xref: /freebsd/contrib/llvm-project/llvm/lib/Target/PowerPC/PPCISelLowering.h (revision 0ad011ececb978e22a9bff2acf76633b094f1ff6)
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/MachineValueType.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SelectionDAGNodes.h"
24 #include "llvm/CodeGen/TargetLowering.h"
25 #include "llvm/CodeGen/ValueTypes.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/CallingConv.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/InlineAsm.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Type.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     /// VEXTS, ByteWidth - takes an input in VSFRC and produces an output in
82     /// VSFRC that is sign-extended from ByteWidth to a 64-byte integer.
83     VEXTS,
84 
85     /// Reciprocal estimate instructions (unary FP ops).
86     FRE,
87     FRSQRTE,
88 
89     /// Test instruction for software square root.
90     FTSQRT,
91 
92     /// Square root instruction.
93     FSQRT,
94 
95     /// VPERM - The PPC VPERM Instruction.
96     ///
97     VPERM,
98 
99     /// XXSPLT - The PPC VSX splat instructions
100     ///
101     XXSPLT,
102 
103     /// XXSPLTI_SP_TO_DP - The PPC VSX splat instructions for immediates for
104     /// converting immediate single precision numbers to double precision
105     /// vector or scalar.
106     XXSPLTI_SP_TO_DP,
107 
108     /// XXSPLTI32DX - The PPC XXSPLTI32DX instruction.
109     ///
110     XXSPLTI32DX,
111 
112     /// VECINSERT - The PPC vector insert instruction
113     ///
114     VECINSERT,
115 
116     /// VECSHL - The PPC vector shift left instruction
117     ///
118     VECSHL,
119 
120     /// XXPERMDI - The PPC XXPERMDI instruction
121     ///
122     XXPERMDI,
123     XXPERM,
124 
125     /// The CMPB instruction (takes two operands of i32 or i64).
126     CMPB,
127 
128     /// Hi/Lo - These represent the high and low 16-bit parts of a global
129     /// address respectively.  These nodes have two operands, the first of
130     /// which must be a TargetGlobalAddress, and the second of which must be a
131     /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
132     /// though these are usually folded into other nodes.
133     Hi,
134     Lo,
135 
136     /// The following two target-specific nodes are used for calls through
137     /// function pointers in the 64-bit SVR4 ABI.
138 
139     /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
140     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
141     /// compute an allocation on the stack.
142     DYNALLOC,
143 
144     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
145     /// compute an offset from native SP to the address  of the most recent
146     /// dynamic alloca.
147     DYNAREAOFFSET,
148 
149     /// To avoid stack clash, allocation is performed by block and each block is
150     /// probed.
151     PROBED_ALLOCA,
152 
153     /// The result of the mflr at function entry, used for PIC code.
154     GlobalBaseReg,
155 
156     /// These nodes represent PPC shifts.
157     ///
158     /// For scalar types, only the last `n + 1` bits of the shift amounts
159     /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
160     /// for exact behaviors.
161     ///
162     /// For vector types, only the last n bits are used. See vsld.
163     SRL,
164     SRA,
165     SHL,
166 
167     /// FNMSUB - Negated multiply-subtract instruction.
168     FNMSUB,
169 
170     /// EXTSWSLI = The PPC extswsli instruction, which does an extend-sign
171     /// word and shift left immediate.
172     EXTSWSLI,
173 
174     /// The combination of sra[wd]i and addze used to implemented signed
175     /// integer division by a power of 2. The first operand is the dividend,
176     /// and the second is the constant shift amount (representing the
177     /// divisor).
178     SRA_ADDZE,
179 
180     /// CALL - A direct function call.
181     /// CALL_NOP is a call with the special NOP which follows 64-bit
182     /// CALL_NOTOC the caller does not use the TOC.
183     /// SVR4 calls and 32-bit/64-bit AIX calls.
184     CALL,
185     CALL_NOP,
186     CALL_NOTOC,
187 
188     /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
189     /// MTCTR instruction.
190     MTCTR,
191 
192     /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
193     /// BCTRL instruction.
194     BCTRL,
195 
196     /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
197     /// instruction and the TOC reload required on 64-bit ELF, 32-bit AIX
198     /// and 64-bit AIX.
