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 "PPCCallingConv.h"
17 #include "PPCMachineFunctionInfo.h"
18 #include "PPCPerfectShuffle.h"
19 #include "PPCTargetMachine.h"
20 #include "PPCTargetObjectFile.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/StringSwitch.h"
23 #include "llvm/ADT/Triple.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineLoopInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
32 #include "llvm/IR/CallingConv.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetOptions.h"
44 // FIXME: Remove this once soft-float is supported.
45 static cl::opt<bool> DisablePPCFloatInVariadic("disable-ppc-float-in-variadic",
46 cl::desc("disable saving float registers for va_start on PPC"), cl::Hidden);
48 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
49 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
51 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
52 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
54 static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned",
55 cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden);
57 // FIXME: Remove this once the bug has been fixed!
58 extern cl::opt<bool> ANDIGlueBug;
60 PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM)
62 Subtarget(*TM.getSubtargetImpl()) {
63 // Use _setjmp/_longjmp instead of setjmp/longjmp.
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(true);
67 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
68 // arguments are at least 4/8 bytes aligned.
69 bool isPPC64 = Subtarget.isPPC64();
70 setMinStackArgumentAlignment(isPPC64 ? 8:4);
72 // Set up the register classes.
73 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
74 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
75 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
77 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
78 for (MVT VT : MVT::integer_valuetypes()) {
79 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
80 setLoadExtAction(ISD::SEXTLOAD, VT, 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 if (Subtarget.useCRBits()) {
98 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
100 if (isPPC64 || Subtarget.hasFPCVT()) {
101 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote);
102 AddPromotedToType (ISD::SINT_TO_FP, MVT::i1,
103 isPPC64 ? MVT::i64 : MVT::i32);
104 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote);
105 AddPromotedToType (ISD::UINT_TO_FP, MVT::i1,
106 isPPC64 ? MVT::i64 : MVT::i32);
108 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Custom);
109 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Custom);
112 // PowerPC does not support direct load / store of condition registers
113 setOperationAction(ISD::LOAD, MVT::i1, Custom);
114 setOperationAction(ISD::STORE, MVT::i1, Custom);
116 // FIXME: Remove this once the ANDI glue bug is fixed:
118 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
120 for (MVT VT : MVT::integer_valuetypes()) {
121 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
122 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
123 setTruncStoreAction(VT, MVT::i1, Expand);
126 addRegisterClass(MVT::i1, &PPC::CRBITRCRegClass);
129 // This is used in the ppcf128->int sequence. Note it has different semantics
130 // from FP_ROUND: that rounds to nearest, this rounds to zero.
131 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
133 // We do not currently implement these libm ops for PowerPC.
134 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
135 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
136 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
137 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
138 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
139 setOperationAction(ISD::FREM, MVT::ppcf128, Expand);
141 // PowerPC has no SREM/UREM instructions
142 setOperationAction(ISD::SREM, MVT::i32, Expand);
143 setOperationAction(ISD::UREM, MVT::i32, Expand);
144 setOperationAction(ISD::SREM, MVT::i64, Expand);
145 setOperationAction(ISD::UREM, MVT::i64, Expand);
147 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
148 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
149 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
150 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
151 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
152 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
153 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
154 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
155 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
157 // We don't support sin/cos/sqrt/fmod/pow
158 setOperationAction(ISD::FSIN , MVT::f64, Expand);
159 setOperationAction(ISD::FCOS , MVT::f64, Expand);
160 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
161 setOperationAction(ISD::FREM , MVT::f64, Expand);
162 setOperationAction(ISD::FPOW , MVT::f64, Expand);
163 setOperationAction(ISD::FMA , MVT::f64, Legal);
164 setOperationAction(ISD::FSIN , MVT::f32, Expand);
165 setOperationAction(ISD::FCOS , MVT::f32, Expand);
166 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
167 setOperationAction(ISD::FREM , MVT::f32, Expand);
168 setOperationAction(ISD::FPOW , MVT::f32, Expand);
169 setOperationAction(ISD::FMA , MVT::f32, Legal);
171 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
173 // If we're enabling GP optimizations, use hardware square root
174 if (!Subtarget.hasFSQRT() &&
175 !(TM.Options.UnsafeFPMath &&
176 Subtarget.hasFRSQRTE() && Subtarget.hasFRE()))
177 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
179 if (!Subtarget.hasFSQRT() &&
180 !(TM.Options.UnsafeFPMath &&
181 Subtarget.hasFRSQRTES() && Subtarget.hasFRES()))
182 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
184 if (Subtarget.hasFCPSGN()) {
185 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Legal);
186 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Legal);
188 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
189 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
192 if (Subtarget.hasFPRND()) {
193 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
194 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
195 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
196 setOperationAction(ISD::FROUND, MVT::f64, Legal);
198 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
199 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
200 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
201 setOperationAction(ISD::FROUND, MVT::f32, Legal);
204 // PowerPC does not have BSWAP, CTPOP or CTTZ
205 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
206 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
207 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
208 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
209 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
210 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
211 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
212 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
214 if (Subtarget.hasPOPCNTD()) {
215 setOperationAction(ISD::CTPOP, MVT::i32 , Legal);
216 setOperationAction(ISD::CTPOP, MVT::i64 , Legal);
218 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
219 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
222 // PowerPC does not have ROTR
223 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
224 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
226 if (!Subtarget.useCRBits()) {
227 // PowerPC does not have Select
228 setOperationAction(ISD::SELECT, MVT::i32, Expand);
229 setOperationAction(ISD::SELECT, MVT::i64, Expand);
230 setOperationAction(ISD::SELECT, MVT::f32, Expand);
231 setOperationAction(ISD::SELECT, MVT::f64, Expand);
234 // PowerPC wants to turn select_cc of FP into fsel when possible.
235 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
236 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
238 // PowerPC wants to optimize integer setcc a bit
239 if (!Subtarget.useCRBits())
240 setOperationAction(ISD::SETCC, MVT::i32, Custom);
242 // PowerPC does not have BRCOND which requires SetCC
243 if (!Subtarget.useCRBits())
244 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
246 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
248 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
249 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
251 // PowerPC does not have [U|S]INT_TO_FP
252 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
253 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
255 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
256 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
257 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
258 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
260 // We cannot sextinreg(i1). Expand to shifts.
261 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
263 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
264 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
265 // support continuation, user-level threading, and etc.. As a result, no
266 // other SjLj exception interfaces are implemented and please don't build
267 // your own exception handling based on them.
268 // LLVM/Clang supports zero-cost DWARF exception handling.
269 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
270 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
272 // We want to legalize GlobalAddress and ConstantPool nodes into the
273 // appropriate instructions to materialize the address.
274 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
275 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
276 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
277 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
278 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
279 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
280 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
281 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
282 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
283 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
286 setOperationAction(ISD::TRAP, MVT::Other, Legal);
288 // TRAMPOLINE is custom lowered.
289 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
290 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
292 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
293 setOperationAction(ISD::VASTART , MVT::Other, Custom);
295 if (Subtarget.isSVR4ABI()) {
297 // VAARG always uses double-word chunks, so promote anything smaller.
298 setOperationAction(ISD::VAARG, MVT::i1, Promote);
299 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
300 setOperationAction(ISD::VAARG, MVT::i8, Promote);
301 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
302 setOperationAction(ISD::VAARG, MVT::i16, Promote);
303 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
304 setOperationAction(ISD::VAARG, MVT::i32, Promote);
305 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
306 setOperationAction(ISD::VAARG, MVT::Other, Expand);
308 // VAARG is custom lowered with the 32-bit SVR4 ABI.
309 setOperationAction(ISD::VAARG, MVT::Other, Custom);
310 setOperationAction(ISD::VAARG, MVT::i64, Custom);
313 setOperationAction(ISD::VAARG, MVT::Other, Expand);
315 if (Subtarget.isSVR4ABI() && !isPPC64)
316 // VACOPY is custom lowered with the 32-bit SVR4 ABI.
317 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
319 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
321 // Use the default implementation.
322 setOperationAction(ISD::VAEND , MVT::Other, Expand);
323 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
324 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
325 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
326 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
328 // We want to custom lower some of our intrinsics.
329 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
331 // To handle counter-based loop conditions.
332 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom);
334 // Comparisons that require checking two conditions.
335 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
336 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
337 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
338 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
339 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
340 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
341 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
342 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
343 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
344 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
345 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
346 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
348 if (Subtarget.has64BitSupport()) {
349 // They also have instructions for converting between i64 and fp.
350 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
351 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
352 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
353 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
354 // This is just the low 32 bits of a (signed) fp->i64 conversion.
355 // We cannot do this with Promote because i64 is not a legal type.
356 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
358 if (Subtarget.hasLFIWAX() || Subtarget.isPPC64())
359 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
361 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
362 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
365 // With the instructions enabled under FPCVT, we can do everything.
366 if (Subtarget.hasFPCVT()) {
367 if (Subtarget.has64BitSupport()) {
368 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
369 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
370 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
371 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
374 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
375 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
376 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
377 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
380 if (Subtarget.use64BitRegs()) {
381 // 64-bit PowerPC implementations can support i64 types directly
382 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
383 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
384 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
385 // 64-bit PowerPC wants to expand i128 shifts itself.
386 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
387 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
388 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
390 // 32-bit PowerPC wants to expand i64 shifts itself.
391 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
392 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
393 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
396 if (Subtarget.hasAltivec()) {
397 // First set operation action for all vector types to expand. Then we
398 // will selectively turn on ones that can be effectively codegen'd.
399 for (MVT VT : MVT::vector_valuetypes()) {
400 // add/sub are legal for all supported vector VT's.
401 setOperationAction(ISD::ADD , VT, Legal);
402 setOperationAction(ISD::SUB , VT, Legal);
404 // We promote all shuffles to v16i8.
405 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
406 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
408 // We promote all non-typed operations to v4i32.
409 setOperationAction(ISD::AND , VT, Promote);
410 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
411 setOperationAction(ISD::OR , VT, Promote);
412 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
413 setOperationAction(ISD::XOR , VT, Promote);
414 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
415 setOperationAction(ISD::LOAD , VT, Promote);
416 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
417 setOperationAction(ISD::SELECT, VT, Promote);
418 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
419 setOperationAction(ISD::STORE, VT, Promote);
420 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
422 // No other operations are legal.
423 setOperationAction(ISD::MUL , VT, Expand);
424 setOperationAction(ISD::SDIV, VT, Expand);
425 setOperationAction(ISD::SREM, VT, Expand);
426 setOperationAction(ISD::UDIV, VT, Expand);
427 setOperationAction(ISD::UREM, VT, Expand);
428 setOperationAction(ISD::FDIV, VT, Expand);
429 setOperationAction(ISD::FREM, VT, Expand);
430 setOperationAction(ISD::FNEG, VT, Expand);
431 setOperationAction(ISD::FSQRT, VT, Expand);
432 setOperationAction(ISD::FLOG, VT, Expand);
433 setOperationAction(ISD::FLOG10, VT, Expand);
434 setOperationAction(ISD::FLOG2, VT, Expand);
435 setOperationAction(ISD::FEXP, VT, Expand);
436 setOperationAction(ISD::FEXP2, VT, Expand);
437 setOperationAction(ISD::FSIN, VT, Expand);
438 setOperationAction(ISD::FCOS, VT, Expand);
439 setOperationAction(ISD::FABS, VT, Expand);
440 setOperationAction(ISD::FPOWI, VT, Expand);
441 setOperationAction(ISD::FFLOOR, VT, Expand);
442 setOperationAction(ISD::FCEIL, VT, Expand);
443 setOperationAction(ISD::FTRUNC, VT, Expand);
444 setOperationAction(ISD::FRINT, VT, Expand);
445 setOperationAction(ISD::FNEARBYINT, VT, Expand);
446 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
447 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
448 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
449 setOperationAction(ISD::MULHU, VT, Expand);
450 setOperationAction(ISD::MULHS, VT, Expand);
451 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
452 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
453 setOperationAction(ISD::UDIVREM, VT, Expand);
454 setOperationAction(ISD::SDIVREM, VT, Expand);
455 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
456 setOperationAction(ISD::FPOW, VT, Expand);
457 setOperationAction(ISD::BSWAP, VT, Expand);
458 setOperationAction(ISD::CTPOP, VT, Expand);
459 setOperationAction(ISD::CTLZ, VT, Expand);
460 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
461 setOperationAction(ISD::CTTZ, VT, Expand);
462 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
463 setOperationAction(ISD::VSELECT, VT, Expand);
464 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
466 for (MVT InnerVT : MVT::vector_valuetypes()) {
467 setTruncStoreAction(VT, InnerVT, Expand);
468 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
469 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
470 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
474 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
475 // with merges, splats, etc.
476 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
478 setOperationAction(ISD::AND , MVT::v4i32, Legal);
479 setOperationAction(ISD::OR , MVT::v4i32, Legal);
480 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
481 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
482 setOperationAction(ISD::SELECT, MVT::v4i32,
483 Subtarget.useCRBits() ? Legal : Expand);
484 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
485 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
486 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
487 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
488 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
489 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
490 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
491 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
492 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
494 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
495 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
496 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
497 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
499 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
500 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
502 if (TM.Options.UnsafeFPMath || Subtarget.hasVSX()) {
503 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
504 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
507 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
508 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
509 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
511 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
512 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
514 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
515 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
516 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
517 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
519 // Altivec does not contain unordered floating-point compare instructions
520 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
521 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
522 setCondCodeAction(ISD::SETO, MVT::v4f32, Expand);
523 setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand);
525 if (Subtarget.hasVSX()) {
526 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
527 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal);
529 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
530 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
531 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
532 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
533 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
535 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
537 setOperationAction(ISD::MUL, MVT::v2f64, Legal);
538 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
540 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
541 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
543 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
544 setOperationAction(ISD::VSELECT, MVT::v8i16, Legal);
545 setOperationAction(ISD::VSELECT, MVT::v4i32, Legal);
546 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
547 setOperationAction(ISD::VSELECT, MVT::v2f64, Legal);
549 // Share the Altivec comparison restrictions.
550 setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand);
551 setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand);
552 setCondCodeAction(ISD::SETO, MVT::v2f64, Expand);
553 setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand);
555 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
556 setOperationAction(ISD::STORE, MVT::v2f64, Legal);
558 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Legal);
560 addRegisterClass(MVT::f64, &PPC::VSFRCRegClass);
562 addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass);
563 addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass);
565 // VSX v2i64 only supports non-arithmetic operations.
566 setOperationAction(ISD::ADD, MVT::v2i64, Expand);
567 setOperationAction(ISD::SUB, MVT::v2i64, Expand);
569 setOperationAction(ISD::SHL, MVT::v2i64, Expand);
570 setOperationAction(ISD::SRA, MVT::v2i64, Expand);
571 setOperationAction(ISD::SRL, MVT::v2i64, Expand);
573 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
575 setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
576 AddPromotedToType (ISD::LOAD, MVT::v2i64, MVT::v2f64);
577 setOperationAction(ISD::STORE, MVT::v2i64, Promote);
578 AddPromotedToType (ISD::STORE, MVT::v2i64, MVT::v2f64);
580 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Legal);
582 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
583 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
584 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
585 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
587 // Vector operation legalization checks the result type of
588 // SIGN_EXTEND_INREG, overall legalization checks the inner type.
589 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Legal);
590 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Legal);
591 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
592 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
594 addRegisterClass(MVT::v2i64, &PPC::VSRCRegClass);
598 if (Subtarget.has64BitSupport())
599 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
601 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, isPPC64 ? Legal : Custom);
604 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
605 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
608 setBooleanContents(ZeroOrOneBooleanContent);
609 // Altivec instructions set fields to all zeros or all ones.
610 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
613 // These libcalls are not available in 32-bit.
614 setLibcallName(RTLIB::SHL_I128, nullptr);
615 setLibcallName(RTLIB::SRL_I128, nullptr);
616 setLibcallName(RTLIB::SRA_I128, nullptr);
620 setStackPointerRegisterToSaveRestore(PPC::X1);
621 setExceptionPointerRegister(PPC::X3);
622 setExceptionSelectorRegister(PPC::X4);
624 setStackPointerRegisterToSaveRestore(PPC::R1);
625 setExceptionPointerRegister(PPC::R3);
626 setExceptionSelectorRegister(PPC::R4);
629 // We have target-specific dag combine patterns for the following nodes:
630 setTargetDAGCombine(ISD::SINT_TO_FP);
631 if (Subtarget.hasFPCVT())
632 setTargetDAGCombine(ISD::UINT_TO_FP);
633 setTargetDAGCombine(ISD::LOAD);
634 setTargetDAGCombine(ISD::STORE);
635 setTargetDAGCombine(ISD::BR_CC);
636 if (Subtarget.useCRBits())
637 setTargetDAGCombine(ISD::BRCOND);
638 setTargetDAGCombine(ISD::BSWAP);
639 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
640 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
641 setTargetDAGCombine(ISD::INTRINSIC_VOID);
643 setTargetDAGCombine(ISD::SIGN_EXTEND);
644 setTargetDAGCombine(ISD::ZERO_EXTEND);
645 setTargetDAGCombine(ISD::ANY_EXTEND);
647 if (Subtarget.useCRBits()) {
648 setTargetDAGCombine(ISD::TRUNCATE);
649 setTargetDAGCombine(ISD::SETCC);
650 setTargetDAGCombine(ISD::SELECT_CC);
653 // Use reciprocal estimates.
654 if (TM.Options.UnsafeFPMath) {
655 setTargetDAGCombine(ISD::FDIV);
656 setTargetDAGCombine(ISD::FSQRT);
659 // Darwin long double math library functions have $LDBL128 appended.
660 if (Subtarget.isDarwin()) {
661 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
662 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
663 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
664 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
665 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
666 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
667 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
668 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
669 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
670 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
673 // With 32 condition bits, we don't need to sink (and duplicate) compares
674 // aggressively in CodeGenPrep.
675 if (Subtarget.useCRBits())
676 setHasMultipleConditionRegisters();
678 setMinFunctionAlignment(2);
679 if (Subtarget.isDarwin())
680 setPrefFunctionAlignment(4);
682 switch (Subtarget.getDarwinDirective()) {
686 case PPC::DIR_E500mc:
695 setPrefFunctionAlignment(4);
696 setPrefLoopAlignment(4);
700 setInsertFencesForAtomic(true);
702 if (Subtarget.enableMachineScheduler())
703 setSchedulingPreference(Sched::Source);
705 setSchedulingPreference(Sched::Hybrid);
707 computeRegisterProperties();
709 // The Freescale cores do better with aggressive inlining of memcpy and
710 // friends. GCC uses same threshold of 128 bytes (= 32 word stores).
711 if (Subtarget.getDarwinDirective() == PPC::DIR_E500mc ||
712 Subtarget.getDarwinDirective() == PPC::DIR_E5500) {
713 MaxStoresPerMemset = 32;
714 MaxStoresPerMemsetOptSize = 16;
715 MaxStoresPerMemcpy = 32;
716 MaxStoresPerMemcpyOptSize = 8;
717 MaxStoresPerMemmove = 32;
718 MaxStoresPerMemmoveOptSize = 8;
722 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
723 /// the desired ByVal argument alignment.
724 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign,
725 unsigned MaxMaxAlign) {
726 if (MaxAlign == MaxMaxAlign)
728 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
729 if (MaxMaxAlign >= 32 && VTy->getBitWidth() >= 256)
731 else if (VTy->getBitWidth() >= 128 && MaxAlign < 16)
733 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
734 unsigned EltAlign = 0;
735 getMaxByValAlign(ATy->getElementType(), EltAlign, MaxMaxAlign);
736 if (EltAlign > MaxAlign)
738 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
739 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
740 unsigned EltAlign = 0;
741 getMaxByValAlign(STy->getElementType(i), EltAlign, MaxMaxAlign);
742 if (EltAlign > MaxAlign)
744 if (MaxAlign == MaxMaxAlign)
750 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
751 /// function arguments in the caller parameter area.
752 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
753 // Darwin passes everything on 4 byte boundary.
754 if (Subtarget.isDarwin())
757 // 16byte and wider vectors are passed on 16byte boundary.
758 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
759 unsigned Align = Subtarget.isPPC64() ? 8 : 4;
760 if (Subtarget.hasAltivec() || Subtarget.hasQPX())
761 getMaxByValAlign(Ty, Align, Subtarget.hasQPX() ? 32 : 16);
765 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
767 default: return nullptr;
768 case PPCISD::FSEL: return "PPCISD::FSEL";
769 case PPCISD::FCFID: return "PPCISD::FCFID";
770 case PPCISD::FCFIDU: return "PPCISD::FCFIDU";
771 case PPCISD::FCFIDS: return "PPCISD::FCFIDS";
772 case PPCISD::FCFIDUS: return "PPCISD::FCFIDUS";
773 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
774 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
775 case PPCISD::FCTIDUZ: return "PPCISD::FCTIDUZ";
776 case PPCISD::FCTIWUZ: return "PPCISD::FCTIWUZ";
777 case PPCISD::FRE: return "PPCISD::FRE";
778 case PPCISD::FRSQRTE: return "PPCISD::FRSQRTE";
779 case PPCISD::STFIWX: return "PPCISD::STFIWX";
780 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
781 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
782 case PPCISD::VPERM: return "PPCISD::VPERM";
783 case PPCISD::CMPB: return "PPCISD::CMPB";
784 case PPCISD::Hi: return "PPCISD::Hi";
785 case PPCISD::Lo: return "PPCISD::Lo";
786 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
787 case PPCISD::LOAD: return "PPCISD::LOAD";
788 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
789 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
790 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
791 case PPCISD::SRL: return "PPCISD::SRL";
792 case PPCISD::SRA: return "PPCISD::SRA";
793 case PPCISD::SHL: return "PPCISD::SHL";
794 case PPCISD::CALL: return "PPCISD::CALL";
795 case PPCISD::CALL_NOP: return "PPCISD::CALL_NOP";
796 case PPCISD::CALL_TLS: return "PPCISD::CALL_TLS";
797 case PPCISD::CALL_NOP_TLS: return "PPCISD::CALL_NOP_TLS";
798 case PPCISD::MTCTR: return "PPCISD::MTCTR";
799 case PPCISD::BCTRL: return "PPCISD::BCTRL";
800 case PPCISD::BCTRL_LOAD_TOC: return "PPCISD::BCTRL_LOAD_TOC";
801 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
802 case PPCISD::READ_TIME_BASE: return "PPCISD::READ_TIME_BASE";
803 case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP";
804 case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP";
805 case PPCISD::MFOCRF: return "PPCISD::MFOCRF";
806 case PPCISD::VCMP: return "PPCISD::VCMP";
807 case PPCISD::VCMPo: return "PPCISD::VCMPo";
808 case PPCISD::LBRX: return "PPCISD::LBRX";
809 case PPCISD::STBRX: return "PPCISD::STBRX";
810 case PPCISD::LFIWAX: return "PPCISD::LFIWAX";
811 case PPCISD::LFIWZX: return "PPCISD::LFIWZX";
812 case PPCISD::LARX: return "PPCISD::LARX";
813 case PPCISD::STCX: return "PPCISD::STCX";
814 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
815 case PPCISD::BDNZ: return "PPCISD::BDNZ";
816 case PPCISD::BDZ: return "PPCISD::BDZ";
817 case PPCISD::MFFS: return "PPCISD::MFFS";
818 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
819 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
820 case PPCISD::CR6SET: return "PPCISD::CR6SET";
821 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
822 case PPCISD::ADDIS_TOC_HA: return "PPCISD::ADDIS_TOC_HA";
823 case PPCISD::LD_TOC_L: return "PPCISD::LD_TOC_L";
824 case PPCISD::ADDI_TOC_L: return "PPCISD::ADDI_TOC_L";
825 case PPCISD::PPC32_GOT: return "PPCISD::PPC32_GOT";
826 case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
827 case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L";
828 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
829 case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA";
830 case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L";
831 case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA";
832 case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L";
833 case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
834 case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L";
835 case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT";
836 case PPCISD::SC: return "PPCISD::SC";
840 EVT PPCTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
842 return Subtarget.useCRBits() ? MVT::i1 : MVT::i32;
843 return VT.changeVectorElementTypeToInteger();
846 bool PPCTargetLowering::enableAggressiveFMAFusion(EVT VT) const {
847 assert(VT.isFloatingPoint() && "Non-floating-point FMA?");
851 //===----------------------------------------------------------------------===//
852 // Node matching predicates, for use by the tblgen matching code.
853 //===----------------------------------------------------------------------===//
855 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
856 static bool isFloatingPointZero(SDValue Op) {
857 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
858 return CFP->getValueAPF().isZero();
859 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
860 // Maybe this has already been legalized into the constant pool?
861 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
862 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
863 return CFP->getValueAPF().isZero();
868 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
869 /// true if Op is undef or if it matches the specified value.
870 static bool isConstantOrUndef(int Op, int Val) {
871 return Op < 0 || Op == Val;
874 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
875 /// VPKUHUM instruction.
876 /// The ShuffleKind distinguishes between big-endian operations with
877 /// two different inputs (0), either-endian operations with two identical
878 /// inputs (1), and little-endian operantion with two different inputs (2).
879 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
880 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
882 bool IsLE = DAG.getSubtarget().getDataLayout()->isLittleEndian();
883 if (ShuffleKind == 0) {
886 for (unsigned i = 0; i != 16; ++i)
887 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
889 } else if (ShuffleKind == 2) {
892 for (unsigned i = 0; i != 16; ++i)
893 if (!isConstantOrUndef(N->getMaskElt(i), i*2))
895 } else if (ShuffleKind == 1) {
896 unsigned j = IsLE ? 0 : 1;
897 for (unsigned i = 0; i != 8; ++i)
898 if (!isConstantOrUndef(N->getMaskElt(i), i*2+j) ||
899 !isConstantOrUndef(N->getMaskElt(i+8), i*2+j))
905 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
906 /// VPKUWUM instruction.
907 /// The ShuffleKind distinguishes between big-endian operations with
908 /// two different inputs (0), either-endian operations with two identical
909 /// inputs (1), and little-endian operantion with two different inputs (2).
910 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
911 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
913 bool IsLE = DAG.getSubtarget().getDataLayout()->isLittleEndian();
914 if (ShuffleKind == 0) {
917 for (unsigned i = 0; i != 16; i += 2)
918 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
919 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
921 } else if (ShuffleKind == 2) {
924 for (unsigned i = 0; i != 16; i += 2)
925 if (!isConstantOrUndef(N->getMaskElt(i ), i*2) ||
926 !isConstantOrUndef(N->getMaskElt(i+1), i*2+1))
928 } else if (ShuffleKind == 1) {
929 unsigned j = IsLE ? 0 : 2;
930 for (unsigned i = 0; i != 8; i += 2)
931 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
932 !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) ||
933 !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) ||
934 !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1))
940 /// isVMerge - Common function, used to match vmrg* shuffles.
942 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
943 unsigned LHSStart, unsigned RHSStart) {
944 if (N->getValueType(0) != MVT::v16i8)
946 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
947 "Unsupported merge size!");
949 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
950 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
951 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
952 LHSStart+j+i*UnitSize) ||
953 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
954 RHSStart+j+i*UnitSize))
960 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
961 /// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes).
962 /// The ShuffleKind distinguishes between big-endian merges with two
963 /// different inputs (0), either-endian merges with two identical inputs (1),
964 /// and little-endian merges with two different inputs (2). For the latter,
965 /// the input operands are swapped (see PPCInstrAltivec.td).
966 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
967 unsigned ShuffleKind, SelectionDAG &DAG) {
968 if (DAG.getSubtarget().getDataLayout()->isLittleEndian()) {
969 if (ShuffleKind == 1) // unary
970 return isVMerge(N, UnitSize, 0, 0);
971 else if (ShuffleKind == 2) // swapped
972 return isVMerge(N, UnitSize, 0, 16);
976 if (ShuffleKind == 1) // unary
977 return isVMerge(N, UnitSize, 8, 8);
978 else if (ShuffleKind == 0) // normal
979 return isVMerge(N, UnitSize, 8, 24);
985 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
986 /// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes).
987 /// The ShuffleKind distinguishes between big-endian merges with two
988 /// different inputs (0), either-endian merges with two identical inputs (1),
989 /// and little-endian merges with two different inputs (2). For the latter,
990 /// the input operands are swapped (see PPCInstrAltivec.td).
991 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
992 unsigned ShuffleKind, SelectionDAG &DAG) {
993 if (DAG.getSubtarget().getDataLayout()->isLittleEndian()) {
994 if (ShuffleKind == 1) // unary
995 return isVMerge(N, UnitSize, 8, 8);
996 else if (ShuffleKind == 2) // swapped
997 return isVMerge(N, UnitSize, 8, 24);
1001 if (ShuffleKind == 1) // unary
1002 return isVMerge(N, UnitSize, 0, 0);
1003 else if (ShuffleKind == 0) // normal
1004 return isVMerge(N, UnitSize, 0, 16);
1011 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
1012 /// amount, otherwise return -1.
1013 /// The ShuffleKind distinguishes between big-endian operations with two
1014 /// different inputs (0), either-endian operations with two identical inputs
1015 /// (1), and little-endian operations with two different inputs (2). For the
1016 /// latter, the input operands are swapped (see PPCInstrAltivec.td).
1017 int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
1018 SelectionDAG &DAG) {
1019 if (N->getValueType(0) != MVT::v16i8)
1022 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1024 // Find the first non-undef value in the shuffle mask.
1026 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
1029 if (i == 16) return -1; // all undef.
1031 // Otherwise, check to see if the rest of the elements are consecutively
1032 // numbered from this value.
1033 unsigned ShiftAmt = SVOp->getMaskElt(i);
1034 if (ShiftAmt < i) return -1;
1037 bool isLE = DAG.getTarget().getSubtargetImpl()->getDataLayout()->
1040 if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) {
1041 // Check the rest of the elements to see if they are consecutive.
1042 for (++i; i != 16; ++i)
1043 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
1045 } else if (ShuffleKind == 1) {
1046 // Check the rest of the elements to see if they are consecutive.
1047 for (++i; i != 16; ++i)
1048 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
1053 if (ShuffleKind == 2 && isLE)
1054 ShiftAmt = 16 - ShiftAmt;
1059 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
1060 /// specifies a splat of a single element that is suitable for input to
1061 /// VSPLTB/VSPLTH/VSPLTW.
1062 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
1063 assert(N->getValueType(0) == MVT::v16i8 &&
1064 (EltSize == 1 || EltSize == 2 || EltSize == 4));
1066 // This is a splat operation if each element of the permute is the same, and
1067 // if the value doesn't reference the second vector.
1068 unsigned ElementBase = N->getMaskElt(0);
1070 // FIXME: Handle UNDEF elements too!
1071 if (ElementBase >= 16)
1074 // Check that the indices are consecutive, in the case of a multi-byte element
1075 // splatted with a v16i8 mask.
1076 for (unsigned i = 1; i != EltSize; ++i)
1077 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
1080 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
1081 if (N->getMaskElt(i) < 0) continue;
1082 for (unsigned j = 0; j != EltSize; ++j)
1083 if (N->getMaskElt(i+j) != N->getMaskElt(j))
1089 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
1091 bool PPC::isAllNegativeZeroVector(SDNode *N) {
1092 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
1094 APInt APVal, APUndef;
1098 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
1099 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
1100 return CFP->getValueAPF().isNegZero();
1105 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
1106 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
1107 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize,
1108 SelectionDAG &DAG) {
1109 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1110 assert(isSplatShuffleMask(SVOp, EltSize));
1111 if (DAG.getSubtarget().getDataLayout()->isLittleEndian())
1112 return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize);
1114 return SVOp->getMaskElt(0) / EltSize;
1117 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
1118 /// by using a vspltis[bhw] instruction of the specified element size, return
1119 /// the constant being splatted. The ByteSize field indicates the number of
1120 /// bytes of each element [124] -> [bhw].
1121 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
1122 SDValue OpVal(nullptr, 0);
1124 // If ByteSize of the splat is bigger than the element size of the
1125 // build_vector, then we have a case where we are checking for a splat where
1126 // multiple elements of the buildvector are folded together into a single
1127 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
1128 unsigned EltSize = 16/N->getNumOperands();
1129 if (EltSize < ByteSize) {
1130 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
1131 SDValue UniquedVals[4];
1132 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
1134 // See if all of the elements in the buildvector agree across.
1135 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1136 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1137 // If the element isn't a constant, bail fully out.
1138 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
1141 if (!UniquedVals[i&(Multiple-1)].getNode())
1142 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
1143 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
1144 return SDValue(); // no match.
1147 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
1148 // either constant or undef values that are identical for each chunk. See
1149 // if these chunks can form into a larger vspltis*.
1151 // Check to see if all of the leading entries are either 0 or -1. If
1152 // neither, then this won't fit into the immediate field.
1153 bool LeadingZero = true;
1154 bool LeadingOnes = true;
1155 for (unsigned i = 0; i != Multiple-1; ++i) {
1156 if (!UniquedVals[i].getNode()) continue; // Must have been undefs.
1158 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
1159 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
1161 // Finally, check the least significant entry.
1163 if (!UniquedVals[Multiple-1].getNode())
1164 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
1165 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
1167 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
1170 if (!UniquedVals[Multiple-1].getNode())
1171 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
1172 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
1173 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
1174 return DAG.getTargetConstant(Val, MVT::i32);
1180 // Check to see if this buildvec has a single non-undef value in its elements.
1181 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1182 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1183 if (!OpVal.getNode())
1184 OpVal = N->getOperand(i);
1185 else if (OpVal != N->getOperand(i))
1189 if (!OpVal.getNode()) return SDValue(); // All UNDEF: use implicit def.
1191 unsigned ValSizeInBytes = EltSize;
1193 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1194 Value = CN->getZExtValue();
1195 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
1196 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
1197 Value = FloatToBits(CN->getValueAPF().convertToFloat());
1200 // If the splat value is larger than the element value, then we can never do
1201 // this splat. The only case that we could fit the replicated bits into our
1202 // immediate field for would be zero, and we prefer to use vxor for it.
1203 if (ValSizeInBytes < ByteSize) return SDValue();
1205 // If the element value is larger than the splat value, cut it in half and
1206 // check to see if the two halves are equal. Continue doing this until we
1207 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
1208 while (ValSizeInBytes > ByteSize) {
1209 ValSizeInBytes >>= 1;
1211 // If the top half equals the bottom half, we're still ok.
1212 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
1213 (Value & ((1 << (8*ValSizeInBytes))-1)))
1217 // Properly sign extend the value.
1218 int MaskVal = SignExtend32(Value, ByteSize * 8);
1220 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
1221 if (MaskVal == 0) return SDValue();
1223 // Finally, if this value fits in a 5 bit sext field, return it
1224 if (SignExtend32<5>(MaskVal) == MaskVal)
1225 return DAG.getTargetConstant(MaskVal, MVT::i32);
1229 //===----------------------------------------------------------------------===//
1230 // Addressing Mode Selection
1231 //===----------------------------------------------------------------------===//
1233 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
1234 /// or 64-bit immediate, and if the value can be accurately represented as a
1235 /// sign extension from a 16-bit value. If so, this returns true and the
1237 static bool isIntS16Immediate(SDNode *N, short &Imm) {
1238 if (!isa<ConstantSDNode>(N))
1241 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
1242 if (N->getValueType(0) == MVT::i32)
1243 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
1245 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
1247 static bool isIntS16Immediate(SDValue Op, short &Imm) {
1248 return isIntS16Immediate(Op.getNode(), Imm);
1252 /// SelectAddressRegReg - Given the specified addressed, check to see if it
1253 /// can be represented as an indexed [r+r] operation. Returns false if it
1254 /// can be more efficiently represented with [r+imm].
