xref: /freebsd/contrib/llvm-project/llvm/lib/Target/PowerPC/PPCISelLowering.h (revision 9f23cbd6cae82fd77edfad7173432fa8dccd0a95)
1 //===-- PPCISelLowering.h - PPC32 DAG Lowering Interface --------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that PPC uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
15 #define LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
16 
17 #include "PPCInstrInfo.h"
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineMemOperand.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGNodes.h"
23 #include "llvm/CodeGen/TargetLowering.h"
24 #include "llvm/CodeGen/ValueTypes.h"
25 #include "llvm/IR/Attributes.h"
26 #include "llvm/IR/CallingConv.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/InlineAsm.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/MachineValueType.h"
32 #include <optional>
33 #include <utility>
34 
35 namespace llvm {
36 
37   namespace PPCISD {
38 
39     // When adding a NEW PPCISD node please add it to the correct position in
40     // the enum. The order of elements in this enum matters!
41     // Values that are added after this entry:
42     //     STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE
43     // are considered memory opcodes and are treated differently than entries
44     // that come before it. For example, ADD or MUL should be placed before
45     // the ISD::FIRST_TARGET_MEMORY_OPCODE while a LOAD or STORE should come
46     // after it.
47   enum NodeType : unsigned {
48     // Start the numbering where the builtin ops and target ops leave off.
49     FIRST_NUMBER = ISD::BUILTIN_OP_END,
50 
51     /// FSEL - Traditional three-operand fsel node.
52     ///
53     FSEL,
54 
55     /// XSMAXC[DQ]P, XSMINC[DQ]P - C-type min/max instructions.
56     XSMAXC,
57     XSMINC,
58 
59     /// FCFID - The FCFID instruction, taking an f64 operand and producing
60     /// and f64 value containing the FP representation of the integer that
61     /// was temporarily in the f64 operand.
62     FCFID,
63 
64     /// Newer FCFID[US] integer-to-floating-point conversion instructions for
65     /// unsigned integers and single-precision outputs.
66     FCFIDU,
67     FCFIDS,
68     FCFIDUS,
69 
70     /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64
71     /// operand, producing an f64 value containing the integer representation
72     /// of that FP value.
73     FCTIDZ,
74     FCTIWZ,
75 
76     /// Newer FCTI[D,W]UZ floating-point-to-integer conversion instructions for
77     /// unsigned integers with round toward zero.
78     FCTIDUZ,
79     FCTIWUZ,
80 
81     /// Floating-point-to-integer conversion instructions
82     FP_TO_UINT_IN_VSR,
83     FP_TO_SINT_IN_VSR,
84 
85     /// VEXTS, ByteWidth - takes an input in VSFRC and produces an output in
86     /// VSFRC that is sign-extended from ByteWidth to a 64-byte integer.
87     VEXTS,
88 
89     /// Reciprocal estimate instructions (unary FP ops).
90     FRE,
91     FRSQRTE,
92 
93     /// Test instruction for software square root.
94     FTSQRT,
95 
96     /// Square root instruction.
97     FSQRT,
98 
99     /// VPERM - The PPC VPERM Instruction.
100     ///
101     VPERM,
102 
103     /// XXSPLT - The PPC VSX splat instructions
104     ///
105     XXSPLT,
106 
107     /// XXSPLTI_SP_TO_DP - The PPC VSX splat instructions for immediates for
108     /// converting immediate single precision numbers to double precision
109     /// vector or scalar.
110     XXSPLTI_SP_TO_DP,
111 
112     /// XXSPLTI32DX - The PPC XXSPLTI32DX instruction.
113     ///
114     XXSPLTI32DX,
115 
116     /// VECINSERT - The PPC vector insert instruction
117     ///
118     VECINSERT,
119 
120     /// VECSHL - The PPC vector shift left instruction
121     ///
122     VECSHL,
123 
124     /// XXPERMDI - The PPC XXPERMDI instruction
125     ///
126     XXPERMDI,
127     XXPERM,
128 
129     /// The CMPB instruction (takes two operands of i32 or i64).
130     CMPB,
131 
132     /// Hi/Lo - These represent the high and low 16-bit parts of a global
133     /// address respectively.  These nodes have two operands, the first of
134     /// which must be a TargetGlobalAddress, and the second of which must be a
135     /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
136     /// though these are usually folded into other nodes.
137     Hi,
138     Lo,
139 
140     /// The following two target-specific nodes are used for calls through
141     /// function pointers in the 64-bit SVR4 ABI.
142 
143     /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
144     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
145     /// compute an allocation on the stack.
146     DYNALLOC,
147 
148     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
149     /// compute an offset from native SP to the address  of the most recent
150     /// dynamic alloca.
151     DYNAREAOFFSET,
152 
153     /// To avoid stack clash, allocation is performed by block and each block is
154     /// probed.
155     PROBED_ALLOCA,
156 
157     /// The result of the mflr at function entry, used for PIC code.
158     GlobalBaseReg,
159 
160     /// These nodes represent PPC shifts.
161     ///
162     /// For scalar types, only the last `n + 1` bits of the shift amounts
163     /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
164     /// for exact behaviors.
165     ///
166     /// For vector types, only the last n bits are used. See vsld.
167     SRL,
168     SRA,
169     SHL,
170 
171     /// FNMSUB - Negated multiply-subtract instruction.
172     FNMSUB,
173 
174     /// EXTSWSLI = The PPC extswsli instruction, which does an extend-sign
175     /// word and shift left immediate.
176     EXTSWSLI,
177 
178     /// The combination of sra[wd]i and addze used to implemented signed
179     /// integer division by a power of 2. The first operand is the dividend,
180     /// and the second is the constant shift amount (representing the
181     /// divisor).
182     SRA_ADDZE,
183 
184     /// CALL - A direct function call.
185     /// CALL_NOP is a call with the special NOP which follows 64-bit
186     /// CALL_NOTOC the caller does not use the TOC.
187     /// SVR4 calls and 32-bit/64-bit AIX calls.
188     CALL,
189     CALL_NOP,
190     CALL_NOTOC,
191 
192     /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
193     /// MTCTR instruction.
194     MTCTR,
195 
196     /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
197     /// BCTRL instruction.
198     BCTRL,
199 
200     /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
201     /// instruction and the TOC reload required on 64-bit ELF, 32-bit AIX
202     /// and 64-bit AIX.
203     BCTRL_LOAD_TOC,
204 
205     /// The variants that implicitly define rounding mode for calls with
206     /// strictfp semantics.
207     CALL_RM,
208     CALL_NOP_RM,
209     CALL_NOTOC_RM,
210     BCTRL_RM,
211     BCTRL_LOAD_TOC_RM,
212 
213     /// Return with a flag operand, matched by 'blr'
214     RET_FLAG,
215 
216     /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
217     /// This copies the bits corresponding to the specified CRREG into the
218     /// resultant GPR.  Bits corresponding to other CR regs are undefined.