199     BCTRL_LOAD_TOC,
200 
201     /// The variants that implicitly define rounding mode for calls with
202     /// strictfp semantics.
203     CALL_RM,
204     CALL_NOP_RM,
205     CALL_NOTOC_RM,
206     BCTRL_RM,
207     BCTRL_LOAD_TOC_RM,
208 
209     /// Return with a glue operand, matched by 'blr'
210     RET_GLUE,
211 
212     /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
213     /// This copies the bits corresponding to the specified CRREG into the
214     /// resultant GPR.  Bits corresponding to other CR regs are undefined.
215     MFOCRF,
216 
217     /// Direct move from a VSX register to a GPR
218     MFVSR,
219 
220     /// Direct move from a GPR to a VSX register (algebraic)
221     MTVSRA,
222 
223     /// Direct move from a GPR to a VSX register (zero)
224     MTVSRZ,
225 
226     /// Direct move of 2 consecutive GPR to a VSX register.
227     BUILD_FP128,
228 
229     /// BUILD_SPE64 and EXTRACT_SPE are analogous to BUILD_PAIR and
230     /// EXTRACT_ELEMENT but take f64 arguments instead of i64, as i64 is
231     /// unsupported for this target.
232     /// Merge 2 GPRs to a single SPE register.
233     BUILD_SPE64,
234 
235     /// Extract SPE register component, second argument is high or low.
236     EXTRACT_SPE,
237 
238     /// Extract a subvector from signed integer vector and convert to FP.
239     /// It is primarily used to convert a (widened) illegal integer vector
240     /// type to a legal floating point vector type.
241     /// For example v2i32 -> widened to v4i32 -> v2f64
242     SINT_VEC_TO_FP,
243 
244     /// Extract a subvector from unsigned integer vector and convert to FP.
245     /// As with SINT_VEC_TO_FP, used for converting illegal types.
246     UINT_VEC_TO_FP,
247 
248     /// PowerPC instructions that have SCALAR_TO_VECTOR semantics tend to
249     /// place the value into the least significant element of the most
250     /// significant doubleword in the vector. This is not element zero for
251     /// anything smaller than a doubleword on either endianness. This node has
252     /// the same semantics as SCALAR_TO_VECTOR except that the value remains in
253     /// the aforementioned location in the vector register.
254     SCALAR_TO_VECTOR_PERMUTED,
255 
256     // FIXME: Remove these once the ANDI glue bug is fixed:
257     /// i1 = ANDI_rec_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
258     /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
259     /// implement truncation of i32 or i64 to i1.
260     ANDI_rec_1_EQ_BIT,
261     ANDI_rec_1_GT_BIT,
262 
263     // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
264     // target (returns (Lo, Hi)). It takes a chain operand.
265     READ_TIME_BASE,
266 
267     // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
268     EH_SJLJ_SETJMP,
269 
270     // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
271     EH_SJLJ_LONGJMP,
272 
273     /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
274     /// instructions.  For lack of better number, we use the opcode number
275     /// encoding for the OPC field to identify the compare.  For example, 838
276     /// is VCMPGTSH.
277     VCMP,
278 
279     /// RESVEC, OUTFLAG = VCMP_rec(LHS, RHS, OPC) - Represents one of the
280     /// altivec VCMP*_rec instructions.  For lack of better number, we use the
281     /// opcode number encoding for the OPC field to identify the compare.  For
282     /// example, 838 is VCMPGTSH.
283     VCMP_rec,
284 
285     /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
286     /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
287     /// condition register to branch on, OPC is the branch opcode to use (e.g.
288     /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
289     /// an optional input flag argument.
290     COND_BRANCH,
291 
292     /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
293     /// loops.
294     BDNZ,
295     BDZ,
296 
297     /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
298     /// towards zero.  Used only as part of the long double-to-int
299     /// conversion sequence.
300     FADDRTZ,
301 
302     /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
303     MFFS,
304 
305     /// TC_RETURN - A tail call return.