1255 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
1257 SelectionDAG &DAG) const {
1259 if (N.getOpcode() == ISD::ADD) {
1260 if (isIntS16Immediate(N.getOperand(1), imm))
1261 return false; // r+i
1262 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
1263 return false; // r+i
1265 Base = N.getOperand(0);
1266 Index = N.getOperand(1);
1268 } else if (N.getOpcode() == ISD::OR) {
1269 if (isIntS16Immediate(N.getOperand(1), imm))
1270 return false; // r+i can fold it if we can.
1272 // If this is an or of disjoint bitfields, we can codegen this as an add
1273 // (for better address arithmetic) if the LHS and RHS of the OR are provably
1275 APInt LHSKnownZero, LHSKnownOne;
1276 APInt RHSKnownZero, RHSKnownOne;
1277 DAG.computeKnownBits(N.getOperand(0),
1278 LHSKnownZero, LHSKnownOne);
1280 if (LHSKnownZero.getBoolValue()) {
1281 DAG.computeKnownBits(N.getOperand(1),
1282 RHSKnownZero, RHSKnownOne);
1283 // If all of the bits are known zero on the LHS or RHS, the add won't
1285 if (~(LHSKnownZero | RHSKnownZero) == 0) {
1286 Base = N.getOperand(0);
1287 Index = N.getOperand(1);
1296 // If we happen to be doing an i64 load or store into a stack slot that has
1297 // less than a 4-byte alignment, then the frame-index elimination may need to
1298 // use an indexed load or store instruction (because the offset may not be a
1299 // multiple of 4). The extra register needed to hold the offset comes from the
1300 // register scavenger, and it is possible that the scavenger will need to use
1301 // an emergency spill slot. As a result, we need to make sure that a spill slot
1302 // is allocated when doing an i64 load/store into a less-than-4-byte-aligned
1304 static void fixupFuncForFI(SelectionDAG &DAG, int FrameIdx, EVT VT) {
1305 // FIXME: This does not handle the LWA case.
1309 // NOTE: We'll exclude negative FIs here, which come from argument
1310 // lowering, because there are no known test cases triggering this problem
1311 // using packed structures (or similar). We can remove this exclusion if
1312 // we find such a test case. The reason why this is so test-case driven is
1313 // because this entire 'fixup' is only to prevent crashes (from the
1314 // register scavenger) on not-really-valid inputs. For example, if we have:
1316 // %b = bitcast i1* %a to i64*
1317 // store i64* a, i64 b
1318 // then the store should really be marked as 'align 1', but is not. If it
1319 // were marked as 'align 1' then the indexed form would have been
1320 // instruction-selected initially, and the problem this 'fixup' is preventing
1321 // won't happen regardless.
1325 MachineFunction &MF = DAG.getMachineFunction();
1326 MachineFrameInfo *MFI = MF.getFrameInfo();
1328 unsigned Align = MFI->getObjectAlignment(FrameIdx);
1332 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1333 FuncInfo->setHasNonRISpills();
1336 /// Returns true if the address N can be represented by a base register plus
1337 /// a signed 16-bit displacement [r+imm], and if it is not better
1338 /// represented as reg+reg. If Aligned is true, only accept displacements
1339 /// suitable for STD and friends, i.e. multiples of 4.
1340 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
1343 bool Aligned) const {
1344 // FIXME dl should come from parent load or store, not from address
1346 // If this can be more profitably realized as r+r, fail.
1347 if (SelectAddressRegReg(N, Disp, Base, DAG))
1350 if (N.getOpcode() == ISD::ADD) {
1352 if (isIntS16Immediate(N.getOperand(1), imm) &&
1353 (!Aligned || (imm & 3) == 0)) {
1354 Disp = DAG.getTargetConstant(imm, N.getValueType());
1355 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1356 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1357 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1359 Base = N.getOperand(0);
1361 return true; // [r+i]
1362 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1363 // Match LOAD (ADD (X, Lo(G))).
1364 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1365 && "Cannot handle constant offsets yet!");
1366 Disp = N.getOperand(1).getOperand(0); // The global address.
1367 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1368 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
1369 Disp.getOpcode() == ISD::TargetConstantPool ||
1370 Disp.getOpcode() == ISD::TargetJumpTable);
1371 Base = N.getOperand(0);
1372 return true; // [&g+r]
1374 } else if (N.getOpcode() == ISD::OR) {
1376 if (isIntS16Immediate(N.getOperand(1), imm) &&
1377 (!Aligned || (imm & 3) == 0)) {
1378 // If this is an or of disjoint bitfields, we can codegen this as an add
1379 // (for better address arithmetic) if the LHS and RHS of the OR are
1380 // provably disjoint.
1381 APInt LHSKnownZero, LHSKnownOne;
1382 DAG.computeKnownBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1384 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1385 // If all of the bits are known zero on the LHS or RHS, the add won't
1387 if (FrameIndexSDNode *FI =
1388 dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1389 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1390 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1392 Base = N.getOperand(0);
1394 Disp = DAG.getTargetConstant(imm, N.getValueType());
1398 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1399 // Loading from a constant address.
1401 // If this address fits entirely in a 16-bit sext immediate field, codegen
1404 if (isIntS16Immediate(CN, Imm) && (!Aligned || (Imm & 3) == 0)) {
1405 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1406 Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1407 CN->getValueType(0));
1411 // Handle 32-bit sext immediates with LIS + addr mode.
1412 if ((CN->getValueType(0) == MVT::i32 ||
1413 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) &&
1414 (!Aligned || (CN->getZExtValue() & 3) == 0)) {
1415 int Addr = (int)CN->getZExtValue();
1417 // Otherwise, break this down into an LIS + disp.
1418 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1420 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1421 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1422 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1427 Disp = DAG.getTargetConstant(0, getPointerTy());
1428 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) {
1429 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1430 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1433 return true; // [r+0]
1436 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1437 /// represented as an indexed [r+r] operation.
1438 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1440 SelectionDAG &DAG) const {
1441 // Check to see if we can easily represent this as an [r+r] address. This
1442 // will fail if it thinks that the address is more profitably represented as
1443 // reg+imm, e.g. where imm = 0.
1444 if (SelectAddressRegReg(N, Base, Index, DAG))
1447 // If the operand is an addition, always emit this as [r+r], since this is
1448 // better (for code size, and execution, as the memop does the add for free)
1449 // than emitting an explicit add.
1450 if (N.getOpcode() == ISD::ADD) {
1451 Base = N.getOperand(0);
1452 Index = N.getOperand(1);
1456 // Otherwise, do it the hard way, using R0 as the base register.
1457 Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1463 /// getPreIndexedAddressParts - returns true by value, base pointer and
1464 /// offset pointer and addressing mode by reference if the node's address
1465 /// can be legally represented as pre-indexed load / store address.
1466 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1468 ISD::MemIndexedMode &AM,
1469 SelectionDAG &DAG) const {
1470 if (DisablePPCPreinc) return false;
1476 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1477 Ptr = LD->getBasePtr();
1478 VT = LD->getMemoryVT();
1479 Alignment = LD->getAlignment();
1480 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1481 Ptr = ST->getBasePtr();
1482 VT = ST->getMemoryVT();
1483 Alignment = ST->getAlignment();
1488 // PowerPC doesn't have preinc load/store instructions for vectors.
1492 if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) {
1494 // Common code will reject creating a pre-inc form if the base pointer
1495 // is a frame index, or if N is a store and the base pointer is either
1496 // the same as or a predecessor of the value being stored. Check for
1497 // those situations here, and try with swapped Base/Offset instead.
1500 if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base))
1503 SDValue Val = cast<StoreSDNode>(N)->getValue();
1504 if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode()))
1509 std::swap(Base, Offset);
1515 // LDU/STU can only handle immediates that are a multiple of 4.
1516 if (VT != MVT::i64) {
1517 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, false))
1520 // LDU/STU need an address with at least 4-byte alignment.
1524 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, true))
1528 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1529 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1530 // sext i32 to i64 when addr mode is r+i.
1531 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1532 LD->getExtensionType() == ISD::SEXTLOAD &&
1533 isa<ConstantSDNode>(Offset))
1541 //===----------------------------------------------------------------------===//
1542 // LowerOperation implementation
1543 //===----------------------------------------------------------------------===//
1545 /// GetLabelAccessInfo - Return true if we should reference labels using a
1546 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1547 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1548 unsigned &LoOpFlags,
1549 const GlobalValue *GV = nullptr) {
1550 HiOpFlags = PPCII::MO_HA;
1551 LoOpFlags = PPCII::MO_LO;
1553 // Don't use the pic base if not in PIC relocation model.
1554 bool isPIC = TM.getRelocationModel() == Reloc::PIC_;
1557 HiOpFlags |= PPCII::MO_PIC_FLAG;
1558 LoOpFlags |= PPCII::MO_PIC_FLAG;
1561 // If this is a reference to a global value that requires a non-lazy-ptr, make
1562 // sure that instruction lowering adds it.
1563 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1564 HiOpFlags |= PPCII::MO_NLP_FLAG;
1565 LoOpFlags |= PPCII::MO_NLP_FLAG;
1567 if (GV->hasHiddenVisibility()) {
1568 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1569 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1576 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1577 SelectionDAG &DAG) {
1578 EVT PtrVT = HiPart.getValueType();
1579 SDValue Zero = DAG.getConstant(0, PtrVT);
1582 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1583 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1585 // With PIC, the first instruction is actually "GR+hi(&G)".
1587 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1588 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1590 // Generate non-pic code that has direct accesses to the constant pool.
1591 // The address of the global is just (hi(&g)+lo(&g)).
1592 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1595 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1596 SelectionDAG &DAG) const {
1597 EVT PtrVT = Op.getValueType();
1598 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1599 const Constant *C = CP->getConstVal();
1601 // 64-bit SVR4 ABI code is always position-independent.
1602 // The actual address of the GlobalValue is stored in the TOC.
1603 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
1604 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1605 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(CP), MVT::i64, GA,
1606 DAG.getRegister(PPC::X2, MVT::i64));
1609 unsigned MOHiFlag, MOLoFlag;
1610 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1612 if (isPIC && Subtarget.isSVR4ABI()) {
1613 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(),
1614 PPCII::MO_PIC_FLAG);
1616 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i32, GA,
1617 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT));
1621 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1623 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1624 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1627 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1628 EVT PtrVT = Op.getValueType();
1629 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1631 // 64-bit SVR4 ABI code is always position-independent.
1632 // The actual address of the GlobalValue is stored in the TOC.
1633 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
1634 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1635 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(JT), MVT::i64, GA,
1636 DAG.getRegister(PPC::X2, MVT::i64));
1639 unsigned MOHiFlag, MOLoFlag;
1640 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1642 if (isPIC && Subtarget.isSVR4ABI()) {
1643 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
1644 PPCII::MO_PIC_FLAG);
1646 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(JT), PtrVT, GA,
1647 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT));
1650 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1651 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1652 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1655 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1656 SelectionDAG &DAG) const {
1657 EVT PtrVT = Op.getValueType();
1658 BlockAddressSDNode *BASDN = cast<BlockAddressSDNode>(Op);
1659 const BlockAddress *BA = BASDN->getBlockAddress();
1661 // 64-bit SVR4 ABI code is always position-independent.
1662 // The actual BlockAddress is stored in the TOC.
1663 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
1664 SDValue GA = DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset());
1665 return DAG.getNode(PPCISD::TOC_ENTRY, SDLoc(BASDN), MVT::i64, GA,
1666 DAG.getRegister(PPC::X2, MVT::i64));
1669 unsigned MOHiFlag, MOLoFlag;
1670 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1671 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1672 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1673 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1676 // Generate a call to __tls_get_addr for the given GOT entry Op.
1677 std::pair<SDValue,SDValue>
1678 PPCTargetLowering::lowerTLSCall(SDValue Op, SDLoc dl,
1679 SelectionDAG &DAG) const {
1681 Type *IntPtrTy = getDataLayout()->getIntPtrType(*DAG.getContext());
1682 TargetLowering::ArgListTy Args;
1683 TargetLowering::ArgListEntry Entry;
1685 Entry.Ty = IntPtrTy;
1686 Args.push_back(Entry);
1688 TargetLowering::CallLoweringInfo CLI(DAG);
1689 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1690 .setCallee(CallingConv::C, IntPtrTy,
1691 DAG.getTargetExternalSymbol("__tls_get_addr", getPointerTy()),
1692 std::move(Args), 0);
1694 return LowerCallTo(CLI);
1697 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1698 SelectionDAG &DAG) const {
1700 // FIXME: TLS addresses currently use medium model code sequences,
1701 // which is the most useful form. Eventually support for small and
1702 // large models could be added if users need it, at the cost of
1703 // additional complexity.
1704 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1706 const GlobalValue *GV = GA->getGlobal();
1707 EVT PtrVT = getPointerTy();
1708 bool is64bit = Subtarget.isPPC64();
1709 const Module *M = DAG.getMachineFunction().getFunction()->getParent();
1710 PICLevel::Level picLevel = M->getPICLevel();
1712 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
1714 if (Model == TLSModel::LocalExec) {
1715 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1716 PPCII::MO_TPREL_HA);
1717 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1718 PPCII::MO_TPREL_LO);
1719 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1720 is64bit ? MVT::i64 : MVT::i32);
1721 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1722 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1725 if (Model == TLSModel::InitialExec) {
1726 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1727 SDValue TGATLS = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1731 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1732 GOTPtr = DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl,
1733 PtrVT, GOTReg, TGA);
1735 GOTPtr = DAG.getNode(PPCISD::PPC32_GOT, dl, PtrVT);
1736 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl,
1737 PtrVT, TGA, GOTPtr);
1738 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS);
1741 if (Model == TLSModel::GeneralDynamic) {
1742 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1746 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1747 GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
1750 if (picLevel == PICLevel::Small)
1751 GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
1753 GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
1755 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSGD_L, dl, PtrVT,
1757 std::pair<SDValue, SDValue> CallResult = lowerTLSCall(GOTEntry, dl, DAG);
1758 return CallResult.first;
1761 if (Model == TLSModel::LocalDynamic) {
1762 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1766 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1767 GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
1770 if (picLevel == PICLevel::Small)
1771 GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
1773 GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
1775 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSLD_L, dl, PtrVT,
1777 std::pair<SDValue, SDValue> CallResult = lowerTLSCall(GOTEntry, dl, DAG);
1778 SDValue TLSAddr = CallResult.first;
1779 SDValue Chain = CallResult.second;
1780 SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, PtrVT,
1781 Chain, TLSAddr, TGA);
1782 return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
1785 llvm_unreachable("Unknown TLS model!");
1788 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1789 SelectionDAG &DAG) const {
1790 EVT PtrVT = Op.getValueType();
1791 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1793 const GlobalValue *GV = GSDN->getGlobal();
1795 // 64-bit SVR4 ABI code is always position-independent.
1796 // The actual address of the GlobalValue is stored in the TOC.
1797 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
1798 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1799 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1800 DAG.getRegister(PPC::X2, MVT::i64));
1803 unsigned MOHiFlag, MOLoFlag;
1804 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1806 if (isPIC && Subtarget.isSVR4ABI()) {
1807 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT,
1809 PPCII::MO_PIC_FLAG);
1810 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i32, GA,
1811 DAG.getNode(PPCISD::GlobalBaseReg, DL, MVT::i32));
1815 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1817 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1819 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1821 // If the global reference is actually to a non-lazy-pointer, we have to do an
1822 // extra load to get the address of the global.
1823 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1824 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1825 false, false, false, 0);
1829 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1830 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1833 if (Op.getValueType() == MVT::v2i64) {
1834 // When the operands themselves are v2i64 values, we need to do something
1835 // special because VSX has no underlying comparison operations for these.
1836 if (Op.getOperand(0).getValueType() == MVT::v2i64) {
1837 // Equality can be handled by casting to the legal type for Altivec
1838 // comparisons, everything else needs to be expanded.
1839 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
1840 return DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
1841 DAG.getSetCC(dl, MVT::v4i32,
1842 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0)),
1843 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(1)),
1850 // We handle most of these in the usual way.
1854 // If we're comparing for equality to zero, expose the fact that this is
1855 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1856 // fold the new nodes.
1857 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1858 if (C->isNullValue() && CC == ISD::SETEQ) {
1859 EVT VT = Op.getOperand(0).getValueType();
1860 SDValue Zext = Op.getOperand(0);
1861 if (VT.bitsLT(MVT::i32)) {
1863 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1865 unsigned Log2b = Log2_32(VT.getSizeInBits());
1866 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1867 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1868 DAG.getConstant(Log2b, MVT::i32));
1869 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1871 // Leave comparisons against 0 and -1 alone for now, since they're usually
1872 // optimized. FIXME: revisit this when we can custom lower all setcc
1874 if (C->isAllOnesValue() || C->isNullValue())
1878 // If we have an integer seteq/setne, turn it into a compare against zero
1879 // by xor'ing the rhs with the lhs, which is faster than setting a
1880 // condition register, reading it back out, and masking the correct bit. The
1881 // normal approach here uses sub to do this instead of xor. Using xor exposes
1882 // the result to other bit-twiddling opportunities.
1883 EVT LHSVT = Op.getOperand(0).getValueType();
1884 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1885 EVT VT = Op.getValueType();
1886 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1888 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1893 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1894 const PPCSubtarget &Subtarget) const {
1895 SDNode *Node = Op.getNode();
1896 EVT VT = Node->getValueType(0);
1897 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1898 SDValue InChain = Node->getOperand(0);
1899 SDValue VAListPtr = Node->getOperand(1);
1900 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1903 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1906 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1907 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1908 false, false, false, 0);
1909 InChain = GprIndex.getValue(1);
1911 if (VT == MVT::i64) {
1912 // Check if GprIndex is even
1913 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1914 DAG.getConstant(1, MVT::i32));
1915 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1916 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1917 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1918 DAG.getConstant(1, MVT::i32));
1919 // Align GprIndex to be even if it isn't
1920 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1924 // fpr index is 1 byte after gpr
1925 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1926 DAG.getConstant(1, MVT::i32));
1929 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1930 FprPtr, MachinePointerInfo(SV), MVT::i8,
1931 false, false, false, 0);
1932 InChain = FprIndex.getValue(1);
1934 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1935 DAG.getConstant(8, MVT::i32));
1937 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1938 DAG.getConstant(4, MVT::i32));
1941 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1942 MachinePointerInfo(), false, false,
1944 InChain = OverflowArea.getValue(1);
1946 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1947 MachinePointerInfo(), false, false,
1949 InChain = RegSaveArea.getValue(1);
1951 // select overflow_area if index > 8
1952 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1953 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1955 // adjustment constant gpr_index * 4/8
1956 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1957 VT.isInteger() ? GprIndex : FprIndex,
1958 DAG.getConstant(VT.isInteger() ? 4 : 8,
1961 // OurReg = RegSaveArea + RegConstant
1962 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1965 // Floating types are 32 bytes into RegSaveArea
1966 if (VT.isFloatingPoint())
1967 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1968 DAG.getConstant(32, MVT::i32));
1970 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1971 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1972 VT.isInteger() ? GprIndex : FprIndex,
1973 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1976 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1977 VT.isInteger() ? VAListPtr : FprPtr,
1978 MachinePointerInfo(SV),
1979 MVT::i8, false, false, 0);
1981 // determine if we should load from reg_save_area or overflow_area
1982 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1984 // increase overflow_area by 4/8 if gpr/fpr > 8
1985 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1986 DAG.getConstant(VT.isInteger() ? 4 : 8,
1989 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1992 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1994 MachinePointerInfo(),
1995 MVT::i32, false, false, 0);
1997 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1998 false, false, false, 0);
2001 SDValue PPCTargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG,
2002 const PPCSubtarget &Subtarget) const {
2003 assert(!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only");
2005 // We have to copy the entire va_list struct:
2006 // 2*sizeof(char) + 2 Byte alignment + 2*sizeof(char*) = 12 Byte
2007 return DAG.getMemcpy(Op.getOperand(0), Op,
2008 Op.getOperand(1), Op.getOperand(2),
2009 DAG.getConstant(12, MVT::i32), 8, false, true,
2010 MachinePointerInfo(), MachinePointerInfo());
2013 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
2014 SelectionDAG &DAG) const {
2015 return Op.getOperand(0);
2018 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
2019 SelectionDAG &DAG) const {
2020 SDValue Chain = Op.getOperand(0);
2021 SDValue Trmp = Op.getOperand(1); // trampoline
2022 SDValue FPtr = Op.getOperand(2); // nested function
2023 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
2026 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2027 bool isPPC64 = (PtrVT == MVT::i64);
2029 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
2032 TargetLowering::ArgListTy Args;
2033 TargetLowering::ArgListEntry Entry;
2035 Entry.Ty = IntPtrTy;
2036 Entry.Node = Trmp; Args.push_back(Entry);
2038 // TrampSize == (isPPC64 ? 48 : 40);
2039 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
2040 isPPC64 ? MVT::i64 : MVT::i32);
2041 Args.push_back(Entry);
2043 Entry.Node = FPtr; Args.push_back(Entry);
2044 Entry.Node = Nest; Args.push_back(Entry);
2046 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
2047 TargetLowering::CallLoweringInfo CLI(DAG);
2048 CLI.setDebugLoc(dl).setChain(Chain)
2049 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2050 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
2051 std::move(Args), 0);
2053 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2054 return CallResult.second;
2057 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
2058 const PPCSubtarget &Subtarget) const {
2059 MachineFunction &MF = DAG.getMachineFunction();
2060 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2064 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
2065 // vastart just stores the address of the VarArgsFrameIndex slot into the
2066 // memory location argument.
2067 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2068 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2069 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2070 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2071 MachinePointerInfo(SV),
2075 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
2076 // We suppose the given va_list is already allocated.
2079 // char gpr; /* index into the array of 8 GPRs
2080 // * stored in the register save area
2081 // * gpr=0 corresponds to r3,
2082 // * gpr=1 to r4, etc.
2084 // char fpr; /* index into the array of 8 FPRs
2085 // * stored in the register save area
2086 // * fpr=0 corresponds to f1,
2087 // * fpr=1 to f2, etc.
2089 // char *overflow_arg_area;
2090 // /* location on stack that holds
2091 // * the next overflow argument
2093 // char *reg_save_area;
2094 // /* where r3:r10 and f1:f8 (if saved)
2100 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
2101 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
2104 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2106 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
2108 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
2111 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
2112 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
2114 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
2115 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
2117 uint64_t FPROffset = 1;
2118 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
2120 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2122 // Store first byte : number of int regs
2123 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
2125 MachinePointerInfo(SV),
2126 MVT::i8, false, false, 0);
2127 uint64_t nextOffset = FPROffset;
2128 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
2131 // Store second byte : number of float regs
2132 SDValue secondStore =
2133 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
2134 MachinePointerInfo(SV, nextOffset), MVT::i8,
2136 nextOffset += StackOffset;
2137 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
2139 // Store second word : arguments given on stack
2140 SDValue thirdStore =
2141 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
2142 MachinePointerInfo(SV, nextOffset),
2144 nextOffset += FrameOffset;
2145 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
2147 // Store third word : arguments given in registers
2148 return DAG.getStore(thirdStore, dl, FR, nextPtr,
2149 MachinePointerInfo(SV, nextOffset),
2154 #include "PPCGenCallingConv.inc"
2156 // Function whose sole purpose is to kill compiler warnings
2157 // stemming from unused functions included from PPCGenCallingConv.inc.
2158 CCAssignFn *PPCTargetLowering::useFastISelCCs(unsigned Flag) const {
2159 return Flag ? CC_PPC64_ELF_FIS : RetCC_PPC64_ELF_FIS;
2162 bool llvm::CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
2163 CCValAssign::LocInfo &LocInfo,
2164 ISD::ArgFlagsTy &ArgFlags,
2169 bool llvm::CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
2171 CCValAssign::LocInfo &LocInfo,
2172 ISD::ArgFlagsTy &ArgFlags,
2174 static const MCPhysReg ArgRegs[] = {
2175 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2176 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2178 const unsigned NumArgRegs = array_lengthof(ArgRegs);
2180 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
2182 // Skip one register if the first unallocated register has an even register
2183 // number and there are still argument registers available which have not been
2184 // allocated yet. RegNum is actually an index into ArgRegs, which means we
2185 // need to skip a register if RegNum is odd.
2186 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
2187 State.AllocateReg(ArgRegs[RegNum]);
2190 // Always return false here, as this function only makes sure that the first
2191 // unallocated register has an odd register number and does not actually
2192 // allocate a register for the current argument.
2196 bool llvm::CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
2198 CCValAssign::LocInfo &LocInfo,
2199 ISD::ArgFlagsTy &ArgFlags,
2201 static const MCPhysReg ArgRegs[] = {
2202 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2206 const unsigned NumArgRegs = array_lengthof(ArgRegs);
2208 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
2210 // If there is only one Floating-point register left we need to put both f64
2211 // values of a split ppc_fp128 value on the stack.
2212 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
2213 State.AllocateReg(ArgRegs[RegNum]);
2216 // Always return false here, as this function only makes sure that the two f64
2217 // values a ppc_fp128 value is split into are both passed in registers or both
2218 // passed on the stack and does not actually allocate a register for the
2219 // current argument.
2223 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
2225 static const MCPhysReg *GetFPR() {
2226 static const MCPhysReg FPR[] = {
2227 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2228 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
2234 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
2236 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
2237 unsigned PtrByteSize) {
2238 unsigned ArgSize = ArgVT.getStoreSize();
2239 if (Flags.isByVal())
2240 ArgSize = Flags.getByValSize();
2242 // Round up to multiples of the pointer size, except for array members,
2243 // which are always packed.
2244 if (!Flags.isInConsecutiveRegs())
2245 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2250 /// CalculateStackSlotAlignment - Calculates the alignment of this argument
2252 static unsigned CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT,
2253 ISD::ArgFlagsTy Flags,
2254 unsigned PtrByteSize) {
2255 unsigned Align = PtrByteSize;
2257 // Altivec parameters are padded to a 16 byte boundary.
2258 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
2259 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
2260 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64)
2263 // ByVal parameters are aligned as requested.
2264 if (Flags.isByVal()) {
2265 unsigned BVAlign = Flags.getByValAlign();
2266 if (BVAlign > PtrByteSize) {
2267 if (BVAlign % PtrByteSize != 0)
2269 "ByVal alignment is not a multiple of the pointer size");
2275 // Array members are always packed to their original alignment.
2276 if (Flags.isInConsecutiveRegs()) {
2277 // If the array member was split into multiple registers, the first
2278 // needs to be aligned to the size of the full type. (Except for
2279 // ppcf128, which is only aligned as its f64 components.)
2280 if (Flags.isSplit() && OrigVT != MVT::ppcf128)
2281 Align = OrigVT.getStoreSize();
2283 Align = ArgVT.getStoreSize();
2289 /// CalculateStackSlotUsed - Return whether this argument will use its
2290 /// stack slot (instead of being passed in registers). ArgOffset,
2291 /// AvailableFPRs, and AvailableVRs must hold the current argument
2292 /// position, and will be updated to account for this argument.
2293 static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT,
2294 ISD::ArgFlagsTy Flags,
2295 unsigned PtrByteSize,
2296 unsigned LinkageSize,
2297 unsigned ParamAreaSize,
2298 unsigned &ArgOffset,
2299 unsigned &AvailableFPRs,
2300 unsigned &AvailableVRs) {
2301 bool UseMemory = false;
2303 // Respect alignment of argument on the stack.
2305 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
2306 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
2307 // If there's no space left in the argument save area, we must
2308 // use memory (this check also catches zero-sized arguments).
2309 if (ArgOffset >= LinkageSize + ParamAreaSize)
2312 // Allocate argument on the stack.
2313 ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2314 if (Flags.isInConsecutiveRegsLast())
2315 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2316 // If we overran the argument save area, we must use memory
2317 // (this check catches arguments passed partially in memory)
2318 if (ArgOffset > LinkageSize + ParamAreaSize)
2321 // However, if the argument is actually passed in an FPR or a VR,
2322 // we don't use memory after all.
2323 if (!Flags.isByVal()) {
2324 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
2325 if (AvailableFPRs > 0) {
2329 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
2330 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
2331 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64)
2332 if (AvailableVRs > 0) {
2341 /// EnsureStackAlignment - Round stack frame size up from NumBytes to
2342 /// ensure minimum alignment required for target.
2343 static unsigned EnsureStackAlignment(const TargetMachine &Target,
2344 unsigned NumBytes) {
2345 unsigned TargetAlign =
2346 Target.getSubtargetImpl()->getFrameLowering()->getStackAlignment();
2347 unsigned AlignMask = TargetAlign - 1;
2348 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2353 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
2354 CallingConv::ID CallConv, bool isVarArg,
2355 const SmallVectorImpl<ISD::InputArg>
2357 SDLoc dl, SelectionDAG &DAG,
2358 SmallVectorImpl<SDValue> &InVals)
2360 if (Subtarget.isSVR4ABI()) {
2361 if (Subtarget.isPPC64())
2362 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
2365 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
2368 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
2374 PPCTargetLowering::LowerFormalArguments_32SVR4(
2376 CallingConv::ID CallConv, bool isVarArg,
2377 const SmallVectorImpl<ISD::InputArg>
2379 SDLoc dl, SelectionDAG &DAG,
2380 SmallVectorImpl<SDValue> &InVals) const {
2382 // 32-bit SVR4 ABI Stack Frame Layout:
2383 // +-----------------------------------+
2384 // +--> | Back chain |
2385 // | +-----------------------------------+
2386 // | | Floating-point register save area |
2387 // | +-----------------------------------+
2388 // | | General register save area |
2389 // | +-----------------------------------+
2390 // | | CR save word |
2391 // | +-----------------------------------+
2392 // | | VRSAVE save word |
2393 // | +-----------------------------------+
2394 // | | Alignment padding |
2395 // | +-----------------------------------+
2396 // | | Vector register save area |
2397 // | +-----------------------------------+
2398 // | | Local variable space |
2399 // | +-----------------------------------+
2400 // | | Parameter list area |
2401 // | +-----------------------------------+
2402 // | | LR save word |
2403 // | +-----------------------------------+
2404 // SP--> +--- | Back chain |
2405 // +-----------------------------------+
2408 // System V Application Binary Interface PowerPC Processor Supplement
2409 // AltiVec Technology Programming Interface Manual
2411 MachineFunction &MF = DAG.getMachineFunction();
2412 MachineFrameInfo *MFI = MF.getFrameInfo();
2413 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2415 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2416 // Potential tail calls could cause overwriting of argument stack slots.
2417 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2418 (CallConv == CallingConv::Fast));
2419 unsigned PtrByteSize = 4;
2421 // Assign locations to all of the incoming arguments.
2422 SmallVector<CCValAssign, 16> ArgLocs;
2423 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2426 // Reserve space for the linkage area on the stack.
2427 unsigned LinkageSize = PPCFrameLowering::getLinkageSize(false, false, false);
2428 CCInfo.AllocateStack(LinkageSize, PtrByteSize);
2430 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
2432 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2433 CCValAssign &VA = ArgLocs[i];
2435 // Arguments stored in registers.
2436 if (VA.isRegLoc()) {
2437 const TargetRegisterClass *RC;
2438 EVT ValVT = VA.getValVT();
2440 switch (ValVT.getSimpleVT().SimpleTy) {
2442 llvm_unreachable("ValVT not supported by formal arguments Lowering");
2445 RC = &PPC::GPRCRegClass;
2448 RC = &PPC::F4RCRegClass;
2451 if (Subtarget.hasVSX())
2452 RC = &PPC::VSFRCRegClass;
2454 RC = &PPC::F8RCRegClass;
2460 RC = &PPC::VRRCRegClass;
2464 RC = &PPC::VSHRCRegClass;
2468 // Transform the arguments stored in physical registers into virtual ones.
2469 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2470 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
2471 ValVT == MVT::i1 ? MVT::i32 : ValVT);
2473 if (ValVT == MVT::i1)
2474 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgValue);
2476 InVals.push_back(ArgValue);
2478 // Argument stored in memory.
2479 assert(VA.isMemLoc());
2481 unsigned ArgSize = VA.getLocVT().getStoreSize();
2482 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2485 // Create load nodes to retrieve arguments from the stack.
2486 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2487 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2488 MachinePointerInfo(),
2489 false, false, false, 0));
2493 // Assign locations to all of the incoming aggregate by value arguments.
2494 // Aggregates passed by value are stored in the local variable space of the
2495 // caller's stack frame, right above the parameter list area.
2496 SmallVector<CCValAssign, 16> ByValArgLocs;
2497 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2498 ByValArgLocs, *DAG.getContext());
2500 // Reserve stack space for the allocations in CCInfo.
2501 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2503 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
2505 // Area that is at least reserved in the caller of this function.
2506 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
2507 MinReservedArea = std::max(MinReservedArea, LinkageSize);
2509 // Set the size that is at least reserved in caller of this function. Tail
2510 // call optimized function's reserved stack space needs to be aligned so that
2511 // taking the difference between two stack areas will result in an aligned
2513 MinReservedArea = EnsureStackAlignment(MF.getTarget(), MinReservedArea);
2514 FuncInfo->setMinReservedArea(MinReservedArea);
2516 SmallVector<SDValue, 8> MemOps;
2518 // If the function takes variable number of arguments, make a frame index for
2519 // the start of the first vararg value... for expansion of llvm.va_start.
2521 static const MCPhysReg GPArgRegs[] = {
2522 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2523 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2525 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
2527 static const MCPhysReg FPArgRegs[] = {
2528 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2531 unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
2532 if (DisablePPCFloatInVariadic)
2535 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
2537 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
2540 // Make room for NumGPArgRegs and NumFPArgRegs.
2541 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
2542 NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8;
2544 FuncInfo->setVarArgsStackOffset(
2545 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2546 CCInfo.getNextStackOffset(), true));
2548 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
2549 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2551 // The fixed integer arguments of a variadic function are stored to the
2552 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
2553 // the result of va_next.
2554 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
2555 // Get an existing live-in vreg, or add a new one.
2556 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
2558 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
2560 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2561 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2562 MachinePointerInfo(), false, false, 0);
2563 MemOps.push_back(Store);
2564 // Increment the address by four for the next argument to store
2565 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2566 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2569 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
2571 // The double arguments are stored to the VarArgsFrameIndex
2573 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
2574 // Get an existing live-in vreg, or add a new one.
2575 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
2577 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
2579 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
2580 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2581 MachinePointerInfo(), false, false, 0);
2582 MemOps.push_back(Store);
2583 // Increment the address by eight for the next argument to store
2584 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
2586 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2590 if (!MemOps.empty())
2591 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2596 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2597 // value to MVT::i64 and then truncate to the correct register size.
2599 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2600 SelectionDAG &DAG, SDValue ArgVal,
2603 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2604 DAG.getValueType(ObjectVT));
2605 else if (Flags.isZExt())
2606 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2607 DAG.getValueType(ObjectVT));
2609 return DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, ArgVal);
2613 PPCTargetLowering::LowerFormalArguments_64SVR4(
2615 CallingConv::ID CallConv, bool isVarArg,
2616 const SmallVectorImpl<ISD::InputArg>
2618 SDLoc dl, SelectionDAG &DAG,
2619 SmallVectorImpl<SDValue> &InVals) const {
2620 // TODO: add description of PPC stack frame format, or at least some docs.
2622 bool isELFv2ABI = Subtarget.isELFv2ABI();
2623 bool isLittleEndian = Subtarget.isLittleEndian();
2624 MachineFunction &MF = DAG.getMachineFunction();
2625 MachineFrameInfo *MFI = MF.getFrameInfo();
2626 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2628 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2629 // Potential tail calls could cause overwriting of argument stack slots.
