or CTTZ
177 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
178 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
179 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
180 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
181 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
182 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
183 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
184 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
186 if (Subtarget->hasPOPCNTD()) {
187 setOperationAction(ISD::CTPOP, MVT::i32 , Legal);
188 setOperationAction(ISD::CTPOP, MVT::i64 , Legal);
190 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
191 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
194 // PowerPC does not have ROTR
195 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
196 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
198 // PowerPC does not have Select
199 setOperationAction(ISD::SELECT, MVT::i32, Expand);
200 setOperationAction(ISD::SELECT, MVT::i64, Expand);
201 setOperationAction(ISD::SELECT, MVT::f32, Expand);
202 setOperationAction(ISD::SELECT, MVT::f64, Expand);
204 // PowerPC wants to turn select_cc of FP into fsel when possible.
205 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
206 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
208 // PowerPC wants to optimize integer setcc a bit
209 setOperationAction(ISD::SETCC, MVT::i32, Custom);
211 // PowerPC does not have BRCOND which requires SetCC
212 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
214 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
216 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
217 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
219 // PowerPC does not have [U|S]INT_TO_FP
220 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
221 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
223 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
224 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
225 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
226 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
228 // We cannot sextinreg(i1). Expand to shifts.
229 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
231 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
232 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
233 // support continuation, user-level threading, and etc.. As a result, no
234 // other SjLj exception interfaces are implemented and please don't build
235 // your own exception handling based on them.
236 // LLVM/Clang supports zero-cost DWARF exception handling.
237 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
238 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
240 // We want to legalize GlobalAddress and ConstantPool nodes into the
241 // appropriate instructions to materialize the address.
242 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
243 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
244 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
245 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
246 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
247 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
248 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
249 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
250 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
251 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
254 setOperationAction(ISD::TRAP, MVT::Other, Legal);
256 // TRAMPOLINE is custom lowered.
257 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
258 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
260 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
261 setOperationAction(ISD::VASTART , MVT::Other, Custom);
263 if (Subtarget->isSVR4ABI()) {
265 // VAARG always uses double-word chunks, so promote anything smaller.
266 setOperationAction(ISD::VAARG, MVT::i1, Promote);
267 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
268 setOperationAction(ISD::VAARG, MVT::i8, Promote);
269 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
270 setOperationAction(ISD::VAARG, MVT::i16, Promote);
271 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
272 setOperationAction(ISD::VAARG, MVT::i32, Promote);
273 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
274 setOperationAction(ISD::VAARG, MVT::Other, Expand);
276 // VAARG is custom lowered with the 32-bit SVR4 ABI.
277 setOperationAction(ISD::VAARG, MVT::Other, Custom);
278 setOperationAction(ISD::VAARG, MVT::i64, Custom);
281 setOperationAction(ISD::VAARG, MVT::Other, Expand);
283 // Use the default implementation.
284 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
285 setOperationAction(ISD::VAEND , MVT::Other, Expand);
286 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
287 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
288 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
289 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
291 // We want to custom lower some of our intrinsics.
292 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
294 // To handle counter-based loop conditions.
295 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom);
297 // Comparisons that require checking two conditions.
298 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
299 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
300 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
301 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
302 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
303 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
304 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
305 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
306 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
307 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
308 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
309 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
311 if (Subtarget->has64BitSupport()) {
312 // They also have instructions for converting between i64 and fp.
313 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
314 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
315 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
316 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
317 // This is just the low 32 bits of a (signed) fp->i64 conversion.
318 // We cannot do this with Promote because i64 is not a legal type.
319 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
321 if (PPCSubTarget.hasLFIWAX() || Subtarget->isPPC64())
322 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
324 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
325 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
328 // With the instructions enabled under FPCVT, we can do everything.
329 if (PPCSubTarget.hasFPCVT()) {
330 if (Subtarget->has64BitSupport()) {
331 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
332 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
333 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
334 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
337 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
338 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
339 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
340 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
343 if (Subtarget->use64BitRegs()) {
344 // 64-bit PowerPC implementations can support i64 types directly
345 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
346 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
347 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
348 // 64-bit PowerPC wants to expand i128 shifts itself.
349 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
350 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
351 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
353 // 32-bit PowerPC wants to expand i64 shifts itself.
354 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
355 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
356 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
359 if (Subtarget->hasAltivec()) {
360 // First set operation action for all vector types to expand. Then we
361 // will selectively turn on ones that can be effectively codegen'd.
362 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
363 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
364 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
366 // add/sub are legal for all supported vector VT's.
367 setOperationAction(ISD::ADD , VT, Legal);
368 setOperationAction(ISD::SUB , VT, Legal);
370 // We promote all shuffles to v16i8.
371 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
372 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
374 // We promote all non-typed operations to v4i32.
375 setOperationAction(ISD::AND , VT, Promote);
376 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
377 setOperationAction(ISD::OR , VT, Promote);
378 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
379 setOperationAction(ISD::XOR , VT, Promote);
380 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
381 setOperationAction(ISD::LOAD , VT, Promote);
382 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
383 setOperationAction(ISD::SELECT, VT, Promote);
384 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
385 setOperationAction(ISD::STORE, VT, Promote);
386 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
388 // No other operations are legal.
389 setOperationAction(ISD::MUL , VT, Expand);
390 setOperationAction(ISD::SDIV, VT, Expand);
391 setOperationAction(ISD::SREM, VT, Expand);
392 setOperationAction(ISD::UDIV, VT, Expand);
393 setOperationAction(ISD::UREM, VT, Expand);
394 setOperationAction(ISD::FDIV, VT, Expand);
395 setOperationAction(ISD::FNEG, VT, Expand);
396 setOperationAction(ISD::FSQRT, VT, Expand);
397 setOperationAction(ISD::FLOG, VT, Expand);
398 setOperationAction(ISD::FLOG10, VT, Expand);
399 setOperationAction(ISD::FLOG2, VT, Expand);
400 setOperationAction(ISD::FEXP, VT, Expand);
401 setOperationAction(ISD::FEXP2, VT, Expand);
402 setOperationAction(ISD::FSIN, VT, Expand);
403 setOperationAction(ISD::FCOS, VT, Expand);
404 setOperationAction(ISD::FABS, VT, Expand);
405 setOperationAction(ISD::FPOWI, VT, Expand);
406 setOperationAction(ISD::FFLOOR, VT, Expand);
407 setOperationAction(ISD::FCEIL, VT, Expand);
408 setOperationAction(ISD::FTRUNC, VT, Expand);
409 setOperationAction(ISD::FRINT, VT, Expand);
410 setOperationAction(ISD::FNEARBYINT, VT, Expand);
411 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
412 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
413 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
414 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
415 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
416 setOperationAction(ISD::UDIVREM, VT, Expand);
417 setOperationAction(ISD::SDIVREM, VT, Expand);
418 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
419 setOperationAction(ISD::FPOW, VT, Expand);
420 setOperationAction(ISD::CTPOP, VT, Expand);
421 setOperationAction(ISD::CTLZ, VT, Expand);
422 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
423 setOperationAction(ISD::CTTZ, VT, Expand);
424 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
425 setOperationAction(ISD::VSELECT, VT, Expand);
426 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
428 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
429 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
430 MVT::SimpleValueType InnerVT = (MVT::SimpleValueType)j;
431 setTruncStoreAction(VT, InnerVT, Expand);
433 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
434 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
435 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
438 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
439 // with merges, splats, etc.
440 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
442 setOperationAction(ISD::AND , MVT::v4i32, Legal);
443 setOperationAction(ISD::OR , MVT::v4i32, Legal);
444 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
445 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
446 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
447 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
448 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
449 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
450 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
451 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
452 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
453 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
454 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
455 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
457 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
458 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
459 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
460 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
462 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
463 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
465 if (TM.Options.UnsafeFPMath) {
466 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
467 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
470 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
471 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
472 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
474 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
475 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
477 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
478 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
479 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
480 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
482 // Altivec does not contain unordered floating-point compare instructions
483 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
484 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
485 setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
486 setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
487 setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
488 setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
491 if (Subtarget->has64BitSupport()) {
492 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
493 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
496 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
497 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
498 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
499 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
501 setBooleanContents(ZeroOrOneBooleanContent);
502 // Altivec instructions set fields to all zeros or all ones.
503 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
506 setStackPointerRegisterToSaveRestore(PPC::X1);
507 setExceptionPointerRegister(PPC::X3);
508 setExceptionSelectorRegister(PPC::X4);
510 setStackPointerRegisterToSaveRestore(PPC::R1);
511 setExceptionPointerRegister(PPC::R3);
512 setExceptionSelectorRegister(PPC::R4);
515 // We have target-specific dag combine patterns for the following nodes:
516 setTargetDAGCombine(ISD::SINT_TO_FP);
517 setTargetDAGCombine(ISD::LOAD);
518 setTargetDAGCombine(ISD::STORE);
519 setTargetDAGCombine(ISD::BR_CC);
520 setTargetDAGCombine(ISD::BSWAP);
521 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
523 // Use reciprocal estimates.
524 if (TM.Options.UnsafeFPMath) {
525 setTargetDAGCombine(ISD::FDIV);
526 setTargetDAGCombine(ISD::FSQRT);
529 // Darwin long double math library functions have $LDBL128 appended.
530 if (Subtarget->isDarwin()) {
531 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
532 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
533 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
534 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
535 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
536 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
537 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
538 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
539 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
540 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
543 setMinFunctionAlignment(2);
544 if (PPCSubTarget.isDarwin())
545 setPrefFunctionAlignment(4);
547 if (isPPC64 && Subtarget->isJITCodeModel())
548 // Temporary workaround for the inability of PPC64 JIT to handle jump
550 setSupportJumpTables(false);
552 setInsertFencesForAtomic(true);
554 setSchedulingPreference(Sched::Hybrid);
556 computeRegisterProperties();
558 // The Freescale cores does better with aggressive inlining of memcpy and
559 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
560 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
561 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
562 MaxStoresPerMemset = 32;
563 MaxStoresPerMemsetOptSize = 16;
564 MaxStoresPerMemcpy = 32;
565 MaxStoresPerMemcpyOptSize = 8;
566 MaxStoresPerMemmove = 32;
567 MaxStoresPerMemmoveOptSize = 8;
569 setPrefFunctionAlignment(4);
573 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
574 /// function arguments in the caller parameter area.
575 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
576 const TargetMachine &TM = getTargetMachine();
577 // Darwin passes everything on 4 byte boundary.
578 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
581 // 16byte and wider vectors are passed on 16byte boundary.
582 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
583 if (VTy->getBitWidth() >= 128)
586 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
587 if (PPCSubTarget.isPPC64())
593 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
596 case PPCISD::FSEL: return "PPCISD::FSEL";
597 case PPCISD::FCFID: return "PPCISD::FCFID";
598 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
599 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
600 case PPCISD::FRE: return "PPCISD::FRE";
601 case PPCISD::FRSQRTE: return "PPCISD::FRSQRTE";
602 case PPCISD::STFIWX: return "PPCISD::STFIWX";
603 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
604 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
605 case PPCISD::VPERM: return "PPCISD::VPERM";
606 case PPCISD::Hi: return "PPCISD::Hi";
607 case PPCISD::Lo: return "PPCISD::Lo";
608 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
609 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
610 case PPCISD::LOAD: return "PPCISD::LOAD";
611 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
612 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
613 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
614 case PPCISD::SRL: return "PPCISD::SRL";
615 case PPCISD::SRA: return "PPCISD::SRA";
616 case PPCISD::SHL: return "PPCISD::SHL";
617 case PPCISD::CALL: return "PPCISD::CALL";
618 case PPCISD::CALL_NOP: return "PPCISD::CALL_NOP";
619 case PPCISD::MTCTR: return "PPCISD::MTCTR";
620 case PPCISD::BCTRL: return "PPCISD::BCTRL";
621 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
622 case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP";
623 case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP";
624 case PPCISD::MFOCRF: return "PPCISD::MFOCRF";
625 case PPCISD::VCMP: return "PPCISD::VCMP";
626 case PPCISD::VCMPo: return "PPCISD::VCMPo";
627 case PPCISD::LBRX: return "PPCISD::LBRX";
628 case PPCISD::STBRX: return "PPCISD::STBRX";
629 case PPCISD::LARX: return "PPCISD::LARX";
630 case PPCISD::STCX: return "PPCISD::STCX";
631 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
632 case PPCISD::BDNZ: return "PPCISD::BDNZ";
633 case PPCISD::BDZ: return "PPCISD::BDZ";
634 case PPCISD::MFFS: return "PPCISD::MFFS";
635 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
636 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
637 case PPCISD::CR6SET: return "PPCISD::CR6SET";
638 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
639 case PPCISD::ADDIS_TOC_HA: return "PPCISD::ADDIS_TOC_HA";
640 case PPCISD::LD_TOC_L: return "PPCISD::LD_TOC_L";
641 case PPCISD::ADDI_TOC_L: return "PPCISD::ADDI_TOC_L";
642 case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
643 case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L";
644 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
645 case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA";
646 case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L";
647 case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR";
648 case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA";
649 case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L";
650 case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR";
651 case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
652 case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L";
653 case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT";
654 case PPCISD::SC: return "PPCISD::SC";
658 EVT PPCTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
661 return VT.changeVectorElementTypeToInteger();
664 //===----------------------------------------------------------------------===//
665 // Node matching predicates, for use by the tblgen matching code.
666 //===----------------------------------------------------------------------===//
668 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
669 static bool isFloatingPointZero(SDValue Op) {
670 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
671 return CFP->getValueAPF().isZero();
672 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
673 // Maybe this has already been legalized into the constant pool?
674 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
675 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
676 return CFP->getValueAPF().isZero();
681 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
682 /// true if Op is undef or if it matches the specified value.
683 static bool isConstantOrUndef(int Op, int Val) {
684 return Op < 0 || Op == Val;
687 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
688 /// VPKUHUM instruction.
689 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
691 for (unsigned i = 0; i != 16; ++i)
692 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
695 for (unsigned i = 0; i != 8; ++i)
696 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
697 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
703 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
704 /// VPKUWUM instruction.
705 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
707 for (unsigned i = 0; i != 16; i += 2)
708 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
709 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
712 for (unsigned i = 0; i != 8; i += 2)
713 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
714 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
715 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
716 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
722 /// isVMerge - Common function, used to match vmrg* shuffles.
724 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
725 unsigned LHSStart, unsigned RHSStart) {
726 assert(N->getValueType(0) == MVT::v16i8 &&
727 "PPC only supports shuffles by bytes!");
728 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
729 "Unsupported merge size!");
731 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
732 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
733 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
734 LHSStart+j+i*UnitSize) ||
735 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
736 RHSStart+j+i*UnitSize))
742 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
743 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
744 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
747 return isVMerge(N, UnitSize, 8, 24);
748 return isVMerge(N, UnitSize, 8, 8);
751 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
752 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
753 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
756 return isVMerge(N, UnitSize, 0, 16);
757 return isVMerge(N, UnitSize, 0, 0);
761 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
762 /// amount, otherwise return -1.
763 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
764 assert(N->getValueType(0) == MVT::v16i8 &&
765 "PPC only supports shuffles by bytes!");
767 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
769 // Find the first non-undef value in the shuffle mask.
771 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
774 if (i == 16) return -1; // all undef.
776 // Otherwise, check to see if the rest of the elements are consecutively
777 // numbered from this value.
778 unsigned ShiftAmt = SVOp->getMaskElt(i);
779 if (ShiftAmt < i) return -1;
783 // Check the rest of the elements to see if they are consecutive.
784 for (++i; i != 16; ++i)
785 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
788 // Check the rest of the elements to see if they are consecutive.
789 for (++i; i != 16; ++i)
790 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
796 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
797 /// specifies a splat of a single element that is suitable for input to
798 /// VSPLTB/VSPLTH/VSPLTW.
799 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
800 assert(N->getValueType(0) == MVT::v16i8 &&
801 (EltSize == 1 || EltSize == 2 || EltSize == 4));
803 // This is a splat operation if each element of the permute is the same, and
804 // if the value doesn't reference the second vector.
805 unsigned ElementBase = N->getMaskElt(0);
807 // FIXME: Handle UNDEF elements too!
808 if (ElementBase >= 16)
811 // Check that the indices are consecutive, in the case of a multi-byte element
812 // splatted with a v16i8 mask.
813 for (unsigned i = 1; i != EltSize; ++i)
814 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
817 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
818 if (N->getMaskElt(i) < 0) continue;
819 for (unsigned j = 0; j != EltSize; ++j)
820 if (N->getMaskElt(i+j) != N->getMaskElt(j))
826 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
828 bool PPC::isAllNegativeZeroVector(SDNode *N) {
829 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
831 APInt APVal, APUndef;
835 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
836 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
837 return CFP->getValueAPF().isNegZero();
842 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
843 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
844 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
845 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
846 assert(isSplatShuffleMask(SVOp, EltSize));
847 return SVOp->getMaskElt(0) / EltSize;
850 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
851 /// by using a vspltis[bhw] instruction of the specified element size, return
852 /// the constant being splatted. The ByteSize field indicates the number of
853 /// bytes of each element [124] -> [bhw].
854 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
857 // If ByteSize of the splat is bigger than the element size of the
858 // build_vector, then we have a case where we are checking for a splat where
859 // multiple elements of the buildvector are folded together into a single
860 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
861 unsigned EltSize = 16/N->getNumOperands();
862 if (EltSize < ByteSize) {
863 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
864 SDValue UniquedVals[4];
865 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
867 // See if all of the elements in the buildvector agree across.
868 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
869 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
870 // If the element isn't a constant, bail fully out.
871 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
874 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
875 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
876 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
877 return SDValue(); // no match.
880 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
881 // either constant or undef values that are identical for each chunk. See
882 // if these chunks can form into a larger vspltis*.
884 // Check to see if all of the leading entries are either 0 or -1. If
885 // neither, then this won't fit into the immediate field.
886 bool LeadingZero = true;
887 bool LeadingOnes = true;
888 for (unsigned i = 0; i != Multiple-1; ++i) {
889 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
891 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
892 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
894 // Finally, check the least significant entry.
896 if (UniquedVals[Multiple-1].getNode() == 0)
897 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
898 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
900 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
903 if (UniquedVals[Multiple-1].getNode() == 0)
904 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
905 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
906 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
907 return DAG.getTargetConstant(Val, MVT::i32);
913 // Check to see if this buildvec has a single non-undef value in its elements.
914 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
915 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
916 if (OpVal.getNode() == 0)
917 OpVal = N->getOperand(i);
918 else if (OpVal != N->getOperand(i))
922 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
924 unsigned ValSizeInBytes = EltSize;
926 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
927 Value = CN->getZExtValue();
928 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
929 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
930 Value = FloatToBits(CN->getValueAPF().convertToFloat());
933 // If the splat value is larger than the element value, then we can never do
934 // this splat. The only case that we could fit the replicated bits into our
935 // immediate field for would be zero, and we prefer to use vxor for it.
936 if (ValSizeInBytes < ByteSize) return SDValue();
938 // If the element value is larger than the splat value, cut it in half and
939 // check to see if the two halves are equal. Continue doing this until we
940 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
941 while (ValSizeInBytes > ByteSize) {
942 ValSizeInBytes >>= 1;
944 // If the top half equals the bottom half, we're still ok.
945 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
946 (Value & ((1 << (8*ValSizeInBytes))-1)))
950 // Properly sign extend the value.
951 int MaskVal = SignExtend32(Value, ByteSize * 8);
953 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
954 if (MaskVal == 0) return SDValue();
956 // Finally, if this value fits in a 5 bit sext field, return it
957 if (SignExtend32<5>(MaskVal) == MaskVal)
958 return DAG.getTargetConstant(MaskVal, MVT::i32);
962 //===----------------------------------------------------------------------===//
963 // Addressing Mode Selection
964 //===----------------------------------------------------------------------===//
966 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
967 /// or 64-bit immediate, and if the value can be accurately represented as a
968 /// sign extension from a 16-bit value. If so, this returns true and the
970 static bool isIntS16Immediate(SDNode *N, short &Imm) {
971 if (N->getOpcode() != ISD::Constant)
974 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
975 if (N->getValueType(0) == MVT::i32)
976 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
978 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
980 static bool isIntS16Immediate(SDValue Op, short &Imm) {
981 return isIntS16Immediate(Op.getNode(), Imm);
985 /// SelectAddressRegReg - Given the specified addressed, check to see if it
986 /// can be represented as an indexed [r+r] operation. Returns false if it
987 /// can be more efficiently represented with [r+imm].
988 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
990 SelectionDAG &DAG) const {
992 if (N.getOpcode() == ISD::ADD) {
993 if (isIntS16Immediate(N.getOperand(1), imm))
995 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
998 Base = N.getOperand(0);
999 Index = N.getOperand(1);
1001 } else if (N.getOpcode() == ISD::OR) {
1002 if (isIntS16Immediate(N.getOperand(1), imm))
1003 return false; // r+i can fold it if we can.
1005 // If this is an or of disjoint bitfields, we can codegen this as an add
1006 // (for better address arithmetic) if the LHS and RHS of the OR are provably
1008 APInt LHSKnownZero, LHSKnownOne;
1009 APInt RHSKnownZero, RHSKnownOne;
1010 DAG.ComputeMaskedBits(N.getOperand(0),
1011 LHSKnownZero, LHSKnownOne);
1013 if (LHSKnownZero.getBoolValue()) {
1014 DAG.ComputeMaskedBits(N.getOperand(1),
1015 RHSKnownZero, RHSKnownOne);
1016 // If all of the bits are known zero on the LHS or RHS, the add won't
1018 if (~(LHSKnownZero | RHSKnownZero) == 0) {
1019 Base = N.getOperand(0);
1020 Index = N.getOperand(1);
1029 /// Returns true if the address N can be represented by a base register plus
1030 /// a signed 16-bit displacement [r+imm], and if it is not better
1031 /// represented as reg+reg. If Aligned is true, only accept displacements
1032 /// suitable for STD and friends, i.e. multiples of 4.
1033 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
1036 bool Aligned) const {
1037 // FIXME dl should come from parent load or store, not from address
1039 // If this can be more profitably realized as r+r, fail.
1040 if (SelectAddressRegReg(N, Disp, Base, DAG))
1043 if (N.getOpcode() == ISD::ADD) {
1045 if (isIntS16Immediate(N.getOperand(1), imm) &&
1046 (!Aligned || (imm & 3) == 0)) {
1047 Disp = DAG.getTargetConstant(imm, N.getValueType());
1048 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1049 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1051 Base = N.getOperand(0);
1053 return true; // [r+i]
1054 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1055 // Match LOAD (ADD (X, Lo(G))).
1056 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1057 && "Cannot handle constant offsets yet!");
1058 Disp = N.getOperand(1).getOperand(0); // The global address.
1059 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1060 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
1061 Disp.getOpcode() == ISD::TargetConstantPool ||
1062 Disp.getOpcode() == ISD::TargetJumpTable);
1063 Base = N.getOperand(0);
1064 return true; // [&g+r]
1066 } else if (N.getOpcode() == ISD::OR) {
1068 if (isIntS16Immediate(N.getOperand(1), imm) &&
1069 (!Aligned || (imm & 3) == 0)) {
1070 // If this is an or of disjoint bitfields, we can codegen this as an add
1071 // (for better address arithmetic) if the LHS and RHS of the OR are
1072 // provably disjoint.
1073 APInt LHSKnownZero, LHSKnownOne;
1074 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1076 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1077 // If all of the bits are known zero on the LHS or RHS, the add won't
1079 Base = N.getOperand(0);
1080 Disp = DAG.getTargetConstant(imm, N.getValueType());
1084 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1085 // Loading from a constant address.
1087 // If this address fits entirely in a 16-bit sext immediate field, codegen
1090 if (isIntS16Immediate(CN, Imm) && (!Aligned || (Imm & 3) == 0)) {
1091 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1092 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1093 CN->getValueType(0));
1097 // Handle 32-bit sext immediates with LIS + addr mode.
1098 if ((CN->getValueType(0) == MVT::i32 ||
1099 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) &&
1100 (!Aligned || (CN->getZExtValue() & 3) == 0)) {
1101 int Addr = (int)CN->getZExtValue();
1103 // Otherwise, break this down into an LIS + disp.
1104 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1106 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1107 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1108 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1113 Disp = DAG.getTargetConstant(0, getPointerTy());
1114 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1115 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1118 return true; // [r+0]
1121 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1122 /// represented as an indexed [r+r] operation.