219     MFOCRF,
220 
221     /// Direct move from a VSX register to a GPR
222     MFVSR,
223 
224     /// Direct move from a GPR to a VSX register (algebraic)
225     MTVSRA,
226 
227     /// Direct move from a GPR to a VSX register (zero)
228     MTVSRZ,
229 
230     /// Direct move of 2 consecutive GPR to a VSX register.
231     BUILD_FP128,
232 
233     /// BUILD_SPE64 and EXTRACT_SPE are analogous to BUILD_PAIR and
234     /// EXTRACT_ELEMENT but take f64 arguments instead of i64, as i64 is
235     /// unsupported for this target.
236     /// Merge 2 GPRs to a single SPE register.
237     BUILD_SPE64,
238 
239     /// Extract SPE register component, second argument is high or low.
240     EXTRACT_SPE,
241 
242     /// Extract a subvector from signed integer vector and convert to FP.
243     /// It is primarily used to convert a (widened) illegal integer vector
244     /// type to a legal floating point vector type.
245     /// For example v2i32 -> widened to v4i32 -> v2f64
246     SINT_VEC_TO_FP,
247 
248     /// Extract a subvector from unsigned integer vector and convert to FP.
249     /// As with SINT_VEC_TO_FP, used for converting illegal types.
250     UINT_VEC_TO_FP,
251 
252     /// PowerPC instructions that have SCALAR_TO_VECTOR semantics tend to
253     /// place the value into the least significant element of the most
254     /// significant doubleword in the vector. This is not element zero for
255     /// anything smaller than a doubleword on either endianness. This node has
256     /// the same semantics as SCALAR_TO_VECTOR except that the value remains in
257     /// the aforementioned location in the vector register.
258     SCALAR_TO_VECTOR_PERMUTED,
259 
260     // FIXME: Remove these once the ANDI glue bug is fixed:
261     /// i1 = ANDI_rec_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
262     /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
263     /// implement truncation of i32 or i64 to i1.
264     ANDI_rec_1_EQ_BIT,
265     ANDI_rec_1_GT_BIT,
266 
267     // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
268     // target (returns (Lo, Hi)). It takes a chain operand.
269     READ_TIME_BASE,
270 
271     // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
272     EH_SJLJ_SETJMP,
273 
274     // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
275     EH_SJLJ_LONGJMP,
276 
277     /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
278     /// instructions.  For lack of better number, we use the opcode number
279     /// encoding for the OPC field to identify the compare.  For example, 838
280     /// is VCMPGTSH.
281     VCMP,
282 
283     /// RESVEC, OUTFLAG = VCMP_rec(LHS, RHS, OPC) - Represents one of the
284     /// altivec VCMP*_rec instructions.  For lack of better number, we use the
285     /// opcode number encoding for the OPC field to identify the compare.  For
286     /// example, 838 is VCMPGTSH.
287     VCMP_rec,
288 
289     /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
290     /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
291     /// condition register to branch on, OPC is the branch opcode to use (e.g.
292     /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
293     /// an optional input flag argument.
294     COND_BRANCH,
295 
296     /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
297     /// loops.
298     BDNZ,
299     BDZ,
300 
301     /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
302     /// towards zero.  Used only as part of the long double-to-int
303     /// conversion sequence.
304     FADDRTZ,
305 
306     /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
307     MFFS,
308 
309     /// TC_RETURN - A tail call return.
310     ///   operand #0 chain
311     ///   operand #1 callee (register or absolute)
312     ///   operand #2 stack adjustment
313     ///   operand #3 optional in flag
314     TC_RETURN,
315 
316     /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
317     CR6SET,
318     CR6UNSET,
319 
320     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
321     /// for non-position independent code on PPC32.
322     PPC32_GOT,
323 
324     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
325     /// local dynamic TLS and position indendepent code on PPC32.
326     PPC32_PICGOT,
327 
328     /// G8RC = ADDIS_GOT_TPREL_HA %x2, Symbol - Used by the initial-exec
329     /// TLS model, produces an ADDIS8 instruction that adds the GOT
330     /// base to sym\@got\@tprel\@ha.
331     ADDIS_GOT_TPREL_HA,
332 
333     /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
334     /// TLS model, produces a LD instruction with base register G8RReg
335     /// and offset sym\@got\@tprel\@l.  This completes the addition that
336     /// finds the offset of "sym" relative to the thread pointer.
337     LD_GOT_TPREL_L,
338 
339     /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
340     /// model, produces an ADD instruction that adds the contents of
341     /// G8RReg to the thread pointer.  Symbol contains a relocation
342     /// sym\@tls which is to be replaced by the thread pointer and
343     /// identifies to the linker that the instruction is part of a
344     /// TLS sequence.
345     ADD_TLS,
346 
347     /// G8RC = ADDIS_TLSGD_HA %x2, Symbol - For the general-dynamic TLS
348     /// model, produces an ADDIS8 instruction that adds the GOT base
349     /// register to sym\@got\@tlsgd\@ha.
350     ADDIS_TLSGD_HA,
351 
352     /// %x3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
353     /// model, produces an ADDI8 instruction that adds G8RReg to
354     /// sym\@got\@tlsgd\@l and stores the result in X3.  Hidden by
355     /// ADDIS_TLSGD_L_ADDR until after register assignment.
356     ADDI_TLSGD_L,
357 
358     /// %x3 = GET_TLS_ADDR %x3, Symbol - For the general-dynamic TLS
359     /// model, produces a call to __tls_get_addr(sym\@tlsgd).  Hidden by
360     /// ADDIS_TLSGD_L_ADDR until after register assignment.
361     GET_TLS_ADDR,
362 
363     /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
364     /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
365     /// register assignment.
366     ADDI_TLSGD_L_ADDR,
367 
368     /// GPRC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
369     /// G8RC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
370     /// Op that combines two register copies of TOC entries
371     /// (region handle into R3 and variable offset into R4) followed by a
372     /// GET_TLS_ADDR node which will be expanded to a call to __get_tls_addr.
373     /// This node is used in 64-bit mode as well (in which case the result is
374     /// G8RC and inputs are X3/X4).
375     TLSGD_AIX,
376 
377     /// G8RC = ADDIS_TLSLD_HA %x2, Symbol - For the local-dynamic TLS
378     /// model, produces an ADDIS8 instruction that adds the GOT base
379     /// register to sym\@got\@tlsld\@ha.
380     ADDIS_TLSLD_HA,
381 
382     /// %x3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
383     /// model, produces an ADDI8 instruction that adds G8RReg to
384     /// sym\@got\@tlsld\@l and stores the result in X3.  Hidden by
385     /// ADDIS_TLSLD_L_ADDR until after register assignment.
386     ADDI_TLSLD_L,
387 
388     /// %x3 = GET_TLSLD_ADDR %x3, Symbol - For the local-dynamic TLS
389     /// model, produces a call to __tls_get_addr(sym\@tlsld).  Hidden by
390     /// ADDIS_TLSLD_L_ADDR until after register assignment.
391     GET_TLSLD_ADDR,
392 
393     /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
394     /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
395     /// following register assignment.
396     ADDI_TLSLD_L_ADDR,
397 
398     /// G8RC = ADDIS_DTPREL_HA %x3, Symbol - For the local-dynamic TLS
399     /// model, produces an ADDIS8 instruction that adds X3 to
400     /// sym\@dtprel\@ha.