306     ///   operand #0 chain
307     ///   operand #1 callee (register or absolute)
308     ///   operand #2 stack adjustment
309     ///   operand #3 optional in flag
310     TC_RETURN,
311 
312     /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
313     CR6SET,
314     CR6UNSET,
315 
316     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
317     /// for non-position independent code on PPC32.
318     PPC32_GOT,
319 
320     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
321     /// local dynamic TLS and position indendepent code on PPC32.
322     PPC32_PICGOT,
323 
324     /// G8RC = ADDIS_GOT_TPREL_HA %x2, Symbol - Used by the initial-exec
325     /// TLS model, produces an ADDIS8 instruction that adds the GOT
326     /// base to sym\@got\@tprel\@ha.
327     ADDIS_GOT_TPREL_HA,
328 
329     /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
330     /// TLS model, produces a LD instruction with base register G8RReg
331     /// and offset sym\@got\@tprel\@l.  This completes the addition that
332     /// finds the offset of "sym" relative to the thread pointer.
333     LD_GOT_TPREL_L,
334 
335     /// G8RC = ADD_TLS G8RReg, Symbol - Can be used by the initial-exec
336     /// and local-exec TLS models, produces an ADD instruction that adds
337     /// the contents of G8RReg to the thread pointer.  Symbol contains a
338     /// relocation sym\@tls which is to be replaced by the thread pointer
339     /// and identifies to the linker that the instruction is part of a
340     /// TLS sequence.
341     ADD_TLS,
342 
343     /// G8RC = ADDIS_TLSGD_HA %x2, Symbol - For the general-dynamic TLS
344     /// model, produces an ADDIS8 instruction that adds the GOT base
345     /// register to sym\@got\@tlsgd\@ha.
346     ADDIS_TLSGD_HA,
347 
348     /// %x3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
349     /// model, produces an ADDI8 instruction that adds G8RReg to
350     /// sym\@got\@tlsgd\@l and stores the result in X3.  Hidden by
351     /// ADDIS_TLSGD_L_ADDR until after register assignment.
352     ADDI_TLSGD_L,
353 
354     /// %x3 = GET_TLS_ADDR %x3, Symbol - For the general-dynamic TLS
355     /// model, produces a call to __tls_get_addr(sym\@tlsgd).  Hidden by
356     /// ADDIS_TLSGD_L_ADDR until after register assignment.
357     GET_TLS_ADDR,
358 
359     /// %x3 = GET_TPOINTER - Used for the local-exec TLS model on 32-bit AIX,
360     /// produces a call to .__get_tpointer to retrieve the thread pointer
361     /// At the end of the call, the thread pointer is found in R3.
362     GET_TPOINTER,
363 
364     /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
365     /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
366     /// register assignment.
367     ADDI_TLSGD_L_ADDR,
368 
369     /// GPRC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
370     /// G8RC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
371     /// Op that combines two register copies of TOC entries
372     /// (region handle into R3 and variable offset into R4) followed by a
373     /// GET_TLS_ADDR node which will be expanded to a call to __get_tls_addr.
374     /// This node is used in 64-bit mode as well (in which case the result is
375     /// G8RC and inputs are X3/X4).
376     TLSGD_AIX,
377 
378     /// G8RC = ADDIS_TLSLD_HA %x2, Symbol - For the local-dynamic TLS
379     /// model, produces an ADDIS8 instruction that adds the GOT base
380     /// register to sym\@got\@tlsld\@ha.
381     ADDIS_TLSLD_HA,
382 
383     /// %x3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
384     /// model, produces an ADDI8 instruction that adds G8RReg to
385     /// sym\@got\@tlsld\@l and stores the result in X3.  Hidden by
386     /// ADDIS_TLSLD_L_ADDR until after register assignment.
387     ADDI_TLSLD_L,
388 
389     /// %x3 = GET_TLSLD_ADDR %x3, Symbol - For the local-dynamic TLS
390     /// model, produces a call to __tls_get_addr(sym\@tlsld).  Hidden by
391     /// ADDIS_TLSLD_L_ADDR until after register assignment.
392     GET_TLSLD_ADDR,
393 
394     /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
395     /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
396     /// following register assignment.