1123 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1125 SelectionDAG &DAG) const {
1126 // Check to see if we can easily represent this as an [r+r] address. This
1127 // will fail if it thinks that the address is more profitably represented as
1128 // reg+imm, e.g. where imm = 0.
1129 if (SelectAddressRegReg(N, Base, Index, DAG))
1132 // If the operand is an addition, always emit this as [r+r], since this is
1133 // better (for code size, and execution, as the memop does the add for free)
1134 // than emitting an explicit add.
1135 if (N.getOpcode() == ISD::ADD) {
1136 Base = N.getOperand(0);
1137 Index = N.getOperand(1);
1141 // Otherwise, do it the hard way, using R0 as the base register.
1142 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1148 /// getPreIndexedAddressParts - returns true by value, base pointer and
1149 /// offset pointer and addressing mode by reference if the node's address
1150 /// can be legally represented as pre-indexed load / store address.
1151 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1153 ISD::MemIndexedMode &AM,
1154 SelectionDAG &DAG) const {
1155 if (DisablePPCPreinc) return false;
1161 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1162 Ptr = LD->getBasePtr();
1163 VT = LD->getMemoryVT();
1164 Alignment = LD->getAlignment();
1165 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1166 Ptr = ST->getBasePtr();
1167 VT = ST->getMemoryVT();
1168 Alignment = ST->getAlignment();
1173 // PowerPC doesn't have preinc load/store instructions for vectors.
1177 if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) {
1179 // Common code will reject creating a pre-inc form if the base pointer
1180 // is a frame index, or if N is a store and the base pointer is either
1181 // the same as or a predecessor of the value being stored. Check for
1182 // those situations here, and try with swapped Base/Offset instead.
1185 if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base))
1188 SDValue Val = cast<StoreSDNode>(N)->getValue();
1189 if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode()))
1194 std::swap(Base, Offset);
1200 // LDU/STU can only handle immediates that are a multiple of 4.
1201 if (VT != MVT::i64) {
1202 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, false))
1205 // LDU/STU need an address with at least 4-byte alignment.
1209 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, true))
1213 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1214 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1215 // sext i32 to i64 when addr mode is r+i.
1216 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1217 LD->getExtensionType() == ISD::SEXTLOAD &&
1218 isa<ConstantSDNode>(Offset))
1226 //===----------------------------------------------------------------------===//
1227 // LowerOperation implementation
1228 //===----------------------------------------------------------------------===//
1230 /// GetLabelAccessInfo - Return true if we should reference labels using a
1231 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1232 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1233 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1234 HiOpFlags = PPCII::MO_HA;
1235 LoOpFlags = PPCII::MO_LO;
1237 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1238 // non-darwin platform. We don't support PIC on other platforms yet.
1239 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1240 TM.getSubtarget<PPCSubtarget>().isDarwin();
1242 HiOpFlags |= PPCII::MO_PIC_FLAG;
1243 LoOpFlags |= PPCII::MO_PIC_FLAG;
1246 // If this is a reference to a global value that requires a non-lazy-ptr, make
1247 // sure that instruction lowering adds it.
1248 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1249 HiOpFlags |= PPCII::MO_NLP_FLAG;
1250 LoOpFlags |= PPCII::MO_NLP_FLAG;
1252 if (GV->hasHiddenVisibility()) {
1253 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1254 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1261 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1262 SelectionDAG &DAG) {
1263 EVT PtrVT = HiPart.getValueType();
1264 SDValue Zero = DAG.getConstant(0, PtrVT);
1267 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1268 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1270 // With PIC, the first instruction is actually "GR+hi(&G)".
1272 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1273 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1275 // Generate non-pic code that has direct accesses to the constant pool.
1276 // The address of the global is just (hi(&g)+lo(&g)).
1277 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1280 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1281 SelectionDAG &DAG) const {
1282 EVT PtrVT = Op.getValueType();
1283 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1284 const Constant *C = CP->getConstVal();
1286 // 64-bit SVR4 ABI code is always position-independent.
1287 // The actual address of the GlobalValue is stored in the TOC.
1288 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1289 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1290 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(CP), MVT::i64, GA,
1291 DAG.getRegister(PPC::X2, MVT::i64));
1294 unsigned MOHiFlag, MOLoFlag;
1295 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1297 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1299 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1300 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1303 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1304 EVT PtrVT = Op.getValueType();
1305 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1307 // 64-bit SVR4 ABI code is always position-independent.
1308 // The actual address of the GlobalValue is stored in the TOC.
1309 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1310 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1311 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(JT), MVT::i64, GA,
1312 DAG.getRegister(PPC::X2, MVT::i64));
1315 unsigned MOHiFlag, MOLoFlag;
1316 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1317 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1318 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1319 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1322 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1323 SelectionDAG &DAG) const {
1324 EVT PtrVT = Op.getValueType();
1326 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1328 unsigned MOHiFlag, MOLoFlag;
1329 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1330 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1331 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1332 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1335 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1336 SelectionDAG &DAG) const {
1338 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1340 const GlobalValue *GV = GA->getGlobal();
1341 EVT PtrVT = getPointerTy();
1342 bool is64bit = PPCSubTarget.isPPC64();
1344 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
1346 if (Model == TLSModel::LocalExec) {
1347 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1348 PPCII::MO_TPREL_HA);
1349 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1350 PPCII::MO_TPREL_LO);
1351 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1352 is64bit ? MVT::i64 : MVT::i32);
1353 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1354 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1358 llvm_unreachable("only local-exec is currently supported for ppc32");
1360 if (Model == TLSModel::InitialExec) {
1361 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1362 SDValue TGATLS = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1364 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1365 SDValue TPOffsetHi = DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl,
1366 PtrVT, GOTReg, TGA);
1367 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl,
1368 PtrVT, TGA, TPOffsetHi);
1369 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS);
1372 if (Model == TLSModel::GeneralDynamic) {
1373 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1374 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1375 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
1377 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSGD_L, dl, PtrVT,
1380 // We need a chain node, and don't have one handy. The underlying
1381 // call has no side effects, so using the function entry node
1383 SDValue Chain = DAG.getEntryNode();
1384 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1385 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1386 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLS_ADDR, dl,
1387 PtrVT, ParmReg, TGA);
1388 // The return value from GET_TLS_ADDR really is in X3 already, but
1389 // some hacks are needed here to tie everything together. The extra
1390 // copies dissolve during subsequent transforms.
1391 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1392 return DAG.getCopyFromReg(Chain, dl, PPC::X3, PtrVT);
1395 if (Model == TLSModel::LocalDynamic) {
1396 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1397 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1398 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
1400 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSLD_L, dl, PtrVT,
1403 // We need a chain node, and don't have one handy. The underlying
1404 // call has no side effects, so using the function entry node
1406 SDValue Chain = DAG.getEntryNode();
1407 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1408 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1409 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLSLD_ADDR, dl,
1410 PtrVT, ParmReg, TGA);
1411 // The return value from GET_TLSLD_ADDR really is in X3 already, but
1412 // some hacks are needed here to tie everything together. The extra
1413 // copies dissolve during subsequent transforms.
1414 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1415 SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, PtrVT,
1416 Chain, ParmReg, TGA);
1417 return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
1420 llvm_unreachable("Unknown TLS model!");
1423 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1424 SelectionDAG &DAG) const {
1425 EVT PtrVT = Op.getValueType();
1426 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1428 const GlobalValue *GV = GSDN->getGlobal();
1430 // 64-bit SVR4 ABI code is always position-independent.
1431 // The actual address of the GlobalValue is stored in the TOC.
1432 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1433 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1434 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1435 DAG.getRegister(PPC::X2, MVT::i64));
1438 unsigned MOHiFlag, MOLoFlag;
1439 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1442 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1444 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1446 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1448 // If the global reference is actually to a non-lazy-pointer, we have to do an
1449 // extra load to get the address of the global.
1450 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1451 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1452 false, false, false, 0);
1456 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1457 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1460 // If we're comparing for equality to zero, expose the fact that this is
1461 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1462 // fold the new nodes.
1463 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1464 if (C->isNullValue() && CC == ISD::SETEQ) {
1465 EVT VT = Op.getOperand(0).getValueType();
1466 SDValue Zext = Op.getOperand(0);
1467 if (VT.bitsLT(MVT::i32)) {
1469 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1471 unsigned Log2b = Log2_32(VT.getSizeInBits());
1472 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1473 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1474 DAG.getConstant(Log2b, MVT::i32));
1475 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1477 // Leave comparisons against 0 and -1 alone for now, since they're usually
1478 // optimized. FIXME: revisit this when we can custom lower all setcc
1480 if (C->isAllOnesValue() || C->isNullValue())
1484 // If we have an integer seteq/setne, turn it into a compare against zero
1485 // by xor'ing the rhs with the lhs, which is faster than setting a
1486 // condition register, reading it back out, and masking the correct bit. The
1487 // normal approach here uses sub to do this instead of xor. Using xor exposes
1488 // the result to other bit-twiddling opportunities.
1489 EVT LHSVT = Op.getOperand(0).getValueType();
1490 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1491 EVT VT = Op.getValueType();
1492 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1494 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1499 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1500 const PPCSubtarget &Subtarget) const {
1501 SDNode *Node = Op.getNode();
1502 EVT VT = Node->getValueType(0);
1503 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1504 SDValue InChain = Node->getOperand(0);
1505 SDValue VAListPtr = Node->getOperand(1);
1506 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1509 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1512 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1513 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1515 InChain = GprIndex.getValue(1);
1517 if (VT == MVT::i64) {
1518 // Check if GprIndex is even
1519 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1520 DAG.getConstant(1, MVT::i32));
1521 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1522 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1523 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1524 DAG.getConstant(1, MVT::i32));
1525 // Align GprIndex to be even if it isn't
1526 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1530 // fpr index is 1 byte after gpr
1531 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1532 DAG.getConstant(1, MVT::i32));
1535 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1536 FprPtr, MachinePointerInfo(SV), MVT::i8,
1538 InChain = FprIndex.getValue(1);
1540 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1541 DAG.getConstant(8, MVT::i32));
1543 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1544 DAG.getConstant(4, MVT::i32));
1547 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1548 MachinePointerInfo(), false, false,
1550 InChain = OverflowArea.getValue(1);
1552 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1553 MachinePointerInfo(), false, false,
1555 InChain = RegSaveArea.getValue(1);
1557 // select overflow_area if index > 8
1558 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1559 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1561 // adjustment constant gpr_index * 4/8
1562 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1563 VT.isInteger() ? GprIndex : FprIndex,
1564 DAG.getConstant(VT.isInteger() ? 4 : 8,
1567 // OurReg = RegSaveArea + RegConstant
1568 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1571 // Floating types are 32 bytes into RegSaveArea
1572 if (VT.isFloatingPoint())
1573 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1574 DAG.getConstant(32, MVT::i32));
1576 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1577 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1578 VT.isInteger() ? GprIndex : FprIndex,
1579 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1582 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1583 VT.isInteger() ? VAListPtr : FprPtr,
1584 MachinePointerInfo(SV),
1585 MVT::i8, false, false, 0);
1587 // determine if we should load from reg_save_area or overflow_area
1588 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1590 // increase overflow_area by 4/8 if gpr/fpr > 8
1591 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1592 DAG.getConstant(VT.isInteger() ? 4 : 8,
1595 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1598 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1600 MachinePointerInfo(),
1601 MVT::i32, false, false, 0);
1603 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1604 false, false, false, 0);
1607 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1608 SelectionDAG &DAG) const {
1609 return Op.getOperand(0);
1612 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1613 SelectionDAG &DAG) const {
1614 SDValue Chain = Op.getOperand(0);
1615 SDValue Trmp = Op.getOperand(1); // trampoline
1616 SDValue FPtr = Op.getOperand(2); // nested function
1617 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1620 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1621 bool isPPC64 = (PtrVT == MVT::i64);
1623 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1626 TargetLowering::ArgListTy Args;
1627 TargetLowering::ArgListEntry Entry;
1629 Entry.Ty = IntPtrTy;
1630 Entry.Node = Trmp; Args.push_back(Entry);
1632 // TrampSize == (isPPC64 ? 48 : 40);
1633 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1634 isPPC64 ? MVT::i64 : MVT::i32);
1635 Args.push_back(Entry);
1637 Entry.Node = FPtr; Args.push_back(Entry);
1638 Entry.Node = Nest; Args.push_back(Entry);
1640 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1641 TargetLowering::CallLoweringInfo CLI(Chain,
1642 Type::getVoidTy(*DAG.getContext()),
1643 false, false, false, false, 0,
1645 /*isTailCall=*/false,
1646 /*doesNotRet=*/false,
1647 /*isReturnValueUsed=*/true,
1648 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1650 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1652 return CallResult.second;
1655 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1656 const PPCSubtarget &Subtarget) const {
1657 MachineFunction &MF = DAG.getMachineFunction();
1658 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1662 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1663 // vastart just stores the address of the VarArgsFrameIndex slot into the
1664 // memory location argument.
1665 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1666 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1667 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1668 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1669 MachinePointerInfo(SV),
1673 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1674 // We suppose the given va_list is already allocated.
1677 // char gpr; /* index into the array of 8 GPRs
1678 // * stored in the register save area
1679 // * gpr=0 corresponds to r3,
1680 // * gpr=1 to r4, etc.
1682 // char fpr; /* index into the array of 8 FPRs
1683 // * stored in the register save area
1684 // * fpr=0 corresponds to f1,
1685 // * fpr=1 to f2, etc.
1687 // char *overflow_arg_area;
1688 // /* location on stack that holds
1689 // * the next overflow argument
1691 // char *reg_save_area;
1692 // /* where r3:r10 and f1:f8 (if saved)
1698 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1699 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1702 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1704 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1706 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1709 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1710 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1712 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1713 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1715 uint64_t FPROffset = 1;
1716 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1718 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1720 // Store first byte : number of int regs
1721 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1723 MachinePointerInfo(SV),
1724 MVT::i8, false, false, 0);
1725 uint64_t nextOffset = FPROffset;
1726 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1729 // Store second byte : number of float regs
1730 SDValue secondStore =
1731 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1732 MachinePointerInfo(SV, nextOffset), MVT::i8,
1734 nextOffset += StackOffset;
1735 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1737 // Store second word : arguments given on stack
1738 SDValue thirdStore =
1739 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1740 MachinePointerInfo(SV, nextOffset),
1742 nextOffset += FrameOffset;
1743 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1745 // Store third word : arguments given in registers
1746 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1747 MachinePointerInfo(SV, nextOffset),
1752 #include "PPCGenCallingConv.inc"
1754 bool llvm::CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1755 CCValAssign::LocInfo &LocInfo,
1756 ISD::ArgFlagsTy &ArgFlags,
1761 bool llvm::CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1763 CCValAssign::LocInfo &LocInfo,
1764 ISD::ArgFlagsTy &ArgFlags,
1766 static const uint16_t ArgRegs[] = {
1767 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1768 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1770 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1772 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1774 // Skip one register if the first unallocated register has an even register
1775 // number and there are still argument registers available which have not been
1776 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1777 // need to skip a register if RegNum is odd.
1778 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1779 State.AllocateReg(ArgRegs[RegNum]);
1782 // Always return false here, as this function only makes sure that the first
1783 // unallocated register has an odd register number and does not actually
1784 // allocate a register for the current argument.
1788 bool llvm::CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1790 CCValAssign::LocInfo &LocInfo,
1791 ISD::ArgFlagsTy &ArgFlags,
1793 static const uint16_t ArgRegs[] = {
1794 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1798 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1800 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1802 // If there is only one Floating-point register left we need to put both f64
1803 // values of a split ppc_fp128 value on the stack.
1804 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1805 State.AllocateReg(ArgRegs[RegNum]);
1808 // Always return false here, as this function only makes sure that the two f64
1809 // values a ppc_fp128 value is split into are both passed in registers or both
1810 // passed on the stack and does not actually allocate a register for the
1811 // current argument.
1815 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1817 static const uint16_t *GetFPR() {
1818 static const uint16_t FPR[] = {
1819 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1820 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1826 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1828 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1829 unsigned PtrByteSize) {
1830 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1831 if (Flags.isByVal())
1832 ArgSize = Flags.getByValSize();
1833 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1839 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1840 CallingConv::ID CallConv, bool isVarArg,
1841 const SmallVectorImpl<ISD::InputArg>
1843 SDLoc dl, SelectionDAG &DAG,
1844 SmallVectorImpl<SDValue> &InVals)
1846 if (PPCSubTarget.isSVR4ABI()) {
1847 if (PPCSubTarget.isPPC64())
1848 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1851 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1854 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1860 PPCTargetLowering::LowerFormalArguments_32SVR4(
1862 CallingConv::ID CallConv, bool isVarArg,
1863 const SmallVectorImpl<ISD::InputArg>
1865 SDLoc dl, SelectionDAG &DAG,
1866 SmallVectorImpl<SDValue> &InVals) const {
1868 // 32-bit SVR4 ABI Stack Frame Layout:
1869 // +-----------------------------------+
1870 // +--> | Back chain |
1871 // | +-----------------------------------+
1872 // | | Floating-point register save area |
1873 // | +-----------------------------------+
1874 // | | General register save area |
1875 // | +-----------------------------------+
1876 // | | CR save word |
1877 // | +-----------------------------------+
1878 // | | VRSAVE save word |
1879 // | +-----------------------------------+
1880 // | | Alignment padding |
1881 // | +-----------------------------------+
1882 // | | Vector register save area |
1883 // | +-----------------------------------+
1884 // | | Local variable space |
1885 // | +-----------------------------------+
1886 // | | Parameter list area |
1887 // | +-----------------------------------+
1888 // | | LR save word |
1889 // | +-----------------------------------+
1890 // SP--> +--- | Back chain |
1891 // +-----------------------------------+
1894 // System V Application Binary Interface PowerPC Processor Supplement
1895 // AltiVec Technology Programming Interface Manual
1897 MachineFunction &MF = DAG.getMachineFunction();
1898 MachineFrameInfo *MFI = MF.getFrameInfo();
1899 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1901 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1902 // Potential tail calls could cause overwriting of argument stack slots.
1903 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1904 (CallConv == CallingConv::Fast));
1905 unsigned PtrByteSize = 4;
1907 // Assign locations to all of the incoming arguments.
1908 SmallVector<CCValAssign, 16> ArgLocs;
1909 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1910 getTargetMachine(), ArgLocs, *DAG.getContext());
1912 // Reserve space for the linkage area on the stack.
1913 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1915 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
1917 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1918 CCValAssign &VA = ArgLocs[i];
1920 // Arguments stored in registers.
1921 if (VA.isRegLoc()) {
1922 const TargetRegisterClass *RC;
1923 EVT ValVT = VA.getValVT();
1925 switch (ValVT.getSimpleVT().SimpleTy) {
1927 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1929 RC = &PPC::GPRCRegClass;
1932 RC = &PPC::F4RCRegClass;
1935 RC = &PPC::F8RCRegClass;
1941 RC = &PPC::VRRCRegClass;
1945 // Transform the arguments stored in physical registers into virtual ones.
1946 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1947 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1949 InVals.push_back(ArgValue);
1951 // Argument stored in memory.
1952 assert(VA.isMemLoc());
1954 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1955 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1958 // Create load nodes to retrieve arguments from the stack.
1959 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1960 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1961 MachinePointerInfo(),
1962 false, false, false, 0));
1966 // Assign locations to all of the incoming aggregate by value arguments.
1967 // Aggregates passed by value are stored in the local variable space of the
1968 // caller's stack frame, right above the parameter list area.
1969 SmallVector<CCValAssign, 16> ByValArgLocs;
1970 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1971 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1973 // Reserve stack space for the allocations in CCInfo.
1974 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1976 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
1978 // Area that is at least reserved in the caller of this function.
1979 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1981 // Set the size that is at least reserved in caller of this function. Tail
1982 // call optimized function's reserved stack space needs to be aligned so that
1983 // taking the difference between two stack areas will result in an aligned
1985 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1988 std::max(MinReservedArea,
1989 PPCFrameLowering::getMinCallFrameSize(false, false));
1991 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1992 getStackAlignment();
1993 unsigned AlignMask = TargetAlign-1;
1994 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1996 FI->setMinReservedArea(MinReservedArea);
1998 SmallVector<SDValue, 8> MemOps;
2000 // If the function takes variable number of arguments, make a frame index for
2001 // the start of the first vararg value... for expansion of llvm.va_start.
2003 static const uint16_t GPArgRegs[] = {
2004 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2005 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2007 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
2009 static const uint16_t FPArgRegs[] = {
2010 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2013 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
2015 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
2017 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
2020 // Make room for NumGPArgRegs and NumFPArgRegs.
2021 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
2022 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
2024 FuncInfo->setVarArgsStackOffset(
2025 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2026 CCInfo.getNextStackOffset(), true));
2028 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
2029 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2031 // The fixed integer arguments of a variadic function are stored to the
2032 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
2033 // the result of va_next.
2034 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
2035 // Get an existing live-in vreg, or add a new one.
2036 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
2038 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
2040 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2041 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2042 MachinePointerInfo(), false, false, 0);
2043 MemOps.push_back(Store);
2044 // Increment the address by four for the next argument to store
2045 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2046 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2049 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
2051 // The double arguments are stored to the VarArgsFrameIndex
2053 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
2054 // Get an existing live-in vreg, or add a new one.
2055 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
2057 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
2059 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
2060 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2061 MachinePointerInfo(), false, false, 0);
2062 MemOps.push_back(Store);
2063 // Increment the address by eight for the next argument to store
2064 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
2066 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2070 if (!MemOps.empty())
2071 Chain = DAG.getNode(ISD::TokenFactor, dl,
2072 MVT::Other, &MemOps[0], MemOps.size());
2077 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2078 // value to MVT::i64 and then truncate to the correct register size.
2080 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2081 SelectionDAG &DAG, SDValue ArgVal,
2084 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2085 DAG.getValueType(ObjectVT));
2086 else if (Flags.isZExt())
2087 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2088 DAG.getValueType(ObjectVT));
2090 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2093 // Set the size that is at least reserved in caller of this function. Tail
2094 // call optimized functions' reserved stack space needs to be aligned so that
2095 // taking the difference between two stack areas will result in an aligned
2098 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2099 unsigned nAltivecParamsAtEnd,
2100 unsigned MinReservedArea,
2101 bool isPPC64) const {
2102 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2103 // Add the Altivec parameters at the end, if needed.
2104 if (nAltivecParamsAtEnd) {
2105 MinReservedArea = ((MinReservedArea+15)/16)*16;
2106 MinReservedArea += 16*nAltivecParamsAtEnd;
2109 std::max(MinReservedArea,
2110 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2111 unsigned TargetAlign
2112 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2113 getStackAlignment();
2114 unsigned AlignMask = TargetAlign-1;
2115 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2116 FI->setMinReservedArea(MinReservedArea);
2120 PPCTargetLowering::LowerFormalArguments_64SVR4(
2122 CallingConv::ID CallConv, bool isVarArg,
2123 const SmallVectorImpl<ISD::InputArg>
2125 SDLoc dl, SelectionDAG &DAG,
2126 SmallVectorImpl<SDValue> &InVals) const {
2127 // TODO: add description of PPC stack frame format, or at least some docs.
2129 MachineFunction &MF = DAG.getMachineFunction();
2130 MachineFrameInfo *MFI = MF.getFrameInfo();
2131 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2133 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2134 // Potential tail calls could cause overwriting of argument stack slots.
2135 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2136 (CallConv == CallingConv::Fast));
2137 unsigned PtrByteSize = 8;
2139 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2140 // Area that is at least reserved in caller of this function.
2141 unsigned MinReservedArea = ArgOffset;
2143 static const uint16_t GPR[] = {
2144 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2145 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2148 static const uint16_t *FPR = GetFPR();
2150 static const uint16_t VR[] = {
2151 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2152 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2155 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2156 const unsigned Num_FPR_Regs = 13;
2157 const unsigned Num_VR_Regs = array_lengthof(VR);
2159 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2161 // Add DAG nodes to load the arguments or copy them out of registers. On
2162 // entry to a function on PPC, the arguments start after the linkage area,
2163 // although the first ones are often in registers.