401     ADDIS_DTPREL_HA,
402 
403     /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
404     /// model, produces an ADDI8 instruction that adds G8RReg to
405     /// sym\@got\@dtprel\@l.
406     ADDI_DTPREL_L,
407 
408     /// G8RC = PADDI_DTPREL %x3, Symbol - For the pc-rel based local-dynamic TLS
409     /// model, produces a PADDI8 instruction that adds X3 to sym\@dtprel.
410     PADDI_DTPREL,
411 
412     /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
413     /// during instruction selection to optimize a BUILD_VECTOR into
414     /// operations on splats.  This is necessary to avoid losing these
415     /// optimizations due to constant folding.
416     VADD_SPLAT,
417 
418     /// CHAIN = SC CHAIN, Imm128 - System call.  The 7-bit unsigned
419     /// operand identifies the operating system entry point.
420     SC,
421 
422     /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
423     CLRBHRB,
424 
425     /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
426     /// history rolling buffer entry.
427     MFBHRBE,
428 
429     /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
430     RFEBB,
431 
432     /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
433     /// endian.  Maps to an xxswapd instruction that corrects an lxvd2x
434     /// or stxvd2x instruction.  The chain is necessary because the
435     /// sequence replaces a load and needs to provide the same number
436     /// of outputs.
437     XXSWAPD,
438 
439     /// An SDNode for swaps that are not associated with any loads/stores
440     /// and thereby have no chain.
441     SWAP_NO_CHAIN,
442 
443     /// FP_EXTEND_HALF(VECTOR, IDX) - Custom extend upper (IDX=0) half or
444     /// lower (IDX=1) half of v4f32 to v2f64.
445     FP_EXTEND_HALF,
446 
447     /// MAT_PCREL_ADDR = Materialize a PC Relative address. This can be done
448     /// either through an add like PADDI or through a PC Relative load like
449     /// PLD.
450     MAT_PCREL_ADDR,
451 
452     /// TLS_DYNAMIC_MAT_PCREL_ADDR = Materialize a PC Relative address for
453     /// TLS global address when using dynamic access models. This can be done
454     /// through an add like PADDI.
455     TLS_DYNAMIC_MAT_PCREL_ADDR,
456 
457     /// TLS_LOCAL_EXEC_MAT_ADDR = Materialize an address for TLS global address
458     /// when using local exec access models, and when prefixed instructions are
459     /// available. This is used with ADD_TLS to produce an add like PADDI.
460     TLS_LOCAL_EXEC_MAT_ADDR,
461 
462     /// ACC_BUILD = Build an accumulator register from 4 VSX registers.
463     ACC_BUILD,
464 
465     /// PAIR_BUILD = Build a vector pair register from 2 VSX registers.
466     PAIR_BUILD,
467 
468     /// EXTRACT_VSX_REG = Extract one of the underlying vsx registers of
469     /// an accumulator or pair register. This node is needed because
470     /// EXTRACT_SUBVECTOR expects the input and output vectors to have the same
471     /// element type.
472     EXTRACT_VSX_REG,
473 
474     /// XXMFACC = This corresponds to the xxmfacc instruction.
475     XXMFACC,
476 
477     // Constrained conversion from floating point to int
478     STRICT_FCTIDZ = ISD::FIRST_TARGET_STRICTFP_OPCODE,
479     STRICT_FCTIWZ,
480     STRICT_FCTIDUZ,
481     STRICT_FCTIWUZ,
482 
483     /// Constrained integer-to-floating-point conversion instructions.
484     STRICT_FCFID,
485     STRICT_FCFIDU,
486     STRICT_FCFIDS,
487     STRICT_FCFIDUS,
488 
489     /// Constrained floating point add in round-to-zero mode.
490     STRICT_FADDRTZ,
491 
492     // NOTE: The nodes below may require PC-Rel specific patterns if the
493     // address could be PC-Relative. When adding new nodes below, consider
494     // whether or not the address can be PC-Relative and add the corresponding
495     // PC-relative patterns and tests.
496 
497     /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
498     /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
499     /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
500     /// i32.
501     STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
502 
503     /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
504     /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
505     /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
506     /// or i32.
507     LBRX,
508 
509     /// STFIWX - The STFIWX instruction.  The first operand is an input token
510     /// chain, then an f64 value to store, then an address to store it to.
511     STFIWX,
512 
513     /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
514     /// load which sign-extends from a 32-bit integer value into the
515     /// destination 64-bit register.
516     LFIWAX,
517 
518     /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
519     /// load which zero-extends from a 32-bit integer value into the
520     /// destination 64-bit register.
521     LFIWZX,
522 
523     /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
524     /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
525     /// This can be used for converting loaded integers to floating point.
526     LXSIZX,
527 
528     /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
529     /// chain, then an f64 value to store, then an address to store it to,
530     /// followed by a byte-width for the store.
531     STXSIX,
532 
533     /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
534     /// Maps directly to an lxvd2x instruction that will be followed by
535     /// an xxswapd.
536     LXVD2X,
537 
538     /// LXVRZX - Load VSX Vector Rightmost and Zero Extend
539     /// This node represents v1i128 BUILD_VECTOR of a zero extending load
540     /// instruction from <byte, halfword, word, or doubleword> to i128.
541     /// Allows utilization of the Load VSX Vector Rightmost Instructions.
542     LXVRZX,
543 
544     /// VSRC, CHAIN = LOAD_VEC_BE CHAIN, Ptr - Occurs only for little endian.
545     /// Maps directly to one of lxvd2x/lxvw4x/lxvh8x/lxvb16x depending on
546     /// the vector type to load vector in big-endian element order.
547     LOAD_VEC_BE,
548 
549     /// VSRC, CHAIN = LD_VSX_LH CHAIN, Ptr - This is a floating-point load of a
550     /// v2f32 value into the lower half of a VSR register.
551     LD_VSX_LH,
552 
553     /// VSRC, CHAIN = LD_SPLAT, CHAIN, Ptr - a splatting load memory
554     /// instructions such as LXVDSX, LXVWSX.
555     LD_SPLAT,
556 
557     /// VSRC, CHAIN = ZEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory
558     /// that zero-extends.
559     ZEXT_LD_SPLAT,
560 
561     /// VSRC, CHAIN = SEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory
562     /// that sign-extends.
563     SEXT_LD_SPLAT,
564 
565     /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
566     /// Maps directly to an stxvd2x instruction that will be preceded by
567     /// an xxswapd.
568     STXVD2X,
569 
570     /// CHAIN = STORE_VEC_BE CHAIN, VSRC, Ptr - Occurs only for little endian.
571     /// Maps directly to one of stxvd2x/stxvw4x/stxvh8x/stxvb16x depending on
572     /// the vector type to store vector in big-endian element order.
573     STORE_VEC_BE,
574 
575     /// Store scalar integers from VSR.
576     ST_VSR_SCAL_INT,
577 
578     /// ATOMIC_CMP_SWAP - the exact same as the target-independent nodes
579     /// except they ensure that the compare input is zero-extended for
580     /// sub-word versions because the atomic loads zero-extend.