2630 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2631 (CallConv == CallingConv::Fast));
2632 unsigned PtrByteSize = 8;
2634 unsigned LinkageSize = PPCFrameLowering::getLinkageSize(true, false,
2637 static const MCPhysReg GPR[] = {
2638 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2639 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2642 static const MCPhysReg *FPR = GetFPR();
2644 static const MCPhysReg VR[] = {
2645 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2646 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2648 static const MCPhysReg VSRH[] = {
2649 PPC::VSH2, PPC::VSH3, PPC::VSH4, PPC::VSH5, PPC::VSH6, PPC::VSH7, PPC::VSH8,
2650 PPC::VSH9, PPC::VSH10, PPC::VSH11, PPC::VSH12, PPC::VSH13
2653 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2654 const unsigned Num_FPR_Regs = 13;
2655 const unsigned Num_VR_Regs = array_lengthof(VR);
2657 // Do a first pass over the arguments to determine whether the ABI
2658 // guarantees that our caller has allocated the parameter save area
2659 // on its stack frame. In the ELFv1 ABI, this is always the case;
2660 // in the ELFv2 ABI, it is true if this is a vararg function or if
2661 // any parameter is located in a stack slot.
2663 bool HasParameterArea = !isELFv2ABI || isVarArg;
2664 unsigned ParamAreaSize = Num_GPR_Regs * PtrByteSize;
2665 unsigned NumBytes = LinkageSize;
2666 unsigned AvailableFPRs = Num_FPR_Regs;
2667 unsigned AvailableVRs = Num_VR_Regs;
2668 for (unsigned i = 0, e = Ins.size(); i != e; ++i)
2669 if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags,
2670 PtrByteSize, LinkageSize, ParamAreaSize,
2671 NumBytes, AvailableFPRs, AvailableVRs))
2672 HasParameterArea = true;
2674 // Add DAG nodes to load the arguments or copy them out of registers. On
2675 // entry to a function on PPC, the arguments start after the linkage area,
2676 // although the first ones are often in registers.
2678 unsigned ArgOffset = LinkageSize;
2679 unsigned GPR_idx, FPR_idx = 0, VR_idx = 0;
2680 SmallVector<SDValue, 8> MemOps;
2681 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2682 unsigned CurArgIdx = 0;
2683 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2685 bool needsLoad = false;
2686 EVT ObjectVT = Ins[ArgNo].VT;
2687 EVT OrigVT = Ins[ArgNo].ArgVT;
2688 unsigned ObjSize = ObjectVT.getStoreSize();
2689 unsigned ArgSize = ObjSize;
2690 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2691 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
2692 CurArgIdx = Ins[ArgNo].OrigArgIndex;
2694 /* Respect alignment of argument on the stack. */
2696 CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize);
2697 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
2698 unsigned CurArgOffset = ArgOffset;
2700 /* Compute GPR index associated with argument offset. */
2701 GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
2702 GPR_idx = std::min(GPR_idx, Num_GPR_Regs);
2704 // FIXME the codegen can be much improved in some cases.
2705 // We do not have to keep everything in memory.
2706 if (Flags.isByVal()) {
2707 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2708 ObjSize = Flags.getByValSize();
2709 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2710 // Empty aggregate parameters do not take up registers. Examples:
2714 // etc. However, we have to provide a place-holder in InVals, so
2715 // pretend we have an 8-byte item at the current address for that
2718 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2719 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2720 InVals.push_back(FIN);
2724 // Create a stack object covering all stack doublewords occupied
2725 // by the argument. If the argument is (fully or partially) on
2726 // the stack, or if the argument is fully in registers but the
2727 // caller has allocated the parameter save anyway, we can refer
2728 // directly to the caller's stack frame. Otherwise, create a
2729 // local copy in our own frame.
2731 if (HasParameterArea ||
2732 ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize)
2733 FI = MFI->CreateFixedObject(ArgSize, ArgOffset, false, true);
2735 FI = MFI->CreateStackObject(ArgSize, Align, false);
2736 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2738 // Handle aggregates smaller than 8 bytes.
2739 if (ObjSize < PtrByteSize) {
2740 // The value of the object is its address, which differs from the
2741 // address of the enclosing doubleword on big-endian systems.
2743 if (!isLittleEndian) {
2744 SDValue ArgOff = DAG.getConstant(PtrByteSize - ObjSize, PtrVT);
2745 Arg = DAG.getNode(ISD::ADD, dl, ArgOff.getValueType(), Arg, ArgOff);
2747 InVals.push_back(Arg);
2749 if (GPR_idx != Num_GPR_Regs) {
2750 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2751 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2754 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2755 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2756 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2757 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, Arg,
2758 MachinePointerInfo(FuncArg),
2759 ObjType, false, false, 0);
2761 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2762 // store the whole register as-is to the parameter save area
2764 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2765 MachinePointerInfo(FuncArg),
2769 MemOps.push_back(Store);
2771 // Whether we copied from a register or not, advance the offset
2772 // into the parameter save area by a full doubleword.
2773 ArgOffset += PtrByteSize;
2777 // The value of the object is its address, which is the address of
2778 // its first stack doubleword.
2779 InVals.push_back(FIN);
2781 // Store whatever pieces of the object are in registers to memory.
2782 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2783 if (GPR_idx == Num_GPR_Regs)
2786 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2787 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2790 SDValue Off = DAG.getConstant(j, PtrVT);
2791 Addr = DAG.getNode(ISD::ADD, dl, Off.getValueType(), Addr, Off);
2793 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, Addr,
2794 MachinePointerInfo(FuncArg, j),
2796 MemOps.push_back(Store);
2799 ArgOffset += ArgSize;
2803 switch (ObjectVT.getSimpleVT().SimpleTy) {
2804 default: llvm_unreachable("Unhandled argument type!");
2808 // These can be scalar arguments or elements of an integer array type
2809 // passed directly. Clang may use those instead of "byval" aggregate
2810 // types to avoid forcing arguments to memory unnecessarily.
2811 if (GPR_idx != Num_GPR_Regs) {
2812 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2813 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2815 if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
2816 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2817 // value to MVT::i64 and then truncate to the correct register size.
2818 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2821 ArgSize = PtrByteSize;
2828 // These can be scalar arguments or elements of a float array type
2829 // passed directly. The latter are used to implement ELFv2 homogenous
2830 // float aggregates.
2831 if (FPR_idx != Num_FPR_Regs) {
2834 if (ObjectVT == MVT::f32)
2835 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2837 VReg = MF.addLiveIn(FPR[FPR_idx], Subtarget.hasVSX() ?
2838 &PPC::VSFRCRegClass :
2839 &PPC::F8RCRegClass);
2841 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2843 } else if (GPR_idx != Num_GPR_Regs) {
2844 // This can only ever happen in the presence of f32 array types,
2845 // since otherwise we never run out of FPRs before running out
2847 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2848 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2850 if (ObjectVT == MVT::f32) {
2851 if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0))
2852 ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal,
2853 DAG.getConstant(32, MVT::i32));
2854 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2857 ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal);
2862 // When passing an array of floats, the array occupies consecutive
2863 // space in the argument area; only round up to the next doubleword
2864 // at the end of the array. Otherwise, each float takes 8 bytes.
2865 ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize;
2866 ArgOffset += ArgSize;
2867 if (Flags.isInConsecutiveRegsLast())
2868 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2876 // These can be scalar arguments or elements of a vector array type
2877 // passed directly. The latter are used to implement ELFv2 homogenous
2878 // vector aggregates.
2879 if (VR_idx != Num_VR_Regs) {
2880 unsigned VReg = (ObjectVT == MVT::v2f64 || ObjectVT == MVT::v2i64) ?
2881 MF.addLiveIn(VSRH[VR_idx], &PPC::VSHRCRegClass) :
2882 MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2883 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2892 // We need to load the argument to a virtual register if we determined
2893 // above that we ran out of physical registers of the appropriate type.
2895 if (ObjSize < ArgSize && !isLittleEndian)
2896 CurArgOffset += ArgSize - ObjSize;
2897 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, isImmutable);
2898 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2899 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2900 false, false, false, 0);
2903 InVals.push_back(ArgVal);
2906 // Area that is at least reserved in the caller of this function.
2907 unsigned MinReservedArea;
2908 if (HasParameterArea)
2909 MinReservedArea = std::max(ArgOffset, LinkageSize + 8 * PtrByteSize);
2911 MinReservedArea = LinkageSize;
2913 // Set the size that is at least reserved in caller of this function. Tail
2914 // call optimized functions' reserved stack space needs to be aligned so that
2915 // taking the difference between two stack areas will result in an aligned
2917 MinReservedArea = EnsureStackAlignment(MF.getTarget(), MinReservedArea);
2918 FuncInfo->setMinReservedArea(MinReservedArea);
2920 // If the function takes variable number of arguments, make a frame index for
2921 // the start of the first vararg value... for expansion of llvm.va_start.
2923 int Depth = ArgOffset;
2925 FuncInfo->setVarArgsFrameIndex(
2926 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2927 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2929 // If this function is vararg, store any remaining integer argument regs
2930 // to their spots on the stack so that they may be loaded by deferencing the
2931 // result of va_next.
2932 for (GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
2933 GPR_idx < Num_GPR_Regs; ++GPR_idx) {
2934 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2935 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2936 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2937 MachinePointerInfo(), false, false, 0);
2938 MemOps.push_back(Store);
2939 // Increment the address by four for the next argument to store
2940 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2941 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2945 if (!MemOps.empty())
2946 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2952 PPCTargetLowering::LowerFormalArguments_Darwin(
2954 CallingConv::ID CallConv, bool isVarArg,
2955 const SmallVectorImpl<ISD::InputArg>
2957 SDLoc dl, SelectionDAG &DAG,
2958 SmallVectorImpl<SDValue> &InVals) const {
2959 // TODO: add description of PPC stack frame format, or at least some docs.
2961 MachineFunction &MF = DAG.getMachineFunction();
2962 MachineFrameInfo *MFI = MF.getFrameInfo();
2963 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2965 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2966 bool isPPC64 = PtrVT == MVT::i64;
2967 // Potential tail calls could cause overwriting of argument stack slots.
2968 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2969 (CallConv == CallingConv::Fast));
2970 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2972 unsigned LinkageSize = PPCFrameLowering::getLinkageSize(isPPC64, true,
2974 unsigned ArgOffset = LinkageSize;
2975 // Area that is at least reserved in caller of this function.
2976 unsigned MinReservedArea = ArgOffset;
2978 static const MCPhysReg GPR_32[] = { // 32-bit registers.
2979 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2980 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2982 static const MCPhysReg GPR_64[] = { // 64-bit registers.
2983 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2984 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2987 static const MCPhysReg *FPR = GetFPR();
2989 static const MCPhysReg VR[] = {
2990 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2991 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2994 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2995 const unsigned Num_FPR_Regs = 13;
2996 const unsigned Num_VR_Regs = array_lengthof( VR);
2998 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3000 const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32;
3002 // In 32-bit non-varargs functions, the stack space for vectors is after the
3003 // stack space for non-vectors. We do not use this space unless we have
3004 // too many vectors to fit in registers, something that only occurs in
3005 // constructed examples:), but we have to walk the arglist to figure
3006 // that out...for the pathological case, compute VecArgOffset as the
3007 // start of the vector parameter area. Computing VecArgOffset is the
3008 // entire point of the following loop.
3009 unsigned VecArgOffset = ArgOffset;
3010 if (!isVarArg && !isPPC64) {
3011 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
3013 EVT ObjectVT = Ins[ArgNo].VT;
3014 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
3016 if (Flags.isByVal()) {
3017 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
3018 unsigned ObjSize = Flags.getByValSize();
3020 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3021 VecArgOffset += ArgSize;
3025 switch(ObjectVT.getSimpleVT().SimpleTy) {
3026 default: llvm_unreachable("Unhandled argument type!");
3032 case MVT::i64: // PPC64
3034 // FIXME: We are guaranteed to be !isPPC64 at this point.
3035 // Does MVT::i64 apply?
3042 // Nothing to do, we're only looking at Nonvector args here.
3047 // We've found where the vector parameter area in memory is. Skip the
3048 // first 12 parameters; these don't use that memory.
3049 VecArgOffset = ((VecArgOffset+15)/16)*16;
3050 VecArgOffset += 12*16;
3052 // Add DAG nodes to load the arguments or copy them out of registers. On
3053 // entry to a function on PPC, the arguments start after the linkage area,
3054 // although the first ones are often in registers.
3056 SmallVector<SDValue, 8> MemOps;
3057 unsigned nAltivecParamsAtEnd = 0;
3058 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
3059 unsigned CurArgIdx = 0;
3060 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
3062 bool needsLoad = false;
3063 EVT ObjectVT = Ins[ArgNo].VT;
3064 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
3065 unsigned ArgSize = ObjSize;
3066 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
3067 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
3068 CurArgIdx = Ins[ArgNo].OrigArgIndex;
3070 unsigned CurArgOffset = ArgOffset;
3072 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
3073 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
3074 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
3075 if (isVarArg || isPPC64) {
3076 MinReservedArea = ((MinReservedArea+15)/16)*16;
3077 MinReservedArea += CalculateStackSlotSize(ObjectVT,
3080 } else nAltivecParamsAtEnd++;
3082 // Calculate min reserved area.
3083 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
3087 // FIXME the codegen can be much improved in some cases.
3088 // We do not have to keep everything in memory.
3089 if (Flags.isByVal()) {
3090 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
3091 ObjSize = Flags.getByValSize();
3092 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3093 // Objects of size 1 and 2 are right justified, everything else is
3094 // left justified. This means the memory address is adjusted forwards.
3095 if (ObjSize==1 || ObjSize==2) {
3096 CurArgOffset = CurArgOffset + (4 - ObjSize);
3098 // The value of the object is its address.
3099 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, false, true);
3100 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3101 InVals.push_back(FIN);
3102 if (ObjSize==1 || ObjSize==2) {
3103 if (GPR_idx != Num_GPR_Regs) {
3106 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3108 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3109 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3110 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
3111 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
3112 MachinePointerInfo(FuncArg),
3113 ObjType, false, false, 0);
3114 MemOps.push_back(Store);
3118 ArgOffset += PtrByteSize;
3122 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
3123 // Store whatever pieces of the object are in registers
3124 // to memory. ArgOffset will be the address of the beginning
3126 if (GPR_idx != Num_GPR_Regs) {
3129 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3131 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3132 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
3133 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3134 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3135 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
3136 MachinePointerInfo(FuncArg, j),
3138 MemOps.push_back(Store);
3140 ArgOffset += PtrByteSize;
3142 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
3149 switch (ObjectVT.getSimpleVT().SimpleTy) {
3150 default: llvm_unreachable("Unhandled argument type!");
3154 if (GPR_idx != Num_GPR_Regs) {
3155 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3156 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3158 if (ObjectVT == MVT::i1)
3159 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgVal);
3164 ArgSize = PtrByteSize;
3166 // All int arguments reserve stack space in the Darwin ABI.
3167 ArgOffset += PtrByteSize;
3171 case MVT::i64: // PPC64
3172 if (GPR_idx != Num_GPR_Regs) {
3173 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3174 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
3176 if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
3177 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
3178 // value to MVT::i64 and then truncate to the correct register size.
3179 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
3184 ArgSize = PtrByteSize;
3186 // All int arguments reserve stack space in the Darwin ABI.
3192 // Every 4 bytes of argument space consumes one of the GPRs available for
3193 // argument passing.
3194 if (GPR_idx != Num_GPR_Regs) {
3196 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
3199 if (FPR_idx != Num_FPR_Regs) {
3202 if (ObjectVT == MVT::f32)
3203 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
3205 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
3207 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3213 // All FP arguments reserve stack space in the Darwin ABI.
3214 ArgOffset += isPPC64 ? 8 : ObjSize;
3220 // Note that vector arguments in registers don't reserve stack space,
3221 // except in varargs functions.
3222 if (VR_idx != Num_VR_Regs) {
3223 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
3224 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3226 while ((ArgOffset % 16) != 0) {
3227 ArgOffset += PtrByteSize;
3228 if (GPR_idx != Num_GPR_Regs)
3232 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
3236 if (!isVarArg && !isPPC64) {
3237 // Vectors go after all the nonvectors.
3238 CurArgOffset = VecArgOffset;
3241 // Vectors are aligned.
3242 ArgOffset = ((ArgOffset+15)/16)*16;
3243 CurArgOffset = ArgOffset;
3251 // We need to load the argument to a virtual register if we determined above
3252 // that we ran out of physical registers of the appropriate type.
3254 int FI = MFI->CreateFixedObject(ObjSize,
3255 CurArgOffset + (ArgSize - ObjSize),
3257 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3258 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
3259 false, false, false, 0);
3262 InVals.push_back(ArgVal);
3265 // Allow for Altivec parameters at the end, if needed.
3266 if (nAltivecParamsAtEnd) {
3267 MinReservedArea = ((MinReservedArea+15)/16)*16;
3268 MinReservedArea += 16*nAltivecParamsAtEnd;
3271 // Area that is at least reserved in the caller of this function.
3272 MinReservedArea = std::max(MinReservedArea, LinkageSize + 8 * PtrByteSize);
3274 // Set the size that is at least reserved in caller of this function. Tail
3275 // call optimized functions' reserved stack space needs to be aligned so that
3276 // taking the difference between two stack areas will result in an aligned
3278 MinReservedArea = EnsureStackAlignment(MF.getTarget(), MinReservedArea);
3279 FuncInfo->setMinReservedArea(MinReservedArea);
3281 // If the function takes variable number of arguments, make a frame index for
3282 // the start of the first vararg value... for expansion of llvm.va_start.
3284 int Depth = ArgOffset;
3286 FuncInfo->setVarArgsFrameIndex(
3287 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
3289 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3291 // If this function is vararg, store any remaining integer argument regs
3292 // to their spots on the stack so that they may be loaded by deferencing the
3293 // result of va_next.
3294 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
3298 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3300 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3302 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3303 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
3304 MachinePointerInfo(), false, false, 0);
3305 MemOps.push_back(Store);
3306 // Increment the address by four for the next argument to store
3307 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
3308 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
3312 if (!MemOps.empty())
3313 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3318 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
3319 /// adjusted to accommodate the arguments for the tailcall.
3320 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
3321 unsigned ParamSize) {
3323 if (!isTailCall) return 0;
3325 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
3326 unsigned CallerMinReservedArea = FI->getMinReservedArea();
3327 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
3328 // Remember only if the new adjustement is bigger.
3329 if (SPDiff < FI->getTailCallSPDelta())
3330 FI->setTailCallSPDelta(SPDiff);
3335 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3336 /// for tail call optimization. Targets which want to do tail call
3337 /// optimization should implement this function.
3339 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3340 CallingConv::ID CalleeCC,
3342 const SmallVectorImpl<ISD::InputArg> &Ins,
3343 SelectionDAG& DAG) const {
3344 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
3347 // Variable argument functions are not supported.
3351 MachineFunction &MF = DAG.getMachineFunction();
3352 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
3353 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
3354 // Functions containing by val parameters are not supported.
3355 for (unsigned i = 0; i != Ins.size(); i++) {
3356 ISD::ArgFlagsTy Flags = Ins[i].Flags;
3357 if (Flags.isByVal()) return false;
3360 // Non-PIC/GOT tail calls are supported.
3361 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
3364 // At the moment we can only do local tail calls (in same module, hidden
3365 // or protected) if we are generating PIC.
3366 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3367 return G->getGlobal()->hasHiddenVisibility()
3368 || G->getGlobal()->hasProtectedVisibility();
3374 /// isCallCompatibleAddress - Return the immediate to use if the specified
3375 /// 32-bit value is representable in the immediate field of a BxA instruction.
3376 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
3377 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3378 if (!C) return nullptr;
3380 int Addr = C->getZExtValue();
3381 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
3382 SignExtend32<26>(Addr) != Addr)
3383 return nullptr; // Top 6 bits have to be sext of immediate.
3385 return DAG.getConstant((int)C->getZExtValue() >> 2,
3386 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
3391 struct TailCallArgumentInfo {
3396 TailCallArgumentInfo() : FrameIdx(0) {}
3401 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
3403 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
3405 const SmallVectorImpl<TailCallArgumentInfo> &TailCallArgs,
3406 SmallVectorImpl<SDValue> &MemOpChains,
3408 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
3409 SDValue Arg = TailCallArgs[i].Arg;
3410 SDValue FIN = TailCallArgs[i].FrameIdxOp;
3411 int FI = TailCallArgs[i].FrameIdx;
3412 // Store relative to framepointer.
3413 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
3414 MachinePointerInfo::getFixedStack(FI),
3419 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
3420 /// the appropriate stack slot for the tail call optimized function call.
3421 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
3422 MachineFunction &MF,
3431 // Calculate the new stack slot for the return address.
3432 int SlotSize = isPPC64 ? 8 : 4;
3433 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
3435 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3436 NewRetAddrLoc, true);
3437 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
3438 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
3439 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
3440 MachinePointerInfo::getFixedStack(NewRetAddr),
3443 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
3444 // slot as the FP is never overwritten.
3447 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
3448 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
3450 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
3451 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
3452 MachinePointerInfo::getFixedStack(NewFPIdx),
3459 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
3460 /// the position of the argument.
3462 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
3463 SDValue Arg, int SPDiff, unsigned ArgOffset,
3464 SmallVectorImpl<TailCallArgumentInfo>& TailCallArguments) {
3465 int Offset = ArgOffset + SPDiff;
3466 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
3467 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3468 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
3469 SDValue FIN = DAG.getFrameIndex(FI, VT);
3470 TailCallArgumentInfo Info;
3472 Info.FrameIdxOp = FIN;
3474 TailCallArguments.push_back(Info);
3477 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
3478 /// stack slot. Returns the chain as result and the loaded frame pointers in
3479 /// LROpOut/FPOpout. Used when tail calling.
3480 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
3488 // Load the LR and FP stack slot for later adjusting.
3489 EVT VT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
3490 LROpOut = getReturnAddrFrameIndex(DAG);
3491 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
3492 false, false, false, 0);
3493 Chain = SDValue(LROpOut.getNode(), 1);
3495 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
3496 // slot as the FP is never overwritten.
3498 FPOpOut = getFramePointerFrameIndex(DAG);
3499 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
3500 false, false, false, 0);
3501 Chain = SDValue(FPOpOut.getNode(), 1);
3507 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
3508 /// by "Src" to address "Dst" of size "Size". Alignment information is
3509 /// specified by the specific parameter attribute. The copy will be passed as
3510 /// a byval function parameter.
3511 /// Sometimes what we are copying is the end of a larger object, the part that
3512 /// does not fit in registers.
3514 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
3515 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
3517 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
3518 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
3519 false, false, MachinePointerInfo(),
3520 MachinePointerInfo());
3523 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
3526 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
3527 SDValue Arg, SDValue PtrOff, int SPDiff,
3528 unsigned ArgOffset, bool isPPC64, bool isTailCall,
3529 bool isVector, SmallVectorImpl<SDValue> &MemOpChains,
3530 SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments,
3532 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3537 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3539 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3540 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3541 DAG.getConstant(ArgOffset, PtrVT));
3543 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3544 MachinePointerInfo(), false, false, 0));
3545 // Calculate and remember argument location.
3546 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
3551 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
3552 SDLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
3553 SDValue LROp, SDValue FPOp, bool isDarwinABI,
3554 SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments) {
3555 MachineFunction &MF = DAG.getMachineFunction();
3557 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
3558 // might overwrite each other in case of tail call optimization.
3559 SmallVector<SDValue, 8> MemOpChains2;
3560 // Do not flag preceding copytoreg stuff together with the following stuff.
3562 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3564 if (!MemOpChains2.empty())
3565 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3567 // Store the return address to the appropriate stack slot.
3568 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3569 isPPC64, isDarwinABI, dl);
3571 // Emit callseq_end just before tailcall node.
3572 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3573 DAG.getIntPtrConstant(0, true), InFlag, dl);
3574 InFlag = Chain.getValue(1);
3577 // Is this global address that of a function that can be called by name? (as
3578 // opposed to something that must hold a descriptor for an indirect call).
3579 static bool isFunctionGlobalAddress(SDValue Callee) {
3580 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3581 if (Callee.getOpcode() == ISD::GlobalTLSAddress ||
3582 Callee.getOpcode() == ISD::TargetGlobalTLSAddress)
3585 return G->getGlobal()->getType()->getElementType()->isFunctionTy();
3592 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3593 SDValue &Chain, SDLoc dl, int SPDiff, bool isTailCall,
3595 SmallVectorImpl<std::pair<unsigned, SDValue> > &RegsToPass,
3596 SmallVectorImpl<SDValue> &Ops, std::vector<EVT> &NodeTys,
3597 const PPCSubtarget &Subtarget) {
3599 bool isPPC64 = Subtarget.isPPC64();
3600 bool isSVR4ABI = Subtarget.isSVR4ABI();
3601 bool isELFv2ABI = Subtarget.isELFv2ABI();
3603 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3604 NodeTys.push_back(MVT::Other); // Returns a chain
3605 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3607 unsigned CallOpc = PPCISD::CALL;
3609 bool needIndirectCall = true;
3610 if (!isSVR4ABI || !isPPC64)
3611 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3612 // If this is an absolute destination address, use the munged value.
3613 Callee = SDValue(Dest, 0);
3614 needIndirectCall = false;
3617 if (isFunctionGlobalAddress(Callee)) {
3618 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Callee);
3619 // A call to a TLS address is actually an indirect call to a
3620 // thread-specific pointer.
3621 unsigned OpFlags = 0;
3622 if ((DAG.getTarget().getRelocationModel() != Reloc::Static &&
3623 (Subtarget.getTargetTriple().isMacOSX() &&
3624 Subtarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3625 (G->getGlobal()->isDeclaration() ||
3626 G->getGlobal()->isWeakForLinker())) ||
3627 (Subtarget.isTargetELF() && !isPPC64 &&
3628 !G->getGlobal()->hasLocalLinkage() &&
3629 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3630 // PC-relative references to external symbols should go through $stub,
3631 // unless we're building with the leopard linker or later, which
3632 // automatically synthesizes these stubs.
3633 OpFlags = PPCII::MO_PLT_OR_STUB;
3636 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3637 // every direct call is) turn it into a TargetGlobalAddress /
3638 // TargetExternalSymbol node so that legalize doesn't hack it.
3639 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3640 Callee.getValueType(), 0, OpFlags);
3641 needIndirectCall = false;
3644 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3645 unsigned char OpFlags = 0;
3647 if ((DAG.getTarget().getRelocationModel() != Reloc::Static &&
3648 (Subtarget.getTargetTriple().isMacOSX() &&
3649 Subtarget.getTargetTriple().isMacOSXVersionLT(10, 5))) ||
3650 (Subtarget.isTargetELF() && !isPPC64 &&
3651 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3652 // PC-relative references to external symbols should go through $stub,
3653 // unless we're building with the leopard linker or later, which
3654 // automatically synthesizes these stubs.
3655 OpFlags = PPCII::MO_PLT_OR_STUB;
3658 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3660 needIndirectCall = false;
3664 // We'll form an invalid direct call when lowering a patchpoint; the full
3665 // sequence for an indirect call is complicated, and many of the
3666 // instructions introduced might have side effects (and, thus, can't be
3667 // removed later). The call itself will be removed as soon as the
3668 // argument/return lowering is complete, so the fact that it has the wrong
3669 // kind of operands should not really matter.
3670 needIndirectCall = false;
3673 if (needIndirectCall) {
3674 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3675 // to do the call, we can't use PPCISD::CALL.
3676 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3678 if (isSVR4ABI && isPPC64 && !isELFv2ABI) {
3679 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3680 // entry point, but to the function descriptor (the function entry point
3681 // address is part of the function descriptor though).
3682 // The function descriptor is a three doubleword structure with the
3683 // following fields: function entry point, TOC base address and
3684 // environment pointer.
3685 // Thus for a call through a function pointer, the following actions need
3687 // 1. Save the TOC of the caller in the TOC save area of its stack
3688 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3689 // 2. Load the address of the function entry point from the function
3691 // 3. Load the TOC of the callee from the function descriptor into r2.
3692 // 4. Load the environment pointer from the function descriptor into
3694 // 5. Branch to the function entry point address.
3695 // 6. On return of the callee, the TOC of the caller needs to be
3696 // restored (this is done in FinishCall()).
3698 // All those operations are flagged together to ensure that no other
3699 // operations can be scheduled in between. E.g. without flagging the
3700 // operations together, a TOC access in the caller could be scheduled
3701 // between the load of the callee TOC and the branch to the callee, which
3702 // results in the TOC access going through the TOC of the callee instead
3703 // of going through the TOC of the caller, which leads to incorrect code.
3705 // Load the address of the function entry point from the function
3707 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3708 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs,
3709 makeArrayRef(MTCTROps, InFlag.getNode() ? 3 : 2));
3710 Chain = LoadFuncPtr.getValue(1);
3711 InFlag = LoadFuncPtr.getValue(2);
3713 // Load environment pointer into r11.
3714 // Offset of the environment pointer within the function descriptor.
3715 SDValue PtrOff = DAG.getIntPtrConstant(16);
3717 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3718 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3720 Chain = LoadEnvPtr.getValue(1);
3721 InFlag = LoadEnvPtr.getValue(2);
3723 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3725 Chain = EnvVal.getValue(0);
3726 InFlag = EnvVal.getValue(1);
3728 // Load TOC of the callee into r2. We are using a target-specific load
3729 // with r2 hard coded, because the result of a target-independent load
3730 // would never go directly into r2, since r2 is a reserved register (which
3731 // prevents the register allocator from allocating it), resulting in an
3732 // additional register being allocated and an unnecessary move instruction
3734 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3735 SDValue TOCOff = DAG.getIntPtrConstant(8);
3736 SDValue AddTOC = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, TOCOff);
3737 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3739 Chain = LoadTOCPtr.getValue(0);
3740 InFlag = LoadTOCPtr.getValue(1);
3742 MTCTROps[0] = Chain;
3743 MTCTROps[1] = LoadFuncPtr;
3744 MTCTROps[2] = InFlag;
3747 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys,
3748 makeArrayRef(MTCTROps, InFlag.getNode() ? 3 : 2));
3749 InFlag = Chain.getValue(1);
3752 NodeTys.push_back(MVT::Other);
3753 NodeTys.push_back(MVT::Glue);
3754 Ops.push_back(Chain);
3755 CallOpc = PPCISD::BCTRL;
3756 Callee.setNode(nullptr);
3757 // Add use of X11 (holding environment pointer)
3758 if (isSVR4ABI && isPPC64 && !isELFv2ABI)
3759 Ops.push_back(DAG.getRegister(PPC::X11, PtrVT));
3760 // Add CTR register as callee so a bctr can be emitted later.
3762 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3765 // If this is a direct call, pass the chain and the callee.
3766 if (Callee.getNode()) {
3767 Ops.push_back(Chain);
3768 Ops.push_back(Callee);
3770 // If this is a call to __tls_get_addr, find the symbol whose address
3771 // is to be taken and add it to the list. This will be used to
3772 // generate __tls_get_addr(<sym>@tlsgd) or __tls_get_addr(<sym>@tlsld).
3773 // We find the symbol by walking the chain to the CopyFromReg, walking
3774 // back from the CopyFromReg to the ADDI_TLSGD_L or ADDI_TLSLD_L, and
3775 // pulling the symbol from that node.
3776 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
3777 if (!strcmp(S->getSymbol(), "__tls_get_addr")) {
3778 assert(!needIndirectCall && "Indirect call to __tls_get_addr???");
3779 SDNode *AddI = Chain.getNode()->getOperand(2).getNode();
3780 SDValue TGTAddr = AddI->getOperand(1);
3781 assert(TGTAddr.getNode()->getOpcode() == ISD::TargetGlobalTLSAddress &&
3782 "Didn't find target global TLS address where we expected one");
3783 Ops.push_back(TGTAddr);
3784 CallOpc = PPCISD::CALL_TLS;
3787 // If this is a tail call add stack pointer delta.
3789 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3791 // Add argument registers to the end of the list so that they are known live
3793 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3794 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3795 RegsToPass[i].second.getValueType()));
3797 // Direct calls in the ELFv2 ABI need the TOC register live into the call.
3798 if (Callee.getNode() && isELFv2ABI && !IsPatchPoint)
3799 Ops.push_back(DAG.getRegister(PPC::X2, PtrVT));
3805 bool isLocalCall(const SDValue &Callee)
3807 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3808 return !G->getGlobal()->isDeclaration() &&
3809 !G->getGlobal()->isWeakForLinker();
3814 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3815 CallingConv::ID CallConv, bool isVarArg,
3816 const SmallVectorImpl<ISD::InputArg> &Ins,
3817 SDLoc dl, SelectionDAG &DAG,
3818 SmallVectorImpl<SDValue> &InVals) const {
3820 SmallVector<CCValAssign, 16> RVLocs;
3821 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3823 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3825 // Copy all of the result registers out of their specified physreg.
3826 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3827 CCValAssign &VA = RVLocs[i];
3828 assert(VA.isRegLoc() && "Can only return in registers!");
3830 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3831 VA.getLocReg(), VA.getLocVT(), InFlag);
3832 Chain = Val.getValue(1);
3833 InFlag = Val.getValue(2);
3835 switch (VA.getLocInfo()) {
3836 default: llvm_unreachable("Unknown loc info!");
3837 case CCValAssign::Full: break;
3838 case CCValAssign::AExt:
3839 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3841 case CCValAssign::ZExt:
3842 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3843 DAG.getValueType(VA.getValVT()));
3844 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3846 case CCValAssign::SExt:
3847 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3848 DAG.getValueType(VA.getValVT()));
3849 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3853 InVals.push_back(Val);
3860 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, SDLoc dl,
3861 bool isTailCall, bool isVarArg, bool IsPatchPoint,
3863 SmallVector<std::pair<unsigned, SDValue>, 8>
3865 SDValue InFlag, SDValue Chain,
3867 int SPDiff, unsigned NumBytes,
3868 const SmallVectorImpl<ISD::InputArg> &Ins,
3869 SmallVectorImpl<SDValue> &InVals) const {
3871 bool isELFv2ABI = Subtarget.isELFv2ABI();
3872 std::vector<EVT> NodeTys;
3873 SmallVector<SDValue, 8> Ops;
3874 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3875 isTailCall, IsPatchPoint, RegsToPass, Ops,
3876 NodeTys, Subtarget);
3878 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3879 if (isVarArg && Subtarget.isSVR4ABI() && !Subtarget.isPPC64())
3880 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3882 // When performing tail call optimization the callee pops its arguments off
3883 // the stack. Account for this here so these bytes can be pushed back on in
3884 // PPCFrameLowering::eliminateCallFramePseudoInstr.
3885 int BytesCalleePops =
3886 (CallConv == CallingConv::Fast &&
3887 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3889 // Add a register mask operand representing the call-preserved registers.
3890 const TargetRegisterInfo *TRI =
3891 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3892 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3893 assert(Mask && "Missing call preserved mask for calling convention");
3894 Ops.push_back(DAG.getRegisterMask(Mask));
3896 if (InFlag.getNode())
3897 Ops.push_back(InFlag);
3901 assert(((Callee.getOpcode() == ISD::Register &&
3902 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3903 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3904 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3905 isa<ConstantSDNode>(Callee)) &&
3906 "Expecting an global address, external symbol, absolute value or register");
3908 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, Ops);
3911 // Add a NOP immediately after the branch instruction when using the 64-bit
3912 // SVR4 ABI. At link time, if caller and callee are in a different module and
3913 // thus have a different TOC, the call will be replaced with a call to a stub
3914 // function which saves the current TOC, loads the TOC of the callee and
3915 // branches to the callee. The NOP will be replaced with a load instruction
3916 // which restores the TOC of the caller from the TOC save slot of the current
3917 // stack frame. If caller and callee belong to the same module (and have the
3918 // same TOC), the NOP will remain unchanged.