2165 SmallVector<SDValue, 8> MemOps;
2166 unsigned nAltivecParamsAtEnd = 0;
2167 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2168 unsigned CurArgIdx = 0;
2169 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2171 bool needsLoad = false;
2172 EVT ObjectVT = Ins[ArgNo].VT;
2173 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2174 unsigned ArgSize = ObjSize;
2175 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2176 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
2177 CurArgIdx = Ins[ArgNo].OrigArgIndex;
2179 unsigned CurArgOffset = ArgOffset;
2181 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2182 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2183 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2185 MinReservedArea = ((MinReservedArea+15)/16)*16;
2186 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2190 nAltivecParamsAtEnd++;
2192 // Calculate min reserved area.
2193 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2197 // FIXME the codegen can be much improved in some cases.
2198 // We do not have to keep everything in memory.
2199 if (Flags.isByVal()) {
2200 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2201 ObjSize = Flags.getByValSize();
2202 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2203 // Empty aggregate parameters do not take up registers. Examples:
2207 // etc. However, we have to provide a place-holder in InVals, so
2208 // pretend we have an 8-byte item at the current address for that
2211 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2212 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2213 InVals.push_back(FIN);
2216 // All aggregates smaller than 8 bytes must be passed right-justified.
2217 if (ObjSize < PtrByteSize)
2218 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2219 // The value of the object is its address.
2220 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2221 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2222 InVals.push_back(FIN);
2225 if (GPR_idx != Num_GPR_Regs) {
2226 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2227 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2230 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2231 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2232 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2233 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2234 MachinePointerInfo(FuncArg, CurArgOffset),
2235 ObjType, false, false, 0);
2237 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2238 // store the whole register as-is to the parameter save area
2239 // slot. The address of the parameter was already calculated
2240 // above (InVals.push_back(FIN)) to be the right-justified
2241 // offset within the slot. For this store, we need a new
2242 // frame index that points at the beginning of the slot.
2243 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2244 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2245 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2246 MachinePointerInfo(FuncArg, ArgOffset),
2250 MemOps.push_back(Store);
2253 // Whether we copied from a register or not, advance the offset
2254 // into the parameter save area by a full doubleword.
2255 ArgOffset += PtrByteSize;
2259 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2260 // Store whatever pieces of the object are in registers
2261 // to memory. ArgOffset will be the address of the beginning
2263 if (GPR_idx != Num_GPR_Regs) {
2265 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2266 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2267 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2268 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2269 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2270 MachinePointerInfo(FuncArg, ArgOffset),
2272 MemOps.push_back(Store);
2274 ArgOffset += PtrByteSize;
2276 ArgOffset += ArgSize - j;
2283 switch (ObjectVT.getSimpleVT().SimpleTy) {
2284 default: llvm_unreachable("Unhandled argument type!");
2287 if (GPR_idx != Num_GPR_Regs) {
2288 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2289 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2291 if (ObjectVT == MVT::i32)
2292 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2293 // value to MVT::i64 and then truncate to the correct register size.
2294 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2299 ArgSize = PtrByteSize;
2306 // Every 8 bytes of argument space consumes one of the GPRs available for
2307 // argument passing.
2308 if (GPR_idx != Num_GPR_Regs) {
2311 if (FPR_idx != Num_FPR_Regs) {
2314 if (ObjectVT == MVT::f32)
2315 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2317 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2319 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2323 ArgSize = PtrByteSize;
2332 // Note that vector arguments in registers don't reserve stack space,
2333 // except in varargs functions.
2334 if (VR_idx != Num_VR_Regs) {
2335 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2336 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2338 while ((ArgOffset % 16) != 0) {
2339 ArgOffset += PtrByteSize;
2340 if (GPR_idx != Num_GPR_Regs)
2344 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2348 // Vectors are aligned.
2349 ArgOffset = ((ArgOffset+15)/16)*16;
2350 CurArgOffset = ArgOffset;
2357 // We need to load the argument to a virtual register if we determined
2358 // above that we ran out of physical registers of the appropriate type.
2360 int FI = MFI->CreateFixedObject(ObjSize,
2361 CurArgOffset + (ArgSize - ObjSize),
2363 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2364 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2365 false, false, false, 0);
2368 InVals.push_back(ArgVal);
2371 // Set the size that is at least reserved in caller of this function. Tail
2372 // call optimized functions' reserved stack space needs to be aligned so that
2373 // taking the difference between two stack areas will result in an aligned
2375 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2377 // If the function takes variable number of arguments, make a frame index for
2378 // the start of the first vararg value... for expansion of llvm.va_start.
2380 int Depth = ArgOffset;
2382 FuncInfo->setVarArgsFrameIndex(
2383 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2384 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2386 // If this function is vararg, store any remaining integer argument regs
2387 // to their spots on the stack so that they may be loaded by deferencing the
2388 // result of va_next.
2389 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2390 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2391 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2392 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2393 MachinePointerInfo(), false, false, 0);
2394 MemOps.push_back(Store);
2395 // Increment the address by four for the next argument to store
2396 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2397 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2401 if (!MemOps.empty())
2402 Chain = DAG.getNode(ISD::TokenFactor, dl,
2403 MVT::Other, &MemOps[0], MemOps.size());
2409 PPCTargetLowering::LowerFormalArguments_Darwin(
2411 CallingConv::ID CallConv, bool isVarArg,
2412 const SmallVectorImpl<ISD::InputArg>
2414 SDLoc dl, SelectionDAG &DAG,
2415 SmallVectorImpl<SDValue> &InVals) const {
2416 // TODO: add description of PPC stack frame format, or at least some docs.
2418 MachineFunction &MF = DAG.getMachineFunction();
2419 MachineFrameInfo *MFI = MF.getFrameInfo();
2420 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2422 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2423 bool isPPC64 = PtrVT == MVT::i64;
2424 // Potential tail calls could cause overwriting of argument stack slots.
2425 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2426 (CallConv == CallingConv::Fast));
2427 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2429 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2430 // Area that is at least reserved in caller of this function.
2431 unsigned MinReservedArea = ArgOffset;
2433 static const uint16_t GPR_32[] = { // 32-bit registers.
2434 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2435 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2437 static const uint16_t GPR_64[] = { // 64-bit registers.
2438 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2439 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2442 static const uint16_t *FPR = GetFPR();
2444 static const uint16_t VR[] = {
2445 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2446 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2449 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2450 const unsigned Num_FPR_Regs = 13;
2451 const unsigned Num_VR_Regs = array_lengthof( VR);
2453 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2455 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2457 // In 32-bit non-varargs functions, the stack space for vectors is after the
2458 // stack space for non-vectors. We do not use this space unless we have
2459 // too many vectors to fit in registers, something that only occurs in
2460 // constructed examples:), but we have to walk the arglist to figure
2461 // that out...for the pathological case, compute VecArgOffset as the
2462 // start of the vector parameter area. Computing VecArgOffset is the
2463 // entire point of the following loop.
2464 unsigned VecArgOffset = ArgOffset;
2465 if (!isVarArg && !isPPC64) {
2466 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2468 EVT ObjectVT = Ins[ArgNo].VT;
2469 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2471 if (Flags.isByVal()) {
2472 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2473 unsigned ObjSize = Flags.getByValSize();
2475 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2476 VecArgOffset += ArgSize;
2480 switch(ObjectVT.getSimpleVT().SimpleTy) {
2481 default: llvm_unreachable("Unhandled argument type!");
2486 case MVT::i64: // PPC64
2488 // FIXME: We are guaranteed to be !isPPC64 at this point.
2489 // Does MVT::i64 apply?
2496 // Nothing to do, we're only looking at Nonvector args here.
2501 // We've found where the vector parameter area in memory is. Skip the
2502 // first 12 parameters; these don't use that memory.
2503 VecArgOffset = ((VecArgOffset+15)/16)*16;
2504 VecArgOffset += 12*16;
2506 // Add DAG nodes to load the arguments or copy them out of registers. On
2507 // entry to a function on PPC, the arguments start after the linkage area,
2508 // although the first ones are often in registers.
2510 SmallVector<SDValue, 8> MemOps;
2511 unsigned nAltivecParamsAtEnd = 0;
2512 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2513 unsigned CurArgIdx = 0;
2514 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2516 bool needsLoad = false;
2517 EVT ObjectVT = Ins[ArgNo].VT;
2518 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2519 unsigned ArgSize = ObjSize;
2520 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2521 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
2522 CurArgIdx = Ins[ArgNo].OrigArgIndex;
2524 unsigned CurArgOffset = ArgOffset;
2526 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2527 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2528 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2529 if (isVarArg || isPPC64) {
2530 MinReservedArea = ((MinReservedArea+15)/16)*16;
2531 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2534 } else nAltivecParamsAtEnd++;
2536 // Calculate min reserved area.
2537 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2541 // FIXME the codegen can be much improved in some cases.
2542 // We do not have to keep everything in memory.
2543 if (Flags.isByVal()) {
2544 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2545 ObjSize = Flags.getByValSize();
2546 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2547 // Objects of size 1 and 2 are right justified, everything else is
2548 // left justified. This means the memory address is adjusted forwards.
2549 if (ObjSize==1 || ObjSize==2) {
2550 CurArgOffset = CurArgOffset + (4 - ObjSize);
2552 // The value of the object is its address.
2553 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2554 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2555 InVals.push_back(FIN);
2556 if (ObjSize==1 || ObjSize==2) {
2557 if (GPR_idx != Num_GPR_Regs) {
2560 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2562 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2563 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2564 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2565 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2566 MachinePointerInfo(FuncArg,
2568 ObjType, false, false, 0);
2569 MemOps.push_back(Store);
2573 ArgOffset += PtrByteSize;
2577 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2578 // Store whatever pieces of the object are in registers
2579 // to memory. ArgOffset will be the address of the beginning
2581 if (GPR_idx != Num_GPR_Regs) {
2584 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2586 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2587 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2588 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2589 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2590 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2591 MachinePointerInfo(FuncArg, ArgOffset),
2593 MemOps.push_back(Store);
2595 ArgOffset += PtrByteSize;
2597 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2604 switch (ObjectVT.getSimpleVT().SimpleTy) {
2605 default: llvm_unreachable("Unhandled argument type!");
2608 if (GPR_idx != Num_GPR_Regs) {
2609 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2610 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2614 ArgSize = PtrByteSize;
2616 // All int arguments reserve stack space in the Darwin ABI.
2617 ArgOffset += PtrByteSize;
2621 case MVT::i64: // PPC64
2622 if (GPR_idx != Num_GPR_Regs) {
2623 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2624 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2626 if (ObjectVT == MVT::i32)
2627 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2628 // value to MVT::i64 and then truncate to the correct register size.
2629 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2634 ArgSize = PtrByteSize;
2636 // All int arguments reserve stack space in the Darwin ABI.
2642 // Every 4 bytes of argument space consumes one of the GPRs available for
2643 // argument passing.
2644 if (GPR_idx != Num_GPR_Regs) {
2646 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2649 if (FPR_idx != Num_FPR_Regs) {
2652 if (ObjectVT == MVT::f32)
2653 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2655 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2657 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2663 // All FP arguments reserve stack space in the Darwin ABI.
2664 ArgOffset += isPPC64 ? 8 : ObjSize;
2670 // Note that vector arguments in registers don't reserve stack space,
2671 // except in varargs functions.
2672 if (VR_idx != Num_VR_Regs) {
2673 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2674 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2676 while ((ArgOffset % 16) != 0) {
2677 ArgOffset += PtrByteSize;
2678 if (GPR_idx != Num_GPR_Regs)
2682 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2686 if (!isVarArg && !isPPC64) {
2687 // Vectors go after all the nonvectors.
2688 CurArgOffset = VecArgOffset;
2691 // Vectors are aligned.
2692 ArgOffset = ((ArgOffset+15)/16)*16;
2693 CurArgOffset = ArgOffset;
2701 // We need to load the argument to a virtual register if we determined above
2702 // that we ran out of physical registers of the appropriate type.
2704 int FI = MFI->CreateFixedObject(ObjSize,
2705 CurArgOffset + (ArgSize - ObjSize),
2707 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2708 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2709 false, false, false, 0);
2712 InVals.push_back(ArgVal);
2715 // Set the size that is at least reserved in caller of this function. Tail
2716 // call optimized functions' reserved stack space needs to be aligned so that
2717 // taking the difference between two stack areas will result in an aligned
2719 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2721 // If the function takes variable number of arguments, make a frame index for
2722 // the start of the first vararg value... for expansion of llvm.va_start.
2724 int Depth = ArgOffset;
2726 FuncInfo->setVarArgsFrameIndex(
2727 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2729 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2731 // If this function is vararg, store any remaining integer argument regs
2732 // to their spots on the stack so that they may be loaded by deferencing the
2733 // result of va_next.
2734 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2738 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2740 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2742 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2743 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2744 MachinePointerInfo(), false, false, 0);
2745 MemOps.push_back(Store);
2746 // Increment the address by four for the next argument to store
2747 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2748 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2752 if (!MemOps.empty())
2753 Chain = DAG.getNode(ISD::TokenFactor, dl,
2754 MVT::Other, &MemOps[0], MemOps.size());
2759 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2760 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2762 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2766 const SmallVectorImpl<ISD::OutputArg>
2768 const SmallVectorImpl<SDValue> &OutVals,
2769 unsigned &nAltivecParamsAtEnd) {
2770 // Count how many bytes are to be pushed on the stack, including the linkage
2771 // area, and parameter passing area. We start with 24/48 bytes, which is
2772 // prereserved space for [SP][CR][LR][3 x unused].
2773 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2774 unsigned NumOps = Outs.size();
2775 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2777 // Add up all the space actually used.
2778 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2779 // they all go in registers, but we must reserve stack space for them for
2780 // possible use by the caller. In varargs or 64-bit calls, parameters are
2781 // assigned stack space in order, with padding so Altivec parameters are
2783 nAltivecParamsAtEnd = 0;
2784 for (unsigned i = 0; i != NumOps; ++i) {
2785 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2786 EVT ArgVT = Outs[i].VT;
2787 // Varargs Altivec parameters are padded to a 16 byte boundary.
2788 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2789 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2790 if (!isVarArg && !isPPC64) {
2791 // Non-varargs Altivec parameters go after all the non-Altivec
2792 // parameters; handle those later so we know how much padding we need.
2793 nAltivecParamsAtEnd++;
2796 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2797 NumBytes = ((NumBytes+15)/16)*16;
2799 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2802 // Allow for Altivec parameters at the end, if needed.
2803 if (nAltivecParamsAtEnd) {
2804 NumBytes = ((NumBytes+15)/16)*16;
2805 NumBytes += 16*nAltivecParamsAtEnd;
2808 // The prolog code of the callee may store up to 8 GPR argument registers to
2809 // the stack, allowing va_start to index over them in memory if its varargs.
2810 // Because we cannot tell if this is needed on the caller side, we have to
2811 // conservatively assume that it is needed. As such, make sure we have at
2812 // least enough stack space for the caller to store the 8 GPRs.
2813 NumBytes = std::max(NumBytes,
2814 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2816 // Tail call needs the stack to be aligned.
2817 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2818 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2819 getFrameLowering()->getStackAlignment();
2820 unsigned AlignMask = TargetAlign-1;
2821 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2827 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2828 /// adjusted to accommodate the arguments for the tailcall.
2829 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2830 unsigned ParamSize) {
2832 if (!isTailCall) return 0;
2834 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2835 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2836 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2837 // Remember only if the new adjustement is bigger.
2838 if (SPDiff < FI->getTailCallSPDelta())
2839 FI->setTailCallSPDelta(SPDiff);
2844 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2845 /// for tail call optimization. Targets which want to do tail call
2846 /// optimization should implement this function.
2848 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2849 CallingConv::ID CalleeCC,
2851 const SmallVectorImpl<ISD::InputArg> &Ins,
2852 SelectionDAG& DAG) const {
2853 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2856 // Variable argument functions are not supported.
2860 MachineFunction &MF = DAG.getMachineFunction();
2861 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2862 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2863 // Functions containing by val parameters are not supported.
2864 for (unsigned i = 0; i != Ins.size(); i++) {
2865 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2866 if (Flags.isByVal()) return false;
2869 // Non PIC/GOT tail calls are supported.
2870 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2873 // At the moment we can only do local tail calls (in same module, hidden
2874 // or protected) if we are generating PIC.
2875 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2876 return G->getGlobal()->hasHiddenVisibility()
2877 || G->getGlobal()->hasProtectedVisibility();
2883 /// isCallCompatibleAddress - Return the immediate to use if the specified
2884 /// 32-bit value is representable in the immediate field of a BxA instruction.
2885 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2886 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2889 int Addr = C->getZExtValue();
2890 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2891 SignExtend32<26>(Addr) != Addr)
2892 return 0; // Top 6 bits have to be sext of immediate.
2894 return DAG.getConstant((int)C->getZExtValue() >> 2,
2895 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2900 struct TailCallArgumentInfo {
2905 TailCallArgumentInfo() : FrameIdx(0) {}
2910 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2912 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2914 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2915 SmallVector<SDValue, 8> &MemOpChains,
2917 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2918 SDValue Arg = TailCallArgs[i].Arg;
2919 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2920 int FI = TailCallArgs[i].FrameIdx;
2921 // Store relative to framepointer.
2922 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2923 MachinePointerInfo::getFixedStack(FI),
2928 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2929 /// the appropriate stack slot for the tail call optimized function call.
2930 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2931 MachineFunction &MF,
2940 // Calculate the new stack slot for the return address.
2941 int SlotSize = isPPC64 ? 8 : 4;
2942 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2944 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2945 NewRetAddrLoc, true);
2946 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2947 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2948 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2949 MachinePointerInfo::getFixedStack(NewRetAddr),
2952 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2953 // slot as the FP is never overwritten.
2956 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2957 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2959 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2960 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2961 MachinePointerInfo::getFixedStack(NewFPIdx),
2968 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2969 /// the position of the argument.
2971 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2972 SDValue Arg, int SPDiff, unsigned ArgOffset,
2973 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2974 int Offset = ArgOffset + SPDiff;
2975 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2976 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2977 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2978 SDValue FIN = DAG.getFrameIndex(FI, VT);
2979 TailCallArgumentInfo Info;
2981 Info.FrameIdxOp = FIN;
2983 TailCallArguments.push_back(Info);
2986 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2987 /// stack slot. Returns the chain as result and the loaded frame pointers in
2988 /// LROpOut/FPOpout. Used when tail calling.
2989 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2997 // Load the LR and FP stack slot for later adjusting.
2998 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2999 LROpOut = getReturnAddrFrameIndex(DAG);
3000 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
3001 false, false, false, 0);
3002 Chain = SDValue(LROpOut.getNode(), 1);
3004 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
3005 // slot as the FP is never overwritten.
3007 FPOpOut = getFramePointerFrameIndex(DAG);
3008 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
3009 false, false, false, 0);
3010 Chain = SDValue(FPOpOut.getNode(), 1);
3016 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
3017 /// by "Src" to address "Dst" of size "Size". Alignment information is
3018 /// specified by the specific parameter attribute. The copy will be passed as
3019 /// a byval function parameter.
3020 /// Sometimes what we are copying is the end of a larger object, the part that
3021 /// does not fit in registers.
3023 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
3024 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
3026 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
3027 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
3028 false, false, MachinePointerInfo(0),
3029 MachinePointerInfo(0));
3032 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
3035 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
3036 SDValue Arg, SDValue PtrOff, int SPDiff,
3037 unsigned ArgOffset, bool isPPC64, bool isTailCall,
3038 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
3039 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
3041 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3046 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3048 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3049 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3050 DAG.getConstant(ArgOffset, PtrVT));
3052 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3053 MachinePointerInfo(), false, false, 0));
3054 // Calculate and remember argument location.
3055 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
3060 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
3061 SDLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
3062 SDValue LROp, SDValue FPOp, bool isDarwinABI,
3063 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
3064 MachineFunction &MF = DAG.getMachineFunction();
3066 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
3067 // might overwrite each other in case of tail call optimization.
3068 SmallVector<SDValue, 8> MemOpChains2;
3069 // Do not flag preceding copytoreg stuff together with the following stuff.
3071 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3073 if (!MemOpChains2.empty())
3074 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3075 &MemOpChains2[0], MemOpChains2.size());
3077 // Store the return address to the appropriate stack slot.
3078 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3079 isPPC64, isDarwinABI, dl);
3081 // Emit callseq_end just before tailcall node.
3082 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3083 DAG.getIntPtrConstant(0, true), InFlag, dl);
3084 InFlag = Chain.getValue(1);
3088 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3089 SDValue &Chain, SDLoc dl, int SPDiff, bool isTailCall,
3090 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
3091 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
3092 const PPCSubtarget &PPCSubTarget) {
3094 bool isPPC64 = PPCSubTarget.isPPC64();
3095 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3097 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3098 NodeTys.push_back(MVT::Other); // Returns a chain
3099 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3101 unsigned CallOpc = PPCISD::CALL;
3103 bool needIndirectCall = true;
3104 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3105 // If this is an absolute destination address, use the munged value.
3106 Callee = SDValue(Dest, 0);
3107 needIndirectCall = false;
3110 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3111 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3112 // Use indirect calls for ALL functions calls in JIT mode, since the
3113 // far-call stubs may be outside relocation limits for a BL instruction.
3114 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3115 unsigned OpFlags = 0;
3116 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3117 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3118 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3119 (G->getGlobal()->isDeclaration() ||
3120 G->getGlobal()->isWeakForLinker())) {
3121 // PC-relative references to external symbols should go through $stub,
3122 // unless we're building with the leopard linker or later, which
3123 // automatically synthesizes these stubs.
3124 OpFlags = PPCII::MO_DARWIN_STUB;
3127 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3128 // every direct call is) turn it into a TargetGlobalAddress /
3129 // TargetExternalSymbol node so that legalize doesn't hack it.
3130 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3131 Callee.getValueType(),
3133 needIndirectCall = false;
3137 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3138 unsigned char OpFlags = 0;
3140 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3141 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3142 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3143 // PC-relative references to external symbols should go through $stub,
3144 // unless we're building with the leopard linker or later, which
3145 // automatically synthesizes these stubs.
3146 OpFlags = PPCII::MO_DARWIN_STUB;
3149 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3151 needIndirectCall = false;
3154 if (needIndirectCall) {
3155 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3156 // to do the call, we can't use PPCISD::CALL.
3157 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3159 if (isSVR4ABI && isPPC64) {
3160 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3161 // entry point, but to the function descriptor (the function entry point
3162 // address is part of the function descriptor though).
3163 // The function descriptor is a three doubleword structure with the
3164 // following fields: function entry point, TOC base address and
3165 // environment pointer.
3166 // Thus for a call through a function pointer, the following actions need
3168 // 1. Save the TOC of the caller in the TOC save area of its stack
3169 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3170 // 2. Load the address of the function entry point from the function
3172 // 3. Load the TOC of the callee from the function descriptor into r2.
3173 // 4. Load the environment pointer from the function descriptor into
3175 // 5. Branch to the function entry point address.
3176 // 6. On return of the callee, the TOC of the caller needs to be
3177 // restored (this is done in FinishCall()).
3179 // All those operations are flagged together to ensure that no other
3180 // operations can be scheduled in between. E.g. without flagging the
3181 // operations together, a TOC access in the caller could be scheduled
3182 // between the load of the callee TOC and the branch to the callee, which
3183 // results in the TOC access going through the TOC of the callee instead
3184 // of going through the TOC of the caller, which leads to incorrect code.
3186 // Load the address of the function entry point from the function
3188 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3189 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3190 InFlag.getNode() ? 3 : 2);
3191 Chain = LoadFuncPtr.getValue(1);
3192 InFlag = LoadFuncPtr.getValue(2);
3194 // Load environment pointer into r11.
3195 // Offset of the environment pointer within the function descriptor.