581     ATOMIC_CMP_SWAP_8,
582     ATOMIC_CMP_SWAP_16,
583 
584     /// CHAIN,Glue = STORE_COND CHAIN, GPR, Ptr
585     /// The store conditional instruction ST[BHWD]ARX that produces a glue
586     /// result to attach it to a conditional branch.
587     STORE_COND,
588 
589     /// GPRC = TOC_ENTRY GA, TOC
590     /// Loads the entry for GA from the TOC, where the TOC base is given by
591     /// the last operand.
592     TOC_ENTRY
593   };
594 
595   } // end namespace PPCISD
596 
597   /// Define some predicates that are used for node matching.
598   namespace PPC {
599 
600     /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
601     /// VPKUHUM instruction.
602     bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
603                               SelectionDAG &DAG);
604 
605     /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
606     /// VPKUWUM instruction.
607     bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
608                               SelectionDAG &DAG);
609 
610     /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
611     /// VPKUDUM instruction.
612     bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
613                               SelectionDAG &DAG);
614 
615     /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
616     /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
617     bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
618                             unsigned ShuffleKind, SelectionDAG &DAG);
619 
620     /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
621     /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
622     bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
623                             unsigned ShuffleKind, SelectionDAG &DAG);
624 
625     /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
626     /// a VMRGEW or VMRGOW instruction
627     bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
628                              unsigned ShuffleKind, SelectionDAG &DAG);
629     /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
630     /// for a XXSLDWI instruction.
631     bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
632                               bool &Swap, bool IsLE);
633 
634     /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable
635     /// for a XXBRH instruction.
636     bool isXXBRHShuffleMask(ShuffleVectorSDNode *N);
637 
638     /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable
639     /// for a XXBRW instruction.
640     bool isXXBRWShuffleMask(ShuffleVectorSDNode *N);
641 
642     /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable
643     /// for a XXBRD instruction.
644     bool isXXBRDShuffleMask(ShuffleVectorSDNode *N);
645 
646     /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable
647     /// for a XXBRQ instruction.
648     bool isXXBRQShuffleMask(ShuffleVectorSDNode *N);
649 
650     /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
651     /// for a XXPERMDI instruction.
652     bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
653                               bool &Swap, bool IsLE);
654 
655     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
656     /// shift amount, otherwise return -1.
657     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
658                             SelectionDAG &DAG);
659 
660     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
661     /// specifies a splat of a single element that is suitable for input to
662     /// VSPLTB/VSPLTH/VSPLTW.
663     bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
664 
665     /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
666     /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
667     /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
668     /// vector into the other. This function will also set a couple of
669     /// output parameters for how much the source vector needs to be shifted and
670     /// what byte number needs to be specified for the instruction to put the
671     /// element in the desired location of the target vector.
672     bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
673                          unsigned &InsertAtByte, bool &Swap, bool IsLE);
674 
675     /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is
676     /// appropriate for PPC mnemonics (which have a big endian bias - namely
677     /// elements are counted from the left of the vector register).
678     unsigned getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize,
679                                         SelectionDAG &DAG);
680 
681     /// get_VSPLTI_elt - If this is a build_vector of constants which can be
682     /// formed by using a vspltis[bhw] instruction of the specified element
683     /// size, return the constant being splatted.  The ByteSize field indicates
684     /// the number of bytes of each element [124] -> [bhw].
685     SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
686 
687     // Flags for computing the optimal addressing mode for loads and stores.
688     enum MemOpFlags {
689       MOF_None = 0,
690 
691       // Extension mode for integer loads.
692       MOF_SExt = 1,
693       MOF_ZExt = 1 << 1,
694       MOF_NoExt = 1 << 2,
695 
696       // Address computation flags.
697       MOF_NotAddNorCst = 1 << 5,      // Not const. or sum of ptr and scalar.
698       MOF_RPlusSImm16 = 1 << 6,       // Reg plus signed 16-bit constant.
699       MOF_RPlusLo = 1 << 7,           // Reg plus signed 16-bit relocation
700       MOF_RPlusSImm16Mult4 = 1 << 8,  // Reg plus 16-bit signed multiple of 4.
701       MOF_RPlusSImm16Mult16 = 1 << 9, // Reg plus 16-bit signed multiple of 16.
702       MOF_RPlusSImm34 = 1 << 10,      // Reg plus 34-bit signed constant.
703       MOF_RPlusR = 1 << 11,           // Sum of two variables.
704       MOF_PCRel = 1 << 12,            // PC-Relative relocation.
705       MOF_AddrIsSImm32 = 1 << 13,     // A simple 32-bit constant.
706 
707       // The in-memory type.
708       MOF_SubWordInt = 1 << 15,
709       MOF_WordInt = 1 << 16,
710       MOF_DoubleWordInt = 1 << 17,
711       MOF_ScalarFloat = 1 << 18, // Scalar single or double precision.
712       MOF_Vector = 1 << 19,      // Vector types and quad precision scalars.
713       MOF_Vector256 = 1 << 20,
714 
715       // Subtarget features.
716       MOF_SubtargetBeforeP9 = 1 << 22,
717       MOF_SubtargetP9 = 1 << 23,
718       MOF_SubtargetP10 = 1 << 24,
719       MOF_SubtargetSPE = 1 << 25
720     };
721 
722     // The addressing modes for loads and stores.
723     enum AddrMode {
724       AM_None,
725       AM_DForm,
726       AM_DSForm,
727       AM_DQForm,
728       AM_PrefixDForm,
729       AM_XForm,
730       AM_PCRel
731     };
732   } // end namespace PPC
733 
734   class PPCTargetLowering : public TargetLowering {
735     const PPCSubtarget &Subtarget;
736 
737   public:
738     explicit PPCTargetLowering(const PPCTargetMachine &TM,
739                                const PPCSubtarget &STI);
740 
741     /// getTargetNodeName() - This method returns the name of a target specific
742     /// DAG node.
743     const char *getTargetNodeName(unsigned Opcode) const override;
744 
745     bool isSelectSupported(SelectSupportKind Kind) const override {
746       // PowerPC does not support scalar condition selects on vectors.
747       return (Kind != SelectSupportKind::ScalarCondVectorVal);
748     }
749 
750     /// getPreferredVectorAction - The code we generate when vector types are
751     /// legalized by promoting the integer element type is often much worse
752     /// than code we generate if we widen the type for applicable vector types.
753     /// The issue with promoting is that the vector is scalaraized, individual
754     /// elements promoted and then the vector is rebuilt. So say we load a pair
755     /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
756     /// loads, moves back into VSR's (or memory ops if we don't have moves) and
757     /// then the VPERM for the shuffle. All in all a very slow sequence.
758     TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT)
759       const override {
760       // Default handling for scalable and single-element vectors.
761       if (VT.isScalableVector() || VT.getVectorNumElements() == 1)
762         return TargetLoweringBase::getPreferredVectorAction(VT);
763 
764       // Split and promote vNi1 vectors so we don't produce v256i1/v512i1
765       // types as those are only for MMA instructions.