3920 if (!isTailCall && Subtarget.isSVR4ABI()&& Subtarget.isPPC64() &&
3922 if (CallOpc == PPCISD::BCTRL) {
3923 // This is a call through a function pointer.
3924 // Restore the caller TOC from the save area into R2.
3925 // See PrepareCall() for more information about calls through function
3926 // pointers in the 64-bit SVR4 ABI.
3927 // We are using a target-specific load with r2 hard coded, because the
3928 // result of a target-independent load would never go directly into r2,
3929 // since r2 is a reserved register (which prevents the register allocator
3930 // from allocating it), resulting in an additional register being
3931 // allocated and an unnecessary move instruction being generated.
3932 CallOpc = PPCISD::BCTRL_LOAD_TOC;
3934 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3935 SDValue StackPtr = DAG.getRegister(PPC::X1, PtrVT);
3936 unsigned TOCSaveOffset = PPCFrameLowering::getTOCSaveOffset(isELFv2ABI);
3937 SDValue TOCOff = DAG.getIntPtrConstant(TOCSaveOffset);
3938 SDValue AddTOC = DAG.getNode(ISD::ADD, dl, MVT::i64, StackPtr, TOCOff);
3940 // The address needs to go after the chain input but before the flag (or
3941 // any other variadic arguments).
3942 Ops.insert(std::next(Ops.begin()), AddTOC);
3943 } else if ((CallOpc == PPCISD::CALL) &&
3944 (!isLocalCall(Callee) ||
3945 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3946 // Otherwise insert NOP for non-local calls.
3947 CallOpc = PPCISD::CALL_NOP;
3948 } else if (CallOpc == PPCISD::CALL_TLS)
3949 // For 64-bit SVR4, TLS calls are always non-local.
3950 CallOpc = PPCISD::CALL_NOP_TLS;
3953 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
3954 InFlag = Chain.getValue(1);
3956 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3957 DAG.getIntPtrConstant(BytesCalleePops, true),
3960 InFlag = Chain.getValue(1);
3962 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3963 Ins, dl, DAG, InVals);
3967 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3968 SmallVectorImpl<SDValue> &InVals) const {
3969 SelectionDAG &DAG = CLI.DAG;
3971 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3972 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3973 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3974 SDValue Chain = CLI.Chain;
3975 SDValue Callee = CLI.Callee;
3976 bool &isTailCall = CLI.IsTailCall;
3977 CallingConv::ID CallConv = CLI.CallConv;
3978 bool isVarArg = CLI.IsVarArg;
3979 bool IsPatchPoint = CLI.IsPatchPoint;
3982 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3985 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
3986 report_fatal_error("failed to perform tail call elimination on a call "
3987 "site marked musttail");
3989 if (Subtarget.isSVR4ABI()) {
3990 if (Subtarget.isPPC64())
3991 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3992 isTailCall, IsPatchPoint, Outs, OutVals, Ins,
3995 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3996 isTailCall, IsPatchPoint, Outs, OutVals, Ins,
4000 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
4001 isTailCall, IsPatchPoint, Outs, OutVals, Ins,
4006 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
4007 CallingConv::ID CallConv, bool isVarArg,
4008 bool isTailCall, bool IsPatchPoint,
4009 const SmallVectorImpl<ISD::OutputArg> &Outs,
4010 const SmallVectorImpl<SDValue> &OutVals,
4011 const SmallVectorImpl<ISD::InputArg> &Ins,
4012 SDLoc dl, SelectionDAG &DAG,
4013 SmallVectorImpl<SDValue> &InVals) const {
4014 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
4015 // of the 32-bit SVR4 ABI stack frame layout.
4017 assert((CallConv == CallingConv::C ||
4018 CallConv == CallingConv::Fast) && "Unknown calling convention!");
4020 unsigned PtrByteSize = 4;
4022 MachineFunction &MF = DAG.getMachineFunction();
4024 // Mark this function as potentially containing a function that contains a
4025 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4026 // and restoring the callers stack pointer in this functions epilog. This is
4027 // done because by tail calling the called function might overwrite the value
4028 // in this function's (MF) stack pointer stack slot 0(SP).
4029 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4030 CallConv == CallingConv::Fast)
4031 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4033 // Count how many bytes are to be pushed on the stack, including the linkage
4034 // area, parameter list area and the part of the local variable space which
4035 // contains copies of aggregates which are passed by value.
4037 // Assign locations to all of the outgoing arguments.
4038 SmallVector<CCValAssign, 16> ArgLocs;
4039 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
4042 // Reserve space for the linkage area on the stack.
4043 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false, false),
4047 // Handle fixed and variable vector arguments differently.
4048 // Fixed vector arguments go into registers as long as registers are
4049 // available. Variable vector arguments always go into memory.
4050 unsigned NumArgs = Outs.size();
4052 for (unsigned i = 0; i != NumArgs; ++i) {
4053 MVT ArgVT = Outs[i].VT;
4054 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
4057 if (Outs[i].IsFixed) {
4058 Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
4061 Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
4067 errs() << "Call operand #" << i << " has unhandled type "
4068 << EVT(ArgVT).getEVTString() << "\n";
4070 llvm_unreachable(nullptr);
4074 // All arguments are treated the same.
4075 CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
4078 // Assign locations to all of the outgoing aggregate by value arguments.
4079 SmallVector<CCValAssign, 16> ByValArgLocs;
4080 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4081 ByValArgLocs, *DAG.getContext());
4083 // Reserve stack space for the allocations in CCInfo.
4084 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
4086 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
4088 // Size of the linkage area, parameter list area and the part of the local
4089 // space variable where copies of aggregates which are passed by value are
4091 unsigned NumBytes = CCByValInfo.getNextStackOffset();
4093 // Calculate by how many bytes the stack has to be adjusted in case of tail
4094 // call optimization.
4095 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4097 // Adjust the stack pointer for the new arguments...
4098 // These operations are automatically eliminated by the prolog/epilog pass
4099 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
4101 SDValue CallSeqStart = Chain;
4103 // Load the return address and frame pointer so it can be moved somewhere else
4106 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
4109 // Set up a copy of the stack pointer for use loading and storing any
4110 // arguments that may not fit in the registers available for argument
4112 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4114 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4115 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4116 SmallVector<SDValue, 8> MemOpChains;
4118 bool seenFloatArg = false;
4119 // Walk the register/memloc assignments, inserting copies/loads.
4120 for (unsigned i = 0, j = 0, e = ArgLocs.size();
4123 CCValAssign &VA = ArgLocs[i];
4124 SDValue Arg = OutVals[i];
4125 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4127 if (Flags.isByVal()) {
4128 // Argument is an aggregate which is passed by value, thus we need to
4129 // create a copy of it in the local variable space of the current stack
4130 // frame (which is the stack frame of the caller) and pass the address of
4131 // this copy to the callee.
4132 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
4133 CCValAssign &ByValVA = ByValArgLocs[j++];
4134 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
4136 // Memory reserved in the local variable space of the callers stack frame.
4137 unsigned LocMemOffset = ByValVA.getLocMemOffset();
4139 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
4140 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
4142 // Create a copy of the argument in the local area of the current
4144 SDValue MemcpyCall =
4145 CreateCopyOfByValArgument(Arg, PtrOff,
4146 CallSeqStart.getNode()->getOperand(0),
4149 // This must go outside the CALLSEQ_START..END.
4150 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
4151 CallSeqStart.getNode()->getOperand(1),
4153 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
4154 NewCallSeqStart.getNode());
4155 Chain = CallSeqStart = NewCallSeqStart;
4157 // Pass the address of the aggregate copy on the stack either in a
4158 // physical register or in the parameter list area of the current stack
4159 // frame to the callee.
4163 if (VA.isRegLoc()) {
4164 if (Arg.getValueType() == MVT::i1)
4165 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Arg);
4167 seenFloatArg |= VA.getLocVT().isFloatingPoint();
4168 // Put argument in a physical register.
4169 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
4171 // Put argument in the parameter list area of the current stack frame.
4172 assert(VA.isMemLoc());
4173 unsigned LocMemOffset = VA.getLocMemOffset();
4176 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
4177 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
4179 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
4180 MachinePointerInfo(),
4183 // Calculate and remember argument location.
4184 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
4190 if (!MemOpChains.empty())
4191 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
4193 // Build a sequence of copy-to-reg nodes chained together with token chain
4194 // and flag operands which copy the outgoing args into the appropriate regs.
4196 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4197 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4198 RegsToPass[i].second, InFlag);
4199 InFlag = Chain.getValue(1);
4202 // Set CR bit 6 to true if this is a vararg call with floating args passed in
4205 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
4206 SDValue Ops[] = { Chain, InFlag };
4208 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
4209 dl, VTs, makeArrayRef(Ops, InFlag.getNode() ? 2 : 1));
4211 InFlag = Chain.getValue(1);
4215 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
4216 false, TailCallArguments);
4218 return FinishCall(CallConv, dl, isTailCall, isVarArg, IsPatchPoint, DAG,
4219 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4223 // Copy an argument into memory, being careful to do this outside the
4224 // call sequence for the call to which the argument belongs.
4226 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
4227 SDValue CallSeqStart,
4228 ISD::ArgFlagsTy Flags,
4231 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
4232 CallSeqStart.getNode()->getOperand(0),
4234 // The MEMCPY must go outside the CALLSEQ_START..END.
4235 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
4236 CallSeqStart.getNode()->getOperand(1),
4238 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
4239 NewCallSeqStart.getNode());
4240 return NewCallSeqStart;
4244 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
4245 CallingConv::ID CallConv, bool isVarArg,
4246 bool isTailCall, bool IsPatchPoint,
4247 const SmallVectorImpl<ISD::OutputArg> &Outs,
4248 const SmallVectorImpl<SDValue> &OutVals,
4249 const SmallVectorImpl<ISD::InputArg> &Ins,
4250 SDLoc dl, SelectionDAG &DAG,
4251 SmallVectorImpl<SDValue> &InVals) const {
4253 bool isELFv2ABI = Subtarget.isELFv2ABI();
4254 bool isLittleEndian = Subtarget.isLittleEndian();
4255 unsigned NumOps = Outs.size();
4257 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4258 unsigned PtrByteSize = 8;
4260 MachineFunction &MF = DAG.getMachineFunction();
4262 // Mark this function as potentially containing a function that contains a
4263 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4264 // and restoring the callers stack pointer in this functions epilog. This is
4265 // done because by tail calling the called function might overwrite the value
4266 // in this function's (MF) stack pointer stack slot 0(SP).
4267 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4268 CallConv == CallingConv::Fast)
4269 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4271 // Count how many bytes are to be pushed on the stack, including the linkage
4272 // area, and parameter passing area. On ELFv1, the linkage area is 48 bytes
4273 // reserved space for [SP][CR][LR][2 x unused][TOC]; on ELFv2, the linkage
4274 // area is 32 bytes reserved space for [SP][CR][LR][TOC].
4275 unsigned LinkageSize = PPCFrameLowering::getLinkageSize(true, false,
4277 unsigned NumBytes = LinkageSize;
4279 // Add up all the space actually used.
4280 for (unsigned i = 0; i != NumOps; ++i) {
4281 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4282 EVT ArgVT = Outs[i].VT;
4283 EVT OrigVT = Outs[i].ArgVT;
4285 /* Respect alignment of argument on the stack. */
4287 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
4288 NumBytes = ((NumBytes + Align - 1) / Align) * Align;
4290 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
4291 if (Flags.isInConsecutiveRegsLast())
4292 NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
4295 unsigned NumBytesActuallyUsed = NumBytes;
4297 // The prolog code of the callee may store up to 8 GPR argument registers to
4298 // the stack, allowing va_start to index over them in memory if its varargs.
4299 // Because we cannot tell if this is needed on the caller side, we have to
4300 // conservatively assume that it is needed. As such, make sure we have at
4301 // least enough stack space for the caller to store the 8 GPRs.
4302 // FIXME: On ELFv2, it may be unnecessary to allocate the parameter area.
4303 NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize);
4305 // Tail call needs the stack to be aligned.
4306 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4307 CallConv == CallingConv::Fast)
4308 NumBytes = EnsureStackAlignment(MF.getTarget(), NumBytes);
4310 // Calculate by how many bytes the stack has to be adjusted in case of tail
4311 // call optimization.
4312 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4314 // To protect arguments on the stack from being clobbered in a tail call,
4315 // force all the loads to happen before doing any other lowering.
4317 Chain = DAG.getStackArgumentTokenFactor(Chain);
4319 // Adjust the stack pointer for the new arguments...
4320 // These operations are automatically eliminated by the prolog/epilog pass
4321 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
4323 SDValue CallSeqStart = Chain;
4325 // Load the return address and frame pointer so it can be move somewhere else
4328 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4331 // Set up a copy of the stack pointer for use loading and storing any
4332 // arguments that may not fit in the registers available for argument
4334 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4336 // Figure out which arguments are going to go in registers, and which in
4337 // memory. Also, if this is a vararg function, floating point operations
4338 // must be stored to our stack, and loaded into integer regs as well, if
4339 // any integer regs are available for argument passing.
4340 unsigned ArgOffset = LinkageSize;
4341 unsigned GPR_idx, FPR_idx = 0, VR_idx = 0;
4343 static const MCPhysReg GPR[] = {
4344 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4345 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4347 static const MCPhysReg *FPR = GetFPR();
4349 static const MCPhysReg VR[] = {
4350 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4351 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4353 static const MCPhysReg VSRH[] = {
4354 PPC::VSH2, PPC::VSH3, PPC::VSH4, PPC::VSH5, PPC::VSH6, PPC::VSH7, PPC::VSH8,
4355 PPC::VSH9, PPC::VSH10, PPC::VSH11, PPC::VSH12, PPC::VSH13
4358 const unsigned NumGPRs = array_lengthof(GPR);
4359 const unsigned NumFPRs = 13;
4360 const unsigned NumVRs = array_lengthof(VR);
4362 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4363 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4365 SmallVector<SDValue, 8> MemOpChains;
4366 for (unsigned i = 0; i != NumOps; ++i) {
4367 SDValue Arg = OutVals[i];
4368 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4369 EVT ArgVT = Outs[i].VT;
4370 EVT OrigVT = Outs[i].ArgVT;
4372 /* Respect alignment of argument on the stack. */
4374 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
4375 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
4377 /* Compute GPR index associated with argument offset. */
4378 GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
4379 GPR_idx = std::min(GPR_idx, NumGPRs);
4381 // PtrOff will be used to store the current argument to the stack if a
4382 // register cannot be found for it.
4385 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4387 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4389 // Promote integers to 64-bit values.
4390 if (Arg.getValueType() == MVT::i32 || Arg.getValueType() == MVT::i1) {
4391 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4392 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4393 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4396 // FIXME memcpy is used way more than necessary. Correctness first.
4397 // Note: "by value" is code for passing a structure by value, not
4399 if (Flags.isByVal()) {
4400 // Note: Size includes alignment padding, so
4401 // struct x { short a; char b; }
4402 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
4403 // These are the proper values we need for right-justifying the
4404 // aggregate in a parameter register.
4405 unsigned Size = Flags.getByValSize();
4407 // An empty aggregate parameter takes up no storage and no
4412 // All aggregates smaller than 8 bytes must be passed right-justified.
4413 if (Size==1 || Size==2 || Size==4) {
4414 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
4415 if (GPR_idx != NumGPRs) {
4416 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4417 MachinePointerInfo(), VT,
4418 false, false, false, 0);
4419 MemOpChains.push_back(Load.getValue(1));
4420 RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Load));
4422 ArgOffset += PtrByteSize;
4427 if (GPR_idx == NumGPRs && Size < 8) {
4428 SDValue AddPtr = PtrOff;
4429 if (!isLittleEndian) {
4430 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4431 PtrOff.getValueType());
4432 AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4434 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4437 ArgOffset += PtrByteSize;
4440 // Copy entire object into memory. There are cases where gcc-generated
4441 // code assumes it is there, even if it could be put entirely into
4442 // registers. (This is not what the doc says.)
4444 // FIXME: The above statement is likely due to a misunderstanding of the
4445 // documents. All arguments must be copied into the parameter area BY
4446 // THE CALLEE in the event that the callee takes the address of any
4447 // formal argument. That has not yet been implemented. However, it is
4448 // reasonable to use the stack area as a staging area for the register
4451 // Skip this for small aggregates, as we will use the same slot for a
4452 // right-justified copy, below.
4454 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4458 // When a register is available, pass a small aggregate right-justified.
4459 if (Size < 8 && GPR_idx != NumGPRs) {
4460 // The easiest way to get this right-justified in a register
4461 // is to copy the structure into the rightmost portion of a
4462 // local variable slot, then load the whole slot into the
4464 // FIXME: The memcpy seems to produce pretty awful code for
4465 // small aggregates, particularly for packed ones.
4466 // FIXME: It would be preferable to use the slot in the
4467 // parameter save area instead of a new local variable.
4468 SDValue AddPtr = PtrOff;
4469 if (!isLittleEndian) {
4470 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
4471 AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4473 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4477 // Load the slot into the register.
4478 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
4479 MachinePointerInfo(),
4480 false, false, false, 0);
4481 MemOpChains.push_back(Load.getValue(1));
4482 RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Load));
4484 // Done with this argument.
4485 ArgOffset += PtrByteSize;
4489 // For aggregates larger than PtrByteSize, copy the pieces of the
4490 // object that fit into registers from the parameter save area.
4491 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4492 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4493 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4494 if (GPR_idx != NumGPRs) {
4495 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4496 MachinePointerInfo(),
4497 false, false, false, 0);
4498 MemOpChains.push_back(Load.getValue(1));
4499 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4500 ArgOffset += PtrByteSize;
4502 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4509 switch (Arg.getSimpleValueType().SimpleTy) {
4510 default: llvm_unreachable("Unexpected ValueType for argument!");
4514 // These can be scalar arguments or elements of an integer array type
4515 // passed directly. Clang may use those instead of "byval" aggregate
4516 // types to avoid forcing arguments to memory unnecessarily.
4517 if (GPR_idx != NumGPRs) {
4518 RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Arg));
4520 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4521 true, isTailCall, false, MemOpChains,
4522 TailCallArguments, dl);
4524 ArgOffset += PtrByteSize;
4528 // These can be scalar arguments or elements of a float array type
4529 // passed directly. The latter are used to implement ELFv2 homogenous
4530 // float aggregates.
4532 // Named arguments go into FPRs first, and once they overflow, the
4533 // remaining arguments go into GPRs and then the parameter save area.
4534 // Unnamed arguments for vararg functions always go to GPRs and
4535 // then the parameter save area. For now, put all arguments to vararg
4536 // routines always in both locations (FPR *and* GPR or stack slot).
4537 bool NeedGPROrStack = isVarArg || FPR_idx == NumFPRs;
4539 // First load the argument into the next available FPR.
4540 if (FPR_idx != NumFPRs)
4541 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4543 // Next, load the argument into GPR or stack slot if needed.
4544 if (!NeedGPROrStack)
4546 else if (GPR_idx != NumGPRs) {
4547 // In the non-vararg case, this can only ever happen in the
4548 // presence of f32 array types, since otherwise we never run
4549 // out of FPRs before running out of GPRs.
4552 // Double values are always passed in a single GPR.
4553 if (Arg.getValueType() != MVT::f32) {
4554 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
4556 // Non-array float values are extended and passed in a GPR.
4557 } else if (!Flags.isInConsecutiveRegs()) {
4558 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
4559 ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
4561 // If we have an array of floats, we collect every odd element
4562 // together with its predecessor into one GPR.
4563 } else if (ArgOffset % PtrByteSize != 0) {
4565 Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]);
4566 Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
4567 if (!isLittleEndian)
4569 ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4571 // The final element, if even, goes into the first half of a GPR.
4572 } else if (Flags.isInConsecutiveRegsLast()) {
4573 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
4574 ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
4575 if (!isLittleEndian)
4576 ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal,
4577 DAG.getConstant(32, MVT::i32));
4579 // Non-final even elements are skipped; they will be handled
4580 // together the with subsequent argument on the next go-around.
4584 if (ArgVal.getNode())
4585 RegsToPass.push_back(std::make_pair(GPR[GPR_idx], ArgVal));
4587 // Single-precision floating-point values are mapped to the
4588 // second (rightmost) word of the stack doubleword.
4589 if (Arg.getValueType() == MVT::f32 &&
4590 !isLittleEndian && !Flags.isInConsecutiveRegs()) {
4591 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4592 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4595 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4596 true, isTailCall, false, MemOpChains,
4597 TailCallArguments, dl);
4599 // When passing an array of floats, the array occupies consecutive
4600 // space in the argument area; only round up to the next doubleword
4601 // at the end of the array. Otherwise, each float takes 8 bytes.
4602 ArgOffset += (Arg.getValueType() == MVT::f32 &&
4603 Flags.isInConsecutiveRegs()) ? 4 : 8;
4604 if (Flags.isInConsecutiveRegsLast())
4605 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
4614 // These can be scalar arguments or elements of a vector array type
4615 // passed directly. The latter are used to implement ELFv2 homogenous
4616 // vector aggregates.
4618 // For a varargs call, named arguments go into VRs or on the stack as
4619 // usual; unnamed arguments always go to the stack or the corresponding
4620 // GPRs when within range. For now, we always put the value in both
4621 // locations (or even all three).
4623 // We could elide this store in the case where the object fits
4624 // entirely in R registers. Maybe later.
4625 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4626 MachinePointerInfo(), false, false, 0);
4627 MemOpChains.push_back(Store);
4628 if (VR_idx != NumVRs) {
4629 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4630 MachinePointerInfo(),
4631 false, false, false, 0);
4632 MemOpChains.push_back(Load.getValue(1));
4634 unsigned VReg = (Arg.getSimpleValueType() == MVT::v2f64 ||
4635 Arg.getSimpleValueType() == MVT::v2i64) ?
4636 VSRH[VR_idx] : VR[VR_idx];
4639 RegsToPass.push_back(std::make_pair(VReg, Load));
4642 for (unsigned i=0; i<16; i+=PtrByteSize) {
4643 if (GPR_idx == NumGPRs)
4645 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4646 DAG.getConstant(i, PtrVT));
4647 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4648 false, false, false, 0);
4649 MemOpChains.push_back(Load.getValue(1));
4650 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4655 // Non-varargs Altivec params go into VRs or on the stack.
4656 if (VR_idx != NumVRs) {
4657 unsigned VReg = (Arg.getSimpleValueType() == MVT::v2f64 ||
4658 Arg.getSimpleValueType() == MVT::v2i64) ?
4659 VSRH[VR_idx] : VR[VR_idx];
4662 RegsToPass.push_back(std::make_pair(VReg, Arg));
4664 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4665 true, isTailCall, true, MemOpChains,
4666 TailCallArguments, dl);
4673 assert(NumBytesActuallyUsed == ArgOffset);
4674 (void)NumBytesActuallyUsed;
4676 if (!MemOpChains.empty())
4677 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
4679 // Check if this is an indirect call (MTCTR/BCTRL).
4680 // See PrepareCall() for more information about calls through function
4681 // pointers in the 64-bit SVR4 ABI.
4682 if (!isTailCall && !IsPatchPoint &&
4683 !isFunctionGlobalAddress(Callee) &&
4684 !isa<ExternalSymbolSDNode>(Callee)) {
4685 // Load r2 into a virtual register and store it to the TOC save area.
4686 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
4687 // TOC save area offset.
4688 unsigned TOCSaveOffset = PPCFrameLowering::getTOCSaveOffset(isELFv2ABI);
4689 SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset);
4690 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4691 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
4693 // In the ELFv2 ABI, R12 must contain the address of an indirect callee.
4694 // This does not mean the MTCTR instruction must use R12; it's easier
4695 // to model this as an extra parameter, so do that.
4696 if (isELFv2ABI && !IsPatchPoint)
4697 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
4700 // Build a sequence of copy-to-reg nodes chained together with token chain
4701 // and flag operands which copy the outgoing args into the appropriate regs.
4703 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4704 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4705 RegsToPass[i].second, InFlag);
4706 InFlag = Chain.getValue(1);
4710 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
4711 FPOp, true, TailCallArguments);
4713 return FinishCall(CallConv, dl, isTailCall, isVarArg, IsPatchPoint, DAG,
4714 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4719 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
4720 CallingConv::ID CallConv, bool isVarArg,
4721 bool isTailCall, bool IsPatchPoint,
4722 const SmallVectorImpl<ISD::OutputArg> &Outs,
4723 const SmallVectorImpl<SDValue> &OutVals,
4724 const SmallVectorImpl<ISD::InputArg> &Ins,
4725 SDLoc dl, SelectionDAG &DAG,
4726 SmallVectorImpl<SDValue> &InVals) const {
4728 unsigned NumOps = Outs.size();
4730 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4731 bool isPPC64 = PtrVT == MVT::i64;
4732 unsigned PtrByteSize = isPPC64 ? 8 : 4;
4734 MachineFunction &MF = DAG.getMachineFunction();
4736 // Mark this function as potentially containing a function that contains a
4737 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4738 // and restoring the callers stack pointer in this functions epilog. This is
4739 // done because by tail calling the called function might overwrite the value
4740 // in this function's (MF) stack pointer stack slot 0(SP).
4741 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4742 CallConv == CallingConv::Fast)
4743 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4745 // Count how many bytes are to be pushed on the stack, including the linkage
4746 // area, and parameter passing area. We start with 24/48 bytes, which is
4747 // prereserved space for [SP][CR][LR][3 x unused].
4748 unsigned LinkageSize = PPCFrameLowering::getLinkageSize(isPPC64, true,
4750 unsigned NumBytes = LinkageSize;
4752 // Add up all the space actually used.
4753 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
4754 // they all go in registers, but we must reserve stack space for them for
4755 // possible use by the caller. In varargs or 64-bit calls, parameters are
4756 // assigned stack space in order, with padding so Altivec parameters are
4758 unsigned nAltivecParamsAtEnd = 0;
4759 for (unsigned i = 0; i != NumOps; ++i) {
4760 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4761 EVT ArgVT = Outs[i].VT;
4762 // Varargs Altivec parameters are padded to a 16 byte boundary.
4763 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
4764 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
4765 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64) {
4766 if (!isVarArg && !isPPC64) {
4767 // Non-varargs Altivec parameters go after all the non-Altivec
4768 // parameters; handle those later so we know how much padding we need.
4769 nAltivecParamsAtEnd++;
4772 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
4773 NumBytes = ((NumBytes+15)/16)*16;
4775 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
4778 // Allow for Altivec parameters at the end, if needed.
4779 if (nAltivecParamsAtEnd) {
4780 NumBytes = ((NumBytes+15)/16)*16;
4781 NumBytes += 16*nAltivecParamsAtEnd;
4784 // The prolog code of the callee may store up to 8 GPR argument registers to
4785 // the stack, allowing va_start to index over them in memory if its varargs.
4786 // Because we cannot tell if this is needed on the caller side, we have to
4787 // conservatively assume that it is needed. As such, make sure we have at
4788 // least enough stack space for the caller to store the 8 GPRs.
4789 NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize);
4791 // Tail call needs the stack to be aligned.
4792 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4793 CallConv == CallingConv::Fast)
4794 NumBytes = EnsureStackAlignment(MF.getTarget(), NumBytes);
4796 // Calculate by how many bytes the stack has to be adjusted in case of tail
4797 // call optimization.
4798 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4800 // To protect arguments on the stack from being clobbered in a tail call,
4801 // force all the loads to happen before doing any other lowering.
4803 Chain = DAG.getStackArgumentTokenFactor(Chain);
4805 // Adjust the stack pointer for the new arguments...
4806 // These operations are automatically eliminated by the prolog/epilog pass
4807 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
4809 SDValue CallSeqStart = Chain;
4811 // Load the return address and frame pointer so it can be move somewhere else
4814 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4817 // Set up a copy of the stack pointer for use loading and storing any
4818 // arguments that may not fit in the registers available for argument
4822 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4824 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4826 // Figure out which arguments are going to go in registers, and which in
4827 // memory. Also, if this is a vararg function, floating point operations
4828 // must be stored to our stack, and loaded into integer regs as well, if
4829 // any integer regs are available for argument passing.
4830 unsigned ArgOffset = LinkageSize;
4831 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4833 static const MCPhysReg GPR_32[] = { // 32-bit registers.
4834 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4835 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4837 static const MCPhysReg GPR_64[] = { // 64-bit registers.
4838 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4839 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4841 static const MCPhysReg *FPR = GetFPR();
4843 static const MCPhysReg VR[] = {
4844 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4845 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4847 const unsigned NumGPRs = array_lengthof(GPR_32);
4848 const unsigned NumFPRs = 13;
4849 const unsigned NumVRs = array_lengthof(VR);
4851 const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32;
4853 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4854 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4856 SmallVector<SDValue, 8> MemOpChains;
4857 for (unsigned i = 0; i != NumOps; ++i) {
4858 SDValue Arg = OutVals[i];
4859 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4861 // PtrOff will be used to store the current argument to the stack if a
4862 // register cannot be found for it.
4865 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4867 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4869 // On PPC64, promote integers to 64-bit values.
4870 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4871 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4872 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4873 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4876 // FIXME memcpy is used way more than necessary. Correctness first.
4877 // Note: "by value" is code for passing a structure by value, not
4879 if (Flags.isByVal()) {
4880 unsigned Size = Flags.getByValSize();
4881 // Very small objects are passed right-justified. Everything else is
4882 // passed left-justified.
4883 if (Size==1 || Size==2) {
4884 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4885 if (GPR_idx != NumGPRs) {
4886 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4887 MachinePointerInfo(), VT,
4888 false, false, false, 0);
4889 MemOpChains.push_back(Load.getValue(1));
4890 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4892 ArgOffset += PtrByteSize;
4894 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4895 PtrOff.getValueType());
4896 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4897 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4900 ArgOffset += PtrByteSize;
4904 // Copy entire object into memory. There are cases where gcc-generated
4905 // code assumes it is there, even if it could be put entirely into
4906 // registers. (This is not what the doc says.)
4907 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4911 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4912 // copy the pieces of the object that fit into registers from the
4913 // parameter save area.
4914 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4915 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4916 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4917 if (GPR_idx != NumGPRs) {
4918 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4919 MachinePointerInfo(),
4920 false, false, false, 0);
4921 MemOpChains.push_back(Load.getValue(1));
4922 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4923 ArgOffset += PtrByteSize;
4925 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4932 switch (Arg.getSimpleValueType().SimpleTy) {
4933 default: llvm_unreachable("Unexpected ValueType for argument!");
4937 if (GPR_idx != NumGPRs) {
4938 if (Arg.getValueType() == MVT::i1)
4939 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, PtrVT, Arg);
4941 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4943 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4944 isPPC64, isTailCall, false, MemOpChains,
4945 TailCallArguments, dl);
4947 ArgOffset += PtrByteSize;
4951 if (FPR_idx != NumFPRs) {
4952 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4955 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4956 MachinePointerInfo(), false, false, 0);
4957 MemOpChains.push_back(Store);
4959 // Float varargs are always shadowed in available integer registers
4960 if (GPR_idx != NumGPRs) {
4961 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4962 MachinePointerInfo(), false, false,
4964 MemOpChains.push_back(Load.getValue(1));
4965 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4967 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4968 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4969 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4970 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4971 MachinePointerInfo(),
4972 false, false, false, 0);
4973 MemOpChains.push_back(Load.getValue(1));
4974 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4977 // If we have any FPRs remaining, we may also have GPRs remaining.
4978 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4980 if (GPR_idx != NumGPRs)
4982 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4983 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4987 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4988 isPPC64, isTailCall, false, MemOpChains,
4989 TailCallArguments, dl);
4993 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
5000 // These go aligned on the stack, or in the corresponding R registers
5001 // when within range. The Darwin PPC ABI doc claims they also go in
5002 // V registers; in fact gcc does this only for arguments that are
5003 // prototyped, not for those that match the ... We do it for all
5004 // arguments, seems to work.
5005 while (ArgOffset % 16 !=0) {
5006 ArgOffset += PtrByteSize;
5007 if (GPR_idx != NumGPRs)
5010 // We could elide this store in the case where the object fits
5011 // entirely in R registers. Maybe later.
5012 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
5013 DAG.getConstant(ArgOffset, PtrVT));
5014 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
5015 MachinePointerInfo(), false, false, 0);
5016 MemOpChains.push_back(Store);
5017 if (VR_idx != NumVRs) {
5018 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
5019 MachinePointerInfo(),
5020 false, false, false, 0);
5021 MemOpChains.push_back(Load.getValue(1));
5022 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
5025 for (unsigned i=0; i<16; i+=PtrByteSize) {
5026 if (GPR_idx == NumGPRs)
5028 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
5029 DAG.getConstant(i, PtrVT));
5030 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
5031 false, false, false, 0);
5032 MemOpChains.push_back(Load.getValue(1));
5033 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5038 // Non-varargs Altivec params generally go in registers, but have
5039 // stack space allocated at the end.
5040 if (VR_idx != NumVRs) {
5041 // Doesn't have GPR space allocated.
5042 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
5043 } else if (nAltivecParamsAtEnd==0) {
5044 // We are emitting Altivec params in order.
5045 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5046 isPPC64, isTailCall, true, MemOpChains,
5047 TailCallArguments, dl);
5053 // If all Altivec parameters fit in registers, as they usually do,
5054 // they get stack space following the non-Altivec parameters. We
5055 // don't track this here because nobody below needs it.
5056 // If there are more Altivec parameters than fit in registers emit
5058 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
5060 // Offset is aligned; skip 1st 12 params which go in V registers.
5061 ArgOffset = ((ArgOffset+15)/16)*16;
5063 for (unsigned i = 0; i != NumOps; ++i) {
5064 SDValue Arg = OutVals[i];
5065 EVT ArgType = Outs[i].VT;
5066 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
5067 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
5070 // We are emitting Altivec params in order.
5071 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5072 isPPC64, isTailCall, true, MemOpChains,
5073 TailCallArguments, dl);
5080 if (!MemOpChains.empty())
5081 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
5083 // On Darwin, R12 must contain the address of an indirect callee. This does
5084 // not mean the MTCTR instruction must use R12; it's easier to model this as
5085 // an extra parameter, so do that.
5087 !isFunctionGlobalAddress(Callee) &&
5088 !isa<ExternalSymbolSDNode>(Callee) &&
5089 !isBLACompatibleAddress(Callee, DAG))
5090 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
5091 PPC::R12), Callee));
5093 // Build a sequence of copy-to-reg nodes chained together with token chain
5094 // and flag operands which copy the outgoing args into the appropriate regs.
5096 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
5097 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
5098 RegsToPass[i].second, InFlag);
5099 InFlag = Chain.getValue(1);
5103 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
5104 FPOp, true, TailCallArguments);
5106 return FinishCall(CallConv, dl, isTailCall, isVarArg, IsPatchPoint, DAG,
5107 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
5112 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
5113 MachineFunction &MF, bool isVarArg,
5114 const SmallVectorImpl<ISD::OutputArg> &Outs,
5115 LLVMContext &Context) const {
5116 SmallVector<CCValAssign, 16> RVLocs;
5117 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
5118 return CCInfo.CheckReturn(Outs, RetCC_PPC);
5122 PPCTargetLowering::LowerReturn(SDValue Chain,
5123 CallingConv::ID CallConv, bool isVarArg,
5124 const SmallVectorImpl<ISD::OutputArg> &Outs,
5125 const SmallVectorImpl<SDValue> &OutVals,
5126 SDLoc dl, SelectionDAG &DAG) const {
5128 SmallVector<CCValAssign, 16> RVLocs;
5129 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
5131 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
5134 SmallVector<SDValue, 4> RetOps(1, Chain);
5136 // Copy the result values into the output registers.