3196 SDValue PtrOff = DAG.getIntPtrConstant(16);
3198 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3199 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3201 Chain = LoadEnvPtr.getValue(1);
3202 InFlag = LoadEnvPtr.getValue(2);
3204 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3206 Chain = EnvVal.getValue(0);
3207 InFlag = EnvVal.getValue(1);
3209 // Load TOC of the callee into r2. We are using a target-specific load
3210 // with r2 hard coded, because the result of a target-independent load
3211 // would never go directly into r2, since r2 is a reserved register (which
3212 // prevents the register allocator from allocating it), resulting in an
3213 // additional register being allocated and an unnecessary move instruction
3215 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3216 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3218 Chain = LoadTOCPtr.getValue(0);
3219 InFlag = LoadTOCPtr.getValue(1);
3221 MTCTROps[0] = Chain;
3222 MTCTROps[1] = LoadFuncPtr;
3223 MTCTROps[2] = InFlag;
3226 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3227 2 + (InFlag.getNode() != 0));
3228 InFlag = Chain.getValue(1);
3231 NodeTys.push_back(MVT::Other);
3232 NodeTys.push_back(MVT::Glue);
3233 Ops.push_back(Chain);
3234 CallOpc = PPCISD::BCTRL;
3236 // Add use of X11 (holding environment pointer)
3237 if (isSVR4ABI && isPPC64)
3238 Ops.push_back(DAG.getRegister(PPC::X11, PtrVT));
3239 // Add CTR register as callee so a bctr can be emitted later.
3241 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3244 // If this is a direct call, pass the chain and the callee.
3245 if (Callee.getNode()) {
3246 Ops.push_back(Chain);
3247 Ops.push_back(Callee);
3249 // If this is a tail call add stack pointer delta.
3251 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3253 // Add argument registers to the end of the list so that they are known live
3255 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3256 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3257 RegsToPass[i].second.getValueType()));
3263 bool isLocalCall(const SDValue &Callee)
3265 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3266 return !G->getGlobal()->isDeclaration() &&
3267 !G->getGlobal()->isWeakForLinker();
3272 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3273 CallingConv::ID CallConv, bool isVarArg,
3274 const SmallVectorImpl<ISD::InputArg> &Ins,
3275 SDLoc dl, SelectionDAG &DAG,
3276 SmallVectorImpl<SDValue> &InVals) const {
3278 SmallVector<CCValAssign, 16> RVLocs;
3279 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3280 getTargetMachine(), RVLocs, *DAG.getContext());
3281 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3283 // Copy all of the result registers out of their specified physreg.
3284 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3285 CCValAssign &VA = RVLocs[i];
3286 assert(VA.isRegLoc() && "Can only return in registers!");
3288 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3289 VA.getLocReg(), VA.getLocVT(), InFlag);
3290 Chain = Val.getValue(1);
3291 InFlag = Val.getValue(2);
3293 switch (VA.getLocInfo()) {
3294 default: llvm_unreachable("Unknown loc info!");
3295 case CCValAssign::Full: break;
3296 case CCValAssign::AExt:
3297 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3299 case CCValAssign::ZExt:
3300 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3301 DAG.getValueType(VA.getValVT()));
3302 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3304 case CCValAssign::SExt:
3305 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3306 DAG.getValueType(VA.getValVT()));
3307 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3311 InVals.push_back(Val);
3318 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, SDLoc dl,
3319 bool isTailCall, bool isVarArg,
3321 SmallVector<std::pair<unsigned, SDValue>, 8>
3323 SDValue InFlag, SDValue Chain,
3325 int SPDiff, unsigned NumBytes,
3326 const SmallVectorImpl<ISD::InputArg> &Ins,
3327 SmallVectorImpl<SDValue> &InVals) const {
3328 std::vector<EVT> NodeTys;
3329 SmallVector<SDValue, 8> Ops;
3330 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3331 isTailCall, RegsToPass, Ops, NodeTys,
3334 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3335 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3336 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3338 // When performing tail call optimization the callee pops its arguments off
3339 // the stack. Account for this here so these bytes can be pushed back on in
3340 // PPCFrameLowering::eliminateCallFramePseudoInstr.
3341 int BytesCalleePops =
3342 (CallConv == CallingConv::Fast &&
3343 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3345 // Add a register mask operand representing the call-preserved registers.
3346 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3347 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3348 assert(Mask && "Missing call preserved mask for calling convention");
3349 Ops.push_back(DAG.getRegisterMask(Mask));
3351 if (InFlag.getNode())
3352 Ops.push_back(InFlag);
3356 assert(((Callee.getOpcode() == ISD::Register &&
3357 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3358 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3359 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3360 isa<ConstantSDNode>(Callee)) &&
3361 "Expecting an global address, external symbol, absolute value or register");
3363 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3366 // Add a NOP immediately after the branch instruction when using the 64-bit
3367 // SVR4 ABI. At link time, if caller and callee are in a different module and
3368 // thus have a different TOC, the call will be replaced with a call to a stub
3369 // function which saves the current TOC, loads the TOC of the callee and
3370 // branches to the callee. The NOP will be replaced with a load instruction
3371 // which restores the TOC of the caller from the TOC save slot of the current
3372 // stack frame. If caller and callee belong to the same module (and have the
3373 // same TOC), the NOP will remain unchanged.
3375 bool needsTOCRestore = false;
3376 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3377 if (CallOpc == PPCISD::BCTRL) {
3378 // This is a call through a function pointer.
3379 // Restore the caller TOC from the save area into R2.
3380 // See PrepareCall() for more information about calls through function
3381 // pointers in the 64-bit SVR4 ABI.
3382 // We are using a target-specific load with r2 hard coded, because the
3383 // result of a target-independent load would never go directly into r2,
3384 // since r2 is a reserved register (which prevents the register allocator
3385 // from allocating it), resulting in an additional register being
3386 // allocated and an unnecessary move instruction being generated.
3387 needsTOCRestore = true;
3388 } else if ((CallOpc == PPCISD::CALL) && !isLocalCall(Callee)) {
3389 // Otherwise insert NOP for non-local calls.
3390 CallOpc = PPCISD::CALL_NOP;
3394 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3395 InFlag = Chain.getValue(1);
3397 if (needsTOCRestore) {
3398 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3399 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3400 InFlag = Chain.getValue(1);
3403 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3404 DAG.getIntPtrConstant(BytesCalleePops, true),
3407 InFlag = Chain.getValue(1);
3409 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3410 Ins, dl, DAG, InVals);
3414 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3415 SmallVectorImpl<SDValue> &InVals) const {
3416 SelectionDAG &DAG = CLI.DAG;
3418 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3419 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3420 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3421 SDValue Chain = CLI.Chain;
3422 SDValue Callee = CLI.Callee;
3423 bool &isTailCall = CLI.IsTailCall;
3424 CallingConv::ID CallConv = CLI.CallConv;
3425 bool isVarArg = CLI.IsVarArg;
3428 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3431 if (PPCSubTarget.isSVR4ABI()) {
3432 if (PPCSubTarget.isPPC64())
3433 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3434 isTailCall, Outs, OutVals, Ins,
3437 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3438 isTailCall, Outs, OutVals, Ins,
3442 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3443 isTailCall, Outs, OutVals, Ins,
3448 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3449 CallingConv::ID CallConv, bool isVarArg,
3451 const SmallVectorImpl<ISD::OutputArg> &Outs,
3452 const SmallVectorImpl<SDValue> &OutVals,
3453 const SmallVectorImpl<ISD::InputArg> &Ins,
3454 SDLoc dl, SelectionDAG &DAG,
3455 SmallVectorImpl<SDValue> &InVals) const {
3456 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3457 // of the 32-bit SVR4 ABI stack frame layout.
3459 assert((CallConv == CallingConv::C ||
3460 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3462 unsigned PtrByteSize = 4;
3464 MachineFunction &MF = DAG.getMachineFunction();
3466 // Mark this function as potentially containing a function that contains a
3467 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3468 // and restoring the callers stack pointer in this functions epilog. This is
3469 // done because by tail calling the called function might overwrite the value
3470 // in this function's (MF) stack pointer stack slot 0(SP).
3471 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3472 CallConv == CallingConv::Fast)
3473 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3475 // Count how many bytes are to be pushed on the stack, including the linkage
3476 // area, parameter list area and the part of the local variable space which
3477 // contains copies of aggregates which are passed by value.
3479 // Assign locations to all of the outgoing arguments.
3480 SmallVector<CCValAssign, 16> ArgLocs;
3481 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3482 getTargetMachine(), ArgLocs, *DAG.getContext());
3484 // Reserve space for the linkage area on the stack.
3485 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3488 // Handle fixed and variable vector arguments differently.
3489 // Fixed vector arguments go into registers as long as registers are
3490 // available. Variable vector arguments always go into memory.
3491 unsigned NumArgs = Outs.size();
3493 for (unsigned i = 0; i != NumArgs; ++i) {
3494 MVT ArgVT = Outs[i].VT;
3495 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3498 if (Outs[i].IsFixed) {
3499 Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3502 Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3508 errs() << "Call operand #" << i << " has unhandled type "
3509 << EVT(ArgVT).getEVTString() << "\n";
3511 llvm_unreachable(0);
3515 // All arguments are treated the same.
3516 CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
3519 // Assign locations to all of the outgoing aggregate by value arguments.
3520 SmallVector<CCValAssign, 16> ByValArgLocs;
3521 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3522 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3524 // Reserve stack space for the allocations in CCInfo.
3525 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3527 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
3529 // Size of the linkage area, parameter list area and the part of the local
3530 // space variable where copies of aggregates which are passed by value are
3532 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3534 // Calculate by how many bytes the stack has to be adjusted in case of tail
3535 // call optimization.
3536 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3538 // Adjust the stack pointer for the new arguments...
3539 // These operations are automatically eliminated by the prolog/epilog pass
3540 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
3542 SDValue CallSeqStart = Chain;
3544 // Load the return address and frame pointer so it can be moved somewhere else
3547 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3550 // Set up a copy of the stack pointer for use loading and storing any
3551 // arguments that may not fit in the registers available for argument
3553 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3555 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3556 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3557 SmallVector<SDValue, 8> MemOpChains;
3559 bool seenFloatArg = false;
3560 // Walk the register/memloc assignments, inserting copies/loads.
3561 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3564 CCValAssign &VA = ArgLocs[i];
3565 SDValue Arg = OutVals[i];
3566 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3568 if (Flags.isByVal()) {
3569 // Argument is an aggregate which is passed by value, thus we need to
3570 // create a copy of it in the local variable space of the current stack
3571 // frame (which is the stack frame of the caller) and pass the address of
3572 // this copy to the callee.
3573 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3574 CCValAssign &ByValVA = ByValArgLocs[j++];
3575 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3577 // Memory reserved in the local variable space of the callers stack frame.
3578 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3580 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3581 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3583 // Create a copy of the argument in the local area of the current
3585 SDValue MemcpyCall =
3586 CreateCopyOfByValArgument(Arg, PtrOff,
3587 CallSeqStart.getNode()->getOperand(0),
3590 // This must go outside the CALLSEQ_START..END.
3591 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3592 CallSeqStart.getNode()->getOperand(1),
3594 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3595 NewCallSeqStart.getNode());
3596 Chain = CallSeqStart = NewCallSeqStart;
3598 // Pass the address of the aggregate copy on the stack either in a
3599 // physical register or in the parameter list area of the current stack
3600 // frame to the callee.
3604 if (VA.isRegLoc()) {
3605 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3606 // Put argument in a physical register.
3607 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3609 // Put argument in the parameter list area of the current stack frame.
3610 assert(VA.isMemLoc());
3611 unsigned LocMemOffset = VA.getLocMemOffset();
3614 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3615 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3617 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3618 MachinePointerInfo(),
3621 // Calculate and remember argument location.
3622 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3628 if (!MemOpChains.empty())
3629 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3630 &MemOpChains[0], MemOpChains.size());
3632 // Build a sequence of copy-to-reg nodes chained together with token chain
3633 // and flag operands which copy the outgoing args into the appropriate regs.
3635 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3636 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3637 RegsToPass[i].second, InFlag);
3638 InFlag = Chain.getValue(1);
3641 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3644 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3645 SDValue Ops[] = { Chain, InFlag };
3647 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3648 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3650 InFlag = Chain.getValue(1);
3654 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3655 false, TailCallArguments);
3657 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3658 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3662 // Copy an argument into memory, being careful to do this outside the
3663 // call sequence for the call to which the argument belongs.
3665 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3666 SDValue CallSeqStart,
3667 ISD::ArgFlagsTy Flags,
3670 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3671 CallSeqStart.getNode()->getOperand(0),
3673 // The MEMCPY must go outside the CALLSEQ_START..END.
3674 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3675 CallSeqStart.getNode()->getOperand(1),
3677 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3678 NewCallSeqStart.getNode());
3679 return NewCallSeqStart;
3683 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3684 CallingConv::ID CallConv, bool isVarArg,
3686 const SmallVectorImpl<ISD::OutputArg> &Outs,
3687 const SmallVectorImpl<SDValue> &OutVals,
3688 const SmallVectorImpl<ISD::InputArg> &Ins,
3689 SDLoc dl, SelectionDAG &DAG,
3690 SmallVectorImpl<SDValue> &InVals) const {
3692 unsigned NumOps = Outs.size();
3694 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3695 unsigned PtrByteSize = 8;
3697 MachineFunction &MF = DAG.getMachineFunction();
3699 // Mark this function as potentially containing a function that contains a
3700 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3701 // and restoring the callers stack pointer in this functions epilog. This is
3702 // done because by tail calling the called function might overwrite the value
3703 // in this function's (MF) stack pointer stack slot 0(SP).
3704 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3705 CallConv == CallingConv::Fast)
3706 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3708 unsigned nAltivecParamsAtEnd = 0;
3710 // Count how many bytes are to be pushed on the stack, including the linkage
3711 // area, and parameter passing area. We start with at least 48 bytes, which
3712 // is reserved space for [SP][CR][LR][3 x unused].
3713 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3716 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3717 Outs, OutVals, nAltivecParamsAtEnd);
3719 // Calculate by how many bytes the stack has to be adjusted in case of tail
3720 // call optimization.
3721 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3723 // To protect arguments on the stack from being clobbered in a tail call,
3724 // force all the loads to happen before doing any other lowering.
3726 Chain = DAG.getStackArgumentTokenFactor(Chain);
3728 // Adjust the stack pointer for the new arguments...
3729 // These operations are automatically eliminated by the prolog/epilog pass
3730 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
3732 SDValue CallSeqStart = Chain;
3734 // Load the return address and frame pointer so it can be move somewhere else
3737 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3740 // Set up a copy of the stack pointer for use loading and storing any
3741 // arguments that may not fit in the registers available for argument
3743 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3745 // Figure out which arguments are going to go in registers, and which in
3746 // memory. Also, if this is a vararg function, floating point operations
3747 // must be stored to our stack, and loaded into integer regs as well, if
3748 // any integer regs are available for argument passing.
3749 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3750 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3752 static const uint16_t GPR[] = {
3753 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3754 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3756 static const uint16_t *FPR = GetFPR();
3758 static const uint16_t VR[] = {
3759 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3760 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3762 const unsigned NumGPRs = array_lengthof(GPR);
3763 const unsigned NumFPRs = 13;
3764 const unsigned NumVRs = array_lengthof(VR);
3766 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3767 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3769 SmallVector<SDValue, 8> MemOpChains;
3770 for (unsigned i = 0; i != NumOps; ++i) {
3771 SDValue Arg = OutVals[i];
3772 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3774 // PtrOff will be used to store the current argument to the stack if a
3775 // register cannot be found for it.
3778 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3780 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3782 // Promote integers to 64-bit values.
3783 if (Arg.getValueType() == MVT::i32) {
3784 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3785 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3786 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3789 // FIXME memcpy is used way more than necessary. Correctness first.
3790 // Note: "by value" is code for passing a structure by value, not
3792 if (Flags.isByVal()) {
3793 // Note: Size includes alignment padding, so
3794 // struct x { short a; char b; }
3795 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3796 // These are the proper values we need for right-justifying the
3797 // aggregate in a parameter register.
3798 unsigned Size = Flags.getByValSize();
3800 // An empty aggregate parameter takes up no storage and no
3805 // All aggregates smaller than 8 bytes must be passed right-justified.
3806 if (Size==1 || Size==2 || Size==4) {
3807 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3808 if (GPR_idx != NumGPRs) {
3809 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3810 MachinePointerInfo(), VT,
3812 MemOpChains.push_back(Load.getValue(1));
3813 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3815 ArgOffset += PtrByteSize;
3820 if (GPR_idx == NumGPRs && Size < 8) {
3821 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3822 PtrOff.getValueType());
3823 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3824 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3827 ArgOffset += PtrByteSize;
3830 // Copy entire object into memory. There are cases where gcc-generated
3831 // code assumes it is there, even if it could be put entirely into
3832 // registers. (This is not what the doc says.)
3834 // FIXME: The above statement is likely due to a misunderstanding of the
3835 // documents. All arguments must be copied into the parameter area BY
3836 // THE CALLEE in the event that the callee takes the address of any
3837 // formal argument. That has not yet been implemented. However, it is
3838 // reasonable to use the stack area as a staging area for the register
3841 // Skip this for small aggregates, as we will use the same slot for a
3842 // right-justified copy, below.
3844 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3848 // When a register is available, pass a small aggregate right-justified.
3849 if (Size < 8 && GPR_idx != NumGPRs) {
3850 // The easiest way to get this right-justified in a register
3851 // is to copy the structure into the rightmost portion of a
3852 // local variable slot, then load the whole slot into the
3854 // FIXME: The memcpy seems to produce pretty awful code for
3855 // small aggregates, particularly for packed ones.
3856 // FIXME: It would be preferable to use the slot in the
3857 // parameter save area instead of a new local variable.
3858 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3859 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3860 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3864 // Load the slot into the register.
3865 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3866 MachinePointerInfo(),
3867 false, false, false, 0);
3868 MemOpChains.push_back(Load.getValue(1));
3869 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3871 // Done with this argument.
3872 ArgOffset += PtrByteSize;
3876 // For aggregates larger than PtrByteSize, copy the pieces of the
3877 // object that fit into registers from the parameter save area.
3878 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3879 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3880 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3881 if (GPR_idx != NumGPRs) {
3882 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3883 MachinePointerInfo(),
3884 false, false, false, 0);
3885 MemOpChains.push_back(Load.getValue(1));
3886 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3887 ArgOffset += PtrByteSize;
3889 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3896 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3897 default: llvm_unreachable("Unexpected ValueType for argument!");
3900 if (GPR_idx != NumGPRs) {
3901 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3903 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3904 true, isTailCall, false, MemOpChains,
3905 TailCallArguments, dl);
3907 ArgOffset += PtrByteSize;
3911 if (FPR_idx != NumFPRs) {
3912 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3915 // A single float or an aggregate containing only a single float
3916 // must be passed right-justified in the stack doubleword, and
3917 // in the GPR, if one is available.
3919 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3920 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3921 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3925 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3926 MachinePointerInfo(), false, false, 0);
3927 MemOpChains.push_back(Store);
3929 // Float varargs are always shadowed in available integer registers
3930 if (GPR_idx != NumGPRs) {
3931 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3932 MachinePointerInfo(), false, false,
3934 MemOpChains.push_back(Load.getValue(1));
3935 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3937 } else if (GPR_idx != NumGPRs)
3938 // If we have any FPRs remaining, we may also have GPRs remaining.
3941 // Single-precision floating-point values are mapped to the
3942 // second (rightmost) word of the stack doubleword.
3943 if (Arg.getValueType() == MVT::f32) {
3944 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3945 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3948 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3949 true, isTailCall, false, MemOpChains,
3950 TailCallArguments, dl);
3959 // These go aligned on the stack, or in the corresponding R registers
3960 // when within range. The Darwin PPC ABI doc claims they also go in
3961 // V registers; in fact gcc does this only for arguments that are
3962 // prototyped, not for those that match the ... We do it for all
3963 // arguments, seems to work.
3964 while (ArgOffset % 16 !=0) {
3965 ArgOffset += PtrByteSize;
3966 if (GPR_idx != NumGPRs)
3969 // We could elide this store in the case where the object fits
3970 // entirely in R registers. Maybe later.
3971 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3972 DAG.getConstant(ArgOffset, PtrVT));
3973 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3974 MachinePointerInfo(), false, false, 0);
3975 MemOpChains.push_back(Store);
3976 if (VR_idx != NumVRs) {
3977 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3978 MachinePointerInfo(),
3979 false, false, false, 0);
3980 MemOpChains.push_back(Load.getValue(1));
3981 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3984 for (unsigned i=0; i<16; i+=PtrByteSize) {
3985 if (GPR_idx == NumGPRs)
3987 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3988 DAG.getConstant(i, PtrVT));
3989 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3990 false, false, false, 0);
3991 MemOpChains.push_back(Load.getValue(1));
3992 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3997 // Non-varargs Altivec params generally go in registers, but have
3998 // stack space allocated at the end.
3999 if (VR_idx != NumVRs) {
4000 // Doesn't have GPR space allocated.
4001 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4003 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4004 true, isTailCall, true, MemOpChains,
4005 TailCallArguments, dl);
4012 if (!MemOpChains.empty())
4013 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4014 &MemOpChains[0], MemOpChains.size());
4016 // Check if this is an indirect call (MTCTR/BCTRL).
4017 // See PrepareCall() for more information about calls through function
4018 // pointers in the 64-bit SVR4 ABI.
4020 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4021 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4022 !isBLACompatibleAddress(Callee, DAG)) {
4023 // Load r2 into a virtual register and store it to the TOC save area.
4024 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
4025 // TOC save area offset.
4026 SDValue PtrOff = DAG.getIntPtrConstant(40);
4027 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4028 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
4030 // R12 must contain the address of an indirect callee. This does not
4031 // mean the MTCTR instruction must use R12; it's easier to model this
4032 // as an extra parameter, so do that.
4033 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
4036 // Build a sequence of copy-to-reg nodes chained together with token chain
4037 // and flag operands which copy the outgoing args into the appropriate regs.
4039 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4040 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4041 RegsToPass[i].second, InFlag);
4042 InFlag = Chain.getValue(1);
4046 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
4047 FPOp, true, TailCallArguments);
4049 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4050 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4055 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
4056 CallingConv::ID CallConv, bool isVarArg,
4058 const SmallVectorImpl<ISD::OutputArg> &Outs,
4059 const SmallVectorImpl<SDValue> &OutVals,
4060 const SmallVectorImpl<ISD::InputArg> &Ins,
4061 SDLoc dl, SelectionDAG &DAG,
4062 SmallVectorImpl<SDValue> &InVals) const {
4064 unsigned NumOps = Outs.size();
4066 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4067 bool isPPC64 = PtrVT == MVT::i64;
4068 unsigned PtrByteSize = isPPC64 ? 8 : 4;
4070 MachineFunction &MF = DAG.getMachineFunction();
4072 // Mark this function as potentially containing a function that contains a
4073 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4074 // and restoring the callers stack pointer in this functions epilog. This is
4075 // done because by tail calling the called function might overwrite the value
4076 // in this function's (MF) stack pointer stack slot 0(SP).
4077 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4078 CallConv == CallingConv::Fast)
4079 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4081 unsigned nAltivecParamsAtEnd = 0;
4083 // Count how many bytes are to be pushed on the stack, including the linkage
4084 // area, and parameter passing area. We start with 24/48 bytes, which is
4085 // prereserved space for [SP][CR][LR][3 x unused].
4087 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
4089 nAltivecParamsAtEnd);
4091 // Calculate by how many bytes the stack has to be adjusted in case of tail
4092 // call optimization.
4093 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4095 // To protect arguments on the stack from being clobbered in a tail call,
4096 // force all the loads to happen before doing any other lowering.
4098 Chain = DAG.getStackArgumentTokenFactor(Chain);
4100 // Adjust the stack pointer for the new arguments...
4101 // These operations are automatically eliminated by the prolog/epilog pass
4102 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
4104 SDValue CallSeqStart = Chain;
4106 // Load the return address and frame pointer so it can be move somewhere else
4109 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4112 // Set up a copy of the stack pointer for use loading and storing any
4113 // arguments that may not fit in the registers available for argument
4117 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4119 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4121 // Figure out which arguments are going to go in registers, and which in
4122 // memory. Also, if this is a vararg function, floating point operations
4123 // must be stored to our stack, and loaded into integer regs as well, if
4124 // any integer regs are available for argument passing.
4125 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4126 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4128 static const uint16_t GPR_32[] = { // 32-bit registers.
4129 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4130 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4132 static const uint16_t GPR_64[] = { // 64-bit registers.
4133 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4134 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4136 static const uint16_t *FPR = GetFPR();
4138 static const uint16_t VR[] = {
4139 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4140 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4142 const unsigned NumGPRs = array_lengthof(GPR_32);
4143 const unsigned NumFPRs = 13;
4144 const unsigned NumVRs = array_lengthof(VR);
4146 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4148 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4149 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4151 SmallVector<SDValue, 8> MemOpChains;
4152 for (unsigned i = 0; i != NumOps; ++i) {
4153 SDValue Arg = OutVals[i];
4154 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4156 // PtrOff will be used to store the current argument to the stack if a
4157 // register cannot be found for it.