766       if (VT.getScalarSizeInBits() == 1 && VT.getSizeInBits() > 16)
767         return TypeSplitVector;
768       if (VT.getScalarSizeInBits() == 1)
769         return TypePromoteInteger;
770 
771       // Widen vectors that have reasonably sized elements.
772       if (VT.getScalarSizeInBits() % 8 == 0)
773         return TypeWidenVector;
774       return TargetLoweringBase::getPreferredVectorAction(VT);
775     }
776 
777     bool useSoftFloat() const override;
778 
779     bool hasSPE() const;
780 
781     MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
782       return MVT::i32;
783     }
784 
785     bool isCheapToSpeculateCttz(Type *Ty) const override {
786       return true;
787     }
788 
789     bool isCheapToSpeculateCtlz(Type *Ty) const override {
790       return true;
791     }
792 
793     bool isCtlzFast() const override {
794       return true;
795     }
796 
797     bool isEqualityCmpFoldedWithSignedCmp() const override {
798       return false;
799     }
800 
801     bool hasAndNotCompare(SDValue) const override {
802       return true;
803     }
804 
805     bool preferIncOfAddToSubOfNot(EVT VT) const override;
806 
807     bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
808       return VT.isScalarInteger();
809     }
810 
811     SDValue getNegatedExpression(SDValue Op, SelectionDAG &DAG, bool LegalOps,
812                                  bool OptForSize, NegatibleCost &Cost,
813                                  unsigned Depth = 0) const override;
814 
815     /// getSetCCResultType - Return the ISD::SETCC ValueType
816     EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
817                            EVT VT) const override;
818 
819     /// Return true if target always benefits from combining into FMA for a
820     /// given value type. This must typically return false on targets where FMA
821     /// takes more cycles to execute than FADD.
822     bool enableAggressiveFMAFusion(EVT VT) const override;
823 
824     /// getPreIndexedAddressParts - returns true by value, base pointer and
825     /// offset pointer and addressing mode by reference if the node's address
826     /// can be legally represented as pre-indexed load / store address.
827     bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
828                                    SDValue &Offset,
829                                    ISD::MemIndexedMode &AM,
830                                    SelectionDAG &DAG) const override;
831 
832     /// SelectAddressEVXRegReg - Given the specified addressed, check to see if
833     /// it can be more efficiently represented as [r+imm].
834     bool SelectAddressEVXRegReg(SDValue N, SDValue &Base, SDValue &Index,
835                                 SelectionDAG &DAG) const;
836 
837     /// SelectAddressRegReg - Given the specified addressed, check to see if it
838     /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment
839     /// is non-zero, only accept displacement which is not suitable for [r+imm].
840     /// Returns false if it can be represented by [r+imm], which are preferred.
841     bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
842                              SelectionDAG &DAG,
843                              MaybeAlign EncodingAlignment = std::nullopt) const;
844 
845     /// SelectAddressRegImm - Returns true if the address N can be represented
846     /// by a base register plus a signed 16-bit displacement [r+imm], and if it
847     /// is not better represented as reg+reg. If \p EncodingAlignment is
848     /// non-zero, only accept displacements suitable for instruction encoding
849     /// requirement, i.e. multiples of 4 for DS form.
850     bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
851                              SelectionDAG &DAG,
852                              MaybeAlign EncodingAlignment) const;
853     bool SelectAddressRegImm34(SDValue N, SDValue &Disp, SDValue &Base,
854                                SelectionDAG &DAG) const;
855 
856     /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
857     /// represented as an indexed [r+r] operation.
858     bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
859                                  SelectionDAG &DAG) const;
860 
861     /// SelectAddressPCRel - Represent the specified address as pc relative to
862     /// be represented as [pc+imm]
863     bool SelectAddressPCRel(SDValue N, SDValue &Base) const;
864 
865     Sched::Preference getSchedulingPreference(SDNode *N) const override;
866 
867     /// LowerOperation - Provide custom lowering hooks for some operations.
868     ///
869     SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
870 
871     /// ReplaceNodeResults - Replace the results of node with an illegal result
872     /// type with new values built out of custom code.
873     ///
874     void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
875                             SelectionDAG &DAG) const override;
876 
877     SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
878     SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
879 
880     SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
881 
882     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
883                           SmallVectorImpl<SDNode *> &Created) const override;
884 
885     Register getRegisterByName(const char* RegName, LLT VT,
886                                const MachineFunction &MF) const override;
887 
888     void computeKnownBitsForTargetNode(const SDValue Op,
889                                        KnownBits &Known,
890                                        const APInt &DemandedElts,
891                                        const SelectionDAG &DAG,
892                                        unsigned Depth = 0) const override;
893 
894     Align getPrefLoopAlignment(MachineLoop *ML) const override;
895 
896     bool shouldInsertFencesForAtomic(const Instruction *I) const override {
897       return true;
898     }
899 
900     Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
901                                   AtomicOrdering Ord) const override;
902     Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
903                                    AtomicOrdering Ord) const override;
904 
905     bool shouldInlineQuadwordAtomics() const;
906 
907     TargetLowering::AtomicExpansionKind
908     shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
909 
910     TargetLowering::AtomicExpansionKind
911     shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
912 
913     Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder,
914                                         AtomicRMWInst *AI, Value *AlignedAddr,
915                                         Value *Incr, Value *Mask,
916                                         Value *ShiftAmt,
917                                         AtomicOrdering Ord) const override;
918     Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder,
919                                             AtomicCmpXchgInst *CI,
920                                             Value *AlignedAddr, Value *CmpVal,
921                                             Value *NewVal, Value *Mask,
922                                             AtomicOrdering Ord) const override;
923 
924     MachineBasicBlock *
925     EmitInstrWithCustomInserter(MachineInstr &MI,
926                                 MachineBasicBlock *MBB) const override;
927     MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
928                                         MachineBasicBlock *MBB,
929                                         unsigned AtomicSize,
930                                         unsigned BinOpcode,
931                                         unsigned CmpOpcode = 0,
932                                         unsigned CmpPred = 0) const;
933     MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
934                                                 MachineBasicBlock *MBB,
935                                                 bool is8bit,
936                                                 unsigned Opcode,
937                                                 unsigned CmpOpcode = 0,
938                                                 unsigned CmpPred = 0) const;
939 
940     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
941                                         MachineBasicBlock *MBB) const;
942 
943     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
944                                          MachineBasicBlock *MBB) const;
945 
946     MachineBasicBlock *emitProbedAlloca(MachineInstr &MI,
947                                         MachineBasicBlock *MBB) const;
948 
949     bool hasInlineStackProbe(const MachineFunction &MF) const override;
950 
951     unsigned getStackProbeSize(const MachineFunction &MF) const;
952 
953     ConstraintType getConstraintType(StringRef Constraint) const override;
954 
955     /// Examine constraint string and operand type and determine a weight value.
956     /// The operand object must already have been set up with the operand type.
957     ConstraintWeight getSingleConstraintMatchWeight(
958       AsmOperandInfo &info, const char *constraint) const override;
959 
960     std::pair<unsigned, const TargetRegisterClass *>
961     getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
962                                  StringRef Constraint, MVT VT) const override;
963 
964     /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
965     /// function arguments in the caller parameter area.  This is the actual
966     /// alignment, not its logarithm.