5137 for (unsigned i = 0; i != RVLocs.size(); ++i) {
5138 CCValAssign &VA = RVLocs[i];
5139 assert(VA.isRegLoc() && "Can only return in registers!");
5141 SDValue Arg = OutVals[i];
5143 switch (VA.getLocInfo()) {
5144 default: llvm_unreachable("Unknown loc info!");
5145 case CCValAssign::Full: break;
5146 case CCValAssign::AExt:
5147 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
5149 case CCValAssign::ZExt:
5150 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
5152 case CCValAssign::SExt:
5153 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
5157 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
5158 Flag = Chain.getValue(1);
5159 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
5162 RetOps[0] = Chain; // Update chain.
5164 // Add the flag if we have it.
5166 RetOps.push_back(Flag);
5168 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, RetOps);
5171 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
5172 const PPCSubtarget &Subtarget) const {
5173 // When we pop the dynamic allocation we need to restore the SP link.
5176 // Get the corect type for pointers.
5177 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5179 // Construct the stack pointer operand.
5180 bool isPPC64 = Subtarget.isPPC64();
5181 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
5182 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
5184 // Get the operands for the STACKRESTORE.
5185 SDValue Chain = Op.getOperand(0);
5186 SDValue SaveSP = Op.getOperand(1);
5188 // Load the old link SP.
5189 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
5190 MachinePointerInfo(),
5191 false, false, false, 0);
5193 // Restore the stack pointer.
5194 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
5196 // Store the old link SP.
5197 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
5204 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
5205 MachineFunction &MF = DAG.getMachineFunction();
5206 bool isPPC64 = Subtarget.isPPC64();
5207 bool isDarwinABI = Subtarget.isDarwinABI();
5208 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5210 // Get current frame pointer save index. The users of this index will be
5211 // primarily DYNALLOC instructions.
5212 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
5213 int RASI = FI->getReturnAddrSaveIndex();
5215 // If the frame pointer save index hasn't been defined yet.
5217 // Find out what the fix offset of the frame pointer save area.
5218 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
5219 // Allocate the frame index for frame pointer save area.
5220 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, false);
5222 FI->setReturnAddrSaveIndex(RASI);
5224 return DAG.getFrameIndex(RASI, PtrVT);
5228 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
5229 MachineFunction &MF = DAG.getMachineFunction();
5230 bool isPPC64 = Subtarget.isPPC64();
5231 bool isDarwinABI = Subtarget.isDarwinABI();
5232 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5234 // Get current frame pointer save index. The users of this index will be
5235 // primarily DYNALLOC instructions.
5236 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
5237 int FPSI = FI->getFramePointerSaveIndex();
5239 // If the frame pointer save index hasn't been defined yet.
5241 // Find out what the fix offset of the frame pointer save area.
5242 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
5245 // Allocate the frame index for frame pointer save area.
5246 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
5248 FI->setFramePointerSaveIndex(FPSI);
5250 return DAG.getFrameIndex(FPSI, PtrVT);
5253 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5255 const PPCSubtarget &Subtarget) const {
5257 SDValue Chain = Op.getOperand(0);
5258 SDValue Size = Op.getOperand(1);
5261 // Get the corect type for pointers.
5262 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5264 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
5265 DAG.getConstant(0, PtrVT), Size);
5266 // Construct a node for the frame pointer save index.
5267 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
5268 // Build a DYNALLOC node.
5269 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
5270 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
5271 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops);
5274 SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
5275 SelectionDAG &DAG) const {
5277 return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL,
5278 DAG.getVTList(MVT::i32, MVT::Other),
5279 Op.getOperand(0), Op.getOperand(1));
5282 SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
5283 SelectionDAG &DAG) const {
5285 return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
5286 Op.getOperand(0), Op.getOperand(1));
5289 SDValue PPCTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
5290 assert(Op.getValueType() == MVT::i1 &&
5291 "Custom lowering only for i1 loads");
5293 // First, load 8 bits into 32 bits, then truncate to 1 bit.
5296 LoadSDNode *LD = cast<LoadSDNode>(Op);
5298 SDValue Chain = LD->getChain();
5299 SDValue BasePtr = LD->getBasePtr();
5300 MachineMemOperand *MMO = LD->getMemOperand();
5302 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, dl, getPointerTy(), Chain,
5303 BasePtr, MVT::i8, MMO);
5304 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewLD);
5306 SDValue Ops[] = { Result, SDValue(NewLD.getNode(), 1) };
5307 return DAG.getMergeValues(Ops, dl);
5310 SDValue PPCTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
5311 assert(Op.getOperand(1).getValueType() == MVT::i1 &&
5312 "Custom lowering only for i1 stores");
5314 // First, zero extend to 32 bits, then use a truncating store to 8 bits.
5317 StoreSDNode *ST = cast<StoreSDNode>(Op);
5319 SDValue Chain = ST->getChain();
5320 SDValue BasePtr = ST->getBasePtr();
5321 SDValue Value = ST->getValue();
5322 MachineMemOperand *MMO = ST->getMemOperand();
5324 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, getPointerTy(), Value);
5325 return DAG.getTruncStore(Chain, dl, Value, BasePtr, MVT::i8, MMO);
5328 // FIXME: Remove this once the ANDI glue bug is fixed:
5329 SDValue PPCTargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
5330 assert(Op.getValueType() == MVT::i1 &&
5331 "Custom lowering only for i1 results");
5334 return DAG.getNode(PPCISD::ANDIo_1_GT_BIT, DL, MVT::i1,
5338 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
5340 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
5341 // Not FP? Not a fsel.
5342 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
5343 !Op.getOperand(2).getValueType().isFloatingPoint())
5346 // We might be able to do better than this under some circumstances, but in
5347 // general, fsel-based lowering of select is a finite-math-only optimization.
5348 // For more information, see section F.3 of the 2.06 ISA specification.
5349 if (!DAG.getTarget().Options.NoInfsFPMath ||
5350 !DAG.getTarget().Options.NoNaNsFPMath)
5353 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
5355 EVT ResVT = Op.getValueType();
5356 EVT CmpVT = Op.getOperand(0).getValueType();
5357 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5358 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
5361 // If the RHS of the comparison is a 0.0, we don't need to do the
5362 // subtraction at all.
5364 if (isFloatingPointZero(RHS))
5366 default: break; // SETUO etc aren't handled by fsel.
5370 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
5371 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
5372 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
5373 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
5374 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
5375 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
5376 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV);
5379 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
5382 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
5383 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
5384 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
5387 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
5390 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
5391 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
5392 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
5393 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
5398 default: break; // SETUO etc aren't handled by fsel.
5402 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
5403 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
5404 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
5405 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
5406 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
5407 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
5408 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
5409 DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV);
5412 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
5413 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
5414 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
5415 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
5418 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
5419 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
5420 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
5421 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
5424 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
5425 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
5426 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
5427 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
5430 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
5431 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
5432 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
5433 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
5438 void PPCTargetLowering::LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
5441 assert(Op.getOperand(0).getValueType().isFloatingPoint());
5442 SDValue Src = Op.getOperand(0);
5443 if (Src.getValueType() == MVT::f32)
5444 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
5447 switch (Op.getSimpleValueType().SimpleTy) {
5448 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
5450 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
5451 (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ :
5456 assert((Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT()) &&
5457 "i64 FP_TO_UINT is supported only with FPCVT");
5458 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
5464 // Convert the FP value to an int value through memory.
5465 bool i32Stack = Op.getValueType() == MVT::i32 && Subtarget.hasSTFIWX() &&
5466 (Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT());
5467 SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64);
5468 int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex();
5469 MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(FI);
5471 // Emit a store to the stack slot.
5474 MachineFunction &MF = DAG.getMachineFunction();
5475 MachineMemOperand *MMO =
5476 MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, 4);
5477 SDValue Ops[] = { DAG.getEntryNode(), Tmp, FIPtr };
5478 Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
5479 DAG.getVTList(MVT::Other), Ops, MVT::i32, MMO);
5481 Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
5482 MPI, false, false, 0);
5484 // Result is a load from the stack slot. If loading 4 bytes, make sure to
5486 if (Op.getValueType() == MVT::i32 && !i32Stack) {
5487 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
5488 DAG.getConstant(4, FIPtr.getValueType()));
5489 MPI = MPI.getWithOffset(4);
5497 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
5500 LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
5502 return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI, false,
5503 false, RLI.IsInvariant, RLI.Alignment, RLI.AAInfo,
5507 // We're trying to insert a regular store, S, and then a load, L. If the
5508 // incoming value, O, is a load, we might just be able to have our load use the
5509 // address used by O. However, we don't know if anything else will store to
5510 // that address before we can load from it. To prevent this situation, we need
5511 // to insert our load, L, into the chain as a peer of O. To do this, we give L
5512 // the same chain operand as O, we create a token factor from the chain results
5513 // of O and L, and we replace all uses of O's chain result with that token
5514 // factor (see spliceIntoChain below for this last part).
5515 bool PPCTargetLowering::canReuseLoadAddress(SDValue Op, EVT MemVT,
5518 ISD::LoadExtType ET) const {
5520 if (ET == ISD::NON_EXTLOAD &&
5521 (Op.getOpcode() == ISD::FP_TO_UINT ||
5522 Op.getOpcode() == ISD::FP_TO_SINT) &&
5523 isOperationLegalOrCustom(Op.getOpcode(),
5524 Op.getOperand(0).getValueType())) {
5526 LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
5530 LoadSDNode *LD = dyn_cast<LoadSDNode>(Op);
5531 if (!LD || LD->getExtensionType() != ET || LD->isVolatile() ||
5532 LD->isNonTemporal())
5534 if (LD->getMemoryVT() != MemVT)
5537 RLI.Ptr = LD->getBasePtr();
5538 if (LD->isIndexed() && LD->getOffset().getOpcode() != ISD::UNDEF) {
5539 assert(LD->getAddressingMode() == ISD::PRE_INC &&
5540 "Non-pre-inc AM on PPC?");
5541 RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr,
5545 RLI.Chain = LD->getChain();
5546 RLI.MPI = LD->getPointerInfo();
5547 RLI.IsInvariant = LD->isInvariant();
5548 RLI.Alignment = LD->getAlignment();
5549 RLI.AAInfo = LD->getAAInfo();
5550 RLI.Ranges = LD->getRanges();
5552 RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1);
5556 // Given the head of the old chain, ResChain, insert a token factor containing
5557 // it and NewResChain, and make users of ResChain now be users of that token
5559 void PPCTargetLowering::spliceIntoChain(SDValue ResChain,
5560 SDValue NewResChain,
5561 SelectionDAG &DAG) const {
5565 SDLoc dl(NewResChain);
5567 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5568 NewResChain, DAG.getUNDEF(MVT::Other));
5569 assert(TF.getNode() != NewResChain.getNode() &&
5570 "A new TF really is required here");
5572 DAG.ReplaceAllUsesOfValueWith(ResChain, TF);
5573 DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain);
5576 SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
5577 SelectionDAG &DAG) const {
5579 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
5580 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
5583 if (Op.getOperand(0).getValueType() == MVT::i1)
5584 return DAG.getNode(ISD::SELECT, dl, Op.getValueType(), Op.getOperand(0),
5585 DAG.getConstantFP(1.0, Op.getValueType()),
5586 DAG.getConstantFP(0.0, Op.getValueType()));
5588 assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
5589 "UINT_TO_FP is supported only with FPCVT");
5591 // If we have FCFIDS, then use it when converting to single-precision.
5592 // Otherwise, convert to double-precision and then round.
5593 unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
5594 (Op.getOpcode() == ISD::UINT_TO_FP ?
5595 PPCISD::FCFIDUS : PPCISD::FCFIDS) :
5596 (Op.getOpcode() == ISD::UINT_TO_FP ?
5597 PPCISD::FCFIDU : PPCISD::FCFID);
5598 MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
5599 MVT::f32 : MVT::f64;
5601 if (Op.getOperand(0).getValueType() == MVT::i64) {
5602 SDValue SINT = Op.getOperand(0);
5603 // When converting to single-precision, we actually need to convert
5604 // to double-precision first and then round to single-precision.
5605 // To avoid double-rounding effects during that operation, we have
5606 // to prepare the input operand. Bits that might be truncated when
5607 // converting to double-precision are replaced by a bit that won't
5608 // be lost at this stage, but is below the single-precision rounding
5611 // However, if -enable-unsafe-fp-math is in effect, accept double
5612 // rounding to avoid the extra overhead.
5613 if (Op.getValueType() == MVT::f32 &&
5614 !Subtarget.hasFPCVT() &&
5615 !DAG.getTarget().Options.UnsafeFPMath) {
5617 // Twiddle input to make sure the low 11 bits are zero. (If this
5618 // is the case, we are guaranteed the value will fit into the 53 bit
5619 // mantissa of an IEEE double-precision value without rounding.)
5620 // If any of those low 11 bits were not zero originally, make sure
5621 // bit 12 (value 2048) is set instead, so that the final rounding
5622 // to single-precision gets the correct result.
5623 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
5624 SINT, DAG.getConstant(2047, MVT::i64));
5625 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
5626 Round, DAG.getConstant(2047, MVT::i64));
5627 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
5628 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
5629 Round, DAG.getConstant(-2048, MVT::i64));
5631 // However, we cannot use that value unconditionally: if the magnitude
5632 // of the input value is small, the bit-twiddling we did above might
5633 // end up visibly changing the output. Fortunately, in that case, we
5634 // don't need to twiddle bits since the original input will convert
5635 // exactly to double-precision floating-point already. Therefore,
5636 // construct a conditional to use the original value if the top 11
5637 // bits are all sign-bit copies, and use the rounded value computed
5639 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
5640 SINT, DAG.getConstant(53, MVT::i32));
5641 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
5642 Cond, DAG.getConstant(1, MVT::i64));
5643 Cond = DAG.getSetCC(dl, MVT::i32,
5644 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
5646 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
5652 MachineFunction &MF = DAG.getMachineFunction();
5653 if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) {
5654 Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI, false,
5655 false, RLI.IsInvariant, RLI.Alignment, RLI.AAInfo,
5657 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
5658 } else if (Subtarget.hasLFIWAX() &&
5659 canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) {
5660 MachineMemOperand *MMO =
5661 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
5662 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
5663 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
5664 Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl,
5665 DAG.getVTList(MVT::f64, MVT::Other),
5666 Ops, MVT::i32, MMO);
5667 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
5668 } else if (Subtarget.hasFPCVT() &&
5669 canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::ZEXTLOAD)) {
5670 MachineMemOperand *MMO =
5671 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
5672 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
5673 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
5674 Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWZX, dl,
5675 DAG.getVTList(MVT::f64, MVT::Other),
5676 Ops, MVT::i32, MMO);
5677 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
5678 } else if (((Subtarget.hasLFIWAX() &&
5679 SINT.getOpcode() == ISD::SIGN_EXTEND) ||
5680 (Subtarget.hasFPCVT() &&
5681 SINT.getOpcode() == ISD::ZERO_EXTEND)) &&
5682 SINT.getOperand(0).getValueType() == MVT::i32) {
5683 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
5684 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5686 int FrameIdx = FrameInfo->CreateStackObject(4, 4, false);
5687 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5690 DAG.getStore(DAG.getEntryNode(), dl, SINT.getOperand(0), FIdx,
5691 MachinePointerInfo::getFixedStack(FrameIdx),
5694 assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
5695 "Expected an i32 store");
5699 RLI.MPI = MachinePointerInfo::getFixedStack(FrameIdx);
5702 MachineMemOperand *MMO =
5703 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
5704 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
5705 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
5706 Bits = DAG.getMemIntrinsicNode(SINT.getOpcode() == ISD::ZERO_EXTEND ?
5707 PPCISD::LFIWZX : PPCISD::LFIWAX,
5708 dl, DAG.getVTList(MVT::f64, MVT::Other),
5709 Ops, MVT::i32, MMO);
5711 Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
5713 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Bits);
5715 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT())
5716 FP = DAG.getNode(ISD::FP_ROUND, dl,
5717 MVT::f32, FP, DAG.getIntPtrConstant(0));
5721 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
5722 "Unhandled INT_TO_FP type in custom expander!");
5723 // Since we only generate this in 64-bit mode, we can take advantage of
5724 // 64-bit registers. In particular, sign extend the input value into the
5725 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
5726 // then lfd it and fcfid it.
5727 MachineFunction &MF = DAG.getMachineFunction();
5728 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
5729 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5732 if (Subtarget.hasLFIWAX() || Subtarget.hasFPCVT()) {
5735 if (!(ReusingLoad = canReuseLoadAddress(Op.getOperand(0), MVT::i32, RLI,
5737 int FrameIdx = FrameInfo->CreateStackObject(4, 4, false);
5738 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5740 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx,
5741 MachinePointerInfo::getFixedStack(FrameIdx),
5744 assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
5745 "Expected an i32 store");
5749 RLI.MPI = MachinePointerInfo::getFixedStack(FrameIdx);
5753 MachineMemOperand *MMO =
5754 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
5755 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
5756 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
5757 Ld = DAG.getMemIntrinsicNode(Op.getOpcode() == ISD::UINT_TO_FP ?
5758 PPCISD::LFIWZX : PPCISD::LFIWAX,
5759 dl, DAG.getVTList(MVT::f64, MVT::Other),
5760 Ops, MVT::i32, MMO);
5762 spliceIntoChain(RLI.ResChain, Ld.getValue(1), DAG);
5764 assert(Subtarget.isPPC64() &&
5765 "i32->FP without LFIWAX supported only on PPC64");
5767 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
5768 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5770 SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64,
5773 // STD the extended value into the stack slot.
5774 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Ext64, FIdx,
5775 MachinePointerInfo::getFixedStack(FrameIdx),
5778 // Load the value as a double.
5779 Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx,
5780 MachinePointerInfo::getFixedStack(FrameIdx),
5781 false, false, false, 0);
5784 // FCFID it and return it.
5785 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Ld);
5786 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT())
5787 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
5791 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
5792 SelectionDAG &DAG) const {
5795 The rounding mode is in bits 30:31 of FPSR, and has the following
5802 FLT_ROUNDS, on the other hand, expects the following:
5809 To perform the conversion, we do:
5810 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
5813 MachineFunction &MF = DAG.getMachineFunction();
5814 EVT VT = Op.getValueType();
5815 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5817 // Save FP Control Word to register
5819 MVT::f64, // return register
5820 MVT::Glue // unused in this context
5822 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, None);
5824 // Save FP register to stack slot
5825 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5826 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
5827 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
5828 StackSlot, MachinePointerInfo(), false, false,0);
5830 // Load FP Control Word from low 32 bits of stack slot.
5831 SDValue Four = DAG.getConstant(4, PtrVT);
5832 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
5833 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
5834 false, false, false, 0);
5836 // Transform as necessary
5838 DAG.getNode(ISD::AND, dl, MVT::i32,
5839 CWD, DAG.getConstant(3, MVT::i32));
5841 DAG.getNode(ISD::SRL, dl, MVT::i32,
5842 DAG.getNode(ISD::AND, dl, MVT::i32,
5843 DAG.getNode(ISD::XOR, dl, MVT::i32,
5844 CWD, DAG.getConstant(3, MVT::i32)),
5845 DAG.getConstant(3, MVT::i32)),
5846 DAG.getConstant(1, MVT::i32));
5849 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
5851 return DAG.getNode((VT.getSizeInBits() < 16 ?
5852 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
5855 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
5856 EVT VT = Op.getValueType();
5857 unsigned BitWidth = VT.getSizeInBits();
5859 assert(Op.getNumOperands() == 3 &&
5860 VT == Op.getOperand(1).getValueType() &&
5863 // Expand into a bunch of logical ops. Note that these ops
5864 // depend on the PPC behavior for oversized shift amounts.
5865 SDValue Lo = Op.getOperand(0);
5866 SDValue Hi = Op.getOperand(1);
5867 SDValue Amt = Op.getOperand(2);
5868 EVT AmtVT = Amt.getValueType();
5870 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
5871 DAG.getConstant(BitWidth, AmtVT), Amt);
5872 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
5873 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
5874 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
5875 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
5876 DAG.getConstant(-BitWidth, AmtVT));
5877 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
5878 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
5879 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
5880 SDValue OutOps[] = { OutLo, OutHi };
5881 return DAG.getMergeValues(OutOps, dl);
5884 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
5885 EVT VT = Op.getValueType();
5887 unsigned BitWidth = VT.getSizeInBits();
5888 assert(Op.getNumOperands() == 3 &&
5889 VT == Op.getOperand(1).getValueType() &&
5892 // Expand into a bunch of logical ops. Note that these ops
5893 // depend on the PPC behavior for oversized shift amounts.
5894 SDValue Lo = Op.getOperand(0);
5895 SDValue Hi = Op.getOperand(1);
5896 SDValue Amt = Op.getOperand(2);
5897 EVT AmtVT = Amt.getValueType();
5899 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
5900 DAG.getConstant(BitWidth, AmtVT), Amt);
5901 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
5902 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
5903 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
5904 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
5905 DAG.getConstant(-BitWidth, AmtVT));
5906 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
5907 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
5908 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
5909 SDValue OutOps[] = { OutLo, OutHi };
5910 return DAG.getMergeValues(OutOps, dl);
5913 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
5915 EVT VT = Op.getValueType();
5916 unsigned BitWidth = VT.getSizeInBits();
5917 assert(Op.getNumOperands() == 3 &&
5918 VT == Op.getOperand(1).getValueType() &&
5921 // Expand into a bunch of logical ops, followed by a select_cc.
5922 SDValue Lo = Op.getOperand(0);
5923 SDValue Hi = Op.getOperand(1);
5924 SDValue Amt = Op.getOperand(2);
5925 EVT AmtVT = Amt.getValueType();
5927 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
5928 DAG.getConstant(BitWidth, AmtVT), Amt);
5929 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
5930 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
5931 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
5932 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
5933 DAG.getConstant(-BitWidth, AmtVT));
5934 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
5935 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
5936 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
5937 Tmp4, Tmp6, ISD::SETLE);
5938 SDValue OutOps[] = { OutLo, OutHi };
5939 return DAG.getMergeValues(OutOps, dl);
5942 //===----------------------------------------------------------------------===//
5943 // Vector related lowering.
5946 /// BuildSplatI - Build a canonical splati of Val with an element size of
5947 /// SplatSize. Cast the result to VT.
5948 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
5949 SelectionDAG &DAG, SDLoc dl) {
5950 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
5952 static const EVT VTys[] = { // canonical VT to use for each size.
5953 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
5956 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
5958 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
5962 EVT CanonicalVT = VTys[SplatSize-1];
5964 // Build a canonical splat for this value.
5965 SDValue Elt = DAG.getConstant(Val, MVT::i32);
5966 SmallVector<SDValue, 8> Ops;
5967 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
5968 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, Ops);
5969 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
5972 /// BuildIntrinsicOp - Return a unary operator intrinsic node with the
5973 /// specified intrinsic ID.
5974 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op,
5975 SelectionDAG &DAG, SDLoc dl,
5976 EVT DestVT = MVT::Other) {
5977 if (DestVT == MVT::Other) DestVT = Op.getValueType();
5978 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5979 DAG.getConstant(IID, MVT::i32), Op);
5982 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
5983 /// specified intrinsic ID.
5984 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
5985 SelectionDAG &DAG, SDLoc dl,
5986 EVT DestVT = MVT::Other) {
5987 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
5988 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5989 DAG.getConstant(IID, MVT::i32), LHS, RHS);
5992 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
5993 /// specified intrinsic ID.
5994 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
5995 SDValue Op2, SelectionDAG &DAG,
5996 SDLoc dl, EVT DestVT = MVT::Other) {
5997 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
5998 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
5999 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
6003 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
6004 /// amount. The result has the specified value type.
6005 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
6006 EVT VT, SelectionDAG &DAG, SDLoc dl) {
6007 // Force LHS/RHS to be the right type.
6008 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
6009 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
6012 for (unsigned i = 0; i != 16; ++i)
6014 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
6015 return DAG.getNode(ISD::BITCAST, dl, VT, T);
6018 // If this is a case we can't handle, return null and let the default
6019 // expansion code take care of it. If we CAN select this case, and if it
6020 // selects to a single instruction, return Op. Otherwise, if we can codegen
6021 // this case more efficiently than a constant pool load, lower it to the
6022 // sequence of ops that should be used.
6023 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
6024 SelectionDAG &DAG) const {
6026 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6027 assert(BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
6029 // Check if this is a splat of a constant value.
6030 APInt APSplatBits, APSplatUndef;
6031 unsigned SplatBitSize;
6033 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
6034 HasAnyUndefs, 0, true) || SplatBitSize > 32)
6037 unsigned SplatBits = APSplatBits.getZExtValue();
6038 unsigned SplatUndef = APSplatUndef.getZExtValue();
6039 unsigned SplatSize = SplatBitSize / 8;
6041 // First, handle single instruction cases.
6044 if (SplatBits == 0) {
6045 // Canonicalize all zero vectors to be v4i32.
6046 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
6047 SDValue Z = DAG.getConstant(0, MVT::i32);
6048 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
6049 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
6054 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
6055 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
6057 if (SextVal >= -16 && SextVal <= 15)
6058 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
6061 // Two instruction sequences.
6063 // If this value is in the range [-32,30] and is even, use:
6064 // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
6065 // If this value is in the range [17,31] and is odd, use:
6066 // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
6067 // If this value is in the range [-31,-17] and is odd, use:
6068 // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
6069 // Note the last two are three-instruction sequences.
6070 if (SextVal >= -32 && SextVal <= 31) {
6071 // To avoid having these optimizations undone by constant folding,
6072 // we convert to a pseudo that will be expanded later into one of
6074 SDValue Elt = DAG.getConstant(SextVal, MVT::i32);
6075 EVT VT = (SplatSize == 1 ? MVT::v16i8 :
6076 (SplatSize == 2 ? MVT::v8i16 : MVT::v4i32));
6077 SDValue EltSize = DAG.getConstant(SplatSize, MVT::i32);
6078 SDValue RetVal = DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
6079 if (VT == Op.getValueType())
6082 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), RetVal);
6085 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
6086 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
6088 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
6089 // Make -1 and vspltisw -1:
6090 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
6092 // Make the VSLW intrinsic, computing 0x8000_0000.
6093 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
6096 // xor by OnesV to invert it.
6097 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
6098 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
6101 // The remaining cases assume either big endian element order or
6102 // a splat-size that equates to the element size of the vector
6103 // to be built. An example that doesn't work for little endian is
6104 // {0, -1, 0, -1, 0, -1, 0, -1} which has a splat size of 32 bits
6105 // and a vector element size of 16 bits. The code below will
6106 // produce the vector in big endian element order, which for little
6107 // endian is {-1, 0, -1, 0, -1, 0, -1, 0}.
6109 // For now, just avoid these optimizations in that case.
6110 // FIXME: Develop correct optimizations for LE with mismatched
6111 // splat and element sizes.
6113 if (Subtarget.isLittleEndian() &&
6114 SplatSize != Op.getValueType().getVectorElementType().getSizeInBits())
6117 // Check to see if this is a wide variety of vsplti*, binop self cases.
6118 static const signed char SplatCsts[] = {
6119 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
6120 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
6123 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
6124 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
6125 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
6126 int i = SplatCsts[idx];
6128 // Figure out what shift amount will be used by altivec if shifted by i in
6130 unsigned TypeShiftAmt = i & (SplatBitSize-1);
6132 // vsplti + shl self.
6133 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
6134 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
6135 static const unsigned IIDs[] = { // Intrinsic to use for each size.
6136 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
6137 Intrinsic::ppc_altivec_vslw
6139 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
6140 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
6143 // vsplti + srl self.
6144 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
6145 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
6146 static const unsigned IIDs[] = { // Intrinsic to use for each size.
6147 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
6148 Intrinsic::ppc_altivec_vsrw
6150 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
6151 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
6154 // vsplti + sra self.
6155 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
6156 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
6157 static const unsigned IIDs[] = { // Intrinsic to use for each size.
6158 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
6159 Intrinsic::ppc_altivec_vsraw
6161 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
6162 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
6165 // vsplti + rol self.
6166 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
6167 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
6168 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
6169 static const unsigned IIDs[] = { // Intrinsic to use for each size.
6170 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
6171 Intrinsic::ppc_altivec_vrlw
6173 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
6174 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
6177 // t = vsplti c, result = vsldoi t, t, 1
6178 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
6179 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
6180 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
6182 // t = vsplti c, result = vsldoi t, t, 2
6183 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
6184 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
6185 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
6187 // t = vsplti c, result = vsldoi t, t, 3
6188 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
6189 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
6190 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
6197 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
6198 /// the specified operations to build the shuffle.
6199 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
6200 SDValue RHS, SelectionDAG &DAG,
6202 unsigned OpNum = (PFEntry >> 26) & 0x0F;
6203 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
6204 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
6207 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
6219 if (OpNum == OP_COPY) {
6220 if (LHSID == (1*9+2)*9+3) return LHS;
6221 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
6225 SDValue OpLHS, OpRHS;
6226 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
6227 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
6231 default: llvm_unreachable("Unknown i32 permute!");
6233 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
6234 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
6235 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
6236 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
6239 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
6240 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
6241 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
6242 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
6245 for (unsigned i = 0; i != 16; ++i)
6246 ShufIdxs[i] = (i&3)+0;
6249 for (unsigned i = 0; i != 16; ++i)
6250 ShufIdxs[i] = (i&3)+4;
6253 for (unsigned i = 0; i != 16; ++i)
6254 ShufIdxs[i] = (i&3)+8;
6257 for (unsigned i = 0; i != 16; ++i)
6258 ShufIdxs[i] = (i&3)+12;
6261 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
6263 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
6265 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
6267 EVT VT = OpLHS.getValueType();
6268 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
6269 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
6270 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
6271 return DAG.getNode(ISD::BITCAST, dl, VT, T);
6274 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
6275 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
6276 /// return the code it can be lowered into. Worst case, it can always be
6277 /// lowered into a vperm.
6278 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
6279 SelectionDAG &DAG) const {
6281 SDValue V1 = Op.getOperand(0);
6282 SDValue V2 = Op.getOperand(1);
6283 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6284 EVT VT = Op.getValueType();
6285 bool isLittleEndian = Subtarget.isLittleEndian();
6287 // Cases that are handled by instructions that take permute immediates
6288 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
6289 // selected by the instruction selector.
6290 if (V2.getOpcode() == ISD::UNDEF) {
6291 if (PPC::isSplatShuffleMask(SVOp, 1) ||
6292 PPC::isSplatShuffleMask(SVOp, 2) ||
6293 PPC::isSplatShuffleMask(SVOp, 4) ||
6294 PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) ||
6295 PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) ||
6296 PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 ||
6297 PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) ||
6298 PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) ||
6299 PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) ||
6300 PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) ||
6301 PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) ||
6302 PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG)) {
6307 // Altivec has a variety of "shuffle immediates" that take two vector inputs
6308 // and produce a fixed permutation. If any of these match, do not lower to
6310 unsigned int ShuffleKind = isLittleEndian ? 2 : 0;
6311 if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) ||
6312 PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) ||
6313 PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 ||
6314 PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
6315 PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
6316 PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
6317 PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
6318 PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
6319 PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG))
6322 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
6323 // perfect shuffle table to emit an optimal matching sequence.
6324 ArrayRef<int> PermMask = SVOp->getMask();
6326 unsigned PFIndexes[4];
6327 bool isFourElementShuffle = true;
6328 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
6329 unsigned EltNo = 8; // Start out undef.
6330 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
6331 if (PermMask[i*4+j] < 0)
6332 continue; // Undef, ignore it.
6334 unsigned ByteSource = PermMask[i*4+j];
6335 if ((ByteSource & 3) != j) {
6336 isFourElementShuffle = false;
6341 EltNo = ByteSource/4;
6342 } else if (EltNo != ByteSource/4) {
6343 isFourElementShuffle = false;
6347 PFIndexes[i] = EltNo;
6350 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
6351 // perfect shuffle vector to determine if it is cost effective to do this as
6352 // discrete instructions, or whether we should use a vperm.
6353 // For now, we skip this for little endian until such time as we have a
6354 // little-endian perfect shuffle table.
6355 if (isFourElementShuffle && !isLittleEndian) {
6356 // Compute the index in the perfect shuffle table.
6357 unsigned PFTableIndex =
6358 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6360 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6361 unsigned Cost = (PFEntry >> 30);
6363 // Determining when to avoid vperm is tricky. Many things affect the cost
6364 // of vperm, particularly how many times the perm mask needs to be computed.
6365 // For example, if the perm mask can be hoisted out of a loop or is already
6366 // used (perhaps because there are multiple permutes with the same shuffle
6367 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
6368 // the loop requires an extra register.
6370 // As a compromise, we only emit discrete instructions if the shuffle can be
6371 // generated in 3 or fewer operations. When we have loop information
6372 // available, if this block is within a loop, we should avoid using vperm
6373 // for 3-operation perms and use a constant pool load instead.
6375 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6378 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
6379 // vector that will get spilled to the constant pool.
6380 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
6382 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
6383 // that it is in input element units, not in bytes. Convert now.
6385 // For little endian, the order of the input vectors is reversed, and
6386 // the permutation mask is complemented with respect to 31. This is
6387 // necessary to produce proper semantics with the big-endian-biased vperm
6389 EVT EltVT = V1.getValueType().getVectorElementType();
6390 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
6392 SmallVector<SDValue, 16> ResultMask;
6393 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
6394 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
6396 for (unsigned j = 0; j != BytesPerElement; ++j)
6398 ResultMask.push_back(DAG.getConstant(31 - (SrcElt*BytesPerElement+j),
6401 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
6405 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
6408 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(),
6411 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(),
6415 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
6416 /// altivec comparison. If it is, return true and fill in Opc/isDot with
6417 /// information about the intrinsic.
6418 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
6420 unsigned IntrinsicID =
6421 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
6424 switch (IntrinsicID) {
6425 default: return false;
6426 // Comparison predicates.
6427 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
6428 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
6429 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
6430 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
6431 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
6432 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
6433 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
6434 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
6435 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
6436 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
6437 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
6438 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
6439 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
6441 // Normal Comparisons.
6442 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
6443 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
6444 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
6445 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
6446 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
6447 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
6448 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
6449 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
6450 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
6451 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
6452 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
6453 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
6454 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
6459 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
6460 /// lower, do it, otherwise return null.
6461 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
6462 SelectionDAG &DAG) const {
6463 // If this is a lowered altivec predicate compare, CompareOpc is set to the
6464 // opcode number of the comparison.
6468 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
6469 return SDValue(); // Don't custom lower most intrinsics.
6471 // If this is a non-dot comparison, make the VCMP node and we are done.
6473 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
6474 Op.getOperand(1), Op.getOperand(2),
6475 DAG.getConstant(CompareOpc, MVT::i32));
6476 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
6479 // Create the PPCISD altivec 'dot' comparison node.
6481 Op.getOperand(2), // LHS
6482 Op.getOperand(3), // RHS
6483 DAG.getConstant(CompareOpc, MVT::i32)
6485 EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
6486 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops);
6488 // Now that we have the comparison, emit a copy from the CR to a GPR.
6489 // This is flagged to the above dot comparison.
6490 SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32,
6491 DAG.getRegister(PPC::CR6, MVT::i32),
6492 CompNode.getValue(1));
6494 // Unpack the result based on how the target uses it.
6495 unsigned BitNo; // Bit # of CR6.
6496 bool InvertBit; // Invert result?