4160 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4162 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4164 // On PPC64, promote integers to 64-bit values.
4165 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4166 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4167 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4168 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4171 // FIXME memcpy is used way more than necessary. Correctness first.
4172 // Note: "by value" is code for passing a structure by value, not
4174 if (Flags.isByVal()) {
4175 unsigned Size = Flags.getByValSize();
4176 // Very small objects are passed right-justified. Everything else is
4177 // passed left-justified.
4178 if (Size==1 || Size==2) {
4179 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4180 if (GPR_idx != NumGPRs) {
4181 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4182 MachinePointerInfo(), VT,
4184 MemOpChains.push_back(Load.getValue(1));
4185 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4187 ArgOffset += PtrByteSize;
4189 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4190 PtrOff.getValueType());
4191 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4192 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4195 ArgOffset += PtrByteSize;
4199 // Copy entire object into memory. There are cases where gcc-generated
4200 // code assumes it is there, even if it could be put entirely into
4201 // registers. (This is not what the doc says.)
4202 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4206 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4207 // copy the pieces of the object that fit into registers from the
4208 // parameter save area.
4209 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4210 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4211 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4212 if (GPR_idx != NumGPRs) {
4213 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4214 MachinePointerInfo(),
4215 false, false, false, 0);
4216 MemOpChains.push_back(Load.getValue(1));
4217 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4218 ArgOffset += PtrByteSize;
4220 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4227 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4228 default: llvm_unreachable("Unexpected ValueType for argument!");
4231 if (GPR_idx != NumGPRs) {
4232 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4234 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4235 isPPC64, isTailCall, false, MemOpChains,
4236 TailCallArguments, dl);
4238 ArgOffset += PtrByteSize;
4242 if (FPR_idx != NumFPRs) {
4243 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4246 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4247 MachinePointerInfo(), false, false, 0);
4248 MemOpChains.push_back(Store);
4250 // Float varargs are always shadowed in available integer registers
4251 if (GPR_idx != NumGPRs) {
4252 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4253 MachinePointerInfo(), false, false,
4255 MemOpChains.push_back(Load.getValue(1));
4256 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4258 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4259 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4260 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4261 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4262 MachinePointerInfo(),
4263 false, false, false, 0);
4264 MemOpChains.push_back(Load.getValue(1));
4265 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4268 // If we have any FPRs remaining, we may also have GPRs remaining.
4269 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4271 if (GPR_idx != NumGPRs)
4273 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4274 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4278 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4279 isPPC64, isTailCall, false, MemOpChains,
4280 TailCallArguments, dl);
4284 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4291 // These go aligned on the stack, or in the corresponding R registers
4292 // when within range. The Darwin PPC ABI doc claims they also go in
4293 // V registers; in fact gcc does this only for arguments that are
4294 // prototyped, not for those that match the ... We do it for all
4295 // arguments, seems to work.
4296 while (ArgOffset % 16 !=0) {
4297 ArgOffset += PtrByteSize;
4298 if (GPR_idx != NumGPRs)
4301 // We could elide this store in the case where the object fits
4302 // entirely in R registers. Maybe later.
4303 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4304 DAG.getConstant(ArgOffset, PtrVT));
4305 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4306 MachinePointerInfo(), false, false, 0);
4307 MemOpChains.push_back(Store);
4308 if (VR_idx != NumVRs) {
4309 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4310 MachinePointerInfo(),
4311 false, false, false, 0);
4312 MemOpChains.push_back(Load.getValue(1));
4313 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4316 for (unsigned i=0; i<16; i+=PtrByteSize) {
4317 if (GPR_idx == NumGPRs)
4319 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4320 DAG.getConstant(i, PtrVT));
4321 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4322 false, false, false, 0);
4323 MemOpChains.push_back(Load.getValue(1));
4324 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4329 // Non-varargs Altivec params generally go in registers, but have
4330 // stack space allocated at the end.
4331 if (VR_idx != NumVRs) {
4332 // Doesn't have GPR space allocated.
4333 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4334 } else if (nAltivecParamsAtEnd==0) {
4335 // We are emitting Altivec params in order.
4336 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4337 isPPC64, isTailCall, true, MemOpChains,
4338 TailCallArguments, dl);
4344 // If all Altivec parameters fit in registers, as they usually do,
4345 // they get stack space following the non-Altivec parameters. We
4346 // don't track this here because nobody below needs it.
4347 // If there are more Altivec parameters than fit in registers emit
4349 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4351 // Offset is aligned; skip 1st 12 params which go in V registers.
4352 ArgOffset = ((ArgOffset+15)/16)*16;
4354 for (unsigned i = 0; i != NumOps; ++i) {
4355 SDValue Arg = OutVals[i];
4356 EVT ArgType = Outs[i].VT;
4357 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4358 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4361 // We are emitting Altivec params in order.
4362 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4363 isPPC64, isTailCall, true, MemOpChains,
4364 TailCallArguments, dl);
4371 if (!MemOpChains.empty())
4372 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4373 &MemOpChains[0], MemOpChains.size());
4375 // On Darwin, R12 must contain the address of an indirect callee. This does
4376 // not mean the MTCTR instruction must use R12; it's easier to model this as
4377 // an extra parameter, so do that.
4379 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4380 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4381 !isBLACompatibleAddress(Callee, DAG))
4382 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4383 PPC::R12), Callee));
4385 // Build a sequence of copy-to-reg nodes chained together with token chain
4386 // and flag operands which copy the outgoing args into the appropriate regs.
4388 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4389 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4390 RegsToPass[i].second, InFlag);
4391 InFlag = Chain.getValue(1);
4395 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4396 FPOp, true, TailCallArguments);
4398 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4399 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4404 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4405 MachineFunction &MF, bool isVarArg,
4406 const SmallVectorImpl<ISD::OutputArg> &Outs,
4407 LLVMContext &Context) const {
4408 SmallVector<CCValAssign, 16> RVLocs;
4409 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4411 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4415 PPCTargetLowering::LowerReturn(SDValue Chain,
4416 CallingConv::ID CallConv, bool isVarArg,
4417 const SmallVectorImpl<ISD::OutputArg> &Outs,
4418 const SmallVectorImpl<SDValue> &OutVals,
4419 SDLoc dl, SelectionDAG &DAG) const {
4421 SmallVector<CCValAssign, 16> RVLocs;
4422 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4423 getTargetMachine(), RVLocs, *DAG.getContext());
4424 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4427 SmallVector<SDValue, 4> RetOps(1, Chain);
4429 // Copy the result values into the output registers.
4430 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4431 CCValAssign &VA = RVLocs[i];
4432 assert(VA.isRegLoc() && "Can only return in registers!");
4434 SDValue Arg = OutVals[i];
4436 switch (VA.getLocInfo()) {
4437 default: llvm_unreachable("Unknown loc info!");
4438 case CCValAssign::Full: break;
4439 case CCValAssign::AExt:
4440 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4442 case CCValAssign::ZExt:
4443 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4445 case CCValAssign::SExt:
4446 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4450 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4451 Flag = Chain.getValue(1);
4452 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
4455 RetOps[0] = Chain; // Update chain.
4457 // Add the flag if we have it.
4459 RetOps.push_back(Flag);
4461 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other,
4462 &RetOps[0], RetOps.size());
4465 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4466 const PPCSubtarget &Subtarget) const {
4467 // When we pop the dynamic allocation we need to restore the SP link.
4470 // Get the corect type for pointers.
4471 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4473 // Construct the stack pointer operand.
4474 bool isPPC64 = Subtarget.isPPC64();
4475 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4476 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4478 // Get the operands for the STACKRESTORE.
4479 SDValue Chain = Op.getOperand(0);
4480 SDValue SaveSP = Op.getOperand(1);
4482 // Load the old link SP.
4483 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4484 MachinePointerInfo(),
4485 false, false, false, 0);
4487 // Restore the stack pointer.
4488 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4490 // Store the old link SP.
4491 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4498 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4499 MachineFunction &MF = DAG.getMachineFunction();
4500 bool isPPC64 = PPCSubTarget.isPPC64();
4501 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4502 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4504 // Get current frame pointer save index. The users of this index will be
4505 // primarily DYNALLOC instructions.
4506 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4507 int RASI = FI->getReturnAddrSaveIndex();
4509 // If the frame pointer save index hasn't been defined yet.
4511 // Find out what the fix offset of the frame pointer save area.
4512 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4513 // Allocate the frame index for frame pointer save area.
4514 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4516 FI->setReturnAddrSaveIndex(RASI);
4518 return DAG.getFrameIndex(RASI, PtrVT);
4522 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4523 MachineFunction &MF = DAG.getMachineFunction();
4524 bool isPPC64 = PPCSubTarget.isPPC64();
4525 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4526 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4528 // Get current frame pointer save index. The users of this index will be
4529 // primarily DYNALLOC instructions.
4530 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4531 int FPSI = FI->getFramePointerSaveIndex();
4533 // If the frame pointer save index hasn't been defined yet.
4535 // Find out what the fix offset of the frame pointer save area.
4536 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4539 // Allocate the frame index for frame pointer save area.
4540 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4542 FI->setFramePointerSaveIndex(FPSI);
4544 return DAG.getFrameIndex(FPSI, PtrVT);
4547 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4549 const PPCSubtarget &Subtarget) const {
4551 SDValue Chain = Op.getOperand(0);
4552 SDValue Size = Op.getOperand(1);
4555 // Get the corect type for pointers.
4556 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4558 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4559 DAG.getConstant(0, PtrVT), Size);
4560 // Construct a node for the frame pointer save index.
4561 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4562 // Build a DYNALLOC node.
4563 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4564 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4565 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4568 SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
4569 SelectionDAG &DAG) const {
4571 return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL,
4572 DAG.getVTList(MVT::i32, MVT::Other),
4573 Op.getOperand(0), Op.getOperand(1));
4576 SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
4577 SelectionDAG &DAG) const {
4579 return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
4580 Op.getOperand(0), Op.getOperand(1));
4583 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4585 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4586 // Not FP? Not a fsel.
4587 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4588 !Op.getOperand(2).getValueType().isFloatingPoint())
4591 // We might be able to do better than this under some circumstances, but in
4592 // general, fsel-based lowering of select is a finite-math-only optimization.
4593 // For more information, see section F.3 of the 2.06 ISA specification.
4594 if (!DAG.getTarget().Options.NoInfsFPMath ||
4595 !DAG.getTarget().Options.NoNaNsFPMath)
4598 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4600 EVT ResVT = Op.getValueType();
4601 EVT CmpVT = Op.getOperand(0).getValueType();
4602 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4603 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4606 // If the RHS of the comparison is a 0.0, we don't need to do the
4607 // subtraction at all.
4609 if (isFloatingPointZero(RHS))
4611 default: break; // SETUO etc aren't handled by fsel.
4615 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4616 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4617 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4618 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
4619 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
4620 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4621 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV);
4624 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4627 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4628 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4629 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4632 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4635 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4636 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4637 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4638 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4643 default: break; // SETUO etc aren't handled by fsel.
4647 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4648 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4649 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4650 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4651 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
4652 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
4653 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4654 DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV);
4657 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4658 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4659 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4660 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4663 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4664 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4665 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4666 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4669 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4670 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4671 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4672 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4675 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4676 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4677 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4678 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4683 // FIXME: Split this code up when LegalizeDAGTypes lands.
4684 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4686 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4687 SDValue Src = Op.getOperand(0);
4688 if (Src.getValueType() == MVT::f32)
4689 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4692 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4693 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4695 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4696 (PPCSubTarget.hasFPCVT() ? PPCISD::FCTIWUZ :
4701 assert((Op.getOpcode() == ISD::FP_TO_SINT || PPCSubTarget.hasFPCVT()) &&
4702 "i64 FP_TO_UINT is supported only with FPCVT");
4703 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
4709 // Convert the FP value to an int value through memory.
4710 bool i32Stack = Op.getValueType() == MVT::i32 && PPCSubTarget.hasSTFIWX() &&
4711 (Op.getOpcode() == ISD::FP_TO_SINT || PPCSubTarget.hasFPCVT());
4712 SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64);
4713 int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex();
4714 MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(FI);
4716 // Emit a store to the stack slot.
4719 MachineFunction &MF = DAG.getMachineFunction();
4720 MachineMemOperand *MMO =
4721 MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, 4);
4722 SDValue Ops[] = { DAG.getEntryNode(), Tmp, FIPtr };
4723 Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
4724 DAG.getVTList(MVT::Other), Ops, array_lengthof(Ops),
4727 Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4728 MPI, false, false, 0);
4730 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4732 if (Op.getValueType() == MVT::i32 && !i32Stack) {
4733 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4734 DAG.getConstant(4, FIPtr.getValueType()));
4735 MPI = MachinePointerInfo();
4738 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MPI,
4739 false, false, false, 0);
4742 SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
4743 SelectionDAG &DAG) const {
4745 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4746 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4749 assert((Op.getOpcode() == ISD::SINT_TO_FP || PPCSubTarget.hasFPCVT()) &&
4750 "UINT_TO_FP is supported only with FPCVT");
4752 // If we have FCFIDS, then use it when converting to single-precision.
4753 // Otherwise, convert to double-precision and then round.
4754 unsigned FCFOp = (PPCSubTarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
4755 (Op.getOpcode() == ISD::UINT_TO_FP ?
4756 PPCISD::FCFIDUS : PPCISD::FCFIDS) :
4757 (Op.getOpcode() == ISD::UINT_TO_FP ?
4758 PPCISD::FCFIDU : PPCISD::FCFID);
4759 MVT FCFTy = (PPCSubTarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
4760 MVT::f32 : MVT::f64;
4762 if (Op.getOperand(0).getValueType() == MVT::i64) {
4763 SDValue SINT = Op.getOperand(0);
4764 // When converting to single-precision, we actually need to convert
4765 // to double-precision first and then round to single-precision.
4766 // To avoid double-rounding effects during that operation, we have
4767 // to prepare the input operand. Bits that might be truncated when
4768 // converting to double-precision are replaced by a bit that won't
4769 // be lost at this stage, but is below the single-precision rounding
4772 // However, if -enable-unsafe-fp-math is in effect, accept double
4773 // rounding to avoid the extra overhead.
4774 if (Op.getValueType() == MVT::f32 &&
4775 !PPCSubTarget.hasFPCVT() &&
4776 !DAG.getTarget().Options.UnsafeFPMath) {
4778 // Twiddle input to make sure the low 11 bits are zero. (If this
4779 // is the case, we are guaranteed the value will fit into the 53 bit
4780 // mantissa of an IEEE double-precision value without rounding.)
4781 // If any of those low 11 bits were not zero originally, make sure
4782 // bit 12 (value 2048) is set instead, so that the final rounding
4783 // to single-precision gets the correct result.
4784 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4785 SINT, DAG.getConstant(2047, MVT::i64));
4786 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4787 Round, DAG.getConstant(2047, MVT::i64));
4788 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4789 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4790 Round, DAG.getConstant(-2048, MVT::i64));
4792 // However, we cannot use that value unconditionally: if the magnitude
4793 // of the input value is small, the bit-twiddling we did above might
4794 // end up visibly changing the output. Fortunately, in that case, we
4795 // don't need to twiddle bits since the original input will convert
4796 // exactly to double-precision floating-point already. Therefore,
4797 // construct a conditional to use the original value if the top 11
4798 // bits are all sign-bit copies, and use the rounded value computed
4800 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4801 SINT, DAG.getConstant(53, MVT::i32));
4802 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4803 Cond, DAG.getConstant(1, MVT::i64));
4804 Cond = DAG.getSetCC(dl, MVT::i32,
4805 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4807 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4810 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4811 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Bits);
4813 if (Op.getValueType() == MVT::f32 && !PPCSubTarget.hasFPCVT())
4814 FP = DAG.getNode(ISD::FP_ROUND, dl,
4815 MVT::f32, FP, DAG.getIntPtrConstant(0));
4819 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4820 "Unhandled INT_TO_FP type in custom expander!");
4821 // Since we only generate this in 64-bit mode, we can take advantage of
4822 // 64-bit registers. In particular, sign extend the input value into the
4823 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4824 // then lfd it and fcfid it.
4825 MachineFunction &MF = DAG.getMachineFunction();
4826 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4827 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4830 if (PPCSubTarget.hasLFIWAX() || PPCSubTarget.hasFPCVT()) {
4831 int FrameIdx = FrameInfo->CreateStackObject(4, 4, false);
4832 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4834 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx,
4835 MachinePointerInfo::getFixedStack(FrameIdx),
4838 assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
4839 "Expected an i32 store");
4840 MachineMemOperand *MMO =
4841 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4842 MachineMemOperand::MOLoad, 4, 4);
4843 SDValue Ops[] = { Store, FIdx };
4844 Ld = DAG.getMemIntrinsicNode(Op.getOpcode() == ISD::UINT_TO_FP ?
4845 PPCISD::LFIWZX : PPCISD::LFIWAX,
4846 dl, DAG.getVTList(MVT::f64, MVT::Other),
4847 Ops, 2, MVT::i32, MMO);
4849 assert(PPCSubTarget.isPPC64() &&
4850 "i32->FP without LFIWAX supported only on PPC64");
4852 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4853 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4855 SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64,
4858 // STD the extended value into the stack slot.
4859 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Ext64, FIdx,
4860 MachinePointerInfo::getFixedStack(FrameIdx),
4863 // Load the value as a double.
4864 Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx,
4865 MachinePointerInfo::getFixedStack(FrameIdx),
4866 false, false, false, 0);
4869 // FCFID it and return it.
4870 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Ld);
4871 if (Op.getValueType() == MVT::f32 && !PPCSubTarget.hasFPCVT())
4872 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4876 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4877 SelectionDAG &DAG) const {
4880 The rounding mode is in bits 30:31 of FPSR, and has the following
4887 FLT_ROUNDS, on the other hand, expects the following:
4894 To perform the conversion, we do:
4895 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4898 MachineFunction &MF = DAG.getMachineFunction();
4899 EVT VT = Op.getValueType();
4900 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4901 SDValue MFFSreg, InFlag;
4903 // Save FP Control Word to register
4905 MVT::f64, // return register
4906 MVT::Glue // unused in this context
4908 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4910 // Save FP register to stack slot
4911 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4912 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4913 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4914 StackSlot, MachinePointerInfo(), false, false,0);
4916 // Load FP Control Word from low 32 bits of stack slot.
4917 SDValue Four = DAG.getConstant(4, PtrVT);
4918 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4919 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4920 false, false, false, 0);
4922 // Transform as necessary
4924 DAG.getNode(ISD::AND, dl, MVT::i32,
4925 CWD, DAG.getConstant(3, MVT::i32));
4927 DAG.getNode(ISD::SRL, dl, MVT::i32,
4928 DAG.getNode(ISD::AND, dl, MVT::i32,
4929 DAG.getNode(ISD::XOR, dl, MVT::i32,
4930 CWD, DAG.getConstant(3, MVT::i32)),
4931 DAG.getConstant(3, MVT::i32)),
4932 DAG.getConstant(1, MVT::i32));
4935 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4937 return DAG.getNode((VT.getSizeInBits() < 16 ?
4938 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4941 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4942 EVT VT = Op.getValueType();
4943 unsigned BitWidth = VT.getSizeInBits();
4945 assert(Op.getNumOperands() == 3 &&
4946 VT == Op.getOperand(1).getValueType() &&
4949 // Expand into a bunch of logical ops. Note that these ops
4950 // depend on the PPC behavior for oversized shift amounts.
4951 SDValue Lo = Op.getOperand(0);
4952 SDValue Hi = Op.getOperand(1);
4953 SDValue Amt = Op.getOperand(2);
4954 EVT AmtVT = Amt.getValueType();
4956 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4957 DAG.getConstant(BitWidth, AmtVT), Amt);
4958 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4959 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4960 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4961 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4962 DAG.getConstant(-BitWidth, AmtVT));
4963 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4964 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4965 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4966 SDValue OutOps[] = { OutLo, OutHi };
4967 return DAG.getMergeValues(OutOps, 2, dl);
4970 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4971 EVT VT = Op.getValueType();
4973 unsigned BitWidth = VT.getSizeInBits();
4974 assert(Op.getNumOperands() == 3 &&
4975 VT == Op.getOperand(1).getValueType() &&
4978 // Expand into a bunch of logical ops. Note that these ops
4979 // depend on the PPC behavior for oversized shift amounts.
4980 SDValue Lo = Op.getOperand(0);
4981 SDValue Hi = Op.getOperand(1);
4982 SDValue Amt = Op.getOperand(2);
4983 EVT AmtVT = Amt.getValueType();
4985 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4986 DAG.getConstant(BitWidth, AmtVT), Amt);
4987 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4988 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4989 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4990 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4991 DAG.getConstant(-BitWidth, AmtVT));
4992 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4993 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4994 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4995 SDValue OutOps[] = { OutLo, OutHi };
4996 return DAG.getMergeValues(OutOps, 2, dl);
4999 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
5001 EVT VT = Op.getValueType();
5002 unsigned BitWidth = VT.getSizeInBits();
5003 assert(Op.getNumOperands() == 3 &&
5004 VT == Op.getOperand(1).getValueType() &&
5007 // Expand into a bunch of logical ops, followed by a select_cc.
5008 SDValue Lo = Op.getOperand(0);
5009 SDValue Hi = Op.getOperand(1);
5010 SDValue Amt = Op.getOperand(2);
5011 EVT AmtVT = Amt.getValueType();
5013 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
5014 DAG.getConstant(BitWidth, AmtVT), Amt);
5015 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
5016 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
5017 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
5018 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
5019 DAG.getConstant(-BitWidth, AmtVT));
5020 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
5021 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
5022 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
5023 Tmp4, Tmp6, ISD::SETLE);
5024 SDValue OutOps[] = { OutLo, OutHi };
5025 return DAG.getMergeValues(OutOps, 2, dl);
5028 //===----------------------------------------------------------------------===//
5029 // Vector related lowering.
5032 /// BuildSplatI - Build a canonical splati of Val with an element size of
5033 /// SplatSize. Cast the result to VT.
5034 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
5035 SelectionDAG &DAG, SDLoc dl) {
5036 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
5038 static const EVT VTys[] = { // canonical VT to use for each size.
5039 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
5042 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
5044 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
5048 EVT CanonicalVT = VTys[SplatSize-1];
5050 // Build a canonical splat for this value.
5051 SDValue Elt = DAG.getConstant(Val, MVT::i32);
5052 SmallVector<SDValue, 8> Ops;
5053 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
5054 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
5055 &Ops[0], Ops.size());
5056 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
5059 /// BuildIntrinsicOp - Return a unary operator intrinsic node with the
5060 /// specified intrinsic ID.
5061 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op,
5062 SelectionDAG &DAG, SDLoc dl,
5063 EVT DestVT = MVT::Other) {
5064 if (DestVT == MVT::Other) DestVT = Op.getValueType();
5065 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5066 DAG.getConstant(IID, MVT::i32), Op);
5069 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
5070 /// specified intrinsic ID.
5071 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
5072 SelectionDAG &DAG, SDLoc dl,
5073 EVT DestVT = MVT::Other) {
5074 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
5075 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5076 DAG.getConstant(IID, MVT::i32), LHS, RHS);
5079 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
5080 /// specified intrinsic ID.
5081 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
5082 SDValue Op2, SelectionDAG &DAG,
5083 SDLoc dl, EVT DestVT = MVT::Other) {
5084 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
5085 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5086 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
5090 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
5091 /// amount. The result has the specified value type.
5092 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
5093 EVT VT, SelectionDAG &DAG, SDLoc dl) {
5094 // Force LHS/RHS to be the right type.
5095 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
5096 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
5099 for (unsigned i = 0; i != 16; ++i)
5101 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
5102 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5105 // If this is a case we can't handle, return null and let the default
5106 // expansion code take care of it. If we CAN select this case, and if it
5107 // selects to a single instruction, return Op. Otherwise, if we can codegen
5108 // this case more efficiently than a constant pool load, lower it to the
5109 // sequence of ops that should be used.
5110 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
5111 SelectionDAG &DAG) const {
5113 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5114 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
5116 // Check if this is a splat of a constant value.