397     ADDI_TLSLD_L_ADDR,
398 
399     /// G8RC = ADDIS_DTPREL_HA %x3, Symbol - For the local-dynamic TLS
400     /// model, produces an ADDIS8 instruction that adds X3 to
401     /// sym\@dtprel\@ha.
402     ADDIS_DTPREL_HA,
403 
404     /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
405     /// model, produces an ADDI8 instruction that adds G8RReg to
406     /// sym\@got\@dtprel\@l.
407     ADDI_DTPREL_L,
408 
409     /// G8RC = PADDI_DTPREL %x3, Symbol - For the pc-rel based local-dynamic TLS
410     /// model, produces a PADDI8 instruction that adds X3 to sym\@dtprel.
411     PADDI_DTPREL,
412 
413     /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
414     /// during instruction selection to optimize a BUILD_VECTOR into
415     /// operations on splats.  This is necessary to avoid losing these
416     /// optimizations due to constant folding.
417     VADD_SPLAT,
418 
419     /// CHAIN = SC CHAIN, Imm128 - System call.  The 7-bit unsigned
420     /// operand identifies the operating system entry point.
421     SC,
422 
423     /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
424     CLRBHRB,
425 
426     /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
427     /// history rolling buffer entry.
428     MFBHRBE,
429 
430     /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
431     RFEBB,
432 
433     /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
434     /// endian.  Maps to an xxswapd instruction that corrects an lxvd2x
435     /// or stxvd2x instruction.  The chain is necessary because the
436     /// sequence replaces a load and needs to provide the same number
437     /// of outputs.
438     XXSWAPD,
439 
440     /// An SDNode for swaps that are not associated with any loads/stores
441     /// and thereby have no chain.
442     SWAP_NO_CHAIN,
443 
444     /// FP_EXTEND_HALF(VECTOR, IDX) - Custom extend upper (IDX=0) half or
445     /// lower (IDX=1) half of v4f32 to v2f64.
446     FP_EXTEND_HALF,
447 
448     /// MAT_PCREL_ADDR = Materialize a PC Relative address. This can be done
449     /// either through an add like PADDI or through a PC Relative load like
450     /// PLD.
451     MAT_PCREL_ADDR,
452 
453     /// TLS_DYNAMIC_MAT_PCREL_ADDR = Materialize a PC Relative address for
454     /// TLS global address when using dynamic access models. This can be done
455     /// through an add like PADDI.
456     TLS_DYNAMIC_MAT_PCREL_ADDR,
457 
458     /// TLS_LOCAL_EXEC_MAT_ADDR = Materialize an address for TLS global address
459     /// when using local exec access models, and when prefixed instructions are
460     /// available. This is used with ADD_TLS to produce an add like PADDI.
461     TLS_LOCAL_EXEC_MAT_ADDR,
462 
463     /// ACC_BUILD = Build an accumulator register from 4 VSX registers.
464     ACC_BUILD,
465 
466     /// PAIR_BUILD = Build a vector pair register from 2 VSX registers.
467     PAIR_BUILD,
468 
469     /// EXTRACT_VSX_REG = Extract one of the underlying vsx registers of
470     /// an accumulator or pair register. This node is needed because
471     /// EXTRACT_SUBVECTOR expects the input and output vectors to have the same
472     /// element type.
473     EXTRACT_VSX_REG,
474 
475     /// XXMFACC = This corresponds to the xxmfacc instruction.
476     XXMFACC,
477 
478     // Constrained conversion from floating point to int
479     STRICT_FCTIDZ = ISD::FIRST_TARGET_STRICTFP_OPCODE,
480     STRICT_FCTIWZ,
481     STRICT_FCTIDUZ,
482     STRICT_FCTIWUZ,
483 
484     /// Constrained integer-to-floating-point conversion instructions.
485     STRICT_FCFID,
486     STRICT_FCFIDU,
487     STRICT_FCFIDS,
488     STRICT_FCFIDUS,
489 
490     /// Constrained floating point add in round-to-zero mode.
491     STRICT_FADDRTZ,
492 
493     // NOTE: The nodes below may require PC-Rel specific patterns if the
494     // address could be PC-Relative. When adding new nodes below, consider
495     // whether or not the address can be PC-Relative and add the corresponding
496     // PC-relative patterns and tests.