6497 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
6498 default: // Can't happen, don't crash on invalid number though.
6499 case 0: // Return the value of the EQ bit of CR6.
6500 BitNo = 0; InvertBit = false;
6502 case 1: // Return the inverted value of the EQ bit of CR6.
6503 BitNo = 0; InvertBit = true;
6505 case 2: // Return the value of the LT bit of CR6.
6506 BitNo = 2; InvertBit = false;
6508 case 3: // Return the inverted value of the LT bit of CR6.
6509 BitNo = 2; InvertBit = true;
6513 // Shift the bit into the low position.
6514 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
6515 DAG.getConstant(8-(3-BitNo), MVT::i32));
6517 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
6518 DAG.getConstant(1, MVT::i32));
6520 // If we are supposed to, toggle the bit.
6522 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
6523 DAG.getConstant(1, MVT::i32));
6527 SDValue PPCTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
6528 SelectionDAG &DAG) const {
6530 // For v2i64 (VSX), we can pattern patch the v2i32 case (using fp <-> int
6531 // instructions), but for smaller types, we need to first extend up to v2i32
6532 // before doing going farther.
6533 if (Op.getValueType() == MVT::v2i64) {
6534 EVT ExtVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
6535 if (ExtVT != MVT::v2i32) {
6536 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0));
6537 Op = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32, Op,
6538 DAG.getValueType(EVT::getVectorVT(*DAG.getContext(),
6539 ExtVT.getVectorElementType(), 4)));
6540 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Op);
6541 Op = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v2i64, Op,
6542 DAG.getValueType(MVT::v2i32));
6551 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
6552 SelectionDAG &DAG) const {
6554 // Create a stack slot that is 16-byte aligned.
6555 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6556 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
6557 EVT PtrVT = getPointerTy();
6558 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
6560 // Store the input value into Value#0 of the stack slot.
6561 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
6562 Op.getOperand(0), FIdx, MachinePointerInfo(),
6565 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
6566 false, false, false, 0);
6569 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
6571 if (Op.getValueType() == MVT::v4i32) {
6572 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
6574 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
6575 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
6577 SDValue RHSSwap = // = vrlw RHS, 16
6578 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
6580 // Shrinkify inputs to v8i16.
6581 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
6582 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
6583 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
6585 // Low parts multiplied together, generating 32-bit results (we ignore the
6587 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
6588 LHS, RHS, DAG, dl, MVT::v4i32);
6590 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
6591 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
6592 // Shift the high parts up 16 bits.
6593 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
6595 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
6596 } else if (Op.getValueType() == MVT::v8i16) {
6597 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
6599 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
6601 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
6602 LHS, RHS, Zero, DAG, dl);
6603 } else if (Op.getValueType() == MVT::v16i8) {
6604 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
6605 bool isLittleEndian = Subtarget.isLittleEndian();
6607 // Multiply the even 8-bit parts, producing 16-bit sums.
6608 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
6609 LHS, RHS, DAG, dl, MVT::v8i16);
6610 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
6612 // Multiply the odd 8-bit parts, producing 16-bit sums.
6613 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
6614 LHS, RHS, DAG, dl, MVT::v8i16);
6615 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
6617 // Merge the results together. Because vmuleub and vmuloub are
6618 // instructions with a big-endian bias, we must reverse the
6619 // element numbering and reverse the meaning of "odd" and "even"
6620 // when generating little endian code.
6622 for (unsigned i = 0; i != 8; ++i) {
6623 if (isLittleEndian) {
6625 Ops[i*2+1] = 2*i+16;
6628 Ops[i*2+1] = 2*i+1+16;
6632 return DAG.getVectorShuffle(MVT::v16i8, dl, OddParts, EvenParts, Ops);
6634 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
6636 llvm_unreachable("Unknown mul to lower!");
6640 /// LowerOperation - Provide custom lowering hooks for some operations.
6642 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6643 switch (Op.getOpcode()) {
6644 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
6645 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6646 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6647 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6648 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6649 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6650 case ISD::SETCC: return LowerSETCC(Op, DAG);
6651 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
6652 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
6654 return LowerVASTART(Op, DAG, Subtarget);
6657 return LowerVAARG(Op, DAG, Subtarget);
6660 return LowerVACOPY(Op, DAG, Subtarget);
6662 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, Subtarget);
6663 case ISD::DYNAMIC_STACKALLOC:
6664 return LowerDYNAMIC_STACKALLOC(Op, DAG, Subtarget);
6666 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
6667 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
6669 case ISD::LOAD: return LowerLOAD(Op, DAG);
6670 case ISD::STORE: return LowerSTORE(Op, DAG);
6671 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
6672 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6673 case ISD::FP_TO_UINT:
6674 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
6676 case ISD::UINT_TO_FP:
6677 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6678 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6680 // Lower 64-bit shifts.
6681 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
6682 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
6683 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
6685 // Vector-related lowering.
6686 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6687 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6688 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6689 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6690 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
6691 case ISD::MUL: return LowerMUL(Op, DAG);
6693 // For counter-based loop handling.
6694 case ISD::INTRINSIC_W_CHAIN: return SDValue();
6696 // Frame & Return address.
6697 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6698 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6702 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
6703 SmallVectorImpl<SDValue>&Results,
6704 SelectionDAG &DAG) const {
6705 const TargetMachine &TM = getTargetMachine();
6707 switch (N->getOpcode()) {
6709 llvm_unreachable("Do not know how to custom type legalize this operation!");
6710 case ISD::READCYCLECOUNTER: {
6711 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6712 SDValue RTB = DAG.getNode(PPCISD::READ_TIME_BASE, dl, VTs, N->getOperand(0));
6714 Results.push_back(RTB);
6715 Results.push_back(RTB.getValue(1));
6716 Results.push_back(RTB.getValue(2));
6719 case ISD::INTRINSIC_W_CHAIN: {
6720 if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() !=
6721 Intrinsic::ppc_is_decremented_ctr_nonzero)
6724 assert(N->getValueType(0) == MVT::i1 &&
6725 "Unexpected result type for CTR decrement intrinsic");
6726 EVT SVT = getSetCCResultType(*DAG.getContext(), N->getValueType(0));
6727 SDVTList VTs = DAG.getVTList(SVT, MVT::Other);
6728 SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0),
6731 Results.push_back(NewInt);
6732 Results.push_back(NewInt.getValue(1));
6736 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
6737 || TM.getSubtarget<PPCSubtarget>().isPPC64())
6740 EVT VT = N->getValueType(0);
6742 if (VT == MVT::i64) {
6743 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, Subtarget);
6745 Results.push_back(NewNode);
6746 Results.push_back(NewNode.getValue(1));
6750 case ISD::FP_ROUND_INREG: {
6751 assert(N->getValueType(0) == MVT::ppcf128);
6752 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
6753 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
6754 MVT::f64, N->getOperand(0),
6755 DAG.getIntPtrConstant(0));
6756 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
6757 MVT::f64, N->getOperand(0),
6758 DAG.getIntPtrConstant(1));
6760 // Add the two halves of the long double in round-to-zero mode.
6761 SDValue FPreg = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi);
6763 // We know the low half is about to be thrown away, so just use something
6765 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
6769 case ISD::FP_TO_SINT:
6770 // LowerFP_TO_INT() can only handle f32 and f64.
6771 if (N->getOperand(0).getValueType() == MVT::ppcf128)
6773 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
6779 //===----------------------------------------------------------------------===//
6780 // Other Lowering Code
6781 //===----------------------------------------------------------------------===//
6783 static Instruction* callIntrinsic(IRBuilder<> &Builder, Intrinsic::ID Id) {
6784 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
6785 Function *Func = Intrinsic::getDeclaration(M, Id);
6786 return Builder.CreateCall(Func);
6789 // The mappings for emitLeading/TrailingFence is taken from
6790 // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
6791 Instruction* PPCTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
6792 AtomicOrdering Ord, bool IsStore,
6793 bool IsLoad) const {
6794 if (Ord == SequentiallyConsistent)
6795 return callIntrinsic(Builder, Intrinsic::ppc_sync);
6796 else if (isAtLeastRelease(Ord))
6797 return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
6802 Instruction* PPCTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
6803 AtomicOrdering Ord, bool IsStore,
6804 bool IsLoad) const {
6805 if (IsLoad && isAtLeastAcquire(Ord))
6806 return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
6807 // FIXME: this is too conservative, a dependent branch + isync is enough.
6808 // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and
6809 // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html
6810 // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification.
6816 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6817 bool is64bit, unsigned BinOpcode) const {
6818 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6819 const TargetInstrInfo *TII =
6820 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6822 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6823 MachineFunction *F = BB->getParent();
6824 MachineFunction::iterator It = BB;
6827 unsigned dest = MI->getOperand(0).getReg();
6828 unsigned ptrA = MI->getOperand(1).getReg();
6829 unsigned ptrB = MI->getOperand(2).getReg();
6830 unsigned incr = MI->getOperand(3).getReg();
6831 DebugLoc dl = MI->getDebugLoc();
6833 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
6834 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6835 F->insert(It, loopMBB);
6836 F->insert(It, exitMBB);
6837 exitMBB->splice(exitMBB->begin(), BB,
6838 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6839 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6841 MachineRegisterInfo &RegInfo = F->getRegInfo();
6842 unsigned TmpReg = (!BinOpcode) ? incr :
6843 RegInfo.createVirtualRegister( is64bit ? &PPC::G8RCRegClass
6844 : &PPC::GPRCRegClass);
6848 // fallthrough --> loopMBB
6849 BB->addSuccessor(loopMBB);
6852 // l[wd]arx dest, ptr
6853 // add r0, dest, incr
6854 // st[wd]cx. r0, ptr
6856 // fallthrough --> exitMBB
6858 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
6859 .addReg(ptrA).addReg(ptrB);
6861 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
6862 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6863 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
6864 BuildMI(BB, dl, TII->get(PPC::BCC))
6865 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
6866 BB->addSuccessor(loopMBB);
6867 BB->addSuccessor(exitMBB);
6876 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
6877 MachineBasicBlock *BB,
6878 bool is8bit, // operation
6879 unsigned BinOpcode) const {
6880 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6881 const TargetInstrInfo *TII =
6882 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6883 // In 64 bit mode we have to use 64 bits for addresses, even though the
6884 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
6885 // registers without caring whether they're 32 or 64, but here we're
6886 // doing actual arithmetic on the addresses.
6887 bool is64bit = Subtarget.isPPC64();
6888 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
6890 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6891 MachineFunction *F = BB->getParent();
6892 MachineFunction::iterator It = BB;
6895 unsigned dest = MI->getOperand(0).getReg();
6896 unsigned ptrA = MI->getOperand(1).getReg();
6897 unsigned ptrB = MI->getOperand(2).getReg();
6898 unsigned incr = MI->getOperand(3).getReg();
6899 DebugLoc dl = MI->getDebugLoc();
6901 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
6902 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6903 F->insert(It, loopMBB);
6904 F->insert(It, exitMBB);
6905 exitMBB->splice(exitMBB->begin(), BB,
6906 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6907 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6909 MachineRegisterInfo &RegInfo = F->getRegInfo();
6910 const TargetRegisterClass *RC = is64bit ? &PPC::G8RCRegClass
6911 : &PPC::GPRCRegClass;
6912 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6913 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6914 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6915 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
6916 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6917 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6918 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6919 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6920 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
6921 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6922 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6924 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
6928 // fallthrough --> loopMBB
6929 BB->addSuccessor(loopMBB);
6931 // The 4-byte load must be aligned, while a char or short may be
6932 // anywhere in the word. Hence all this nasty bookkeeping code.
6933 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6934 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6935 // xori shift, shift1, 24 [16]
6936 // rlwinm ptr, ptr1, 0, 0, 29
6937 // slw incr2, incr, shift
6938 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6939 // slw mask, mask2, shift
6941 // lwarx tmpDest, ptr
6942 // add tmp, tmpDest, incr2
6943 // andc tmp2, tmpDest, mask
6944 // and tmp3, tmp, mask
6945 // or tmp4, tmp3, tmp2
6948 // fallthrough --> exitMBB
6949 // srw dest, tmpDest, shift
6950 if (ptrA != ZeroReg) {
6951 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6952 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6953 .addReg(ptrA).addReg(ptrB);
6957 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6958 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6959 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6960 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6962 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6963 .addReg(Ptr1Reg).addImm(0).addImm(61);
6965 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6966 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6967 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
6968 .addReg(incr).addReg(ShiftReg);
6970 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6972 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6973 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
6975 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6976 .addReg(Mask2Reg).addReg(ShiftReg);
6979 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6980 .addReg(ZeroReg).addReg(PtrReg);
6982 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
6983 .addReg(Incr2Reg).addReg(TmpDestReg);
6984 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
6985 .addReg(TmpDestReg).addReg(MaskReg);
6986 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
6987 .addReg(TmpReg).addReg(MaskReg);
6988 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
6989 .addReg(Tmp3Reg).addReg(Tmp2Reg);
6990 BuildMI(BB, dl, TII->get(PPC::STWCX))
6991 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
6992 BuildMI(BB, dl, TII->get(PPC::BCC))
6993 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
6994 BB->addSuccessor(loopMBB);
6995 BB->addSuccessor(exitMBB);
7000 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
7005 llvm::MachineBasicBlock*
7006 PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
7007 MachineBasicBlock *MBB) const {
7008 DebugLoc DL = MI->getDebugLoc();
7009 const TargetInstrInfo *TII =
7010 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7012 MachineFunction *MF = MBB->getParent();
7013 MachineRegisterInfo &MRI = MF->getRegInfo();
7015 const BasicBlock *BB = MBB->getBasicBlock();
7016 MachineFunction::iterator I = MBB;
7020 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
7021 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
7023 unsigned DstReg = MI->getOperand(0).getReg();
7024 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
7025 assert(RC->hasType(MVT::i32) && "Invalid destination!");
7026 unsigned mainDstReg = MRI.createVirtualRegister(RC);
7027 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
7029 MVT PVT = getPointerTy();
7030 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
7031 "Invalid Pointer Size!");
7032 // For v = setjmp(buf), we generate
7035 // SjLjSetup mainMBB
7041 // buf[LabelOffset] = LR
7045 // v = phi(main, restore)
7048 MachineBasicBlock *thisMBB = MBB;
7049 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
7050 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
7051 MF->insert(I, mainMBB);
7052 MF->insert(I, sinkMBB);
7054 MachineInstrBuilder MIB;
7056 // Transfer the remainder of BB and its successor edges to sinkMBB.
7057 sinkMBB->splice(sinkMBB->begin(), MBB,
7058 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
7059 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
7061 // Note that the structure of the jmp_buf used here is not compatible
7062 // with that used by libc, and is not designed to be. Specifically, it
7063 // stores only those 'reserved' registers that LLVM does not otherwise
7064 // understand how to spill. Also, by convention, by the time this
7065 // intrinsic is called, Clang has already stored the frame address in the
7066 // first slot of the buffer and stack address in the third. Following the
7067 // X86 target code, we'll store the jump address in the second slot. We also
7068 // need to save the TOC pointer (R2) to handle jumps between shared
7069 // libraries, and that will be stored in the fourth slot. The thread
7070 // identifier (R13) is not affected.
7073 const int64_t LabelOffset = 1 * PVT.getStoreSize();
7074 const int64_t TOCOffset = 3 * PVT.getStoreSize();
7075 const int64_t BPOffset = 4 * PVT.getStoreSize();
7077 // Prepare IP either in reg.
7078 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
7079 unsigned LabelReg = MRI.createVirtualRegister(PtrRC);
7080 unsigned BufReg = MI->getOperand(1).getReg();
7082 if (Subtarget.isPPC64() && Subtarget.isSVR4ABI()) {
7083 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD))
7087 MIB.setMemRefs(MMOBegin, MMOEnd);
7090 // Naked functions never have a base pointer, and so we use r1. For all
7091 // other functions, this decision must be delayed until during PEI.
7093 if (MF->getFunction()->getAttributes().hasAttribute(
7094 AttributeSet::FunctionIndex, Attribute::Naked))
7095 BaseReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1;
7097 BaseReg = Subtarget.isPPC64() ? PPC::BP8 : PPC::BP;
7099 MIB = BuildMI(*thisMBB, MI, DL,
7100 TII->get(Subtarget.isPPC64() ? PPC::STD : PPC::STW))
7104 MIB.setMemRefs(MMOBegin, MMOEnd);
7107 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB);
7108 const PPCRegisterInfo *TRI =
7109 getTargetMachine().getSubtarget<PPCSubtarget>().getRegisterInfo();
7110 MIB.addRegMask(TRI->getNoPreservedMask());
7112 BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1);
7114 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup))
7116 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB);
7118 thisMBB->addSuccessor(mainMBB, /* weight */ 0);
7119 thisMBB->addSuccessor(sinkMBB, /* weight */ 1);
7123 MIB = BuildMI(mainMBB, DL,
7124 TII->get(Subtarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg);
7127 if (Subtarget.isPPC64()) {
7128 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD))
7130 .addImm(LabelOffset)
7133 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW))
7135 .addImm(LabelOffset)
7139 MIB.setMemRefs(MMOBegin, MMOEnd);
7141 BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0);
7142 mainMBB->addSuccessor(sinkMBB);
7145 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
7146 TII->get(PPC::PHI), DstReg)
7147 .addReg(mainDstReg).addMBB(mainMBB)
7148 .addReg(restoreDstReg).addMBB(thisMBB);
7150 MI->eraseFromParent();
7155 PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
7156 MachineBasicBlock *MBB) const {
7157 DebugLoc DL = MI->getDebugLoc();
7158 const TargetInstrInfo *TII =
7159 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7161 MachineFunction *MF = MBB->getParent();
7162 MachineRegisterInfo &MRI = MF->getRegInfo();
7165 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
7166 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
7168 MVT PVT = getPointerTy();
7169 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
7170 "Invalid Pointer Size!");
7172 const TargetRegisterClass *RC =
7173 (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
7174 unsigned Tmp = MRI.createVirtualRegister(RC);
7175 // Since FP is only updated here but NOT referenced, it's treated as GPR.
7176 unsigned FP = (PVT == MVT::i64) ? PPC::X31 : PPC::R31;
7177 unsigned SP = (PVT == MVT::i64) ? PPC::X1 : PPC::R1;
7178 unsigned BP = (PVT == MVT::i64) ? PPC::X30 :
7179 (Subtarget.isSVR4ABI() &&
7180 MF->getTarget().getRelocationModel() == Reloc::PIC_ ?
7181 PPC::R29 : PPC::R30);
7183 MachineInstrBuilder MIB;
7185 const int64_t LabelOffset = 1 * PVT.getStoreSize();
7186 const int64_t SPOffset = 2 * PVT.getStoreSize();
7187 const int64_t TOCOffset = 3 * PVT.getStoreSize();
7188 const int64_t BPOffset = 4 * PVT.getStoreSize();
7190 unsigned BufReg = MI->getOperand(0).getReg();
7192 // Reload FP (the jumped-to function may not have had a
7193 // frame pointer, and if so, then its r31 will be restored
7195 if (PVT == MVT::i64) {
7196 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP)
7200 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP)
7204 MIB.setMemRefs(MMOBegin, MMOEnd);
7207 if (PVT == MVT::i64) {
7208 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp)
7209 .addImm(LabelOffset)
7212 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp)
7213 .addImm(LabelOffset)
7216 MIB.setMemRefs(MMOBegin, MMOEnd);
7219 if (PVT == MVT::i64) {
7220 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP)
7224 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP)
7228 MIB.setMemRefs(MMOBegin, MMOEnd);
7231 if (PVT == MVT::i64) {
7232 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), BP)
7236 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), BP)
7240 MIB.setMemRefs(MMOBegin, MMOEnd);
7243 if (PVT == MVT::i64 && Subtarget.isSVR4ABI()) {
7244 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2)
7248 MIB.setMemRefs(MMOBegin, MMOEnd);
7252 BuildMI(*MBB, MI, DL,
7253 TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp);
7254 BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR));
7256 MI->eraseFromParent();
7261 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7262 MachineBasicBlock *BB) const {
7263 if (MI->getOpcode() == TargetOpcode::STACKMAP ||
7264 MI->getOpcode() == TargetOpcode::PATCHPOINT)
7265 return emitPatchPoint(MI, BB);
7267 if (MI->getOpcode() == PPC::EH_SjLj_SetJmp32 ||
7268 MI->getOpcode() == PPC::EH_SjLj_SetJmp64) {
7269 return emitEHSjLjSetJmp(MI, BB);
7270 } else if (MI->getOpcode() == PPC::EH_SjLj_LongJmp32 ||
7271 MI->getOpcode() == PPC::EH_SjLj_LongJmp64) {
7272 return emitEHSjLjLongJmp(MI, BB);
7275 const TargetInstrInfo *TII =
7276 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7278 // To "insert" these instructions we actually have to insert their
7279 // control-flow patterns.
7280 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7281 MachineFunction::iterator It = BB;
7284 MachineFunction *F = BB->getParent();
7286 if (Subtarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
7287 MI->getOpcode() == PPC::SELECT_CC_I8 ||
7288 MI->getOpcode() == PPC::SELECT_I4 ||
7289 MI->getOpcode() == PPC::SELECT_I8)) {
7290 SmallVector<MachineOperand, 2> Cond;
7291 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
7292 MI->getOpcode() == PPC::SELECT_CC_I8)
7293 Cond.push_back(MI->getOperand(4));
7295 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
7296 Cond.push_back(MI->getOperand(1));
7298 DebugLoc dl = MI->getDebugLoc();
7299 const TargetInstrInfo *TII =
7300 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7301 TII->insertSelect(*BB, MI, dl, MI->getOperand(0).getReg(),
7302 Cond, MI->getOperand(2).getReg(),
7303 MI->getOperand(3).getReg());
7304 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
7305 MI->getOpcode() == PPC::SELECT_CC_I8 ||
7306 MI->getOpcode() == PPC::SELECT_CC_F4 ||
7307 MI->getOpcode() == PPC::SELECT_CC_F8 ||
7308 MI->getOpcode() == PPC::SELECT_CC_VRRC ||
7309 MI->getOpcode() == PPC::SELECT_CC_VSFRC ||
7310 MI->getOpcode() == PPC::SELECT_CC_VSRC ||
7311 MI->getOpcode() == PPC::SELECT_I4 ||
7312 MI->getOpcode() == PPC::SELECT_I8 ||
7313 MI->getOpcode() == PPC::SELECT_F4 ||
7314 MI->getOpcode() == PPC::SELECT_F8 ||
7315 MI->getOpcode() == PPC::SELECT_VRRC ||
7316 MI->getOpcode() == PPC::SELECT_VSFRC ||
7317 MI->getOpcode() == PPC::SELECT_VSRC) {
7318 // The incoming instruction knows the destination vreg to set, the
7319 // condition code register to branch on, the true/false values to
7320 // select between, and a branch opcode to use.
7325 // cmpTY ccX, r1, r2
7327 // fallthrough --> copy0MBB
7328 MachineBasicBlock *thisMBB = BB;
7329 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7330 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7331 DebugLoc dl = MI->getDebugLoc();
7332 F->insert(It, copy0MBB);
7333 F->insert(It, sinkMBB);
7335 // Transfer the remainder of BB and its successor edges to sinkMBB.
7336 sinkMBB->splice(sinkMBB->begin(), BB,
7337 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7338 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7340 // Next, add the true and fallthrough blocks as its successors.
7341 BB->addSuccessor(copy0MBB);
7342 BB->addSuccessor(sinkMBB);
7344 if (MI->getOpcode() == PPC::SELECT_I4 ||
7345 MI->getOpcode() == PPC::SELECT_I8 ||
7346 MI->getOpcode() == PPC::SELECT_F4 ||
7347 MI->getOpcode() == PPC::SELECT_F8 ||
7348 MI->getOpcode() == PPC::SELECT_VRRC ||
7349 MI->getOpcode() == PPC::SELECT_VSFRC ||
7350 MI->getOpcode() == PPC::SELECT_VSRC) {
7351 BuildMI(BB, dl, TII->get(PPC::BC))
7352 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
7354 unsigned SelectPred = MI->getOperand(4).getImm();
7355 BuildMI(BB, dl, TII->get(PPC::BCC))
7356 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
7360 // %FalseValue = ...
7361 // # fallthrough to sinkMBB
7364 // Update machine-CFG edges
7365 BB->addSuccessor(sinkMBB);
7368 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7371 BuildMI(*BB, BB->begin(), dl,
7372 TII->get(PPC::PHI), MI->getOperand(0).getReg())
7373 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
7374 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7375 } else if (MI->getOpcode() == PPC::ReadTB) {
7376 // To read the 64-bit time-base register on a 32-bit target, we read the
7377 // two halves. Should the counter have wrapped while it was being read, we
7378 // need to try again.
7381 // mfspr Rx,TBU # load from TBU
7382 // mfspr Ry,TB # load from TB
7383 // mfspr Rz,TBU # load from TBU
7384 // cmpw crX,Rx,Rz # check if ‘old’=’new’
7385 // bne readLoop # branch if they're not equal
7388 MachineBasicBlock *readMBB = F->CreateMachineBasicBlock(LLVM_BB);
7389 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7390 DebugLoc dl = MI->getDebugLoc();
7391 F->insert(It, readMBB);
7392 F->insert(It, sinkMBB);
7394 // Transfer the remainder of BB and its successor edges to sinkMBB.
7395 sinkMBB->splice(sinkMBB->begin(), BB,
7396 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7397 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7399 BB->addSuccessor(readMBB);
7402 MachineRegisterInfo &RegInfo = F->getRegInfo();
7403 unsigned ReadAgainReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
7404 unsigned LoReg = MI->getOperand(0).getReg();
7405 unsigned HiReg = MI->getOperand(1).getReg();
7407 BuildMI(BB, dl, TII->get(PPC::MFSPR), HiReg).addImm(269);
7408 BuildMI(BB, dl, TII->get(PPC::MFSPR), LoReg).addImm(268);
7409 BuildMI(BB, dl, TII->get(PPC::MFSPR), ReadAgainReg).addImm(269);
7411 unsigned CmpReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
7413 BuildMI(BB, dl, TII->get(PPC::CMPW), CmpReg)
7414 .addReg(HiReg).addReg(ReadAgainReg);
7415 BuildMI(BB, dl, TII->get(PPC::BCC))
7416 .addImm(PPC::PRED_NE).addReg(CmpReg).addMBB(readMBB);
7418 BB->addSuccessor(readMBB);
7419 BB->addSuccessor(sinkMBB);
7421 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
7422 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
7423 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
7424 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
7425 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
7426 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
7427 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
7428 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
7430 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
7431 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
7432 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
7433 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
7434 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
7435 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
7436 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
7437 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
7439 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
7440 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
7441 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
7442 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
7443 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
7444 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
7445 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
7446 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
7448 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
7449 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
7450 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
7451 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
7452 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
7453 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
7454 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
7455 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
7457 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
7458 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::NAND);
7459 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
7460 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::NAND);
7461 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
7462 BB = EmitAtomicBinary(MI, BB, false, PPC::NAND);
7463 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
7464 BB = EmitAtomicBinary(MI, BB, true, PPC::NAND8);
7466 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
7467 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
7468 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
7469 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
7470 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
7471 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
7472 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
7473 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
7475 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
7476 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
7477 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
7478 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
7479 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
7480 BB = EmitAtomicBinary(MI, BB, false, 0);
7481 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
7482 BB = EmitAtomicBinary(MI, BB, true, 0);
7484 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
7485 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
7486 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
7488 unsigned dest = MI->getOperand(0).getReg();
7489 unsigned ptrA = MI->getOperand(1).getReg();
7490 unsigned ptrB = MI->getOperand(2).getReg();
7491 unsigned oldval = MI->getOperand(3).getReg();
7492 unsigned newval = MI->getOperand(4).getReg();
7493 DebugLoc dl = MI->getDebugLoc();
7495 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
7496 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
7497 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
7498 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
7499 F->insert(It, loop1MBB);
7500 F->insert(It, loop2MBB);
7501 F->insert(It, midMBB);
7502 F->insert(It, exitMBB);
7503 exitMBB->splice(exitMBB->begin(), BB,
7504 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7505 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7509 // fallthrough --> loopMBB
7510 BB->addSuccessor(loop1MBB);
7513 // l[wd]arx dest, ptr
7514 // cmp[wd] dest, oldval
7517 // st[wd]cx. newval, ptr
7521 // st[wd]cx. dest, ptr
7524 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
7525 .addReg(ptrA).addReg(ptrB);
7526 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
7527 .addReg(oldval).addReg(dest);
7528 BuildMI(BB, dl, TII->get(PPC::BCC))
7529 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
7530 BB->addSuccessor(loop2MBB);
7531 BB->addSuccessor(midMBB);
7534 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
7535 .addReg(newval).addReg(ptrA).addReg(ptrB);
7536 BuildMI(BB, dl, TII->get(PPC::BCC))
7537 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
7538 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
7539 BB->addSuccessor(loop1MBB);
7540 BB->addSuccessor(exitMBB);
7543 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
7544 .addReg(dest).addReg(ptrA).addReg(ptrB);
7545 BB->addSuccessor(exitMBB);
7550 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
7551 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
7552 // We must use 64-bit registers for addresses when targeting 64-bit,
7553 // since we're actually doing arithmetic on them. Other registers
7555 bool is64bit = Subtarget.isPPC64();
7556 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
7558 unsigned dest = MI->getOperand(0).getReg();
7559 unsigned ptrA = MI->getOperand(1).getReg();
7560 unsigned ptrB = MI->getOperand(2).getReg();
7561 unsigned oldval = MI->getOperand(3).getReg();
7562 unsigned newval = MI->getOperand(4).getReg();
7563 DebugLoc dl = MI->getDebugLoc();
7565 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
7566 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
7567 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
7568 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
7569 F->insert(It, loop1MBB);
7570 F->insert(It, loop2MBB);
7571 F->insert(It, midMBB);
7572 F->insert(It, exitMBB);
7573 exitMBB->splice(exitMBB->begin(), BB,
7574 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7575 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7577 MachineRegisterInfo &RegInfo = F->getRegInfo();
7578 const TargetRegisterClass *RC = is64bit ? &PPC::G8RCRegClass
7579 : &PPC::GPRCRegClass;
7580 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
7581 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
7582 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
7583 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
7584 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
7585 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
7586 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
7587 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
7588 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
7589 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
7590 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
7591 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
7592 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
7594 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
7595 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
7598 // fallthrough --> loopMBB
7599 BB->addSuccessor(loop1MBB);
7601 // The 4-byte load must be aligned, while a char or short may be
7602 // anywhere in the word. Hence all this nasty bookkeeping code.
7603 // add ptr1, ptrA, ptrB [copy if ptrA==0]
7604 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
7605 // xori shift, shift1, 24 [16]
7606 // rlwinm ptr, ptr1, 0, 0, 29
7607 // slw newval2, newval, shift
7608 // slw oldval2, oldval,shift
7609 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
7610 // slw mask, mask2, shift
7611 // and newval3, newval2, mask
7612 // and oldval3, oldval2, mask
7614 // lwarx tmpDest, ptr
7615 // and tmp, tmpDest, mask
7616 // cmpw tmp, oldval3
7619 // andc tmp2, tmpDest, mask
7620 // or tmp4, tmp2, newval3
7625 // stwcx. tmpDest, ptr
7627 // srw dest, tmpDest, shift
7628 if (ptrA != ZeroReg) {
7629 Ptr1Reg = RegInfo.createVirtualRegister(RC);
7630 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
7631 .addReg(ptrA).addReg(ptrB);
7635 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
7636 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
7637 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
7638 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
7640 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
7641 .addReg(Ptr1Reg).addImm(0).addImm(61);
7643 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
7644 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
7645 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
7646 .addReg(newval).addReg(ShiftReg);
7647 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
7648 .addReg(oldval).addReg(ShiftReg);
7650 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
7652 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
7653 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
7654 .addReg(Mask3Reg).addImm(65535);
7656 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
7657 .addReg(Mask2Reg).addReg(ShiftReg);
7658 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
7659 .addReg(NewVal2Reg).addReg(MaskReg);
7660 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
7661 .addReg(OldVal2Reg).addReg(MaskReg);
7664 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
7665 .addReg(ZeroReg).addReg(PtrReg);
7666 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
7667 .addReg(TmpDestReg).addReg(MaskReg);
7668 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
7669 .addReg(TmpReg).addReg(OldVal3Reg);
7670 BuildMI(BB, dl, TII->get(PPC::BCC))
7671 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
7672 BB->addSuccessor(loop2MBB);
7673 BB->addSuccessor(midMBB);
7676 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
7677 .addReg(TmpDestReg).addReg(MaskReg);
7678 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
7679 .addReg(Tmp2Reg).addReg(NewVal3Reg);
7680 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
7681 .addReg(ZeroReg).addReg(PtrReg);
7682 BuildMI(BB, dl, TII->get(PPC::BCC))
7683 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
7684 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
7685 BB->addSuccessor(loop1MBB);
7686 BB->addSuccessor(exitMBB);
7689 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
7690 .addReg(ZeroReg).addReg(PtrReg);
7691 BB->addSuccessor(exitMBB);
7696 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
7698 } else if (MI->getOpcode() == PPC::FADDrtz) {
7699 // This pseudo performs an FADD with rounding mode temporarily forced
7700 // to round-to-zero. We emit this via custom inserter since the FPSCR
7701 // is not modeled at the SelectionDAG level.
7702 unsigned Dest = MI->getOperand(0).getReg();
7703 unsigned Src1 = MI->getOperand(1).getReg();
7704 unsigned Src2 = MI->getOperand(2).getReg();
7705 DebugLoc dl = MI->getDebugLoc();
7707 MachineRegisterInfo &RegInfo = F->getRegInfo();
7708 unsigned MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
7710 // Save FPSCR value.
7711 BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg);
7713 // Set rounding mode to round-to-zero.
7714 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1)).addImm(31);
7715 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0)).addImm(30);
7717 // Perform addition.
7718 BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest).addReg(Src1).addReg(Src2);
7720 // Restore FPSCR value.
7721 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF)).addImm(1).addReg(MFFSReg);
7722 } else if (MI->getOpcode() == PPC::ANDIo_1_EQ_BIT ||
7723 MI->getOpcode() == PPC::ANDIo_1_GT_BIT ||
7724 MI->getOpcode() == PPC::ANDIo_1_EQ_BIT8 ||
7725 MI->getOpcode() == PPC::ANDIo_1_GT_BIT8) {
7726 unsigned Opcode = (MI->getOpcode() == PPC::ANDIo_1_EQ_BIT8 ||
7727 MI->getOpcode() == PPC::ANDIo_1_GT_BIT8) ?
7728 PPC::ANDIo8 : PPC::ANDIo;
7729 bool isEQ = (MI->getOpcode() == PPC::ANDIo_1_EQ_BIT ||
7730 MI->getOpcode() == PPC::ANDIo_1_EQ_BIT8);
7732 MachineRegisterInfo &RegInfo = F->getRegInfo();
7733 unsigned Dest = RegInfo.createVirtualRegister(Opcode == PPC::ANDIo ?