5117 APInt APSplatBits, APSplatUndef;
5118 unsigned SplatBitSize;
5120 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
5121 HasAnyUndefs, 0, true) || SplatBitSize > 32)
5124 unsigned SplatBits = APSplatBits.getZExtValue();
5125 unsigned SplatUndef = APSplatUndef.getZExtValue();
5126 unsigned SplatSize = SplatBitSize / 8;
5128 // First, handle single instruction cases.
5131 if (SplatBits == 0) {
5132 // Canonicalize all zero vectors to be v4i32.
5133 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
5134 SDValue Z = DAG.getConstant(0, MVT::i32);
5135 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
5136 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
5141 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
5142 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
5144 if (SextVal >= -16 && SextVal <= 15)
5145 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
5148 // Two instruction sequences.
5150 // If this value is in the range [-32,30] and is even, use:
5151 // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
5152 // If this value is in the range [17,31] and is odd, use:
5153 // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
5154 // If this value is in the range [-31,-17] and is odd, use:
5155 // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
5156 // Note the last two are three-instruction sequences.
5157 if (SextVal >= -32 && SextVal <= 31) {
5158 // To avoid having these optimizations undone by constant folding,
5159 // we convert to a pseudo that will be expanded later into one of
5161 SDValue Elt = DAG.getConstant(SextVal, MVT::i32);
5162 EVT VT = Op.getValueType();
5163 int Size = VT == MVT::v16i8 ? 1 : (VT == MVT::v8i16 ? 2 : 4);
5164 SDValue EltSize = DAG.getConstant(Size, MVT::i32);
5165 return DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
5168 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
5169 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
5171 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
5172 // Make -1 and vspltisw -1:
5173 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
5175 // Make the VSLW intrinsic, computing 0x8000_0000.
5176 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
5179 // xor by OnesV to invert it.
5180 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
5181 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5184 // Check to see if this is a wide variety of vsplti*, binop self cases.
5185 static const signed char SplatCsts[] = {
5186 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
5187 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
5190 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
5191 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
5192 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
5193 int i = SplatCsts[idx];
5195 // Figure out what shift amount will be used by altivec if shifted by i in
5197 unsigned TypeShiftAmt = i & (SplatBitSize-1);
5199 // vsplti + shl self.
5200 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
5201 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5202 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5203 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
5204 Intrinsic::ppc_altivec_vslw
5206 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5207 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5210 // vsplti + srl self.
5211 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5212 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5213 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5214 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
5215 Intrinsic::ppc_altivec_vsrw
5217 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5218 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5221 // vsplti + sra self.
5222 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5223 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5224 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5225 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5226 Intrinsic::ppc_altivec_vsraw
5228 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5229 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5232 // vsplti + rol self.
5233 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5234 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5235 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5236 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5237 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5238 Intrinsic::ppc_altivec_vrlw
5240 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5241 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5244 // t = vsplti c, result = vsldoi t, t, 1
5245 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5246 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5247 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5249 // t = vsplti c, result = vsldoi t, t, 2
5250 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5251 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5252 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5254 // t = vsplti c, result = vsldoi t, t, 3
5255 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5256 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5257 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5264 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5265 /// the specified operations to build the shuffle.
5266 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5267 SDValue RHS, SelectionDAG &DAG,
5269 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5270 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5271 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5274 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5286 if (OpNum == OP_COPY) {
5287 if (LHSID == (1*9+2)*9+3) return LHS;
5288 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5292 SDValue OpLHS, OpRHS;
5293 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5294 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5298 default: llvm_unreachable("Unknown i32 permute!");
5300 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5301 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5302 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5303 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5306 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5307 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5308 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5309 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5312 for (unsigned i = 0; i != 16; ++i)
5313 ShufIdxs[i] = (i&3)+0;
5316 for (unsigned i = 0; i != 16; ++i)
5317 ShufIdxs[i] = (i&3)+4;
5320 for (unsigned i = 0; i != 16; ++i)
5321 ShufIdxs[i] = (i&3)+8;
5324 for (unsigned i = 0; i != 16; ++i)
5325 ShufIdxs[i] = (i&3)+12;
5328 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5330 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5332 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5334 EVT VT = OpLHS.getValueType();
5335 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5336 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5337 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5338 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5341 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5342 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5343 /// return the code it can be lowered into. Worst case, it can always be
5344 /// lowered into a vperm.
5345 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5346 SelectionDAG &DAG) const {
5348 SDValue V1 = Op.getOperand(0);
5349 SDValue V2 = Op.getOperand(1);
5350 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5351 EVT VT = Op.getValueType();
5353 // Cases that are handled by instructions that take permute immediates
5354 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5355 // selected by the instruction selector.
5356 if (V2.getOpcode() == ISD::UNDEF) {
5357 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5358 PPC::isSplatShuffleMask(SVOp, 2) ||
5359 PPC::isSplatShuffleMask(SVOp, 4) ||
5360 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5361 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5362 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5363 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5364 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5365 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5366 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5367 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5368 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5373 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5374 // and produce a fixed permutation. If any of these match, do not lower to
5376 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5377 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5378 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5379 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5380 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5381 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5382 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5383 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5384 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5387 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5388 // perfect shuffle table to emit an optimal matching sequence.
5389 ArrayRef<int> PermMask = SVOp->getMask();
5391 unsigned PFIndexes[4];
5392 bool isFourElementShuffle = true;
5393 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5394 unsigned EltNo = 8; // Start out undef.
5395 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5396 if (PermMask[i*4+j] < 0)
5397 continue; // Undef, ignore it.
5399 unsigned ByteSource = PermMask[i*4+j];
5400 if ((ByteSource & 3) != j) {
5401 isFourElementShuffle = false;
5406 EltNo = ByteSource/4;
5407 } else if (EltNo != ByteSource/4) {
5408 isFourElementShuffle = false;
5412 PFIndexes[i] = EltNo;
5415 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5416 // perfect shuffle vector to determine if it is cost effective to do this as
5417 // discrete instructions, or whether we should use a vperm.
5418 if (isFourElementShuffle) {
5419 // Compute the index in the perfect shuffle table.
5420 unsigned PFTableIndex =
5421 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5423 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5424 unsigned Cost = (PFEntry >> 30);
5426 // Determining when to avoid vperm is tricky. Many things affect the cost
5427 // of vperm, particularly how many times the perm mask needs to be computed.
5428 // For example, if the perm mask can be hoisted out of a loop or is already
5429 // used (perhaps because there are multiple permutes with the same shuffle
5430 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5431 // the loop requires an extra register.
5433 // As a compromise, we only emit discrete instructions if the shuffle can be
5434 // generated in 3 or fewer operations. When we have loop information
5435 // available, if this block is within a loop, we should avoid using vperm
5436 // for 3-operation perms and use a constant pool load instead.
5438 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5441 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5442 // vector that will get spilled to the constant pool.
5443 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5445 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5446 // that it is in input element units, not in bytes. Convert now.
5447 EVT EltVT = V1.getValueType().getVectorElementType();
5448 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5450 SmallVector<SDValue, 16> ResultMask;
5451 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5452 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5454 for (unsigned j = 0; j != BytesPerElement; ++j)
5455 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5459 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5460 &ResultMask[0], ResultMask.size());
5461 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5464 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5465 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5466 /// information about the intrinsic.
5467 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5469 unsigned IntrinsicID =
5470 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5473 switch (IntrinsicID) {
5474 default: return false;
5475 // Comparison predicates.
5476 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5477 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5478 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5479 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5480 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5481 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5482 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5483 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5484 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5485 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5486 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5487 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5488 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5490 // Normal Comparisons.
5491 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5492 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5493 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5494 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5495 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5496 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5497 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5498 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5499 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5500 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5501 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5502 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5503 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5508 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5509 /// lower, do it, otherwise return null.
5510 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5511 SelectionDAG &DAG) const {
5512 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5513 // opcode number of the comparison.
5517 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5518 return SDValue(); // Don't custom lower most intrinsics.
5520 // If this is a non-dot comparison, make the VCMP node and we are done.
5522 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5523 Op.getOperand(1), Op.getOperand(2),
5524 DAG.getConstant(CompareOpc, MVT::i32));
5525 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5528 // Create the PPCISD altivec 'dot' comparison node.
5530 Op.getOperand(2), // LHS
5531 Op.getOperand(3), // RHS
5532 DAG.getConstant(CompareOpc, MVT::i32)
5534 EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
5535 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5537 // Now that we have the comparison, emit a copy from the CR to a GPR.
5538 // This is flagged to the above dot comparison.
5539 SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32,
5540 DAG.getRegister(PPC::CR6, MVT::i32),
5541 CompNode.getValue(1));
5543 // Unpack the result based on how the target uses it.
5544 unsigned BitNo; // Bit # of CR6.
5545 bool InvertBit; // Invert result?
5546 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5547 default: // Can't happen, don't crash on invalid number though.
5548 case 0: // Return the value of the EQ bit of CR6.
5549 BitNo = 0; InvertBit = false;
5551 case 1: // Return the inverted value of the EQ bit of CR6.
5552 BitNo = 0; InvertBit = true;
5554 case 2: // Return the value of the LT bit of CR6.
5555 BitNo = 2; InvertBit = false;
5557 case 3: // Return the inverted value of the LT bit of CR6.
5558 BitNo = 2; InvertBit = true;
5562 // Shift the bit into the low position.
5563 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5564 DAG.getConstant(8-(3-BitNo), MVT::i32));
5566 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5567 DAG.getConstant(1, MVT::i32));
5569 // If we are supposed to, toggle the bit.
5571 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5572 DAG.getConstant(1, MVT::i32));
5576 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5577 SelectionDAG &DAG) const {
5579 // Create a stack slot that is 16-byte aligned.
5580 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5581 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5582 EVT PtrVT = getPointerTy();
5583 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5585 // Store the input value into Value#0 of the stack slot.
5586 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5587 Op.getOperand(0), FIdx, MachinePointerInfo(),
5590 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5591 false, false, false, 0);
5594 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5596 if (Op.getValueType() == MVT::v4i32) {
5597 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5599 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5600 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5602 SDValue RHSSwap = // = vrlw RHS, 16
5603 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5605 // Shrinkify inputs to v8i16.
5606 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5607 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5608 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5610 // Low parts multiplied together, generating 32-bit results (we ignore the
5612 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5613 LHS, RHS, DAG, dl, MVT::v4i32);
5615 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5616 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5617 // Shift the high parts up 16 bits.
5618 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5620 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5621 } else if (Op.getValueType() == MVT::v8i16) {
5622 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5624 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5626 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5627 LHS, RHS, Zero, DAG, dl);
5628 } else if (Op.getValueType() == MVT::v16i8) {
5629 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5631 // Multiply the even 8-bit parts, producing 16-bit sums.
5632 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5633 LHS, RHS, DAG, dl, MVT::v8i16);
5634 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5636 // Multiply the odd 8-bit parts, producing 16-bit sums.
5637 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5638 LHS, RHS, DAG, dl, MVT::v8i16);
5639 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5641 // Merge the results together.
5643 for (unsigned i = 0; i != 8; ++i) {
5645 Ops[i*2+1] = 2*i+1+16;
5647 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5649 llvm_unreachable("Unknown mul to lower!");
5653 /// LowerOperation - Provide custom lowering hooks for some operations.
5655 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5656 switch (Op.getOpcode()) {
5657 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5658 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5659 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5660 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5661 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5662 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5663 case ISD::SETCC: return LowerSETCC(Op, DAG);
5664 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5665 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5667 return LowerVASTART(Op, DAG, PPCSubTarget);
5670 return LowerVAARG(Op, DAG, PPCSubTarget);
5672 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5673 case ISD::DYNAMIC_STACKALLOC:
5674 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5676 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
5677 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
5679 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5680 case ISD::FP_TO_UINT:
5681 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5683 case ISD::UINT_TO_FP:
5684 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5685 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5687 // Lower 64-bit shifts.
5688 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5689 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5690 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5692 // Vector-related lowering.
5693 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5694 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5695 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5696 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5697 case ISD::MUL: return LowerMUL(Op, DAG);
5699 // For counter-based loop handling.
5700 case ISD::INTRINSIC_W_CHAIN: return SDValue();
5702 // Frame & Return address.
5703 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5704 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5708 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5709 SmallVectorImpl<SDValue>&Results,
5710 SelectionDAG &DAG) const {
5711 const TargetMachine &TM = getTargetMachine();
5713 switch (N->getOpcode()) {
5715 llvm_unreachable("Do not know how to custom type legalize this operation!");
5716 case ISD::INTRINSIC_W_CHAIN: {
5717 if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() !=
5718 Intrinsic::ppc_is_decremented_ctr_nonzero)
5721 assert(N->getValueType(0) == MVT::i1 &&
5722 "Unexpected result type for CTR decrement intrinsic");
5723 EVT SVT = getSetCCResultType(*DAG.getContext(), N->getValueType(0));
5724 SDVTList VTs = DAG.getVTList(SVT, MVT::Other);
5725 SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0),
5728 Results.push_back(NewInt);
5729 Results.push_back(NewInt.getValue(1));
5733 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5734 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5737 EVT VT = N->getValueType(0);
5739 if (VT == MVT::i64) {
5740 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5742 Results.push_back(NewNode);
5743 Results.push_back(NewNode.getValue(1));
5747 case ISD::FP_ROUND_INREG: {
5748 assert(N->getValueType(0) == MVT::ppcf128);
5749 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5750 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5751 MVT::f64, N->getOperand(0),
5752 DAG.getIntPtrConstant(0));
5753 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5754 MVT::f64, N->getOperand(0),
5755 DAG.getIntPtrConstant(1));
5757 // Add the two halves of the long double in round-to-zero mode.
5758 SDValue FPreg = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi);
5760 // We know the low half is about to be thrown away, so just use something
5762 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5766 case ISD::FP_TO_SINT:
5767 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5773 //===----------------------------------------------------------------------===//
5774 // Other Lowering Code
5775 //===----------------------------------------------------------------------===//
5778 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5779 bool is64bit, unsigned BinOpcode) const {
5780 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5781 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5783 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5784 MachineFunction *F = BB->getParent();
5785 MachineFunction::iterator It = BB;
5788 unsigned dest = MI->getOperand(0).getReg();
5789 unsigned ptrA = MI->getOperand(1).getReg();
5790 unsigned ptrB = MI->getOperand(2).getReg();
5791 unsigned incr = MI->getOperand(3).getReg();
5792 DebugLoc dl = MI->getDebugLoc();
5794 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5795 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5796 F->insert(It, loopMBB);
5797 F->insert(It, exitMBB);
5798 exitMBB->splice(exitMBB->begin(), BB,
5799 llvm::next(MachineBasicBlock::iterator(MI)),
5801 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5803 MachineRegisterInfo &RegInfo = F->getRegInfo();
5804 unsigned TmpReg = (!BinOpcode) ? incr :
5805 RegInfo.createVirtualRegister(
5806 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5807 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5811 // fallthrough --> loopMBB
5812 BB->addSuccessor(loopMBB);
5815 // l[wd]arx dest, ptr
5816 // add r0, dest, incr
5817 // st[wd]cx. r0, ptr
5819 // fallthrough --> exitMBB
5821 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5822 .addReg(ptrA).addReg(ptrB);
5824 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5825 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5826 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5827 BuildMI(BB, dl, TII->get(PPC::BCC))
5828 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5829 BB->addSuccessor(loopMBB);
5830 BB->addSuccessor(exitMBB);
5839 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5840 MachineBasicBlock *BB,
5841 bool is8bit, // operation
5842 unsigned BinOpcode) const {
5843 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5844 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5845 // In 64 bit mode we have to use 64 bits for addresses, even though the
5846 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5847 // registers without caring whether they're 32 or 64, but here we're
5848 // doing actual arithmetic on the addresses.
5849 bool is64bit = PPCSubTarget.isPPC64();
5850 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
5852 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5853 MachineFunction *F = BB->getParent();
5854 MachineFunction::iterator It = BB;
5857 unsigned dest = MI->getOperand(0).getReg();
5858 unsigned ptrA = MI->getOperand(1).getReg();
5859 unsigned ptrB = MI->getOperand(2).getReg();
5860 unsigned incr = MI->getOperand(3).getReg();
5861 DebugLoc dl = MI->getDebugLoc();
5863 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5864 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5865 F->insert(It, loopMBB);
5866 F->insert(It, exitMBB);
5867 exitMBB->splice(exitMBB->begin(), BB,
5868 llvm::next(MachineBasicBlock::iterator(MI)),
5870 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5872 MachineRegisterInfo &RegInfo = F->getRegInfo();
5873 const TargetRegisterClass *RC =
5874 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5875 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5876 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5877 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5878 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5879 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5880 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5881 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5882 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5883 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5884 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5885 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5886 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5888 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5892 // fallthrough --> loopMBB
5893 BB->addSuccessor(loopMBB);
5895 // The 4-byte load must be aligned, while a char or short may be
5896 // anywhere in the word. Hence all this nasty bookkeeping code.
5897 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5898 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5899 // xori shift, shift1, 24 [16]
5900 // rlwinm ptr, ptr1, 0, 0, 29
5901 // slw incr2, incr, shift
5902 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5903 // slw mask, mask2, shift
5905 // lwarx tmpDest, ptr
5906 // add tmp, tmpDest, incr2
5907 // andc tmp2, tmpDest, mask
5908 // and tmp3, tmp, mask
5909 // or tmp4, tmp3, tmp2
5912 // fallthrough --> exitMBB
5913 // srw dest, tmpDest, shift
5914 if (ptrA != ZeroReg) {
5915 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5916 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5917 .addReg(ptrA).addReg(ptrB);
5921 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5922 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5923 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5924 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5926 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5927 .addReg(Ptr1Reg).addImm(0).addImm(61);
5929 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5930 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5931 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5932 .addReg(incr).addReg(ShiftReg);
5934 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5936 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5937 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5939 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5940 .addReg(Mask2Reg).addReg(ShiftReg);
5943 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5944 .addReg(ZeroReg).addReg(PtrReg);
5946 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5947 .addReg(Incr2Reg).addReg(TmpDestReg);
5948 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5949 .addReg(TmpDestReg).addReg(MaskReg);
5950 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5951 .addReg(TmpReg).addReg(MaskReg);
5952 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5953 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5954 BuildMI(BB, dl, TII->get(PPC::STWCX))
5955 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5956 BuildMI(BB, dl, TII->get(PPC::BCC))
5957 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5958 BB->addSuccessor(loopMBB);
5959 BB->addSuccessor(exitMBB);
5964 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5969 llvm::MachineBasicBlock*
5970 PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
5971 MachineBasicBlock *MBB) const {
5972 DebugLoc DL = MI->getDebugLoc();
5973 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5975 MachineFunction *MF = MBB->getParent();
5976 MachineRegisterInfo &MRI = MF->getRegInfo();
5978 const BasicBlock *BB = MBB->getBasicBlock();
5979 MachineFunction::iterator I = MBB;
5983 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
5984 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
5986 unsigned DstReg = MI->getOperand(0).getReg();
5987 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
5988 assert(RC->hasType(MVT::i32) && "Invalid destination!");
5989 unsigned mainDstReg = MRI.createVirtualRegister(RC);
5990 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
5992 MVT PVT = getPointerTy();
5993 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
5994 "Invalid Pointer Size!");
5995 // For v = setjmp(buf), we generate
5998 // SjLjSetup mainMBB
6004 // buf[LabelOffset] = LR
6008 // v = phi(main, restore)
6011 MachineBasicBlock *thisMBB = MBB;
6012 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
6013 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
6014 MF->insert(I, mainMBB);
6015 MF->insert(I, sinkMBB);
6017 MachineInstrBuilder MIB;
6019 // Transfer the remainder of BB and its successor edges to sinkMBB.
6020 sinkMBB->splice(sinkMBB->begin(), MBB,
6021 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
6022 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
6024 // Note that the structure of the jmp_buf used here is not compatible
6025 // with that used by libc, and is not designed to be. Specifically, it
6026 // stores only those 'reserved' registers that LLVM does not otherwise
6027 // understand how to spill. Also, by convention, by the time this
6028 // intrinsic is called, Clang has already stored the frame address in the
6029 // first slot of the buffer and stack address in the third. Following the
6030 // X86 target code, we'll store the jump address in the second slot. We also
6031 // need to save the TOC pointer (R2) to handle jumps between shared
6032 // libraries, and that will be stored in the fourth slot. The thread
6033 // identifier (R13) is not affected.
6036 const int64_t LabelOffset = 1 * PVT.getStoreSize();
6037 const int64_t TOCOffset = 3 * PVT.getStoreSize();
6039 // Prepare IP either in reg.
6040 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
6041 unsigned LabelReg = MRI.createVirtualRegister(PtrRC);
6042 unsigned BufReg = MI->getOperand(1).getReg();
6044 if (PPCSubTarget.isPPC64() && PPCSubTarget.isSVR4ABI()) {
6045 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD))
6050 MIB.setMemRefs(MMOBegin, MMOEnd);
6054 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB);
6055 const PPCRegisterInfo *TRI =
6056 static_cast<const PPCRegisterInfo*>(getTargetMachine().getRegisterInfo());
6057 MIB.addRegMask(TRI->getNoPreservedMask());
6059 BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1);
6061 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup))
6063 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB);
6065 thisMBB->addSuccessor(mainMBB, /* weight */ 0);
6066 thisMBB->addSuccessor(sinkMBB, /* weight */ 1);
6070 MIB = BuildMI(mainMBB, DL,
6071 TII->get(PPCSubTarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg);
6074 if (PPCSubTarget.isPPC64()) {
6075 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD))
6077 .addImm(LabelOffset)
6080 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW))
6082 .addImm(LabelOffset)
6086 MIB.setMemRefs(MMOBegin, MMOEnd);
6088 BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0);
6089 mainMBB->addSuccessor(sinkMBB);
6092 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
6093 TII->get(PPC::PHI), DstReg)
6094 .addReg(mainDstReg).addMBB(mainMBB)
6095 .addReg(restoreDstReg).addMBB(thisMBB);
6097 MI->eraseFromParent();
6102 PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
6103 MachineBasicBlock *MBB) const {
6104 DebugLoc DL = MI->getDebugLoc();
6105 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6107 MachineFunction *MF = MBB->getParent();
6108 MachineRegisterInfo &MRI = MF->getRegInfo();
6111 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
6112 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
6114 MVT PVT = getPointerTy();
6115 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
6116 "Invalid Pointer Size!");
6118 const TargetRegisterClass *RC =
6119 (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
6120 unsigned Tmp = MRI.createVirtualRegister(RC);
6121 // Since FP is only updated here but NOT referenced, it's treated as GPR.
6122 unsigned FP = (PVT == MVT::i64) ? PPC::X31 : PPC::R31;
6123 unsigned SP = (PVT == MVT::i64) ? PPC::X1 : PPC::R1;
6125 MachineInstrBuilder MIB;
6127 const int64_t LabelOffset = 1 * PVT.getStoreSize();
6128 const int64_t SPOffset = 2 * PVT.getStoreSize();
6129 const int64_t TOCOffset = 3 * PVT.getStoreSize();
6131 unsigned BufReg = MI->getOperand(0).getReg();
6133 // Reload FP (the jumped-to function may not have had a
6134 // frame pointer, and if so, then its r31 will be restored
6136 if (PVT == MVT::i64) {
6137 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP)
6141 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP)
6145 MIB.setMemRefs(MMOBegin, MMOEnd);
6148 if (PVT == MVT::i64) {
6149 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp)
6150 .addImm(LabelOffset)
6153 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp)
6154 .addImm(LabelOffset)
6157 MIB.setMemRefs(MMOBegin, MMOEnd);
6160 if (PVT == MVT::i64) {
6161 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP)
6165 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP)
6169 MIB.setMemRefs(MMOBegin, MMOEnd);
6171 // FIXME: When we also support base pointers, that register must also be
6175 if (PVT == MVT::i64 && PPCSubTarget.isSVR4ABI()) {
6176 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2)
6180 MIB.setMemRefs(MMOBegin, MMOEnd);
6184 BuildMI(*MBB, MI, DL,
6185 TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp);
6186 BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR));
6188 MI->eraseFromParent();
6193 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6194 MachineBasicBlock *BB) const {
6195 if (MI->getOpcode() == PPC::EH_SjLj_SetJmp32 ||
6196 MI->getOpcode() == PPC::EH_SjLj_SetJmp64) {
6197 return emitEHSjLjSetJmp(MI, BB);
6198 } else if (MI->getOpcode() == PPC::EH_SjLj_LongJmp32 ||
6199 MI->getOpcode() == PPC::EH_SjLj_LongJmp64) {
6200 return emitEHSjLjLongJmp(MI, BB);
6203 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6205 // To "insert" these instructions we actually have to insert their
6206 // control-flow patterns.