967     uint64_t getByValTypeAlignment(Type *Ty,
968                                    const DataLayout &DL) const override;
969 
970     /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
971     /// vector.  If it is invalid, don't add anything to Ops.
972     void LowerAsmOperandForConstraint(SDValue Op,
973                                       std::string &Constraint,
974                                       std::vector<SDValue> &Ops,
975                                       SelectionDAG &DAG) const override;
976 
977     unsigned
978     getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
979       if (ConstraintCode == "es")
980         return InlineAsm::Constraint_es;
981       else if (ConstraintCode == "Q")
982         return InlineAsm::Constraint_Q;
983       else if (ConstraintCode == "Z")
984         return InlineAsm::Constraint_Z;
985       else if (ConstraintCode == "Zy")
986         return InlineAsm::Constraint_Zy;
987       return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
988     }
989 
990     void CollectTargetIntrinsicOperands(const CallInst &I,
991                                  SmallVectorImpl<SDValue> &Ops,
992                                  SelectionDAG &DAG) const override;
993 
994     /// isLegalAddressingMode - Return true if the addressing mode represented
995     /// by AM is legal for this target, for a load/store of the specified type.
996     bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
997                                Type *Ty, unsigned AS,
998                                Instruction *I = nullptr) const override;
999 
1000     /// isLegalICmpImmediate - Return true if the specified immediate is legal
1001     /// icmp immediate, that is the target has icmp instructions which can
1002     /// compare a register against the immediate without having to materialize
1003     /// the immediate into a register.
1004     bool isLegalICmpImmediate(int64_t Imm) const override;
1005 
1006     /// isLegalAddImmediate - Return true if the specified immediate is legal
1007     /// add immediate, that is the target has add instructions which can
1008     /// add a register and the immediate without having to materialize
1009     /// the immediate into a register.
1010     bool isLegalAddImmediate(int64_t Imm) const override;
1011 
1012     /// isTruncateFree - Return true if it's free to truncate a value of
1013     /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
1014     /// register X1 to i32 by referencing its sub-register R1.
1015     bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
1016     bool isTruncateFree(EVT VT1, EVT VT2) const override;
1017 
1018     bool isZExtFree(SDValue Val, EVT VT2) const override;
1019 
1020     bool isFPExtFree(EVT DestVT, EVT SrcVT) const override;
1021 
1022     /// Returns true if it is beneficial to convert a load of a constant
1023     /// to just the constant itself.
1024     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1025                                            Type *Ty) const override;
1026 
1027     bool convertSelectOfConstantsToMath(EVT VT) const override {
1028       return true;
1029     }
1030 
1031     bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
1032                                 SDValue C) const override;
1033 
1034     bool isDesirableToTransformToIntegerOp(unsigned Opc,
1035                                            EVT VT) const override {
1036       // Only handle float load/store pair because float(fpr) load/store
1037       // instruction has more cycles than integer(gpr) load/store in PPC.
1038       if (Opc != ISD::LOAD && Opc != ISD::STORE)
1039         return false;
1040       if (VT != MVT::f32 && VT != MVT::f64)
1041         return false;
1042 
1043       return true;
1044     }
1045 
1046     // Returns true if the address of the global is stored in TOC entry.
1047     bool isAccessedAsGotIndirect(SDValue N) const;
1048 
1049     bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
1050 
1051     bool getTgtMemIntrinsic(IntrinsicInfo &Info,
1052                             const CallInst &I,
1053                             MachineFunction &MF,
1054                             unsigned Intrinsic) const override;
1055 
1056     /// It returns EVT::Other if the type should be determined using generic
1057     /// target-independent logic.
1058     EVT getOptimalMemOpType(const MemOp &Op,
1059                             const AttributeList &FuncAttributes) const override;
1060 
1061     /// Is unaligned memory access allowed for the given type, and is it fast
1062     /// relative to software emulation.
1063     bool allowsMisalignedMemoryAccesses(
1064         EVT VT, unsigned AddrSpace, Align Alignment = Align(1),
1065         MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
1066         unsigned *Fast = nullptr) const override;
1067 
1068     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
1069     /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
1070     /// expanded to FMAs when this method returns true, otherwise fmuladd is
1071     /// expanded to fmul + fadd.
1072     bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
1073                                     EVT VT) const override;
1074 
1075     bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override;
1076 
1077     /// isProfitableToHoist - Check if it is profitable to hoist instruction
1078     /// \p I to its dominator block.
1079     /// For example, it is not profitable if \p I and it's only user can form a
1080     /// FMA instruction, because Powerpc prefers FMADD.
1081     bool isProfitableToHoist(Instruction *I) const override;
1082 
1083     const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1084 
1085     // Should we expand the build vector with shuffles?
1086     bool
1087     shouldExpandBuildVectorWithShuffles(EVT VT,
1088                                         unsigned DefinedValues) const override;
1089 
1090     // Keep the zero-extensions for arguments to libcalls.
1091     bool shouldKeepZExtForFP16Conv() const override { return true; }
1092 
1093     /// createFastISel - This method returns a target-specific FastISel object,
1094     /// or null if the target does not support "fast" instruction selection.
1095     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1096                              const TargetLibraryInfo *LibInfo) const override;
1097 
1098     /// Returns true if an argument of type Ty needs to be passed in a
1099     /// contiguous block of registers in calling convention CallConv.
1100     bool functionArgumentNeedsConsecutiveRegisters(
1101         Type *Ty, CallingConv::ID CallConv, bool isVarArg,
1102         const DataLayout &DL) const override {
1103       // We support any array type as "consecutive" block in the parameter
1104       // save area.  The element type defines the alignment requirement and
1105       // whether the argument should go in GPRs, FPRs, or VRs if available.
1106       //
1107       // Note that clang uses this capability both to implement the ELFv2
1108       // homogeneous float/vector aggregate ABI, and to avoid having to use
1109       // "byval" when passing aggregates that might fully fit in registers.
1110       return Ty->isArrayTy();
1111     }
1112 
1113     /// If a physical register, this returns the register that receives the
1114     /// exception address on entry to an EH pad.
1115     Register
1116     getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1117 
1118     /// If a physical register, this returns the register that receives the
1119     /// exception typeid on entry to a landing pad.
1120     Register
1121     getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1122 
1123     /// Override to support customized stack guard loading.
1124     bool useLoadStackGuardNode() const override;
1125     void insertSSPDeclarations(Module &M) const override;
1126     Value *getSDagStackGuard(const Module &M) const override;
1127 
1128     bool isFPImmLegal(const APFloat &Imm, EVT VT,
1129                       bool ForCodeSize) const override;
1130 
1131     unsigned getJumpTableEncoding() const override;
1132     bool isJumpTableRelative() const override;
1133     SDValue getPICJumpTableRelocBase(SDValue Table,
1134                                      SelectionDAG &DAG) const override;
1135     const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1136                                                unsigned JTI,
1137                                                MCContext &Ctx) const override;
1138 
1139     /// SelectOptimalAddrMode - Based on a node N and it's Parent (a MemSDNode),
1140     /// compute the address flags of the node, get the optimal address mode
1141     /// based on the flags, and set the Base and Disp based on the address mode.