7734 &PPC::GPRCRegClass :
7735 &PPC::G8RCRegClass);
7737 DebugLoc dl = MI->getDebugLoc();
7738 BuildMI(*BB, MI, dl, TII->get(Opcode), Dest)
7739 .addReg(MI->getOperand(1).getReg()).addImm(1);
7740 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY),
7741 MI->getOperand(0).getReg())
7742 .addReg(isEQ ? PPC::CR0EQ : PPC::CR0GT);
7744 llvm_unreachable("Unexpected instr type to insert");
7747 MI->eraseFromParent(); // The pseudo instruction is gone now.
7751 //===----------------------------------------------------------------------===//
7752 // Target Optimization Hooks
7753 //===----------------------------------------------------------------------===//
7755 SDValue PPCTargetLowering::getRsqrtEstimate(SDValue Operand,
7756 DAGCombinerInfo &DCI,
7757 unsigned &RefinementSteps,
7758 bool &UseOneConstNR) const {
7759 EVT VT = Operand.getValueType();
7760 if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) ||
7761 (VT == MVT::f64 && Subtarget.hasFRSQRTE()) ||
7762 (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
7763 (VT == MVT::v2f64 && Subtarget.hasVSX())) {
7764 // Convergence is quadratic, so we essentially double the number of digits
7765 // correct after every iteration. For both FRE and FRSQRTE, the minimum
7766 // architected relative accuracy is 2^-5. When hasRecipPrec(), this is
7767 // 2^-14. IEEE float has 23 digits and double has 52 digits.
7768 RefinementSteps = Subtarget.hasRecipPrec() ? 1 : 3;
7769 if (VT.getScalarType() == MVT::f64)
7771 UseOneConstNR = true;
7772 return DCI.DAG.getNode(PPCISD::FRSQRTE, SDLoc(Operand), VT, Operand);
7777 SDValue PPCTargetLowering::getRecipEstimate(SDValue Operand,
7778 DAGCombinerInfo &DCI,
7779 unsigned &RefinementSteps) const {
7780 EVT VT = Operand.getValueType();
7781 if ((VT == MVT::f32 && Subtarget.hasFRES()) ||
7782 (VT == MVT::f64 && Subtarget.hasFRE()) ||
7783 (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
7784 (VT == MVT::v2f64 && Subtarget.hasVSX())) {
7785 // Convergence is quadratic, so we essentially double the number of digits
7786 // correct after every iteration. For both FRE and FRSQRTE, the minimum
7787 // architected relative accuracy is 2^-5. When hasRecipPrec(), this is
7788 // 2^-14. IEEE float has 23 digits and double has 52 digits.
7789 RefinementSteps = Subtarget.hasRecipPrec() ? 1 : 3;
7790 if (VT.getScalarType() == MVT::f64)
7792 return DCI.DAG.getNode(PPCISD::FRE, SDLoc(Operand), VT, Operand);
7797 bool PPCTargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
7798 // Note: This functionality is used only when unsafe-fp-math is enabled, and
7799 // on cores with reciprocal estimates (which are used when unsafe-fp-math is
7800 // enabled for division), this functionality is redundant with the default
7801 // combiner logic (once the division -> reciprocal/multiply transformation
7802 // has taken place). As a result, this matters more for older cores than for
7805 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7806 // reciprocal if there are two or more FDIVs (for embedded cores with only
7807 // one FP pipeline) for three or more FDIVs (for generic OOO cores).
7808 switch (Subtarget.getDarwinDirective()) {
7810 return NumUsers > 2;
7813 case PPC::DIR_E500mc:
7814 case PPC::DIR_E5500:
7815 return NumUsers > 1;
7819 static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base,
7820 unsigned Bytes, int Dist,
7821 SelectionDAG &DAG) {
7822 if (VT.getSizeInBits() / 8 != Bytes)
7825 SDValue BaseLoc = Base->getBasePtr();
7826 if (Loc.getOpcode() == ISD::FrameIndex) {
7827 if (BaseLoc.getOpcode() != ISD::FrameIndex)
7829 const MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7830 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
7831 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
7832 int FS = MFI->getObjectSize(FI);
7833 int BFS = MFI->getObjectSize(BFI);
7834 if (FS != BFS || FS != (int)Bytes) return false;
7835 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
7839 if (DAG.isBaseWithConstantOffset(Loc) && Loc.getOperand(0) == BaseLoc &&
7840 cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue() == Dist*Bytes)
7843 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7844 const GlobalValue *GV1 = nullptr;
7845 const GlobalValue *GV2 = nullptr;
7846 int64_t Offset1 = 0;
7847 int64_t Offset2 = 0;
7848 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
7849 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
7850 if (isGA1 && isGA2 && GV1 == GV2)
7851 return Offset1 == (Offset2 + Dist*Bytes);
7855 // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
7856 // not enforce equality of the chain operands.
7857 static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base,
7858 unsigned Bytes, int Dist,
7859 SelectionDAG &DAG) {
7860 if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) {
7861 EVT VT = LS->getMemoryVT();
7862 SDValue Loc = LS->getBasePtr();
7863 return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG);
7866 if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
7868 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7869 default: return false;
7870 case Intrinsic::ppc_altivec_lvx:
7871 case Intrinsic::ppc_altivec_lvxl:
7872 case Intrinsic::ppc_vsx_lxvw4x:
7875 case Intrinsic::ppc_vsx_lxvd2x:
7878 case Intrinsic::ppc_altivec_lvebx:
7881 case Intrinsic::ppc_altivec_lvehx:
7884 case Intrinsic::ppc_altivec_lvewx:
7889 return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG);
7892 if (N->getOpcode() == ISD::INTRINSIC_VOID) {
7894 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7895 default: return false;
7896 case Intrinsic::ppc_altivec_stvx:
7897 case Intrinsic::ppc_altivec_stvxl:
7898 case Intrinsic::ppc_vsx_stxvw4x:
7901 case Intrinsic::ppc_vsx_stxvd2x:
7904 case Intrinsic::ppc_altivec_stvebx:
7907 case Intrinsic::ppc_altivec_stvehx:
7910 case Intrinsic::ppc_altivec_stvewx:
7915 return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG);
7921 // Return true is there is a nearyby consecutive load to the one provided
7922 // (regardless of alignment). We search up and down the chain, looking though
7923 // token factors and other loads (but nothing else). As a result, a true result
7924 // indicates that it is safe to create a new consecutive load adjacent to the
7926 static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
7927 SDValue Chain = LD->getChain();
7928 EVT VT = LD->getMemoryVT();
7930 SmallSet<SDNode *, 16> LoadRoots;
7931 SmallVector<SDNode *, 8> Queue(1, Chain.getNode());
7932 SmallSet<SDNode *, 16> Visited;
7934 // First, search up the chain, branching to follow all token-factor operands.
7935 // If we find a consecutive load, then we're done, otherwise, record all
7936 // nodes just above the top-level loads and token factors.
7937 while (!Queue.empty()) {
7938 SDNode *ChainNext = Queue.pop_back_val();
7939 if (!Visited.insert(ChainNext).second)
7942 if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) {
7943 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
7946 if (!Visited.count(ChainLD->getChain().getNode()))
7947 Queue.push_back(ChainLD->getChain().getNode());
7948 } else if (ChainNext->getOpcode() == ISD::TokenFactor) {
7949 for (const SDUse &O : ChainNext->ops())
7950 if (!Visited.count(O.getNode()))
7951 Queue.push_back(O.getNode());
7953 LoadRoots.insert(ChainNext);
7956 // Second, search down the chain, starting from the top-level nodes recorded
7957 // in the first phase. These top-level nodes are the nodes just above all
7958 // loads and token factors. Starting with their uses, recursively look though
7959 // all loads (just the chain uses) and token factors to find a consecutive
7964 for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(),
7965 IE = LoadRoots.end(); I != IE; ++I) {
7966 Queue.push_back(*I);
7968 while (!Queue.empty()) {
7969 SDNode *LoadRoot = Queue.pop_back_val();
7970 if (!Visited.insert(LoadRoot).second)
7973 if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot))
7974 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
7977 for (SDNode::use_iterator UI = LoadRoot->use_begin(),
7978 UE = LoadRoot->use_end(); UI != UE; ++UI)
7979 if (((isa<MemSDNode>(*UI) &&
7980 cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
7981 UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI))
7982 Queue.push_back(*UI);
7989 SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N,
7990 DAGCombinerInfo &DCI) const {
7991 SelectionDAG &DAG = DCI.DAG;
7994 assert(Subtarget.useCRBits() &&
7995 "Expecting to be tracking CR bits");
7996 // If we're tracking CR bits, we need to be careful that we don't have:
7997 // trunc(binary-ops(zext(x), zext(y)))
7999 // trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
8000 // such that we're unnecessarily moving things into GPRs when it would be
8001 // better to keep them in CR bits.
8003 // Note that trunc here can be an actual i1 trunc, or can be the effective
8004 // truncation that comes from a setcc or select_cc.
8005 if (N->getOpcode() == ISD::TRUNCATE &&
8006 N->getValueType(0) != MVT::i1)
8009 if (N->getOperand(0).getValueType() != MVT::i32 &&
8010 N->getOperand(0).getValueType() != MVT::i64)
8013 if (N->getOpcode() == ISD::SETCC ||
8014 N->getOpcode() == ISD::SELECT_CC) {
8015 // If we're looking at a comparison, then we need to make sure that the
8016 // high bits (all except for the first) don't matter the result.
8018 cast<CondCodeSDNode>(N->getOperand(
8019 N->getOpcode() == ISD::SETCC ? 2 : 4))->get();
8020 unsigned OpBits = N->getOperand(0).getValueSizeInBits();
8022 if (ISD::isSignedIntSetCC(CC)) {
8023 if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits ||
8024 DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits)
8026 } else if (ISD::isUnsignedIntSetCC(CC)) {
8027 if (!DAG.MaskedValueIsZero(N->getOperand(0),
8028 APInt::getHighBitsSet(OpBits, OpBits-1)) ||
8029 !DAG.MaskedValueIsZero(N->getOperand(1),
8030 APInt::getHighBitsSet(OpBits, OpBits-1)))
8033 // This is neither a signed nor an unsigned comparison, just make sure
8034 // that the high bits are equal.
8035 APInt Op1Zero, Op1One;
8036 APInt Op2Zero, Op2One;
8037 DAG.computeKnownBits(N->getOperand(0), Op1Zero, Op1One);
8038 DAG.computeKnownBits(N->getOperand(1), Op2Zero, Op2One);
8040 // We don't really care about what is known about the first bit (if
8041 // anything), so clear it in all masks prior to comparing them.
8042 Op1Zero.clearBit(0); Op1One.clearBit(0);
8043 Op2Zero.clearBit(0); Op2One.clearBit(0);
8045 if (Op1Zero != Op2Zero || Op1One != Op2One)
8050 // We now know that the higher-order bits are irrelevant, we just need to
8051 // make sure that all of the intermediate operations are bit operations, and
8052 // all inputs are extensions.
8053 if (N->getOperand(0).getOpcode() != ISD::AND &&
8054 N->getOperand(0).getOpcode() != ISD::OR &&
8055 N->getOperand(0).getOpcode() != ISD::XOR &&
8056 N->getOperand(0).getOpcode() != ISD::SELECT &&
8057 N->getOperand(0).getOpcode() != ISD::SELECT_CC &&
8058 N->getOperand(0).getOpcode() != ISD::TRUNCATE &&
8059 N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND &&
8060 N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND &&
8061 N->getOperand(0).getOpcode() != ISD::ANY_EXTEND)
8064 if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) &&
8065 N->getOperand(1).getOpcode() != ISD::AND &&
8066 N->getOperand(1).getOpcode() != ISD::OR &&
8067 N->getOperand(1).getOpcode() != ISD::XOR &&
8068 N->getOperand(1).getOpcode() != ISD::SELECT &&
8069 N->getOperand(1).getOpcode() != ISD::SELECT_CC &&
8070 N->getOperand(1).getOpcode() != ISD::TRUNCATE &&
8071 N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND &&
8072 N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND &&
8073 N->getOperand(1).getOpcode() != ISD::ANY_EXTEND)
8076 SmallVector<SDValue, 4> Inputs;
8077 SmallVector<SDValue, 8> BinOps, PromOps;
8078 SmallPtrSet<SDNode *, 16> Visited;
8080 for (unsigned i = 0; i < 2; ++i) {
8081 if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
8082 N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
8083 N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
8084 N->getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
8085 isa<ConstantSDNode>(N->getOperand(i)))
8086 Inputs.push_back(N->getOperand(i));
8088 BinOps.push_back(N->getOperand(i));
8090 if (N->getOpcode() == ISD::TRUNCATE)
8094 // Visit all inputs, collect all binary operations (and, or, xor and
8095 // select) that are all fed by extensions.
8096 while (!BinOps.empty()) {
8097 SDValue BinOp = BinOps.back();
8100 if (!Visited.insert(BinOp.getNode()).second)
8103 PromOps.push_back(BinOp);
8105 for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
8106 // The condition of the select is not promoted.
8107 if (BinOp.getOpcode() == ISD::SELECT && i == 0)
8109 if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
8112 if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
8113 BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
8114 BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
8115 BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
8116 isa<ConstantSDNode>(BinOp.getOperand(i))) {
8117 Inputs.push_back(BinOp.getOperand(i));
8118 } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
8119 BinOp.getOperand(i).getOpcode() == ISD::OR ||
8120 BinOp.getOperand(i).getOpcode() == ISD::XOR ||
8121 BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
8122 BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC ||
8123 BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
8124 BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
8125 BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
8126 BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) {
8127 BinOps.push_back(BinOp.getOperand(i));
8129 // We have an input that is not an extension or another binary
8130 // operation; we'll abort this transformation.
8136 // Make sure that this is a self-contained cluster of operations (which
8137 // is not quite the same thing as saying that everything has only one
8139 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
8140 if (isa<ConstantSDNode>(Inputs[i]))
8143 for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
8144 UE = Inputs[i].getNode()->use_end();
8147 if (User != N && !Visited.count(User))
8150 // Make sure that we're not going to promote the non-output-value
8151 // operand(s) or SELECT or SELECT_CC.
8152 // FIXME: Although we could sometimes handle this, and it does occur in
8153 // practice that one of the condition inputs to the select is also one of
8154 // the outputs, we currently can't deal with this.
8155 if (User->getOpcode() == ISD::SELECT) {
8156 if (User->getOperand(0) == Inputs[i])
8158 } else if (User->getOpcode() == ISD::SELECT_CC) {
8159 if (User->getOperand(0) == Inputs[i] ||
8160 User->getOperand(1) == Inputs[i])
8166 for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
8167 for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
8168 UE = PromOps[i].getNode()->use_end();
8171 if (User != N && !Visited.count(User))
8174 // Make sure that we're not going to promote the non-output-value
8175 // operand(s) or SELECT or SELECT_CC.
8176 // FIXME: Although we could sometimes handle this, and it does occur in
8177 // practice that one of the condition inputs to the select is also one of
8178 // the outputs, we currently can't deal with this.
8179 if (User->getOpcode() == ISD::SELECT) {
8180 if (User->getOperand(0) == PromOps[i])
8182 } else if (User->getOpcode() == ISD::SELECT_CC) {
8183 if (User->getOperand(0) == PromOps[i] ||
8184 User->getOperand(1) == PromOps[i])
8190 // Replace all inputs with the extension operand.
8191 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
8192 // Constants may have users outside the cluster of to-be-promoted nodes,
8193 // and so we need to replace those as we do the promotions.
8194 if (isa<ConstantSDNode>(Inputs[i]))
8197 DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0));
8200 // Replace all operations (these are all the same, but have a different
8201 // (i1) return type). DAG.getNode will validate that the types of
8202 // a binary operator match, so go through the list in reverse so that
8203 // we've likely promoted both operands first. Any intermediate truncations or
8204 // extensions disappear.
8205 while (!PromOps.empty()) {
8206 SDValue PromOp = PromOps.back();
8209 if (PromOp.getOpcode() == ISD::TRUNCATE ||
8210 PromOp.getOpcode() == ISD::SIGN_EXTEND ||
8211 PromOp.getOpcode() == ISD::ZERO_EXTEND ||
8212 PromOp.getOpcode() == ISD::ANY_EXTEND) {
8213 if (!isa<ConstantSDNode>(PromOp.getOperand(0)) &&
8214 PromOp.getOperand(0).getValueType() != MVT::i1) {
8215 // The operand is not yet ready (see comment below).
8216 PromOps.insert(PromOps.begin(), PromOp);
8220 SDValue RepValue = PromOp.getOperand(0);
8221 if (isa<ConstantSDNode>(RepValue))
8222 RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue);
8224 DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue);
8229 switch (PromOp.getOpcode()) {
8230 default: C = 0; break;
8231 case ISD::SELECT: C = 1; break;
8232 case ISD::SELECT_CC: C = 2; break;
8235 if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
8236 PromOp.getOperand(C).getValueType() != MVT::i1) ||
8237 (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
8238 PromOp.getOperand(C+1).getValueType() != MVT::i1)) {
8239 // The to-be-promoted operands of this node have not yet been
8240 // promoted (this should be rare because we're going through the
8241 // list backward, but if one of the operands has several users in
8242 // this cluster of to-be-promoted nodes, it is possible).
8243 PromOps.insert(PromOps.begin(), PromOp);
8247 SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
8248 PromOp.getNode()->op_end());
8250 // If there are any constant inputs, make sure they're replaced now.
8251 for (unsigned i = 0; i < 2; ++i)
8252 if (isa<ConstantSDNode>(Ops[C+i]))
8253 Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]);
8255 DAG.ReplaceAllUsesOfValueWith(PromOp,
8256 DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops));
8259 // Now we're left with the initial truncation itself.
8260 if (N->getOpcode() == ISD::TRUNCATE)
8261 return N->getOperand(0);
8263 // Otherwise, this is a comparison. The operands to be compared have just
8264 // changed type (to i1), but everything else is the same.
8265 return SDValue(N, 0);
8268 SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N,
8269 DAGCombinerInfo &DCI) const {
8270 SelectionDAG &DAG = DCI.DAG;
8273 // If we're tracking CR bits, we need to be careful that we don't have:
8274 // zext(binary-ops(trunc(x), trunc(y)))
8276 // zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
8277 // such that we're unnecessarily moving things into CR bits that can more
8278 // efficiently stay in GPRs. Note that if we're not certain that the high
8279 // bits are set as required by the final extension, we still may need to do
8280 // some masking to get the proper behavior.
8282 // This same functionality is important on PPC64 when dealing with
8283 // 32-to-64-bit extensions; these occur often when 32-bit values are used as
8284 // the return values of functions. Because it is so similar, it is handled
8287 if (N->getValueType(0) != MVT::i32 &&
8288 N->getValueType(0) != MVT::i64)
8291 if (!((N->getOperand(0).getValueType() == MVT::i1 &&
8292 Subtarget.useCRBits()) ||
8293 (N->getOperand(0).getValueType() == MVT::i32 &&
8294 Subtarget.isPPC64())))
8297 if (N->getOperand(0).getOpcode() != ISD::AND &&
8298 N->getOperand(0).getOpcode() != ISD::OR &&
8299 N->getOperand(0).getOpcode() != ISD::XOR &&
8300 N->getOperand(0).getOpcode() != ISD::SELECT &&
8301 N->getOperand(0).getOpcode() != ISD::SELECT_CC)
8304 SmallVector<SDValue, 4> Inputs;
8305 SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps;
8306 SmallPtrSet<SDNode *, 16> Visited;
8308 // Visit all inputs, collect all binary operations (and, or, xor and
8309 // select) that are all fed by truncations.
8310 while (!BinOps.empty()) {
8311 SDValue BinOp = BinOps.back();
8314 if (!Visited.insert(BinOp.getNode()).second)
8317 PromOps.push_back(BinOp);
8319 for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
8320 // The condition of the select is not promoted.
8321 if (BinOp.getOpcode() == ISD::SELECT && i == 0)
8323 if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
8326 if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
8327 isa<ConstantSDNode>(BinOp.getOperand(i))) {
8328 Inputs.push_back(BinOp.getOperand(i));
8329 } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
8330 BinOp.getOperand(i).getOpcode() == ISD::OR ||
8331 BinOp.getOperand(i).getOpcode() == ISD::XOR ||
8332 BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
8333 BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) {
8334 BinOps.push_back(BinOp.getOperand(i));
8336 // We have an input that is not a truncation or another binary
8337 // operation; we'll abort this transformation.
8343 // The operands of a select that must be truncated when the select is
8344 // promoted because the operand is actually part of the to-be-promoted set.
8345 DenseMap<SDNode *, EVT> SelectTruncOp[2];
8347 // Make sure that this is a self-contained cluster of operations (which
8348 // is not quite the same thing as saying that everything has only one
8350 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
8351 if (isa<ConstantSDNode>(Inputs[i]))
8354 for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
8355 UE = Inputs[i].getNode()->use_end();
8358 if (User != N && !Visited.count(User))
8361 // If we're going to promote the non-output-value operand(s) or SELECT or
8362 // SELECT_CC, record them for truncation.
8363 if (User->getOpcode() == ISD::SELECT) {
8364 if (User->getOperand(0) == Inputs[i])
8365 SelectTruncOp[0].insert(std::make_pair(User,
8366 User->getOperand(0).getValueType()));
8367 } else if (User->getOpcode() == ISD::SELECT_CC) {
8368 if (User->getOperand(0) == Inputs[i])
8369 SelectTruncOp[0].insert(std::make_pair(User,
8370 User->getOperand(0).getValueType()));
8371 if (User->getOperand(1) == Inputs[i])
8372 SelectTruncOp[1].insert(std::make_pair(User,
8373 User->getOperand(1).getValueType()));
8378 for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
8379 for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
8380 UE = PromOps[i].getNode()->use_end();
8383 if (User != N && !Visited.count(User))
8386 // If we're going to promote the non-output-value operand(s) or SELECT or
8387 // SELECT_CC, record them for truncation.
8388 if (User->getOpcode() == ISD::SELECT) {
8389 if (User->getOperand(0) == PromOps[i])
8390 SelectTruncOp[0].insert(std::make_pair(User,
8391 User->getOperand(0).getValueType()));
8392 } else if (User->getOpcode() == ISD::SELECT_CC) {
8393 if (User->getOperand(0) == PromOps[i])
8394 SelectTruncOp[0].insert(std::make_pair(User,
8395 User->getOperand(0).getValueType()));
8396 if (User->getOperand(1) == PromOps[i])
8397 SelectTruncOp[1].insert(std::make_pair(User,
8398 User->getOperand(1).getValueType()));
8403 unsigned PromBits = N->getOperand(0).getValueSizeInBits();
8404 bool ReallyNeedsExt = false;
8405 if (N->getOpcode() != ISD::ANY_EXTEND) {
8406 // If all of the inputs are not already sign/zero extended, then
8407 // we'll still need to do that at the end.
8408 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
8409 if (isa<ConstantSDNode>(Inputs[i]))
8413 Inputs[i].getOperand(0).getValueSizeInBits();
8414 assert(PromBits < OpBits && "Truncation not to a smaller bit count?");
8416 if ((N->getOpcode() == ISD::ZERO_EXTEND &&
8417 !DAG.MaskedValueIsZero(Inputs[i].getOperand(0),
8418 APInt::getHighBitsSet(OpBits,
8419 OpBits-PromBits))) ||
8420 (N->getOpcode() == ISD::SIGN_EXTEND &&
8421 DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) <
8422 (OpBits-(PromBits-1)))) {
8423 ReallyNeedsExt = true;
8429 // Replace all inputs, either with the truncation operand, or a
8430 // truncation or extension to the final output type.
8431 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
8432 // Constant inputs need to be replaced with the to-be-promoted nodes that
8433 // use them because they might have users outside of the cluster of
8435 if (isa<ConstantSDNode>(Inputs[i]))
8438 SDValue InSrc = Inputs[i].getOperand(0);
8439 if (Inputs[i].getValueType() == N->getValueType(0))
8440 DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc);
8441 else if (N->getOpcode() == ISD::SIGN_EXTEND)
8442 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
8443 DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0)));
8444 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8445 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
8446 DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0)));
8448 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
8449 DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0)));
8452 // Replace all operations (these are all the same, but have a different
8453 // (promoted) return type). DAG.getNode will validate that the types of
8454 // a binary operator match, so go through the list in reverse so that
8455 // we've likely promoted both operands first.
8456 while (!PromOps.empty()) {
8457 SDValue PromOp = PromOps.back();
8461 switch (PromOp.getOpcode()) {
8462 default: C = 0; break;
8463 case ISD::SELECT: C = 1; break;
8464 case ISD::SELECT_CC: C = 2; break;
8467 if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
8468 PromOp.getOperand(C).getValueType() != N->getValueType(0)) ||
8469 (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
8470 PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) {
8471 // The to-be-promoted operands of this node have not yet been
8472 // promoted (this should be rare because we're going through the
8473 // list backward, but if one of the operands has several users in
8474 // this cluster of to-be-promoted nodes, it is possible).
8475 PromOps.insert(PromOps.begin(), PromOp);
8479 // For SELECT and SELECT_CC nodes, we do a similar check for any
8480 // to-be-promoted comparison inputs.
8481 if (PromOp.getOpcode() == ISD::SELECT ||
8482 PromOp.getOpcode() == ISD::SELECT_CC) {
8483 if ((SelectTruncOp[0].count(PromOp.getNode()) &&
8484 PromOp.getOperand(0).getValueType() != N->getValueType(0)) ||
8485 (SelectTruncOp[1].count(PromOp.getNode()) &&
8486 PromOp.getOperand(1).getValueType() != N->getValueType(0))) {
8487 PromOps.insert(PromOps.begin(), PromOp);
8492 SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
8493 PromOp.getNode()->op_end());
8495 // If this node has constant inputs, then they'll need to be promoted here.
8496 for (unsigned i = 0; i < 2; ++i) {
8497 if (!isa<ConstantSDNode>(Ops[C+i]))
8499 if (Ops[C+i].getValueType() == N->getValueType(0))
8502 if (N->getOpcode() == ISD::SIGN_EXTEND)
8503 Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
8504 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8505 Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
8507 Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
8510 // If we've promoted the comparison inputs of a SELECT or SELECT_CC,
8511 // truncate them again to the original value type.
8512 if (PromOp.getOpcode() == ISD::SELECT ||
8513 PromOp.getOpcode() == ISD::SELECT_CC) {
8514 auto SI0 = SelectTruncOp[0].find(PromOp.getNode());
8515 if (SI0 != SelectTruncOp[0].end())
8516 Ops[0] = DAG.getNode(ISD::TRUNCATE, dl, SI0->second, Ops[0]);
8517 auto SI1 = SelectTruncOp[1].find(PromOp.getNode());
8518 if (SI1 != SelectTruncOp[1].end())
8519 Ops[1] = DAG.getNode(ISD::TRUNCATE, dl, SI1->second, Ops[1]);
8522 DAG.ReplaceAllUsesOfValueWith(PromOp,
8523 DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops));
8526 // Now we're left with the initial extension itself.
8527 if (!ReallyNeedsExt)
8528 return N->getOperand(0);
8530 // To zero extend, just mask off everything except for the first bit (in the
8532 if (N->getOpcode() == ISD::ZERO_EXTEND)
8533 return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0),
8534 DAG.getConstant(APInt::getLowBitsSet(
8535 N->getValueSizeInBits(0), PromBits),
8536 N->getValueType(0)));
8538 assert(N->getOpcode() == ISD::SIGN_EXTEND &&
8539 "Invalid extension type");
8540 EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0));
8542 DAG.getConstant(N->getValueSizeInBits(0)-PromBits, ShiftAmountTy);
8543 return DAG.getNode(ISD::SRA, dl, N->getValueType(0),
8544 DAG.getNode(ISD::SHL, dl, N->getValueType(0),
8545 N->getOperand(0), ShiftCst), ShiftCst);
8548 SDValue PPCTargetLowering::combineFPToIntToFP(SDNode *N,
8549 DAGCombinerInfo &DCI) const {
8550 assert((N->getOpcode() == ISD::SINT_TO_FP ||
8551 N->getOpcode() == ISD::UINT_TO_FP) &&
8552 "Need an int -> FP conversion node here");
8554 if (!Subtarget.has64BitSupport())
8557 SelectionDAG &DAG = DCI.DAG;
8561 // Don't handle ppc_fp128 here or i1 conversions.
8562 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
8564 if (Op.getOperand(0).getValueType() == MVT::i1)
8567 // For i32 intermediate values, unfortunately, the conversion functions
8568 // leave the upper 32 bits of the value are undefined. Within the set of
8569 // scalar instructions, we have no method for zero- or sign-extending the
8570 // value. Thus, we cannot handle i32 intermediate values here.
8571 if (Op.getOperand(0).getValueType() == MVT::i32)
8574 assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
8575 "UINT_TO_FP is supported only with FPCVT");
8577 // If we have FCFIDS, then use it when converting to single-precision.
8578 // Otherwise, convert to double-precision and then round.
8579 unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
8580 (Op.getOpcode() == ISD::UINT_TO_FP ?
8581 PPCISD::FCFIDUS : PPCISD::FCFIDS) :
8582 (Op.getOpcode() == ISD::UINT_TO_FP ?
8583 PPCISD::FCFIDU : PPCISD::FCFID);
8584 MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) ?
8585 MVT::f32 : MVT::f64;
8587 // If we're converting from a float, to an int, and back to a float again,
8588 // then we don't need the store/load pair at all.
8589 if ((Op.getOperand(0).getOpcode() == ISD::FP_TO_UINT &&
8590 Subtarget.hasFPCVT()) ||
8591 (Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT)) {
8592 SDValue Src = Op.getOperand(0).getOperand(0);
8593 if (Src.getValueType() == MVT::f32) {
8594 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
8595 DCI.AddToWorklist(Src.getNode());
8599 Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
8602 SDValue Tmp = DAG.getNode(FCTOp, dl, MVT::f64, Src);
8603 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Tmp);
8605 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) {
8606 FP = DAG.getNode(ISD::FP_ROUND, dl,
8607 MVT::f32, FP, DAG.getIntPtrConstant(0));
8608 DCI.AddToWorklist(FP.getNode());
8617 // expandVSXLoadForLE - Convert VSX loads (which may be intrinsics for
8618 // builtins) into loads with swaps.
8619 SDValue PPCTargetLowering::expandVSXLoadForLE(SDNode *N,
8620 DAGCombinerInfo &DCI) const {
8621 SelectionDAG &DAG = DCI.DAG;
8625 MachineMemOperand *MMO;
8627 switch (N->getOpcode()) {
8629 llvm_unreachable("Unexpected opcode for little endian VSX load");
8631 LoadSDNode *LD = cast<LoadSDNode>(N);
8632 Chain = LD->getChain();
8633 Base = LD->getBasePtr();
8634 MMO = LD->getMemOperand();
8635 // If the MMO suggests this isn't a load of a full vector, leave
8636 // things alone. For a built-in, we have to make the change for
8637 // correctness, so if there is a size problem that will be a bug.
8638 if (MMO->getSize() < 16)
8642 case ISD::INTRINSIC_W_CHAIN: {
8643 MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
8644 Chain = Intrin->getChain();
8645 Base = Intrin->getBasePtr();
8646 MMO = Intrin->getMemOperand();
8651 MVT VecTy = N->getValueType(0).getSimpleVT();
8652 SDValue LoadOps[] = { Chain, Base };
8653 SDValue Load = DAG.getMemIntrinsicNode(PPCISD::LXVD2X, dl,
8654 DAG.getVTList(VecTy, MVT::Other),
8655 LoadOps, VecTy, MMO);
8656 DCI.AddToWorklist(Load.getNode());
8657 Chain = Load.getValue(1);
8658 SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl,
8659 DAG.getVTList(VecTy, MVT::Other), Chain, Load);
8660 DCI.AddToWorklist(Swap.getNode());
8664 // expandVSXStoreForLE - Convert VSX stores (which may be intrinsics for
8665 // builtins) into stores with swaps.
8666 SDValue PPCTargetLowering::expandVSXStoreForLE(SDNode *N,
8667 DAGCombinerInfo &DCI) const {
8668 SelectionDAG &DAG = DCI.DAG;
8673 MachineMemOperand *MMO;
8675 switch (N->getOpcode()) {
8677 llvm_unreachable("Unexpected opcode for little endian VSX store");
8679 StoreSDNode *ST = cast<StoreSDNode>(N);
8680 Chain = ST->getChain();
8681 Base = ST->getBasePtr();
8682 MMO = ST->getMemOperand();
8684 // If the MMO suggests this isn't a store of a full vector, leave
8685 // things alone. For a built-in, we have to make the change for
8686 // correctness, so if there is a size problem that will be a bug.
8687 if (MMO->getSize() < 16)
8691 case ISD::INTRINSIC_VOID: {
8692 MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
8693 Chain = Intrin->getChain();
8694 // Intrin->getBasePtr() oddly does not get what we want.
8695 Base = Intrin->getOperand(3);
8696 MMO = Intrin->getMemOperand();
8702 SDValue Src = N->getOperand(SrcOpnd);
8703 MVT VecTy = Src.getValueType().getSimpleVT();
8704 SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl,
8705 DAG.getVTList(VecTy, MVT::Other), Chain, Src);
8706 DCI.AddToWorklist(Swap.getNode());
8707 Chain = Swap.getValue(1);
8708 SDValue StoreOps[] = { Chain, Swap, Base };
8709 SDValue Store = DAG.getMemIntrinsicNode(PPCISD::STXVD2X, dl,
8710 DAG.getVTList(MVT::Other),
8711 StoreOps, VecTy, MMO);
8712 DCI.AddToWorklist(Store.getNode());
8716 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
8717 DAGCombinerInfo &DCI) const {
8718 const TargetMachine &TM = getTargetMachine();
8719 SelectionDAG &DAG = DCI.DAG;
8721 switch (N->getOpcode()) {
8724 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8725 if (C->isNullValue()) // 0 << V -> 0.
8726 return N->getOperand(0);
8730 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8731 if (C->isNullValue()) // 0 >>u V -> 0.
8732 return N->getOperand(0);
8736 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8737 if (C->isNullValue() || // 0 >>s V -> 0.
8738 C->isAllOnesValue()) // -1 >>s V -> -1.
8739 return N->getOperand(0);
8742 case ISD::SIGN_EXTEND:
8743 case ISD::ZERO_EXTEND:
8744 case ISD::ANY_EXTEND:
8745 return DAGCombineExtBoolTrunc(N, DCI);
8748 case ISD::SELECT_CC:
8749 return DAGCombineTruncBoolExt(N, DCI);
8750 case ISD::SINT_TO_FP:
8751 case ISD::UINT_TO_FP:
8752 return combineFPToIntToFP(N, DCI);
8754 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
8755 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
8756 !cast<StoreSDNode>(N)->isTruncatingStore() &&
8757 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
8758 N->getOperand(1).getValueType() == MVT::i32 &&
8759 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
8760 SDValue Val = N->getOperand(1).getOperand(0);
8761 if (Val.getValueType() == MVT::f32) {
8762 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
8763 DCI.AddToWorklist(Val.getNode());
8765 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
8766 DCI.AddToWorklist(Val.getNode());
8769 N->getOperand(0), Val, N->getOperand(2),
8770 DAG.getValueType(N->getOperand(1).getValueType())
8773 Val = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
8774 DAG.getVTList(MVT::Other), Ops,
8775 cast<StoreSDNode>(N)->getMemoryVT(),
8776 cast<StoreSDNode>(N)->getMemOperand());
8777 DCI.AddToWorklist(Val.getNode());
8781 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
8782 if (cast<StoreSDNode>(N)->isUnindexed() &&
8783 N->getOperand(1).getOpcode() == ISD::BSWAP &&
8784 N->getOperand(1).getNode()->hasOneUse() &&
8785 (N->getOperand(1).getValueType() == MVT::i32 ||
8786 N->getOperand(1).getValueType() == MVT::i16 ||
8787 (TM.getSubtarget<PPCSubtarget>().hasLDBRX() &&
8788 TM.getSubtarget<PPCSubtarget>().isPPC64() &&
8789 N->getOperand(1).getValueType() == MVT::i64))) {
8790 SDValue BSwapOp = N->getOperand(1).getOperand(0);
8791 // Do an any-extend to 32-bits if this is a half-word input.