6207 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6208 MachineFunction::iterator It = BB;
6211 MachineFunction *F = BB->getParent();
6213 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
6214 MI->getOpcode() == PPC::SELECT_CC_I8)) {
6215 SmallVector<MachineOperand, 2> Cond;
6216 Cond.push_back(MI->getOperand(4));
6217 Cond.push_back(MI->getOperand(1));
6219 DebugLoc dl = MI->getDebugLoc();
6220 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6221 TII->insertSelect(*BB, MI, dl, MI->getOperand(0).getReg(),
6222 Cond, MI->getOperand(2).getReg(),
6223 MI->getOperand(3).getReg());
6224 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
6225 MI->getOpcode() == PPC::SELECT_CC_I8 ||
6226 MI->getOpcode() == PPC::SELECT_CC_F4 ||
6227 MI->getOpcode() == PPC::SELECT_CC_F8 ||
6228 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
6231 // The incoming instruction knows the destination vreg to set, the
6232 // condition code register to branch on, the true/false values to
6233 // select between, and a branch opcode to use.
6238 // cmpTY ccX, r1, r2
6240 // fallthrough --> copy0MBB
6241 MachineBasicBlock *thisMBB = BB;
6242 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6243 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6244 unsigned SelectPred = MI->getOperand(4).getImm();
6245 DebugLoc dl = MI->getDebugLoc();
6246 F->insert(It, copy0MBB);
6247 F->insert(It, sinkMBB);
6249 // Transfer the remainder of BB and its successor edges to sinkMBB.
6250 sinkMBB->splice(sinkMBB->begin(), BB,
6251 llvm::next(MachineBasicBlock::iterator(MI)),
6253 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6255 // Next, add the true and fallthrough blocks as its successors.
6256 BB->addSuccessor(copy0MBB);
6257 BB->addSuccessor(sinkMBB);
6259 BuildMI(BB, dl, TII->get(PPC::BCC))
6260 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
6263 // %FalseValue = ...
6264 // # fallthrough to sinkMBB
6267 // Update machine-CFG edges
6268 BB->addSuccessor(sinkMBB);
6271 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6274 BuildMI(*BB, BB->begin(), dl,
6275 TII->get(PPC::PHI), MI->getOperand(0).getReg())
6276 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
6277 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6279 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
6280 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
6281 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
6282 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
6283 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
6284 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
6285 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
6286 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
6288 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
6289 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
6290 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
6291 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
6292 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
6293 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
6294 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
6295 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
6297 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
6298 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
6299 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
6300 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
6301 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
6302 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
6303 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
6304 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
6306 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
6307 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
6308 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
6309 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
6310 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
6311 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
6312 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
6313 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
6315 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
6316 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
6317 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
6318 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
6319 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
6320 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
6321 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
6322 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
6324 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
6325 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
6326 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
6327 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
6328 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
6329 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
6330 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
6331 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
6333 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
6334 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
6335 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
6336 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
6337 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
6338 BB = EmitAtomicBinary(MI, BB, false, 0);
6339 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
6340 BB = EmitAtomicBinary(MI, BB, true, 0);
6342 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
6343 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
6344 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
6346 unsigned dest = MI->getOperand(0).getReg();
6347 unsigned ptrA = MI->getOperand(1).getReg();
6348 unsigned ptrB = MI->getOperand(2).getReg();
6349 unsigned oldval = MI->getOperand(3).getReg();
6350 unsigned newval = MI->getOperand(4).getReg();
6351 DebugLoc dl = MI->getDebugLoc();
6353 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6354 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6355 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6356 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6357 F->insert(It, loop1MBB);
6358 F->insert(It, loop2MBB);
6359 F->insert(It, midMBB);
6360 F->insert(It, exitMBB);
6361 exitMBB->splice(exitMBB->begin(), BB,
6362 llvm::next(MachineBasicBlock::iterator(MI)),
6364 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6368 // fallthrough --> loopMBB
6369 BB->addSuccessor(loop1MBB);
6372 // l[wd]arx dest, ptr
6373 // cmp[wd] dest, oldval
6376 // st[wd]cx. newval, ptr
6380 // st[wd]cx. dest, ptr
6383 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
6384 .addReg(ptrA).addReg(ptrB);
6385 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
6386 .addReg(oldval).addReg(dest);
6387 BuildMI(BB, dl, TII->get(PPC::BCC))
6388 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6389 BB->addSuccessor(loop2MBB);
6390 BB->addSuccessor(midMBB);
6393 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6394 .addReg(newval).addReg(ptrA).addReg(ptrB);
6395 BuildMI(BB, dl, TII->get(PPC::BCC))
6396 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6397 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6398 BB->addSuccessor(loop1MBB);
6399 BB->addSuccessor(exitMBB);
6402 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6403 .addReg(dest).addReg(ptrA).addReg(ptrB);
6404 BB->addSuccessor(exitMBB);
6409 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6410 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6411 // We must use 64-bit registers for addresses when targeting 64-bit,
6412 // since we're actually doing arithmetic on them. Other registers
6414 bool is64bit = PPCSubTarget.isPPC64();
6415 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6417 unsigned dest = MI->getOperand(0).getReg();
6418 unsigned ptrA = MI->getOperand(1).getReg();
6419 unsigned ptrB = MI->getOperand(2).getReg();
6420 unsigned oldval = MI->getOperand(3).getReg();
6421 unsigned newval = MI->getOperand(4).getReg();
6422 DebugLoc dl = MI->getDebugLoc();
6424 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6425 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6426 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6427 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6428 F->insert(It, loop1MBB);
6429 F->insert(It, loop2MBB);
6430 F->insert(It, midMBB);
6431 F->insert(It, exitMBB);
6432 exitMBB->splice(exitMBB->begin(), BB,
6433 llvm::next(MachineBasicBlock::iterator(MI)),
6435 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6437 MachineRegisterInfo &RegInfo = F->getRegInfo();
6438 const TargetRegisterClass *RC =
6439 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6440 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6441 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6442 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6443 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6444 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6445 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6446 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6447 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6448 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6449 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6450 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6451 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6452 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6453 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6455 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6456 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
6459 // fallthrough --> loopMBB
6460 BB->addSuccessor(loop1MBB);
6462 // The 4-byte load must be aligned, while a char or short may be
6463 // anywhere in the word. Hence all this nasty bookkeeping code.
6464 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6465 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6466 // xori shift, shift1, 24 [16]
6467 // rlwinm ptr, ptr1, 0, 0, 29
6468 // slw newval2, newval, shift
6469 // slw oldval2, oldval,shift
6470 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6471 // slw mask, mask2, shift
6472 // and newval3, newval2, mask
6473 // and oldval3, oldval2, mask
6475 // lwarx tmpDest, ptr
6476 // and tmp, tmpDest, mask
6477 // cmpw tmp, oldval3
6480 // andc tmp2, tmpDest, mask
6481 // or tmp4, tmp2, newval3
6486 // stwcx. tmpDest, ptr
6488 // srw dest, tmpDest, shift
6489 if (ptrA != ZeroReg) {
6490 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6491 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6492 .addReg(ptrA).addReg(ptrB);
6496 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6497 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6498 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6499 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6501 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6502 .addReg(Ptr1Reg).addImm(0).addImm(61);
6504 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6505 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6506 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6507 .addReg(newval).addReg(ShiftReg);
6508 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6509 .addReg(oldval).addReg(ShiftReg);
6511 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6513 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6514 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6515 .addReg(Mask3Reg).addImm(65535);
6517 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6518 .addReg(Mask2Reg).addReg(ShiftReg);
6519 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6520 .addReg(NewVal2Reg).addReg(MaskReg);
6521 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6522 .addReg(OldVal2Reg).addReg(MaskReg);
6525 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6526 .addReg(ZeroReg).addReg(PtrReg);
6527 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6528 .addReg(TmpDestReg).addReg(MaskReg);
6529 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6530 .addReg(TmpReg).addReg(OldVal3Reg);
6531 BuildMI(BB, dl, TII->get(PPC::BCC))
6532 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6533 BB->addSuccessor(loop2MBB);
6534 BB->addSuccessor(midMBB);
6537 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6538 .addReg(TmpDestReg).addReg(MaskReg);
6539 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6540 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6541 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6542 .addReg(ZeroReg).addReg(PtrReg);
6543 BuildMI(BB, dl, TII->get(PPC::BCC))
6544 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6545 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6546 BB->addSuccessor(loop1MBB);
6547 BB->addSuccessor(exitMBB);
6550 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6551 .addReg(ZeroReg).addReg(PtrReg);
6552 BB->addSuccessor(exitMBB);
6557 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6559 } else if (MI->getOpcode() == PPC::FADDrtz) {
6560 // This pseudo performs an FADD with rounding mode temporarily forced
6561 // to round-to-zero. We emit this via custom inserter since the FPSCR
6562 // is not modeled at the SelectionDAG level.
6563 unsigned Dest = MI->getOperand(0).getReg();
6564 unsigned Src1 = MI->getOperand(1).getReg();
6565 unsigned Src2 = MI->getOperand(2).getReg();
6566 DebugLoc dl = MI->getDebugLoc();
6568 MachineRegisterInfo &RegInfo = F->getRegInfo();
6569 unsigned MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
6571 // Save FPSCR value.
6572 BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg);
6574 // Set rounding mode to round-to-zero.
6575 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1)).addImm(31);
6576 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0)).addImm(30);
6578 // Perform addition.
6579 BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest).addReg(Src1).addReg(Src2);
6581 // Restore FPSCR value.
6582 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF)).addImm(1).addReg(MFFSReg);
6583 } else if (MI->getOpcode() == PPC::FRINDrint ||
6584 MI->getOpcode() == PPC::FRINSrint) {
6585 bool isf32 = MI->getOpcode() == PPC::FRINSrint;
6586 unsigned Dest = MI->getOperand(0).getReg();
6587 unsigned Src = MI->getOperand(1).getReg();
6588 DebugLoc dl = MI->getDebugLoc();
6590 MachineRegisterInfo &RegInfo = F->getRegInfo();
6591 unsigned CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
6593 // Perform the rounding.
6594 BuildMI(*BB, MI, dl, TII->get(isf32 ? PPC::FRINS : PPC::FRIND), Dest)
6597 // Compare the results.
6598 BuildMI(*BB, MI, dl, TII->get(isf32 ? PPC::FCMPUS : PPC::FCMPUD), CRReg)
6599 .addReg(Dest).addReg(Src);
6601 // If the results were not equal, then set the FPSCR XX bit.
6602 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6603 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6604 F->insert(It, midMBB);
6605 F->insert(It, exitMBB);
6606 exitMBB->splice(exitMBB->begin(), BB,
6607 llvm::next(MachineBasicBlock::iterator(MI)),
6609 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6611 BuildMI(*BB, MI, dl, TII->get(PPC::BCC))
6612 .addImm(PPC::PRED_EQ).addReg(CRReg).addMBB(exitMBB);
6614 BB->addSuccessor(midMBB);
6615 BB->addSuccessor(exitMBB);
6619 // Set the FPSCR XX bit (FE_INEXACT). Note that we cannot just set
6620 // the FI bit here because that will not automatically set XX also,
6621 // and XX is what libm interprets as the FE_INEXACT flag.
6622 BuildMI(BB, dl, TII->get(PPC::MTFSB1)).addImm(/* 38 - 32 = */ 6);
6623 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6625 BB->addSuccessor(exitMBB);
6629 llvm_unreachable("Unexpected instr type to insert");
6632 MI->eraseFromParent(); // The pseudo instruction is gone now.
6636 //===----------------------------------------------------------------------===//
6637 // Target Optimization Hooks
6638 //===----------------------------------------------------------------------===//
6640 SDValue PPCTargetLowering::DAGCombineFastRecip(SDValue Op,
6641 DAGCombinerInfo &DCI) const {
6642 if (DCI.isAfterLegalizeVectorOps())
6645 EVT VT = Op.getValueType();
6647 if ((VT == MVT::f32 && PPCSubTarget.hasFRES()) ||
6648 (VT == MVT::f64 && PPCSubTarget.hasFRE()) ||
6649 (VT == MVT::v4f32 && PPCSubTarget.hasAltivec())) {
6651 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
6652 // For the reciprocal, we need to find the zero of the function:
6653 // F(X) = A X - 1 [which has a zero at X = 1/A]
6655 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
6656 // does not require additional intermediate precision]
6658 // Convergence is quadratic, so we essentially double the number of digits
6659 // correct after every iteration. The minimum architected relative
6660 // accuracy is 2^-5. When hasRecipPrec(), this is 2^-14. IEEE float has
6661 // 23 digits and double has 52 digits.
6662 int Iterations = PPCSubTarget.hasRecipPrec() ? 1 : 3;
6663 if (VT.getScalarType() == MVT::f64)
6666 SelectionDAG &DAG = DCI.DAG;
6670 DAG.getConstantFP(1.0, VT.getScalarType());
6671 if (VT.isVector()) {
6672 assert(VT.getVectorNumElements() == 4 &&
6673 "Unknown vector type");
6674 FPOne = DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
6675 FPOne, FPOne, FPOne, FPOne);
6678 SDValue Est = DAG.getNode(PPCISD::FRE, dl, VT, Op);
6679 DCI.AddToWorklist(Est.getNode());
6681 // Newton iterations: Est = Est + Est (1 - Arg * Est)
6682 for (int i = 0; i < Iterations; ++i) {
6683 SDValue NewEst = DAG.getNode(ISD::FMUL, dl, VT, Op, Est);
6684 DCI.AddToWorklist(NewEst.getNode());
6686 NewEst = DAG.getNode(ISD::FSUB, dl, VT, FPOne, NewEst);
6687 DCI.AddToWorklist(NewEst.getNode());
6689 NewEst = DAG.getNode(ISD::FMUL, dl, VT, Est, NewEst);
6690 DCI.AddToWorklist(NewEst.getNode());
6692 Est = DAG.getNode(ISD::FADD, dl, VT, Est, NewEst);
6693 DCI.AddToWorklist(Est.getNode());
6702 SDValue PPCTargetLowering::DAGCombineFastRecipFSQRT(SDValue Op,
6703 DAGCombinerInfo &DCI) const {
6704 if (DCI.isAfterLegalizeVectorOps())
6707 EVT VT = Op.getValueType();
6709 if ((VT == MVT::f32 && PPCSubTarget.hasFRSQRTES()) ||
6710 (VT == MVT::f64 && PPCSubTarget.hasFRSQRTE()) ||
6711 (VT == MVT::v4f32 && PPCSubTarget.hasAltivec())) {
6713 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
6714 // For the reciprocal sqrt, we need to find the zero of the function:
6715 // F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
6717 // X_{i+1} = X_i (1.5 - A X_i^2 / 2)
6718 // As a result, we precompute A/2 prior to the iteration loop.
6720 // Convergence is quadratic, so we essentially double the number of digits
6721 // correct after every iteration. The minimum architected relative
6722 // accuracy is 2^-5. When hasRecipPrec(), this is 2^-14. IEEE float has
6723 // 23 digits and double has 52 digits.
6724 int Iterations = PPCSubTarget.hasRecipPrec() ? 1 : 3;
6725 if (VT.getScalarType() == MVT::f64)
6728 SelectionDAG &DAG = DCI.DAG;
6731 SDValue FPThreeHalves =
6732 DAG.getConstantFP(1.5, VT.getScalarType());
6733 if (VT.isVector()) {
6734 assert(VT.getVectorNumElements() == 4 &&
6735 "Unknown vector type");
6736 FPThreeHalves = DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
6737 FPThreeHalves, FPThreeHalves,
6738 FPThreeHalves, FPThreeHalves);
6741 SDValue Est = DAG.getNode(PPCISD::FRSQRTE, dl, VT, Op);
6742 DCI.AddToWorklist(Est.getNode());
6744 // We now need 0.5*Arg which we can write as (1.5*Arg - Arg) so that
6745 // this entire sequence requires only one FP constant.
6746 SDValue HalfArg = DAG.getNode(ISD::FMUL, dl, VT, FPThreeHalves, Op);
6747 DCI.AddToWorklist(HalfArg.getNode());
6749 HalfArg = DAG.getNode(ISD::FSUB, dl, VT, HalfArg, Op);
6750 DCI.AddToWorklist(HalfArg.getNode());
6752 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
6753 for (int i = 0; i < Iterations; ++i) {
6754 SDValue NewEst = DAG.getNode(ISD::FMUL, dl, VT, Est, Est);
6755 DCI.AddToWorklist(NewEst.getNode());
6757 NewEst = DAG.getNode(ISD::FMUL, dl, VT, HalfArg, NewEst);
6758 DCI.AddToWorklist(NewEst.getNode());
6760 NewEst = DAG.getNode(ISD::FSUB, dl, VT, FPThreeHalves, NewEst);
6761 DCI.AddToWorklist(NewEst.getNode());
6763 Est = DAG.getNode(ISD::FMUL, dl, VT, Est, NewEst);
6764 DCI.AddToWorklist(Est.getNode());
6773 // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
6774 // not enforce equality of the chain operands.
6775 static bool isConsecutiveLS(LSBaseSDNode *LS, LSBaseSDNode *Base,
6776 unsigned Bytes, int Dist,
6777 SelectionDAG &DAG) {
6778 EVT VT = LS->getMemoryVT();
6779 if (VT.getSizeInBits() / 8 != Bytes)
6782 SDValue Loc = LS->getBasePtr();
6783 SDValue BaseLoc = Base->getBasePtr();
6784 if (Loc.getOpcode() == ISD::FrameIndex) {
6785 if (BaseLoc.getOpcode() != ISD::FrameIndex)
6787 const MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6788 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
6789 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6790 int FS = MFI->getObjectSize(FI);
6791 int BFS = MFI->getObjectSize(BFI);
6792 if (FS != BFS || FS != (int)Bytes) return false;
6793 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
6797 if (DAG.isBaseWithConstantOffset(Loc) && Loc.getOperand(0) == BaseLoc &&
6798 cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue() == Dist*Bytes)
6801 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6802 const GlobalValue *GV1 = NULL;
6803 const GlobalValue *GV2 = NULL;
6804 int64_t Offset1 = 0;
6805 int64_t Offset2 = 0;
6806 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
6807 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
6808 if (isGA1 && isGA2 && GV1 == GV2)
6809 return Offset1 == (Offset2 + Dist*Bytes);
6813 // Return true is there is a nearyby consecutive load to the one provided
6814 // (regardless of alignment). We search up and down the chain, looking though
6815 // token factors and other loads (but nothing else). As a result, a true
6816 // results indicates that it is safe to create a new consecutive load adjacent
6817 // to the load provided.
6818 static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
6819 SDValue Chain = LD->getChain();
6820 EVT VT = LD->getMemoryVT();
6822 SmallSet<SDNode *, 16> LoadRoots;
6823 SmallVector<SDNode *, 8> Queue(1, Chain.getNode());
6824 SmallSet<SDNode *, 16> Visited;
6826 // First, search up the chain, branching to follow all token-factor operands.
6827 // If we find a consecutive load, then we're done, otherwise, record all
6828 // nodes just above the top-level loads and token factors.
6829 while (!Queue.empty()) {
6830 SDNode *ChainNext = Queue.pop_back_val();
6831 if (!Visited.insert(ChainNext))
6834 if (LoadSDNode *ChainLD = dyn_cast<LoadSDNode>(ChainNext)) {
6835 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
6838 if (!Visited.count(ChainLD->getChain().getNode()))
6839 Queue.push_back(ChainLD->getChain().getNode());
6840 } else if (ChainNext->getOpcode() == ISD::TokenFactor) {
6841 for (SDNode::op_iterator O = ChainNext->op_begin(),
6842 OE = ChainNext->op_end(); O != OE; ++O)
6843 if (!Visited.count(O->getNode()))
6844 Queue.push_back(O->getNode());
6846 LoadRoots.insert(ChainNext);
6849 // Second, search down the chain, starting from the top-level nodes recorded
6850 // in the first phase. These top-level nodes are the nodes just above all
6851 // loads and token factors. Starting with their uses, recursively look though
6852 // all loads (just the chain uses) and token factors to find a consecutive
6857 for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(),
6858 IE = LoadRoots.end(); I != IE; ++I) {
6859 Queue.push_back(*I);
6861 while (!Queue.empty()) {
6862 SDNode *LoadRoot = Queue.pop_back_val();
6863 if (!Visited.insert(LoadRoot))
6866 if (LoadSDNode *ChainLD = dyn_cast<LoadSDNode>(LoadRoot))
6867 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
6870 for (SDNode::use_iterator UI = LoadRoot->use_begin(),
6871 UE = LoadRoot->use_end(); UI != UE; ++UI)
6872 if (((isa<LoadSDNode>(*UI) &&
6873 cast<LoadSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
6874 UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI))
6875 Queue.push_back(*UI);
6882 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6883 DAGCombinerInfo &DCI) const {
6884 const TargetMachine &TM = getTargetMachine();
6885 SelectionDAG &DAG = DCI.DAG;
6887 switch (N->getOpcode()) {
6890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6891 if (C->isNullValue()) // 0 << V -> 0.
6892 return N->getOperand(0);
6896 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6897 if (C->isNullValue()) // 0 >>u V -> 0.
6898 return N->getOperand(0);
6902 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6903 if (C->isNullValue() || // 0 >>s V -> 0.
6904 C->isAllOnesValue()) // -1 >>s V -> -1.
6905 return N->getOperand(0);
6909 assert(TM.Options.UnsafeFPMath &&
6910 "Reciprocal estimates require UnsafeFPMath");
6912 if (N->getOperand(1).getOpcode() == ISD::FSQRT) {
6914 DAGCombineFastRecipFSQRT(N->getOperand(1).getOperand(0), DCI);
6915 if (RV.getNode() != 0) {
6916 DCI.AddToWorklist(RV.getNode());
6917 return DAG.getNode(ISD::FMUL, dl, N->getValueType(0),
6918 N->getOperand(0), RV);
6920 } else if (N->getOperand(1).getOpcode() == ISD::FP_EXTEND &&
6921 N->getOperand(1).getOperand(0).getOpcode() == ISD::FSQRT) {
6923 DAGCombineFastRecipFSQRT(N->getOperand(1).getOperand(0).getOperand(0),
6925 if (RV.getNode() != 0) {
6926 DCI.AddToWorklist(RV.getNode());
6927 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N->getOperand(1)),
6928 N->getValueType(0), RV);
6929 DCI.AddToWorklist(RV.getNode());
6930 return DAG.getNode(ISD::FMUL, dl, N->getValueType(0),
6931 N->getOperand(0), RV);
6933 } else if (N->getOperand(1).getOpcode() == ISD::FP_ROUND &&
6934 N->getOperand(1).getOperand(0).getOpcode() == ISD::FSQRT) {
6936 DAGCombineFastRecipFSQRT(N->getOperand(1).getOperand(0).getOperand(0),
6938 if (RV.getNode() != 0) {
6939 DCI.AddToWorklist(RV.getNode());
6940 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N->getOperand(1)),
6941 N->getValueType(0), RV,
6942 N->getOperand(1).getOperand(1));
6943 DCI.AddToWorklist(RV.getNode());
6944 return DAG.getNode(ISD::FMUL, dl, N->getValueType(0),
6945 N->getOperand(0), RV);
6949 SDValue RV = DAGCombineFastRecip(N->getOperand(1), DCI);
6950 if (RV.getNode() != 0) {
6951 DCI.AddToWorklist(RV.getNode());
6952 return DAG.getNode(ISD::FMUL, dl, N->getValueType(0),
6953 N->getOperand(0), RV);
6959 assert(TM.Options.UnsafeFPMath &&
6960 "Reciprocal estimates require UnsafeFPMath");
6962 // Compute this as 1/(1/sqrt(X)), which is the reciprocal of the
6964 SDValue RV = DAGCombineFastRecipFSQRT(N->getOperand(0), DCI);
6965 if (RV.getNode() != 0) {
6966 DCI.AddToWorklist(RV.getNode());
6967 RV = DAGCombineFastRecip(RV, DCI);
6968 if (RV.getNode() != 0)
6974 case ISD::SINT_TO_FP:
6975 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6976 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6977 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6978 // We allow the src/dst to be either f32/f64, but the intermediate
6979 // type must be i64.