1142     PPC::AddrMode SelectOptimalAddrMode(const SDNode *Parent, SDValue N,
1143                                         SDValue &Disp, SDValue &Base,
1144                                         SelectionDAG &DAG,
1145                                         MaybeAlign Align) const;
1146     /// SelectForceXFormMode - Given the specified address, force it to be
1147     /// represented as an indexed [r+r] operation (an XForm instruction).
1148     PPC::AddrMode SelectForceXFormMode(SDValue N, SDValue &Disp, SDValue &Base,
1149                                        SelectionDAG &DAG) const;
1150 
1151     bool splitValueIntoRegisterParts(
1152         SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
1153         unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC)
1154         const override;
1155     /// Structure that collects some common arguments that get passed around
1156     /// between the functions for call lowering.
1157     struct CallFlags {
1158       const CallingConv::ID CallConv;
1159       const bool IsTailCall : 1;
1160       const bool IsVarArg : 1;
1161       const bool IsPatchPoint : 1;
1162       const bool IsIndirect : 1;
1163       const bool HasNest : 1;
1164       const bool NoMerge : 1;
1165 
1166       CallFlags(CallingConv::ID CC, bool IsTailCall, bool IsVarArg,
1167                 bool IsPatchPoint, bool IsIndirect, bool HasNest, bool NoMerge)
1168           : CallConv(CC), IsTailCall(IsTailCall), IsVarArg(IsVarArg),
1169             IsPatchPoint(IsPatchPoint), IsIndirect(IsIndirect),
1170             HasNest(HasNest), NoMerge(NoMerge) {}
1171     };
1172 
1173     CCAssignFn *ccAssignFnForCall(CallingConv::ID CC, bool Return,
1174                                   bool IsVarArg) const;
1175 
1176   private:
1177     struct ReuseLoadInfo {
1178       SDValue Ptr;
1179       SDValue Chain;
1180       SDValue ResChain;
1181       MachinePointerInfo MPI;
1182       bool IsDereferenceable = false;
1183       bool IsInvariant = false;
1184       Align Alignment;
1185       AAMDNodes AAInfo;
1186       const MDNode *Ranges = nullptr;
1187 
1188       ReuseLoadInfo() = default;
1189 
1190       MachineMemOperand::Flags MMOFlags() const {
1191         MachineMemOperand::Flags F = MachineMemOperand::MONone;
1192         if (IsDereferenceable)
1193           F |= MachineMemOperand::MODereferenceable;
1194         if (IsInvariant)
1195           F |= MachineMemOperand::MOInvariant;
1196         return F;
1197       }
1198     };
1199 
1200     // Map that relates a set of common address flags to PPC addressing modes.
1201     std::map<PPC::AddrMode, SmallVector<unsigned, 16>> AddrModesMap;
1202     void initializeAddrModeMap();
1203 
1204     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
1205                              SelectionDAG &DAG,
1206                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
1207     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
1208                          SelectionDAG &DAG) const;
1209 
1210     void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
1211                                 SelectionDAG &DAG, const SDLoc &dl) const;
1212     SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
1213                                      const SDLoc &dl) const;
1214 
1215     bool directMoveIsProfitable(const SDValue &Op) const;
1216     SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
1217                                      const SDLoc &dl) const;
1218 
1219     SDValue LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG,
1220                                  const SDLoc &dl) const;
1221 
1222     SDValue LowerTRUNCATEVector(SDValue Op, SelectionDAG &DAG) const;
1223 
1224     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
1225     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
1226 
1227     bool
1228     IsEligibleForTailCallOptimization(SDValue Callee,
1229                                       CallingConv::ID CalleeCC,
1230                                       bool isVarArg,
1231                                       const SmallVectorImpl<ISD::InputArg> &Ins,
1232                                       SelectionDAG& DAG) const;
1233 
1234     bool IsEligibleForTailCallOptimization_64SVR4(
1235         SDValue Callee, CallingConv::ID CalleeCC, const CallBase *CB,
1236         bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs,
1237         const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
1238 
1239     SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
1240                                          SDValue Chain, SDValue &LROpOut,
1241                                          SDValue &FPOpOut,
1242                                          const SDLoc &dl) const;
1243 
1244     SDValue getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, SDValue GA) const;
1245 
1246     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1247     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1248     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1249     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1250     SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1251     SDValue LowerGlobalTLSAddressAIX(SDValue Op, SelectionDAG &DAG) const;
1252     SDValue LowerGlobalTLSAddressLinux(SDValue Op, SelectionDAG &DAG) const;
1253     SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1254     SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1255     SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1256     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1257     SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1258     SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const;
1259     SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1260     SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1261     SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
1262     SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
1263     SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1264     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1265     SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
1266     SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
1267     SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
1268     SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1269     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
1270     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
1271                            const SDLoc &dl) const;
1272     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1273     SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
1274     SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1275     SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1276     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
1277     SDValue LowerFunnelShift(SDValue Op, SelectionDAG &DAG) const;
1278     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1279     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
1280     SDValue LowerVPERM(SDValue Op, SelectionDAG &DAG, ArrayRef<int> PermMask,
1281                        EVT VT, SDValue V1, SDValue V2) const;
1282     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1283     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1284     SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
1285     SDValue LowerBSWAP(SDValue Op, SelectionDAG &DAG) const;
1286     SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
1287     SDValue lowerToLibCall(const char *LibCallName, SDValue Op,
1288                            SelectionDAG &DAG) const;
1289     SDValue lowerLibCallBasedOnType(const char *LibCallFloatName,
1290                                     const char *LibCallDoubleName, SDValue Op,
1291                                     SelectionDAG &DAG) const;
1292     bool isLowringToMASSFiniteSafe(SDValue Op) const;
1293     bool isLowringToMASSSafe(SDValue Op) const;
1294     bool isScalarMASSConversionEnabled() const;
1295     SDValue lowerLibCallBase(const char *LibCallDoubleName,
1296                              const char *LibCallFloatName,
1297                              const char *LibCallDoubleNameFinite,
1298                              const char *LibCallFloatNameFinite, SDValue Op,
1299                              SelectionDAG &DAG) const;
1300     SDValue lowerPow(SDValue Op, SelectionDAG &DAG) const;
1301     SDValue lowerSin(SDValue Op, SelectionDAG &DAG) const;
1302     SDValue lowerCos(SDValue Op, SelectionDAG &DAG) const;
1303     SDValue lowerLog(SDValue Op, SelectionDAG &DAG) const;
1304     SDValue lowerLog10(SDValue Op, SelectionDAG &DAG) const;
1305     SDValue lowerExp(SDValue Op, SelectionDAG &DAG) const;
1306     SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) const;
1307     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1308     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
1309     SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
1310     SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
1311     SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const;
1312 
1313     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
1314     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
1315 
1316     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1317                             CallingConv::ID CallConv, bool isVarArg,
1318                             const SmallVectorImpl<ISD::InputArg> &Ins,
1319                             const SDLoc &dl, SelectionDAG &DAG,