8792 if (BSwapOp.getValueType() == MVT::i16)
8793 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
8796 N->getOperand(0), BSwapOp, N->getOperand(2),
8797 DAG.getValueType(N->getOperand(1).getValueType())
8800 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
8801 Ops, cast<StoreSDNode>(N)->getMemoryVT(),
8802 cast<StoreSDNode>(N)->getMemOperand());
8805 // For little endian, VSX stores require generating xxswapd/lxvd2x.
8806 EVT VT = N->getOperand(1).getValueType();
8807 if (VT.isSimple()) {
8808 MVT StoreVT = VT.getSimpleVT();
8809 if (TM.getSubtarget<PPCSubtarget>().hasVSX() &&
8810 TM.getSubtarget<PPCSubtarget>().isLittleEndian() &&
8811 (StoreVT == MVT::v2f64 || StoreVT == MVT::v2i64 ||
8812 StoreVT == MVT::v4f32 || StoreVT == MVT::v4i32))
8813 return expandVSXStoreForLE(N, DCI);
8818 LoadSDNode *LD = cast<LoadSDNode>(N);
8819 EVT VT = LD->getValueType(0);
8821 // For little endian, VSX loads require generating lxvd2x/xxswapd.
8822 if (VT.isSimple()) {
8823 MVT LoadVT = VT.getSimpleVT();
8824 if (TM.getSubtarget<PPCSubtarget>().hasVSX() &&
8825 TM.getSubtarget<PPCSubtarget>().isLittleEndian() &&
8826 (LoadVT == MVT::v2f64 || LoadVT == MVT::v2i64 ||
8827 LoadVT == MVT::v4f32 || LoadVT == MVT::v4i32))
8828 return expandVSXLoadForLE(N, DCI);
8831 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
8832 unsigned ABIAlignment = getDataLayout()->getABITypeAlignment(Ty);
8833 if (ISD::isNON_EXTLoad(N) && VT.isVector() &&
8834 TM.getSubtarget<PPCSubtarget>().hasAltivec() &&
8835 // P8 and later hardware should just use LOAD.
8836 !TM.getSubtarget<PPCSubtarget>().hasP8Vector() &&
8837 (VT == MVT::v16i8 || VT == MVT::v8i16 ||
8838 VT == MVT::v4i32 || VT == MVT::v4f32) &&
8839 LD->getAlignment() < ABIAlignment) {
8840 // This is a type-legal unaligned Altivec load.
8841 SDValue Chain = LD->getChain();
8842 SDValue Ptr = LD->getBasePtr();
8843 bool isLittleEndian = Subtarget.isLittleEndian();
8845 // This implements the loading of unaligned vectors as described in
8846 // the venerable Apple Velocity Engine overview. Specifically:
8847 // https://developer.apple.com/hardwaredrivers/ve/alignment.html
8848 // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html
8850 // The general idea is to expand a sequence of one or more unaligned
8851 // loads into an alignment-based permutation-control instruction (lvsl
8852 // or lvsr), a series of regular vector loads (which always truncate
8853 // their input address to an aligned address), and a series of
8854 // permutations. The results of these permutations are the requested
8855 // loaded values. The trick is that the last "extra" load is not taken
8856 // from the address you might suspect (sizeof(vector) bytes after the
8857 // last requested load), but rather sizeof(vector) - 1 bytes after the
8858 // last requested vector. The point of this is to avoid a page fault if
8859 // the base address happened to be aligned. This works because if the
8860 // base address is aligned, then adding less than a full vector length
8861 // will cause the last vector in the sequence to be (re)loaded.
8862 // Otherwise, the next vector will be fetched as you might suspect was
8865 // We might be able to reuse the permutation generation from
8866 // a different base address offset from this one by an aligned amount.
8867 // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this
8868 // optimization later.
8869 Intrinsic::ID Intr = (isLittleEndian ?
8870 Intrinsic::ppc_altivec_lvsr :
8871 Intrinsic::ppc_altivec_lvsl);
8872 SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, MVT::v16i8);
8874 // Create the new MMO for the new base load. It is like the original MMO,
8875 // but represents an area in memory almost twice the vector size centered
8876 // on the original address. If the address is unaligned, we might start
8877 // reading up to (sizeof(vector)-1) bytes below the address of the
8878 // original unaligned load.
8879 MachineFunction &MF = DAG.getMachineFunction();
8880 MachineMemOperand *BaseMMO =
8881 MF.getMachineMemOperand(LD->getMemOperand(),
8882 -LD->getMemoryVT().getStoreSize()+1,
8883 2*LD->getMemoryVT().getStoreSize()-1);
8885 // Create the new base load.
8886 SDValue LDXIntID = DAG.getTargetConstant(Intrinsic::ppc_altivec_lvx,
8888 SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr };
8890 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
8891 DAG.getVTList(MVT::v4i32, MVT::Other),
8892 BaseLoadOps, MVT::v4i32, BaseMMO);
8894 // Note that the value of IncOffset (which is provided to the next
8895 // load's pointer info offset value, and thus used to calculate the
8896 // alignment), and the value of IncValue (which is actually used to
8897 // increment the pointer value) are different! This is because we
8898 // require the next load to appear to be aligned, even though it
8899 // is actually offset from the base pointer by a lesser amount.
8900 int IncOffset = VT.getSizeInBits() / 8;
8901 int IncValue = IncOffset;
8903 // Walk (both up and down) the chain looking for another load at the real
8904 // (aligned) offset (the alignment of the other load does not matter in
8905 // this case). If found, then do not use the offset reduction trick, as
8906 // that will prevent the loads from being later combined (as they would
8907 // otherwise be duplicates).
8908 if (!findConsecutiveLoad(LD, DAG))
8911 SDValue Increment = DAG.getConstant(IncValue, getPointerTy());
8912 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
8914 MachineMemOperand *ExtraMMO =
8915 MF.getMachineMemOperand(LD->getMemOperand(),
8916 1, 2*LD->getMemoryVT().getStoreSize()-1);
8917 SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr };
8919 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
8920 DAG.getVTList(MVT::v4i32, MVT::Other),
8921 ExtraLoadOps, MVT::v4i32, ExtraMMO);
8923 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
8924 BaseLoad.getValue(1), ExtraLoad.getValue(1));
8926 // Because vperm has a big-endian bias, we must reverse the order
8927 // of the input vectors and complement the permute control vector
8928 // when generating little endian code. We have already handled the
8929 // latter by using lvsr instead of lvsl, so just reverse BaseLoad
8930 // and ExtraLoad here.
8933 Perm = BuildIntrinsicOp(Intrinsic::ppc_altivec_vperm,
8934 ExtraLoad, BaseLoad, PermCntl, DAG, dl);
8936 Perm = BuildIntrinsicOp(Intrinsic::ppc_altivec_vperm,
8937 BaseLoad, ExtraLoad, PermCntl, DAG, dl);
8939 if (VT != MVT::v4i32)
8940 Perm = DAG.getNode(ISD::BITCAST, dl, VT, Perm);
8942 // The output of the permutation is our loaded result, the TokenFactor is
8944 DCI.CombineTo(N, Perm, TF);
8945 return SDValue(N, 0);
8949 case ISD::INTRINSIC_WO_CHAIN: {
8950 bool isLittleEndian = Subtarget.isLittleEndian();
8951 Intrinsic::ID Intr = (isLittleEndian ?
8952 Intrinsic::ppc_altivec_lvsr :
8953 Intrinsic::ppc_altivec_lvsl);
8954 if (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() == Intr &&
8955 N->getOperand(1)->getOpcode() == ISD::ADD) {
8956 SDValue Add = N->getOperand(1);
8958 if (DAG.MaskedValueIsZero(Add->getOperand(1),
8959 APInt::getAllOnesValue(4 /* 16 byte alignment */).zext(
8960 Add.getValueType().getScalarType().getSizeInBits()))) {
8961 SDNode *BasePtr = Add->getOperand(0).getNode();
8962 for (SDNode::use_iterator UI = BasePtr->use_begin(),
8963 UE = BasePtr->use_end(); UI != UE; ++UI) {
8964 if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
8965 cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() ==
8967 // We've found another LVSL/LVSR, and this address is an aligned
8968 // multiple of that one. The results will be the same, so use the
8969 // one we've just found instead.
8971 return SDValue(*UI, 0);
8979 case ISD::INTRINSIC_W_CHAIN: {
8980 // For little endian, VSX loads require generating lxvd2x/xxswapd.
8981 if (TM.getSubtarget<PPCSubtarget>().hasVSX() &&
8982 TM.getSubtarget<PPCSubtarget>().isLittleEndian()) {
8983 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8986 case Intrinsic::ppc_vsx_lxvw4x:
8987 case Intrinsic::ppc_vsx_lxvd2x:
8988 return expandVSXLoadForLE(N, DCI);
8993 case ISD::INTRINSIC_VOID: {
8994 // For little endian, VSX stores require generating xxswapd/stxvd2x.
8995 if (TM.getSubtarget<PPCSubtarget>().hasVSX() &&
8996 TM.getSubtarget<PPCSubtarget>().isLittleEndian()) {
8997 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9000 case Intrinsic::ppc_vsx_stxvw4x:
9001 case Intrinsic::ppc_vsx_stxvd2x:
9002 return expandVSXStoreForLE(N, DCI);
9008 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
9009 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
9010 N->getOperand(0).hasOneUse() &&
9011 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 ||
9012 (TM.getSubtarget<PPCSubtarget>().hasLDBRX() &&
9013 TM.getSubtarget<PPCSubtarget>().isPPC64() &&
9014 N->getValueType(0) == MVT::i64))) {
9015 SDValue Load = N->getOperand(0);
9016 LoadSDNode *LD = cast<LoadSDNode>(Load);
9017 // Create the byte-swapping load.
9019 LD->getChain(), // Chain
9020 LD->getBasePtr(), // Ptr
9021 DAG.getValueType(N->getValueType(0)) // VT
9024 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
9025 DAG.getVTList(N->getValueType(0) == MVT::i64 ?
9026 MVT::i64 : MVT::i32, MVT::Other),
9027 Ops, LD->getMemoryVT(), LD->getMemOperand());
9029 // If this is an i16 load, insert the truncate.
9030 SDValue ResVal = BSLoad;
9031 if (N->getValueType(0) == MVT::i16)
9032 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
9034 // First, combine the bswap away. This makes the value produced by the
9036 DCI.CombineTo(N, ResVal);
9038 // Next, combine the load away, we give it a bogus result value but a real
9039 // chain result. The result value is dead because the bswap is dead.
9040 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
9042 // Return N so it doesn't get rechecked!
9043 return SDValue(N, 0);
9047 case PPCISD::VCMP: {
9048 // If a VCMPo node already exists with exactly the same operands as this
9049 // node, use its result instead of this node (VCMPo computes both a CR6 and
9050 // a normal output).
9052 if (!N->getOperand(0).hasOneUse() &&
9053 !N->getOperand(1).hasOneUse() &&
9054 !N->getOperand(2).hasOneUse()) {
9056 // Scan all of the users of the LHS, looking for VCMPo's that match.
9057 SDNode *VCMPoNode = nullptr;
9059 SDNode *LHSN = N->getOperand(0).getNode();
9060 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
9062 if (UI->getOpcode() == PPCISD::VCMPo &&
9063 UI->getOperand(1) == N->getOperand(1) &&
9064 UI->getOperand(2) == N->getOperand(2) &&
9065 UI->getOperand(0) == N->getOperand(0)) {
9070 // If there is no VCMPo node, or if the flag value has a single use, don't
9072 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
9075 // Look at the (necessarily single) use of the flag value. If it has a
9076 // chain, this transformation is more complex. Note that multiple things
9077 // could use the value result, which we should ignore.
9078 SDNode *FlagUser = nullptr;
9079 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
9080 FlagUser == nullptr; ++UI) {
9081 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
9083 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
9084 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
9091 // If the user is a MFOCRF instruction, we know this is safe.
9092 // Otherwise we give up for right now.
9093 if (FlagUser->getOpcode() == PPCISD::MFOCRF)
9094 return SDValue(VCMPoNode, 0);
9099 SDValue Cond = N->getOperand(1);
9100 SDValue Target = N->getOperand(2);
9102 if (Cond.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
9103 cast<ConstantSDNode>(Cond.getOperand(1))->getZExtValue() ==
9104 Intrinsic::ppc_is_decremented_ctr_nonzero) {
9106 // We now need to make the intrinsic dead (it cannot be instruction
9108 DAG.ReplaceAllUsesOfValueWith(Cond.getValue(1), Cond.getOperand(0));
9109 assert(Cond.getNode()->hasOneUse() &&
9110 "Counter decrement has more than one use");
9112 return DAG.getNode(PPCISD::BDNZ, dl, MVT::Other,
9113 N->getOperand(0), Target);
9118 // If this is a branch on an altivec predicate comparison, lower this so
9119 // that we don't have to do a MFOCRF: instead, branch directly on CR6. This
9120 // lowering is done pre-legalize, because the legalizer lowers the predicate
9121 // compare down to code that is difficult to reassemble.
9122 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
9123 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
9125 // Sometimes the promoted value of the intrinsic is ANDed by some non-zero
9126 // value. If so, pass-through the AND to get to the intrinsic.
9127 if (LHS.getOpcode() == ISD::AND &&
9128 LHS.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN &&
9129 cast<ConstantSDNode>(LHS.getOperand(0).getOperand(1))->getZExtValue() ==
9130 Intrinsic::ppc_is_decremented_ctr_nonzero &&
9131 isa<ConstantSDNode>(LHS.getOperand(1)) &&
9132 !cast<ConstantSDNode>(LHS.getOperand(1))->getConstantIntValue()->
9134 LHS = LHS.getOperand(0);
9136 if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
9137 cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() ==
9138 Intrinsic::ppc_is_decremented_ctr_nonzero &&
9139 isa<ConstantSDNode>(RHS)) {
9140 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
9141 "Counter decrement comparison is not EQ or NE");
9143 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
9144 bool isBDNZ = (CC == ISD::SETEQ && Val) ||
9145 (CC == ISD::SETNE && !Val);
9147 // We now need to make the intrinsic dead (it cannot be instruction
9149 DAG.ReplaceAllUsesOfValueWith(LHS.getValue(1), LHS.getOperand(0));
9150 assert(LHS.getNode()->hasOneUse() &&
9151 "Counter decrement has more than one use");
9153 return DAG.getNode(isBDNZ ? PPCISD::BDNZ : PPCISD::BDZ, dl, MVT::Other,
9154 N->getOperand(0), N->getOperand(4));
9160 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
9161 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
9162 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
9163 assert(isDot && "Can't compare against a vector result!");
9165 // If this is a comparison against something other than 0/1, then we know
9166 // that the condition is never/always true.
9167 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
9168 if (Val != 0 && Val != 1) {
9169 if (CC == ISD::SETEQ) // Cond never true, remove branch.
9170 return N->getOperand(0);
9171 // Always !=, turn it into an unconditional branch.
9172 return DAG.getNode(ISD::BR, dl, MVT::Other,
9173 N->getOperand(0), N->getOperand(4));
9176 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
9178 // Create the PPCISD altivec 'dot' comparison node.
9180 LHS.getOperand(2), // LHS of compare
9181 LHS.getOperand(3), // RHS of compare
9182 DAG.getConstant(CompareOpc, MVT::i32)
9184 EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
9185 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops);
9187 // Unpack the result based on how the target uses it.
9188 PPC::Predicate CompOpc;
9189 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
9190 default: // Can't happen, don't crash on invalid number though.
9191 case 0: // Branch on the value of the EQ bit of CR6.
9192 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
9194 case 1: // Branch on the inverted value of the EQ bit of CR6.
9195 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
9197 case 2: // Branch on the value of the LT bit of CR6.
9198 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
9200 case 3: // Branch on the inverted value of the LT bit of CR6.
9201 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
9205 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
9206 DAG.getConstant(CompOpc, MVT::i32),
9207 DAG.getRegister(PPC::CR6, MVT::i32),
9208 N->getOperand(4), CompNode.getValue(1));
9218 PPCTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
9220 std::vector<SDNode *> *Created) const {
9221 // fold (sdiv X, pow2)
9222 EVT VT = N->getValueType(0);
9223 if (VT == MVT::i64 && !Subtarget.isPPC64())
9225 if ((VT != MVT::i32 && VT != MVT::i64) ||
9226 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
9230 SDValue N0 = N->getOperand(0);
9232 bool IsNegPow2 = (-Divisor).isPowerOf2();
9233 unsigned Lg2 = (IsNegPow2 ? -Divisor : Divisor).countTrailingZeros();
9234 SDValue ShiftAmt = DAG.getConstant(Lg2, VT);
9236 SDValue Op = DAG.getNode(PPCISD::SRA_ADDZE, DL, VT, N0, ShiftAmt);
9238 Created->push_back(Op.getNode());
9241 Op = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), Op);
9243 Created->push_back(Op.getNode());
9249 //===----------------------------------------------------------------------===//
9250 // Inline Assembly Support
9251 //===----------------------------------------------------------------------===//
9253 void PPCTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
9256 const SelectionDAG &DAG,
9257 unsigned Depth) const {
9258 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
9259 switch (Op.getOpcode()) {
9261 case PPCISD::LBRX: {
9262 // lhbrx is known to have the top bits cleared out.
9263 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
9264 KnownZero = 0xFFFF0000;
9267 case ISD::INTRINSIC_WO_CHAIN: {
9268 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
9270 case Intrinsic::ppc_altivec_vcmpbfp_p:
9271 case Intrinsic::ppc_altivec_vcmpeqfp_p:
9272 case Intrinsic::ppc_altivec_vcmpequb_p:
9273 case Intrinsic::ppc_altivec_vcmpequh_p:
9274 case Intrinsic::ppc_altivec_vcmpequw_p:
9275 case Intrinsic::ppc_altivec_vcmpgefp_p:
9276 case Intrinsic::ppc_altivec_vcmpgtfp_p:
9277 case Intrinsic::ppc_altivec_vcmpgtsb_p:
9278 case Intrinsic::ppc_altivec_vcmpgtsh_p:
9279 case Intrinsic::ppc_altivec_vcmpgtsw_p:
9280 case Intrinsic::ppc_altivec_vcmpgtub_p:
9281 case Intrinsic::ppc_altivec_vcmpgtuh_p:
9282 case Intrinsic::ppc_altivec_vcmpgtuw_p:
9283 KnownZero = ~1U; // All bits but the low one are known to be zero.
9290 unsigned PPCTargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {
9291 switch (Subtarget.getDarwinDirective()) {
9296 case PPC::DIR_PWR5X:
9298 case PPC::DIR_PWR6X:
9300 case PPC::DIR_PWR8: {
9304 const PPCInstrInfo *TII =
9305 static_cast<const PPCInstrInfo *>(getTargetMachine().getSubtargetImpl()->
9308 // For small loops (between 5 and 8 instructions), align to a 32-byte
9309 // boundary so that the entire loop fits in one instruction-cache line.
9310 uint64_t LoopSize = 0;
9311 for (auto I = ML->block_begin(), IE = ML->block_end(); I != IE; ++I)
9312 for (auto J = (*I)->begin(), JE = (*I)->end(); J != JE; ++J)
9313 LoopSize += TII->GetInstSizeInBytes(J);
9315 if (LoopSize > 16 && LoopSize <= 32)
9322 return TargetLowering::getPrefLoopAlignment(ML);
9325 /// getConstraintType - Given a constraint, return the type of
9326 /// constraint it is for this target.
9327 PPCTargetLowering::ConstraintType
9328 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
9329 if (Constraint.size() == 1) {
9330 switch (Constraint[0]) {
9337 return C_RegisterClass;
9339 // FIXME: While Z does indicate a memory constraint, it specifically
9340 // indicates an r+r address (used in conjunction with the 'y' modifier
9341 // in the replacement string). Currently, we're forcing the base
9342 // register to be r0 in the asm printer (which is interpreted as zero)
9343 // and forming the complete address in the second register. This is
9347 } else if (Constraint == "wc") { // individual CR bits.
9348 return C_RegisterClass;
9349 } else if (Constraint == "wa" || Constraint == "wd" ||
9350 Constraint == "wf" || Constraint == "ws") {
9351 return C_RegisterClass; // VSX registers.
9353 return TargetLowering::getConstraintType(Constraint);
9356 /// Examine constraint type and operand type and determine a weight value.
9357 /// This object must already have been set up with the operand type
9358 /// and the current alternative constraint selected.
9359 TargetLowering::ConstraintWeight
9360 PPCTargetLowering::getSingleConstraintMatchWeight(
9361 AsmOperandInfo &info, const char *constraint) const {
9362 ConstraintWeight weight = CW_Invalid;
9363 Value *CallOperandVal = info.CallOperandVal;
9364 // If we don't have a value, we can't do a match,
9365 // but allow it at the lowest weight.
9366 if (!CallOperandVal)
9368 Type *type = CallOperandVal->getType();
9370 // Look at the constraint type.
9371 if (StringRef(constraint) == "wc" && type->isIntegerTy(1))
9372 return CW_Register; // an individual CR bit.
9373 else if ((StringRef(constraint) == "wa" ||
9374 StringRef(constraint) == "wd" ||
9375 StringRef(constraint) == "wf") &&
9378 else if (StringRef(constraint) == "ws" && type->isDoubleTy())
9381 switch (*constraint) {
9383 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9386 if (type->isIntegerTy())
9387 weight = CW_Register;
9390 if (type->isFloatTy())
9391 weight = CW_Register;
9394 if (type->isDoubleTy())
9395 weight = CW_Register;
9398 if (type->isVectorTy())
9399 weight = CW_Register;
9402 weight = CW_Register;
9411 std::pair<unsigned, const TargetRegisterClass*>
9412 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9414 if (Constraint.size() == 1) {
9415 // GCC RS6000 Constraint Letters
9416 switch (Constraint[0]) {
9418 if (VT == MVT::i64 && Subtarget.isPPC64())
9419 return std::make_pair(0U, &PPC::G8RC_NOX0RegClass);
9420 return std::make_pair(0U, &PPC::GPRC_NOR0RegClass);
9422 if (VT == MVT::i64 && Subtarget.isPPC64())
9423 return std::make_pair(0U, &PPC::G8RCRegClass);
9424 return std::make_pair(0U, &PPC::GPRCRegClass);
9426 if (VT == MVT::f32 || VT == MVT::i32)
9427 return std::make_pair(0U, &PPC::F4RCRegClass);
9428 if (VT == MVT::f64 || VT == MVT::i64)
9429 return std::make_pair(0U, &PPC::F8RCRegClass);
9432 return std::make_pair(0U, &PPC::VRRCRegClass);
9434 return std::make_pair(0U, &PPC::CRRCRegClass);
9436 } else if (Constraint == "wc") { // an individual CR bit.
9437 return std::make_pair(0U, &PPC::CRBITRCRegClass);
9438 } else if (Constraint == "wa" || Constraint == "wd" ||
9439 Constraint == "wf") {
9440 return std::make_pair(0U, &PPC::VSRCRegClass);
9441 } else if (Constraint == "ws") {
9442 return std::make_pair(0U, &PPC::VSFRCRegClass);
9445 std::pair<unsigned, const TargetRegisterClass*> R =
9446 TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9448 // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers
9449 // (which we call X[0-9]+). If a 64-bit value has been requested, and a
9450 // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent
9452 // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use
9453 // the AsmName field from *RegisterInfo.td, then this would not be necessary.
9454 if (R.first && VT == MVT::i64 && Subtarget.isPPC64() &&
9455 PPC::GPRCRegClass.contains(R.first)) {
9456 const TargetRegisterInfo *TRI =
9457 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
9458 return std::make_pair(TRI->getMatchingSuperReg(R.first,
9459 PPC::sub_32, &PPC::G8RCRegClass),
9460 &PPC::G8RCRegClass);
9463 // GCC accepts 'cc' as an alias for 'cr0', and we need to do the same.
9464 if (!R.second && StringRef("{cc}").equals_lower(Constraint)) {
9466 R.second = &PPC::CRRCRegClass;
9473 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9474 /// vector. If it is invalid, don't add anything to Ops.
9475 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9476 std::string &Constraint,
9477 std::vector<SDValue>&Ops,
9478 SelectionDAG &DAG) const {
9481 // Only support length 1 constraints.
9482 if (Constraint.length() > 1) return;
9484 char Letter = Constraint[0];
9495 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
9496 if (!CST) return; // Must be an immediate to match.
9497 int64_t Value = CST->getSExtValue();
9498 EVT TCVT = MVT::i64; // All constants taken to be 64 bits so that negative
9499 // numbers are printed as such.
9501 default: llvm_unreachable("Unknown constraint letter!");
9502 case 'I': // "I" is a signed 16-bit constant.
9503 if (isInt<16>(Value))
9504 Result = DAG.getTargetConstant(Value, TCVT);
9506 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
9507 if (isShiftedUInt<16, 16>(Value))
9508 Result = DAG.getTargetConstant(Value, TCVT);
9510 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
9511 if (isShiftedInt<16, 16>(Value))
9512 Result = DAG.getTargetConstant(Value, TCVT);
9514 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
9515 if (isUInt<16>(Value))
9516 Result = DAG.getTargetConstant(Value, TCVT);
9518 case 'M': // "M" is a constant that is greater than 31.
9520 Result = DAG.getTargetConstant(Value, TCVT);
9522 case 'N': // "N" is a positive constant that is an exact power of two.
9523 if (Value > 0 && isPowerOf2_64(Value))
9524 Result = DAG.getTargetConstant(Value, TCVT);
9526 case 'O': // "O" is the constant zero.
9528 Result = DAG.getTargetConstant(Value, TCVT);
9530 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
9531 if (isInt<16>(-Value))
9532 Result = DAG.getTargetConstant(Value, TCVT);
9539 if (Result.getNode()) {
9540 Ops.push_back(Result);
9544 // Handle standard constraint letters.
9545 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9548 // isLegalAddressingMode - Return true if the addressing mode represented
9549 // by AM is legal for this target, for a load/store of the specified type.
9550 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9552 // FIXME: PPC does not allow r+i addressing modes for vectors!
9554 // PPC allows a sign-extended 16-bit immediate field.
9555 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
9558 // No global is ever allowed as a base.
9562 // PPC only support r+r,
9564 case 0: // "r+i" or just "i", depending on HasBaseReg.
9567 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
9569 // Otherwise we have r+r or r+i.
9572 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
9574 // Allow 2*r as r+r.
9577 // No other scales are supported.
9584 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
9585 SelectionDAG &DAG) const {
9586 MachineFunction &MF = DAG.getMachineFunction();
9587 MachineFrameInfo *MFI = MF.getFrameInfo();
9588 MFI->setReturnAddressIsTaken(true);
9590 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
9594 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9596 // Make sure the function does not optimize away the store of the RA to
9598 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
9599 FuncInfo->setLRStoreRequired();
9600 bool isPPC64 = Subtarget.isPPC64();
9601 bool isDarwinABI = Subtarget.isDarwinABI();
9604 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
9607 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
9608 isPPC64? MVT::i64 : MVT::i32);
9609 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9610 DAG.getNode(ISD::ADD, dl, getPointerTy(),
9612 MachinePointerInfo(), false, false, false, 0);
9615 // Just load the return address off the stack.
9616 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
9617 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9618 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
9621 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
9622 SelectionDAG &DAG) const {
9624 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9626 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
9627 bool isPPC64 = PtrVT == MVT::i64;
9629 MachineFunction &MF = DAG.getMachineFunction();
9630 MachineFrameInfo *MFI = MF.getFrameInfo();
9631 MFI->setFrameAddressIsTaken(true);
9633 // Naked functions never have a frame pointer, and so we use r1. For all
9634 // other functions, this decision must be delayed until during PEI.
9636 if (MF.getFunction()->getAttributes().hasAttribute(
9637 AttributeSet::FunctionIndex, Attribute::Naked))
9638 FrameReg = isPPC64 ? PPC::X1 : PPC::R1;
9640 FrameReg = isPPC64 ? PPC::FP8 : PPC::FP;
9642 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
9645 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
9646 FrameAddr, MachinePointerInfo(), false, false,
9651 // FIXME? Maybe this could be a TableGen attribute on some registers and
9652 // this table could be generated automatically from RegInfo.
9653 unsigned PPCTargetLowering::getRegisterByName(const char* RegName,
9655 bool isPPC64 = Subtarget.isPPC64();
9656 bool isDarwinABI = Subtarget.isDarwinABI();
9658 if ((isPPC64 && VT != MVT::i64 && VT != MVT::i32) ||
9659 (!isPPC64 && VT != MVT::i32))
9660 report_fatal_error("Invalid register global variable type");
9662 bool is64Bit = isPPC64 && VT == MVT::i64;
9663 unsigned Reg = StringSwitch<unsigned>(RegName)
9664 .Case("r1", is64Bit ? PPC::X1 : PPC::R1)
9665 .Case("r2", isDarwinABI ? 0 : (is64Bit ? PPC::X2 : PPC::R2))
9666 .Case("r13", (!isPPC64 && isDarwinABI) ? 0 :
9667 (is64Bit ? PPC::X13 : PPC::R13))
9672 report_fatal_error("Invalid register name global variable");
9676 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9677 // The PowerPC target isn't yet aware of offsets.
9681 bool PPCTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9683 unsigned Intrinsic) const {
9685 switch (Intrinsic) {
9686 case Intrinsic::ppc_altivec_lvx:
9687 case Intrinsic::ppc_altivec_lvxl:
9688 case Intrinsic::ppc_altivec_lvebx:
9689 case Intrinsic::ppc_altivec_lvehx:
9690 case Intrinsic::ppc_altivec_lvewx:
9691 case Intrinsic::ppc_vsx_lxvd2x:
9692 case Intrinsic::ppc_vsx_lxvw4x: {
9694 switch (Intrinsic) {
9695 case Intrinsic::ppc_altivec_lvebx:
9698 case Intrinsic::ppc_altivec_lvehx:
9701 case Intrinsic::ppc_altivec_lvewx:
9704 case Intrinsic::ppc_vsx_lxvd2x:
9712 Info.opc = ISD::INTRINSIC_W_CHAIN;
9714 Info.ptrVal = I.getArgOperand(0);
9715 Info.offset = -VT.getStoreSize()+1;
9716 Info.size = 2*VT.getStoreSize()-1;
9719 Info.readMem = true;
9720 Info.writeMem = false;
9723 case Intrinsic::ppc_altivec_stvx:
9724 case Intrinsic::ppc_altivec_stvxl:
9725 case Intrinsic::ppc_altivec_stvebx:
9726 case Intrinsic::ppc_altivec_stvehx:
9727 case Intrinsic::ppc_altivec_stvewx:
9728 case Intrinsic::ppc_vsx_stxvd2x:
9729 case Intrinsic::ppc_vsx_stxvw4x: {
9731 switch (Intrinsic) {
9732 case Intrinsic::ppc_altivec_stvebx:
9735 case Intrinsic::ppc_altivec_stvehx:
9738 case Intrinsic::ppc_altivec_stvewx:
9741 case Intrinsic::ppc_vsx_stxvd2x:
9749 Info.opc = ISD::INTRINSIC_VOID;
9751 Info.ptrVal = I.getArgOperand(1);
9752 Info.offset = -VT.getStoreSize()+1;
9753 Info.size = 2*VT.getStoreSize()-1;
9756 Info.readMem = false;
9757 Info.writeMem = true;
9767 /// getOptimalMemOpType - Returns the target specific optimal type for load
9768 /// and store operations as a result of memset, memcpy, and memmove
9769 /// lowering. If DstAlign is zero that means it's safe to destination
9770 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
9771 /// means there isn't a need to check it against alignment requirement,
9772 /// probably because the source does not need to be loaded. If 'IsMemset' is
9773 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
9774 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
9775 /// source is constant so it does not need to be loaded.
9776 /// It returns EVT::Other if the type should be determined using generic
9777 /// target-independent logic.
9778 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
9779 unsigned DstAlign, unsigned SrcAlign,
9780 bool IsMemset, bool ZeroMemset,
9782 MachineFunction &MF) const {
9783 if (Subtarget.isPPC64()) {
9790 /// \brief Returns true if it is beneficial to convert a load of a constant
9791 /// to just the constant itself.
9792 bool PPCTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
9794 assert(Ty->isIntegerTy());
9796 unsigned BitSize = Ty->getPrimitiveSizeInBits();
9797 if (BitSize == 0 || BitSize > 64)
9802 bool PPCTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
9803 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9805 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9806 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9807 return NumBits1 == 64 && NumBits2 == 32;
9810 bool PPCTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9811 if (!VT1.isInteger() || !VT2.isInteger())
9813 unsigned NumBits1 = VT1.getSizeInBits();
9814 unsigned NumBits2 = VT2.getSizeInBits();
9815 return NumBits1 == 64 && NumBits2 == 32;
9818 bool PPCTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9819 // Generally speaking, zexts are not free, but they are free when they can be
9820 // folded with other operations.
9821 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val)) {
9822 EVT MemVT = LD->getMemoryVT();
9823 if ((MemVT == MVT::i1 || MemVT == MVT::i8 || MemVT == MVT::i16 ||
9824 (Subtarget.isPPC64() && MemVT == MVT::i32)) &&
9825 (LD->getExtensionType() == ISD::NON_EXTLOAD ||
9826 LD->getExtensionType() == ISD::ZEXTLOAD))
9830 // FIXME: Add other cases...
9831 // - 32-bit shifts with a zext to i64
9832 // - zext after ctlz, bswap, etc.
9833 // - zext after and by a constant mask
9835 return TargetLowering::isZExtFree(Val, VT2);
9838 bool PPCTargetLowering::isFPExtFree(EVT VT) const {
9839 assert(VT.isFloatingPoint());
9843 bool PPCTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9844 return isInt<16>(Imm) || isUInt<16>(Imm);
9847 bool PPCTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9848 return isInt<16>(Imm) || isUInt<16>(Imm);
9851 bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9855 if (DisablePPCUnaligned)
9858 // PowerPC supports unaligned memory access for simple non-vector types.
9859 // Although accessing unaligned addresses is not as efficient as accessing
9860 // aligned addresses, it is generally more efficient than manual expansion,
9861 // and generally only traps for software emulation when crossing page
9867 if (VT.getSimpleVT().isVector()) {
9868 if (Subtarget.hasVSX()) {
9869 if (VT != MVT::v2f64 && VT != MVT::v2i64 &&
9870 VT != MVT::v4f32 && VT != MVT::v4i32)
9877 if (VT == MVT::ppcf128)
9886 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
9887 VT = VT.getScalarType();
9892 switch (VT.getSimpleVT().SimpleTy) {
9904 PPCTargetLowering::getScratchRegisters(CallingConv::ID) const {
9905 // LR is a callee-save register, but we must treat it as clobbered by any call
9906 // site. Hence we include LR in the scratch registers, which are in turn added
9907 // as implicit-defs for stackmaps and patchpoints. The same reasoning applies
9908 // to CTR, which is used by any indirect call.
9909 static const MCPhysReg ScratchRegs[] = {
9910 PPC::X11, PPC::X12, PPC::LR8, PPC::CTR8, 0
9917 PPCTargetLowering::shouldExpandBuildVectorWithShuffles(
9918 EVT VT , unsigned DefinedValues) const {
9919 if (VT == MVT::v2i64)
9922 return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
9925 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
9926 if (DisableILPPref || Subtarget.enableMachineScheduler())
9927 return TargetLowering::getSchedulingPreference(N);
9932 // Create a fast isel object.
9934 PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo,
9935 const TargetLibraryInfo *LibInfo) const {
9936 return PPC::createFastISel(FuncInfo, LibInfo);