6980 if (N->getOperand(0).getValueType() == MVT::i64 &&
6981 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6982 SDValue Val = N->getOperand(0).getOperand(0);
6983 if (Val.getValueType() == MVT::f32) {
6984 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6985 DCI.AddToWorklist(Val.getNode());
6988 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6989 DCI.AddToWorklist(Val.getNode());
6990 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6991 DCI.AddToWorklist(Val.getNode());
6992 if (N->getValueType(0) == MVT::f32) {
6993 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6994 DAG.getIntPtrConstant(0));
6995 DCI.AddToWorklist(Val.getNode());
6998 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6999 // If the intermediate type is i32, we can avoid the load/store here
7006 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
7007 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
7008 !cast<StoreSDNode>(N)->isTruncatingStore() &&
7009 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
7010 N->getOperand(1).getValueType() == MVT::i32 &&
7011 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
7012 SDValue Val = N->getOperand(1).getOperand(0);
7013 if (Val.getValueType() == MVT::f32) {
7014 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
7015 DCI.AddToWorklist(Val.getNode());
7017 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
7018 DCI.AddToWorklist(Val.getNode());
7021 N->getOperand(0), Val, N->getOperand(2),
7022 DAG.getValueType(N->getOperand(1).getValueType())
7025 Val = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
7026 DAG.getVTList(MVT::Other), Ops, array_lengthof(Ops),
7027 cast<StoreSDNode>(N)->getMemoryVT(),
7028 cast<StoreSDNode>(N)->getMemOperand());
7029 DCI.AddToWorklist(Val.getNode());
7033 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
7034 if (cast<StoreSDNode>(N)->isUnindexed() &&
7035 N->getOperand(1).getOpcode() == ISD::BSWAP &&
7036 N->getOperand(1).getNode()->hasOneUse() &&
7037 (N->getOperand(1).getValueType() == MVT::i32 ||
7038 N->getOperand(1).getValueType() == MVT::i16 ||
7039 (TM.getSubtarget<PPCSubtarget>().hasLDBRX() &&
7040 TM.getSubtarget<PPCSubtarget>().isPPC64() &&
7041 N->getOperand(1).getValueType() == MVT::i64))) {
7042 SDValue BSwapOp = N->getOperand(1).getOperand(0);
7043 // Do an any-extend to 32-bits if this is a half-word input.
7044 if (BSwapOp.getValueType() == MVT::i16)
7045 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
7048 N->getOperand(0), BSwapOp, N->getOperand(2),
7049 DAG.getValueType(N->getOperand(1).getValueType())
7052 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
7053 Ops, array_lengthof(Ops),
7054 cast<StoreSDNode>(N)->getMemoryVT(),
7055 cast<StoreSDNode>(N)->getMemOperand());
7059 LoadSDNode *LD = cast<LoadSDNode>(N);
7060 EVT VT = LD->getValueType(0);
7061 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
7062 unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
7063 if (ISD::isNON_EXTLoad(N) && VT.isVector() &&
7064 TM.getSubtarget<PPCSubtarget>().hasAltivec() &&
7065 DCI.getDAGCombineLevel() == AfterLegalizeTypes &&
7066 LD->getAlignment() < ABIAlignment) {
7067 // This is a type-legal unaligned Altivec load.
7068 SDValue Chain = LD->getChain();
7069 SDValue Ptr = LD->getBasePtr();
7071 // This implements the loading of unaligned vectors as described in
7072 // the venerable Apple Velocity Engine overview. Specifically:
7073 // https://developer.apple.com/hardwaredrivers/ve/alignment.html
7074 // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html
7076 // The general idea is to expand a sequence of one or more unaligned
7077 // loads into a alignment-based permutation-control instruction (lvsl),
7078 // a series of regular vector loads (which always truncate their
7079 // input address to an aligned address), and a series of permutations.
7080 // The results of these permutations are the requested loaded values.
7081 // The trick is that the last "extra" load is not taken from the address
7082 // you might suspect (sizeof(vector) bytes after the last requested
7083 // load), but rather sizeof(vector) - 1 bytes after the last
7084 // requested vector. The point of this is to avoid a page fault if the
7085 // base address happend to be aligned. This works because if the base
7086 // address is aligned, then adding less than a full vector length will
7087 // cause the last vector in the sequence to be (re)loaded. Otherwise,
7088 // the next vector will be fetched as you might suspect was necessary.
7090 // We might be able to reuse the permutation generation from
7091 // a different base address offset from this one by an aligned amount.
7092 // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this
7093 // optimization later.
7094 SDValue PermCntl = BuildIntrinsicOp(Intrinsic::ppc_altivec_lvsl, Ptr,
7095 DAG, dl, MVT::v16i8);
7097 // Refine the alignment of the original load (a "new" load created here
7098 // which was identical to the first except for the alignment would be
7099 // merged with the existing node regardless).
7100 MachineFunction &MF = DAG.getMachineFunction();
7101 MachineMemOperand *MMO =
7102 MF.getMachineMemOperand(LD->getPointerInfo(),
7103 LD->getMemOperand()->getFlags(),
7104 LD->getMemoryVT().getStoreSize(),
7106 LD->refineAlignment(MMO);
7107 SDValue BaseLoad = SDValue(LD, 0);
7109 // Note that the value of IncOffset (which is provided to the next
7110 // load's pointer info offset value, and thus used to calculate the
7111 // alignment), and the value of IncValue (which is actually used to
7112 // increment the pointer value) are different! This is because we
7113 // require the next load to appear to be aligned, even though it
7114 // is actually offset from the base pointer by a lesser amount.
7115 int IncOffset = VT.getSizeInBits() / 8;
7116 int IncValue = IncOffset;
7118 // Walk (both up and down) the chain looking for another load at the real
7119 // (aligned) offset (the alignment of the other load does not matter in
7120 // this case). If found, then do not use the offset reduction trick, as
7121 // that will prevent the loads from being later combined (as they would
7122 // otherwise be duplicates).
7123 if (!findConsecutiveLoad(LD, DAG))
7126 SDValue Increment = DAG.getConstant(IncValue, getPointerTy());
7127 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
7130 DAG.getLoad(VT, dl, Chain, Ptr,
7131 LD->getPointerInfo().getWithOffset(IncOffset),
7132 LD->isVolatile(), LD->isNonTemporal(),
7133 LD->isInvariant(), ABIAlignment);
7135 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7136 BaseLoad.getValue(1), ExtraLoad.getValue(1));
7138 if (BaseLoad.getValueType() != MVT::v4i32)
7139 BaseLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, BaseLoad);
7141 if (ExtraLoad.getValueType() != MVT::v4i32)
7142 ExtraLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ExtraLoad);
7144 SDValue Perm = BuildIntrinsicOp(Intrinsic::ppc_altivec_vperm,
7145 BaseLoad, ExtraLoad, PermCntl, DAG, dl);
7147 if (VT != MVT::v4i32)
7148 Perm = DAG.getNode(ISD::BITCAST, dl, VT, Perm);
7150 // Now we need to be really careful about how we update the users of the
7151 // original load. We cannot just call DCI.CombineTo (or
7152 // DAG.ReplaceAllUsesWith for that matter), because the load still has
7153 // uses created here (the permutation for example) that need to stay.
7154 SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
7156 SDUse &Use = UI.getUse();
7158 // Note: BaseLoad is checked here because it might not be N, but a
7160 if (User == Perm.getNode() || User == BaseLoad.getNode() ||
7161 User == TF.getNode() || Use.getResNo() > 1) {
7166 SDValue To = Use.getResNo() ? TF : Perm;
7169 SmallVector<SDValue, 8> Ops;
7170 for (SDNode::op_iterator O = User->op_begin(),
7171 OE = User->op_end(); O != OE; ++O) {
7178 DAG.UpdateNodeOperands(User, Ops.data(), Ops.size());
7181 return SDValue(N, 0);
7185 case ISD::INTRINSIC_WO_CHAIN:
7186 if (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() ==
7187 Intrinsic::ppc_altivec_lvsl &&
7188 N->getOperand(1)->getOpcode() == ISD::ADD) {
7189 SDValue Add = N->getOperand(1);
7191 if (DAG.MaskedValueIsZero(Add->getOperand(1),
7192 APInt::getAllOnesValue(4 /* 16 byte alignment */).zext(
7193 Add.getValueType().getScalarType().getSizeInBits()))) {
7194 SDNode *BasePtr = Add->getOperand(0).getNode();
7195 for (SDNode::use_iterator UI = BasePtr->use_begin(),
7196 UE = BasePtr->use_end(); UI != UE; ++UI) {
7197 if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
7198 cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() ==
7199 Intrinsic::ppc_altivec_lvsl) {
7200 // We've found another LVSL, and this address if an aligned
7201 // multiple of that one. The results will be the same, so use the
7202 // one we've just found instead.
7204 return SDValue(*UI, 0);
7210 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
7211 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
7212 N->getOperand(0).hasOneUse() &&
7213 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 ||
7214 (TM.getSubtarget<PPCSubtarget>().hasLDBRX() &&
7215 TM.getSubtarget<PPCSubtarget>().isPPC64() &&
7216 N->getValueType(0) == MVT::i64))) {
7217 SDValue Load = N->getOperand(0);
7218 LoadSDNode *LD = cast<LoadSDNode>(Load);
7219 // Create the byte-swapping load.
7221 LD->getChain(), // Chain
7222 LD->getBasePtr(), // Ptr
7223 DAG.getValueType(N->getValueType(0)) // VT
7226 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
7227 DAG.getVTList(N->getValueType(0) == MVT::i64 ?
7228 MVT::i64 : MVT::i32, MVT::Other),
7229 Ops, 3, LD->getMemoryVT(), LD->getMemOperand());
7231 // If this is an i16 load, insert the truncate.
7232 SDValue ResVal = BSLoad;
7233 if (N->getValueType(0) == MVT::i16)
7234 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
7236 // First, combine the bswap away. This makes the value produced by the
7238 DCI.CombineTo(N, ResVal);
7240 // Next, combine the load away, we give it a bogus result value but a real
7241 // chain result. The result value is dead because the bswap is dead.
7242 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
7244 // Return N so it doesn't get rechecked!
7245 return SDValue(N, 0);
7249 case PPCISD::VCMP: {
7250 // If a VCMPo node already exists with exactly the same operands as this
7251 // node, use its result instead of this node (VCMPo computes both a CR6 and
7252 // a normal output).
7254 if (!N->getOperand(0).hasOneUse() &&
7255 !N->getOperand(1).hasOneUse() &&
7256 !N->getOperand(2).hasOneUse()) {
7258 // Scan all of the users of the LHS, looking for VCMPo's that match.
7259 SDNode *VCMPoNode = 0;
7261 SDNode *LHSN = N->getOperand(0).getNode();
7262 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
7264 if (UI->getOpcode() == PPCISD::VCMPo &&
7265 UI->getOperand(1) == N->getOperand(1) &&
7266 UI->getOperand(2) == N->getOperand(2) &&
7267 UI->getOperand(0) == N->getOperand(0)) {
7272 // If there is no VCMPo node, or if the flag value has a single use, don't
7274 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
7277 // Look at the (necessarily single) use of the flag value. If it has a
7278 // chain, this transformation is more complex. Note that multiple things
7279 // could use the value result, which we should ignore.
7280 SDNode *FlagUser = 0;
7281 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
7282 FlagUser == 0; ++UI) {
7283 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
7285 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
7286 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
7293 // If the user is a MFOCRF instruction, we know this is safe.
7294 // Otherwise we give up for right now.
7295 if (FlagUser->getOpcode() == PPCISD::MFOCRF)
7296 return SDValue(VCMPoNode, 0);
7301 // If this is a branch on an altivec predicate comparison, lower this so
7302 // that we don't have to do a MFOCRF: instead, branch directly on CR6. This
7303 // lowering is done pre-legalize, because the legalizer lowers the predicate
7304 // compare down to code that is difficult to reassemble.
7305 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
7306 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
7308 // Sometimes the promoted value of the intrinsic is ANDed by some non-zero
7309 // value. If so, pass-through the AND to get to the intrinsic.
7310 if (LHS.getOpcode() == ISD::AND &&
7311 LHS.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN &&
7312 cast<ConstantSDNode>(LHS.getOperand(0).getOperand(1))->getZExtValue() ==
7313 Intrinsic::ppc_is_decremented_ctr_nonzero &&
7314 isa<ConstantSDNode>(LHS.getOperand(1)) &&
7315 !cast<ConstantSDNode>(LHS.getOperand(1))->getConstantIntValue()->
7317 LHS = LHS.getOperand(0);
7319 if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
7320 cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() ==
7321 Intrinsic::ppc_is_decremented_ctr_nonzero &&
7322 isa<ConstantSDNode>(RHS)) {
7323 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
7324 "Counter decrement comparison is not EQ or NE");
7326 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
7327 bool isBDNZ = (CC == ISD::SETEQ && Val) ||
7328 (CC == ISD::SETNE && !Val);
7330 // We now need to make the intrinsic dead (it cannot be instruction
7332 DAG.ReplaceAllUsesOfValueWith(LHS.getValue(1), LHS.getOperand(0));
7333 assert(LHS.getNode()->hasOneUse() &&
7334 "Counter decrement has more than one use");
7336 return DAG.getNode(isBDNZ ? PPCISD::BDNZ : PPCISD::BDZ, dl, MVT::Other,
7337 N->getOperand(0), N->getOperand(4));
7343 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
7344 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
7345 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
7346 assert(isDot && "Can't compare against a vector result!");
7348 // If this is a comparison against something other than 0/1, then we know
7349 // that the condition is never/always true.
7350 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
7351 if (Val != 0 && Val != 1) {
7352 if (CC == ISD::SETEQ) // Cond never true, remove branch.
7353 return N->getOperand(0);
7354 // Always !=, turn it into an unconditional branch.
7355 return DAG.getNode(ISD::BR, dl, MVT::Other,
7356 N->getOperand(0), N->getOperand(4));
7359 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
7361 // Create the PPCISD altivec 'dot' comparison node.
7363 LHS.getOperand(2), // LHS of compare
7364 LHS.getOperand(3), // RHS of compare
7365 DAG.getConstant(CompareOpc, MVT::i32)
7367 EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
7368 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
7370 // Unpack the result based on how the target uses it.
7371 PPC::Predicate CompOpc;
7372 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
7373 default: // Can't happen, don't crash on invalid number though.
7374 case 0: // Branch on the value of the EQ bit of CR6.
7375 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
7377 case 1: // Branch on the inverted value of the EQ bit of CR6.
7378 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
7380 case 2: // Branch on the value of the LT bit of CR6.
7381 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
7383 case 3: // Branch on the inverted value of the LT bit of CR6.
7384 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
7388 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
7389 DAG.getConstant(CompOpc, MVT::i32),
7390 DAG.getRegister(PPC::CR6, MVT::i32),
7391 N->getOperand(4), CompNode.getValue(1));
7400 //===----------------------------------------------------------------------===//
7401 // Inline Assembly Support
7402 //===----------------------------------------------------------------------===//
7404 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7407 const SelectionDAG &DAG,
7408 unsigned Depth) const {
7409 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
7410 switch (Op.getOpcode()) {
7412 case PPCISD::LBRX: {
7413 // lhbrx is known to have the top bits cleared out.
7414 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
7415 KnownZero = 0xFFFF0000;
7418 case ISD::INTRINSIC_WO_CHAIN: {
7419 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
7421 case Intrinsic::ppc_altivec_vcmpbfp_p:
7422 case Intrinsic::ppc_altivec_vcmpeqfp_p:
7423 case Intrinsic::ppc_altivec_vcmpequb_p:
7424 case Intrinsic::ppc_altivec_vcmpequh_p:
7425 case Intrinsic::ppc_altivec_vcmpequw_p:
7426 case Intrinsic::ppc_altivec_vcmpgefp_p:
7427 case Intrinsic::ppc_altivec_vcmpgtfp_p:
7428 case Intrinsic::ppc_altivec_vcmpgtsb_p:
7429 case Intrinsic::ppc_altivec_vcmpgtsh_p:
7430 case Intrinsic::ppc_altivec_vcmpgtsw_p:
7431 case Intrinsic::ppc_altivec_vcmpgtub_p:
7432 case Intrinsic::ppc_altivec_vcmpgtuh_p:
7433 case Intrinsic::ppc_altivec_vcmpgtuw_p:
7434 KnownZero = ~1U; // All bits but the low one are known to be zero.
7442 /// getConstraintType - Given a constraint, return the type of
7443 /// constraint it is for this target.
7444 PPCTargetLowering::ConstraintType
7445 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
7446 if (Constraint.size() == 1) {
7447 switch (Constraint[0]) {
7454 return C_RegisterClass;
7456 // FIXME: While Z does indicate a memory constraint, it specifically
7457 // indicates an r+r address (used in conjunction with the 'y' modifier
7458 // in the replacement string). Currently, we're forcing the base
7459 // register to be r0 in the asm printer (which is interpreted as zero)
7460 // and forming the complete address in the second register. This is
7465 return TargetLowering::getConstraintType(Constraint);
7468 /// Examine constraint type and operand type and determine a weight value.
7469 /// This object must already have been set up with the operand type
7470 /// and the current alternative constraint selected.
7471 TargetLowering::ConstraintWeight
7472 PPCTargetLowering::getSingleConstraintMatchWeight(
7473 AsmOperandInfo &info, const char *constraint) const {
7474 ConstraintWeight weight = CW_Invalid;
7475 Value *CallOperandVal = info.CallOperandVal;
7476 // If we don't have a value, we can't do a match,
7477 // but allow it at the lowest weight.
7478 if (CallOperandVal == NULL)
7480 Type *type = CallOperandVal->getType();
7481 // Look at the constraint type.
7482 switch (*constraint) {
7484 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7487 if (type->isIntegerTy())
7488 weight = CW_Register;
7491 if (type->isFloatTy())
7492 weight = CW_Register;
7495 if (type->isDoubleTy())
7496 weight = CW_Register;
7499 if (type->isVectorTy())
7500 weight = CW_Register;
7503 weight = CW_Register;
7512 std::pair<unsigned, const TargetRegisterClass*>
7513 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7515 if (Constraint.size() == 1) {
7516 // GCC RS6000 Constraint Letters
7517 switch (Constraint[0]) {
7519 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
7520 return std::make_pair(0U, &PPC::G8RC_NOX0RegClass);
7521 return std::make_pair(0U, &PPC::GPRC_NOR0RegClass);
7523 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
7524 return std::make_pair(0U, &PPC::G8RCRegClass);
7525 return std::make_pair(0U, &PPC::GPRCRegClass);
7527 if (VT == MVT::f32 || VT == MVT::i32)
7528 return std::make_pair(0U, &PPC::F4RCRegClass);
7529 if (VT == MVT::f64 || VT == MVT::i64)
7530 return std::make_pair(0U, &PPC::F8RCRegClass);
7533 return std::make_pair(0U, &PPC::VRRCRegClass);
7535 return std::make_pair(0U, &PPC::CRRCRegClass);
7539 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7543 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7544 /// vector. If it is invalid, don't add anything to Ops.
7545 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7546 std::string &Constraint,
7547 std::vector<SDValue>&Ops,
7548 SelectionDAG &DAG) const {
7549 SDValue Result(0,0);
7551 // Only support length 1 constraints.
7552 if (Constraint.length() > 1) return;
7554 char Letter = Constraint[0];
7565 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
7566 if (!CST) return; // Must be an immediate to match.
7567 unsigned Value = CST->getZExtValue();
7569 default: llvm_unreachable("Unknown constraint letter!");
7570 case 'I': // "I" is a signed 16-bit constant.
7571 if ((short)Value == (int)Value)
7572 Result = DAG.getTargetConstant(Value, Op.getValueType());
7574 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
7575 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
7576 if ((short)Value == 0)
7577 Result = DAG.getTargetConstant(Value, Op.getValueType());
7579 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
7580 if ((Value >> 16) == 0)
7581 Result = DAG.getTargetConstant(Value, Op.getValueType());
7583 case 'M': // "M" is a constant that is greater than 31.
7585 Result = DAG.getTargetConstant(Value, Op.getValueType());
7587 case 'N': // "N" is a positive constant that is an exact power of two.
7588 if ((int)Value > 0 && isPowerOf2_32(Value))
7589 Result = DAG.getTargetConstant(Value, Op.getValueType());
7591 case 'O': // "O" is the constant zero.
7593 Result = DAG.getTargetConstant(Value, Op.getValueType());
7595 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
7596 if ((short)-Value == (int)-Value)
7597 Result = DAG.getTargetConstant(Value, Op.getValueType());
7604 if (Result.getNode()) {
7605 Ops.push_back(Result);
7609 // Handle standard constraint letters.
7610 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7613 // isLegalAddressingMode - Return true if the addressing mode represented
7614 // by AM is legal for this target, for a load/store of the specified type.
7615 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
7617 // FIXME: PPC does not allow r+i addressing modes for vectors!
7619 // PPC allows a sign-extended 16-bit immediate field.
7620 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
7623 // No global is ever allowed as a base.
7627 // PPC only support r+r,
7629 case 0: // "r+i" or just "i", depending on HasBaseReg.
7632 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
7634 // Otherwise we have r+r or r+i.
7637 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
7639 // Allow 2*r as r+r.
7642 // No other scales are supported.
7649 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
7650 SelectionDAG &DAG) const {
7651 MachineFunction &MF = DAG.getMachineFunction();
7652 MachineFrameInfo *MFI = MF.getFrameInfo();
7653 MFI->setReturnAddressIsTaken(true);
7656 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7658 // Make sure the function does not optimize away the store of the RA to
7660 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
7661 FuncInfo->setLRStoreRequired();
7662 bool isPPC64 = PPCSubTarget.isPPC64();
7663 bool isDarwinABI = PPCSubTarget.isDarwinABI();
7666 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7669 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
7670 isPPC64? MVT::i64 : MVT::i32);
7671 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7672 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7674 MachinePointerInfo(), false, false, false, 0);
7677 // Just load the return address off the stack.
7678 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
7679 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7680 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
7683 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
7684 SelectionDAG &DAG) const {
7686 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7688 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
7689 bool isPPC64 = PtrVT == MVT::i64;
7691 MachineFunction &MF = DAG.getMachineFunction();
7692 MachineFrameInfo *MFI = MF.getFrameInfo();
7693 MFI->setFrameAddressIsTaken(true);
7695 // Naked functions never have a frame pointer, and so we use r1. For all
7696 // other functions, this decision must be delayed until during PEI.
7698 if (MF.getFunction()->getAttributes().hasAttribute(
7699 AttributeSet::FunctionIndex, Attribute::Naked))
7700 FrameReg = isPPC64 ? PPC::X1 : PPC::R1;
7702 FrameReg = isPPC64 ? PPC::FP8 : PPC::FP;
7704 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
7707 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
7708 FrameAddr, MachinePointerInfo(), false, false,
7714 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
7715 // The PowerPC target isn't yet aware of offsets.
7719 /// getOptimalMemOpType - Returns the target specific optimal type for load
7720 /// and store operations as a result of memset, memcpy, and memmove
7721 /// lowering. If DstAlign is zero that means it's safe to destination
7722 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
7723 /// means there isn't a need to check it against alignment requirement,
7724 /// probably because the source does not need to be loaded. If 'IsMemset' is
7725 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
7726 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
7727 /// source is constant so it does not need to be loaded.
7728 /// It returns EVT::Other if the type should be determined using generic
7729 /// target-independent logic.
7730 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
7731 unsigned DstAlign, unsigned SrcAlign,
7732 bool IsMemset, bool ZeroMemset,
7734 MachineFunction &MF) const {
7735 if (this->PPCSubTarget.isPPC64()) {
7742 bool PPCTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
7744 if (DisablePPCUnaligned)
7747 // PowerPC supports unaligned memory access for simple non-vector types.
7748 // Although accessing unaligned addresses is not as efficient as accessing
7749 // aligned addresses, it is generally more efficient than manual expansion,
7750 // and generally only traps for software emulation when crossing page
7756 if (VT.getSimpleVT().isVector())
7759 if (VT == MVT::ppcf128)
7768 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
7769 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
7770 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
7771 /// is expanded to mul + add.
7772 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
7776 switch (VT.getSimpleVT().SimpleTy) {
7788 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
7790 return TargetLowering::getSchedulingPreference(N);