497 
498     /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
499     /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
500     /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
501     /// i32.
502     STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
503 
504     /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
505     /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
506     /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
507     /// or i32.
508     LBRX,
509 
510     /// STFIWX - The STFIWX instruction.  The first operand is an input token
511     /// chain, then an f64 value to store, then an address to store it to.
512     STFIWX,
513 
514     /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
515     /// load which sign-extends from a 32-bit integer value into the
516     /// destination 64-bit register.
517     LFIWAX,
518 
519     /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
520     /// load which zero-extends from a 32-bit integer value into the
521     /// destination 64-bit register.
522     LFIWZX,
523 
524     /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
525     /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
526     /// This can be used for converting loaded integers to floating point.
527     LXSIZX,
528 
529     /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
530     /// chain, then an f64 value to store, then an address to store it to,
531     /// followed by a byte-width for the store.
532     STXSIX,
533 
534     /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
535     /// Maps directly to an lxvd2x instruction that will be followed by
536     /// an xxswapd.
537     LXVD2X,
538 
539     /// LXVRZX - Load VSX Vector Rightmost and Zero Extend
540     /// This node represents v1i128 BUILD_VECTOR of a zero extending load
541     /// instruction from <byte, halfword, word, or doubleword> to i128.
542     /// Allows utilization of the Load VSX Vector Rightmost Instructions.
543     LXVRZX,
544 
545     /// VSRC, CHAIN = LOAD_VEC_BE CHAIN, Ptr - Occurs only for little endian.
546     /// Maps directly to one of lxvd2x/lxvw4x/lxvh8x/lxvb16x depending on
547     /// the vector type to load vector in big-endian element order.
548     LOAD_VEC_BE,
549 
550     /// VSRC, CHAIN = LD_VSX_LH CHAIN, Ptr - This is a floating-point load of a
551     /// v2f32 value into the lower half of a VSR register.
552     LD_VSX_LH,
553 
554     /// VSRC, CHAIN = LD_SPLAT, CHAIN, Ptr - a splatting load memory
555     /// instructions such as LXVDSX, LXVWSX.
556     LD_SPLAT,
557 
558     /// VSRC, CHAIN = ZEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory
559     /// that zero-extends.
560     ZEXT_LD_SPLAT,
561 
562     /// VSRC, CHAIN = SEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory
563     /// that sign-extends.
564     SEXT_LD_SPLAT,
565 
566     /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
567     /// Maps directly to an stxvd2x instruction that will be preceded by
568     /// an xxswapd.
569     STXVD2X,
570 
571     /// CHAIN = STORE_VEC_BE CHAIN, VSRC, Ptr - Occurs only for little endian.
572     /// Maps directly to one of stxvd2x/stxvw4x/stxvh8x/stxvb16x depending on
573     /// the vector type to store vector in big-endian element order.
574     STORE_VEC_BE,
575 
576     /// Store scalar integers from VSR.
577     ST_VSR_SCAL_INT,
578 
579     /// ATOMIC_CMP_SWAP - the exact same as the target-independent nodes
580     /// except they ensure that the compare input is zero-extended for
581     /// sub-word versions because the atomic loads zero-extend.
582     ATOMIC_CMP_SWAP_8,
583     ATOMIC_CMP_SWAP_16,
584 
585     /// CHAIN,Glue = STORE_COND CHAIN, GPR, Ptr
586     /// The store conditional instruction ST[BHWD]ARX that produces a glue
587     /// result to attach it to a conditional branch.
588     STORE_COND,
589 
590     /// GPRC = TOC_ENTRY GA, TOC
591     /// Loads the entry for GA from the TOC, where the TOC base is given by
592     /// the last operand.
593     TOC_ENTRY
594   };
595 
596   } // end namespace PPCISD
597 
598   /// Define some predicates that are used for node matching.
599   namespace PPC {
600 
601     /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
602     /// VPKUHUM instruction.
603     bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
604                               SelectionDAG &DAG);
605 
606     /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
607     /// VPKUWUM instruction.
608     bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
609                               SelectionDAG &DAG);
610 
611     /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
612     /// VPKUDUM instruction.