1320                             SmallVectorImpl<SDValue> &InVals) const;
1321 
1322     SDValue FinishCall(CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG,
1323                        SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
1324                        SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
1325                        SDValue &Callee, int SPDiff, unsigned NumBytes,
1326                        const SmallVectorImpl<ISD::InputArg> &Ins,
1327                        SmallVectorImpl<SDValue> &InVals,
1328                        const CallBase *CB) const;
1329 
1330     SDValue
1331     LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1332                          const SmallVectorImpl<ISD::InputArg> &Ins,
1333                          const SDLoc &dl, SelectionDAG &DAG,
1334                          SmallVectorImpl<SDValue> &InVals) const override;
1335 
1336     SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
1337                       SmallVectorImpl<SDValue> &InVals) const override;
1338 
1339     bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1340                         bool isVarArg,
1341                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1342                         LLVMContext &Context) const override;
1343 
1344     SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1345                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1346                         const SmallVectorImpl<SDValue> &OutVals,
1347                         const SDLoc &dl, SelectionDAG &DAG) const override;
1348 
1349     SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1350                               SelectionDAG &DAG, SDValue ArgVal,
1351                               const SDLoc &dl) const;
1352 
1353     SDValue LowerFormalArguments_AIX(
1354         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1355         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1356         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1357     SDValue LowerFormalArguments_64SVR4(
1358         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1359         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1360         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1361     SDValue LowerFormalArguments_32SVR4(
1362         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1363         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1364         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1365 
1366     SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
1367                                        SDValue CallSeqStart,
1368                                        ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1369                                        const SDLoc &dl) const;
1370 
1371     SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1372                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1373                              const SmallVectorImpl<SDValue> &OutVals,
1374                              const SmallVectorImpl<ISD::InputArg> &Ins,
1375                              const SDLoc &dl, SelectionDAG &DAG,
1376                              SmallVectorImpl<SDValue> &InVals,
1377                              const CallBase *CB) const;
1378     SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1379                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1380                              const SmallVectorImpl<SDValue> &OutVals,
1381                              const SmallVectorImpl<ISD::InputArg> &Ins,
1382                              const SDLoc &dl, SelectionDAG &DAG,
1383                              SmallVectorImpl<SDValue> &InVals,
1384                              const CallBase *CB) const;
1385     SDValue LowerCall_AIX(SDValue Chain, SDValue Callee, CallFlags CFlags,
1386                           const SmallVectorImpl<ISD::OutputArg> &Outs,
1387                           const SmallVectorImpl<SDValue> &OutVals,
1388                           const SmallVectorImpl<ISD::InputArg> &Ins,
1389                           const SDLoc &dl, SelectionDAG &DAG,
1390                           SmallVectorImpl<SDValue> &InVals,
1391                           const CallBase *CB) const;
1392 
1393     SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1394     SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1395     SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
1396 
1397     SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1398     SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1399     SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1400     SDValue combineStoreFPToInt(SDNode *N, DAGCombinerInfo &DCI) const;
1401     SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1402     SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1403     SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1404     SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1405     SDValue combineMUL(SDNode *N, DAGCombinerInfo &DCI) const;
1406     SDValue combineADD(SDNode *N, DAGCombinerInfo &DCI) const;
1407     SDValue combineFMALike(SDNode *N, DAGCombinerInfo &DCI) const;
1408     SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const;
1409     SDValue combineSetCC(SDNode *N, DAGCombinerInfo &DCI) const;
1410     SDValue combineVSelect(SDNode *N, DAGCombinerInfo &DCI) const;
1411     SDValue combineVectorShuffle(ShuffleVectorSDNode *SVN,
1412                                  SelectionDAG &DAG) const;
1413     SDValue combineVReverseMemOP(ShuffleVectorSDNode *SVN, LSBaseSDNode *LSBase,
1414                                  DAGCombinerInfo &DCI) const;
1415 
1416     /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1417     /// SETCC with integer subtraction when (1) there is a legal way of doing it
1418     /// (2) keeping the result of comparison in GPR has performance benefit.
1419     SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1420 
1421     SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1422                             int &RefinementSteps, bool &UseOneConstNR,
1423                             bool Reciprocal) const override;
1424     SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1425                              int &RefinementSteps) const override;
1426     SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG,
1427                              const DenormalMode &Mode) const override;
1428     SDValue getSqrtResultForDenormInput(SDValue Operand,
1429                                         SelectionDAG &DAG) const override;
1430     unsigned combineRepeatedFPDivisors() const override;
1431 
1432     SDValue
1433     combineElementTruncationToVectorTruncation(SDNode *N,
1434                                                DAGCombinerInfo &DCI) const;
1435 
1436     /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be
1437     /// handled by the VINSERTH instruction introduced in ISA 3.0. This is
1438     /// essentially any shuffle of v8i16 vectors that just inserts one element
1439     /// from one vector into the other.
1440     SDValue lowerToVINSERTH(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1441 
1442     /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be
1443     /// handled by the VINSERTB instruction introduced in ISA 3.0. This is
1444     /// essentially v16i8 vector version of VINSERTH.
1445     SDValue lowerToVINSERTB(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1446 
1447     /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be
1448     /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1.
1449     SDValue lowerToXXSPLTI32DX(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1450 
1451     // Return whether the call instruction can potentially be optimized to a
1452     // tail call. This will cause the optimizers to attempt to move, or
1453     // duplicate return instructions to help enable tail call optimizations.
1454     bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1455     bool hasBitPreservingFPLogic(EVT VT) const override;
1456     bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
1457 
1458     /// getAddrModeForFlags - Based on the set of address flags, select the most
1459     /// optimal instruction format to match by.
1460     PPC::AddrMode getAddrModeForFlags(unsigned Flags) const;
1461 
1462     /// computeMOFlags - Given a node N and it's Parent (a MemSDNode), compute
1463     /// the address flags of the load/store instruction that is to be matched.
1464     /// The address flags are stored in a map, which is then searched
1465     /// through to determine the optimal load/store instruction format.
1466     unsigned computeMOFlags(const SDNode *Parent, SDValue N,
1467                             SelectionDAG &DAG) const;
1468   }; // end class PPCTargetLowering
1469 
1470   namespace PPC {
1471 
1472     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1473                              const TargetLibraryInfo *LibInfo);
1474 
1475   } // end namespace PPC
1476 
1477   bool isIntS16Immediate(SDNode *N, int16_t &Imm);
1478   bool isIntS16Immediate(SDValue Op, int16_t &Imm);
1479   bool isIntS34Immediate(SDNode *N, int64_t &Imm);
1480   bool isIntS34Immediate(SDValue Op, int64_t &Imm);
1481 
1482   bool convertToNonDenormSingle(APInt &ArgAPInt);
1483   bool convertToNonDenormSingle(APFloat &ArgAPFloat);
1484   bool checkConvertToNonDenormSingle(APFloat &ArgAPFloat);
1485 
1486 } // end namespace llvm
1487 
1488 #endif // LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
1489