613     bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
614                               SelectionDAG &DAG);
615 
616     /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
617     /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
618     bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
619                             unsigned ShuffleKind, SelectionDAG &DAG);
620 
621     /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
622     /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
623     bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
624                             unsigned ShuffleKind, SelectionDAG &DAG);
625 
626     /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
627     /// a VMRGEW or VMRGOW instruction
628     bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
629                              unsigned ShuffleKind, SelectionDAG &DAG);
630     /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
631     /// for a XXSLDWI instruction.
632     bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
633                               bool &Swap, bool IsLE);
634 
635     /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable
636     /// for a XXBRH instruction.
637     bool isXXBRHShuffleMask(ShuffleVectorSDNode *N);
638 
639     /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable
640     /// for a XXBRW instruction.
641     bool isXXBRWShuffleMask(ShuffleVectorSDNode *N);
642 
643     /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable
644     /// for a XXBRD instruction.
645     bool isXXBRDShuffleMask(ShuffleVectorSDNode *N);
646 
647     /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable
648     /// for a XXBRQ instruction.
649     bool isXXBRQShuffleMask(ShuffleVectorSDNode *N);
650 
651     /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
652     /// for a XXPERMDI instruction.
653     bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
654                               bool &Swap, bool IsLE);
655 
656     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
657     /// shift amount, otherwise return -1.
658     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
659                             SelectionDAG &DAG);
660 
661     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
662     /// specifies a splat of a single element that is suitable for input to
663     /// VSPLTB/VSPLTH/VSPLTW.
664     bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
665 
666     /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
667     /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
668     /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
669     /// vector into the other. This function will also set a couple of
670     /// output parameters for how much the source vector needs to be shifted and
671     /// what byte number needs to be specified for the instruction to put the
672     /// element in the desired location of the target vector.
673     bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
674                          unsigned &InsertAtByte, bool &Swap, bool IsLE);
675 
676     /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is
677     /// appropriate for PPC mnemonics (which have a big endian bias - namely
678     /// elements are counted from the left of the vector register).
679     unsigned getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize,
680                                         SelectionDAG &DAG);
681 
682     /// get_VSPLTI_elt - If this is a build_vector of constants which can be
683     /// formed by using a vspltis[bhw] instruction of the specified element
684     /// size, return the constant being splatted.  The ByteSize field indicates
685     /// the number of bytes of each element [124] -> [bhw].
686     SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
687 
688     // Flags for computing the optimal addressing mode for loads and stores.
689     enum MemOpFlags {
690       MOF_None = 0,
691 
692       // Extension mode for integer loads.
693       MOF_SExt = 1,
694       MOF_ZExt = 1 << 1,
695       MOF_NoExt = 1 << 2,
696 
697       // Address computation flags.
698       MOF_NotAddNorCst = 1 << 5,      // Not const. or sum of ptr and scalar.
699       MOF_RPlusSImm16 = 1 << 6,       // Reg plus signed 16-bit constant.
700       MOF_RPlusLo = 1 << 7,           // Reg plus signed 16-bit relocation
701       MOF_RPlusSImm16Mult4 = 1 << 8,  // Reg plus 16-bit signed multiple of 4.
702       MOF_RPlusSImm16Mult16 = 1 << 9, // Reg plus 16-bit signed multiple of 16.
703       MOF_RPlusSImm34 = 1 << 10,      // Reg plus 34-bit signed constant.
704       MOF_RPlusR = 1 << 11,           // Sum of two variables.
705       MOF_PCRel = 1 << 12,            // PC-Relative relocation.
706       MOF_AddrIsSImm32 = 1 << 13,     // A simple 32-bit constant.
707 
708       // The in-memory type.
709       MOF_SubWordInt = 1 << 15,
710       MOF_WordInt = 1 << 16,
711       MOF_DoubleWordInt = 1 << 17,
712       MOF_ScalarFloat = 1 << 18, // Scalar single or double precision.
713       MOF_Vector = 1 << 19,      // Vector types and quad precision scalars.
714       MOF_Vector256 = 1 << 20,
715 
716       // Subtarget features.
717       MOF_SubtargetBeforeP9 = 1 << 22,
718       MOF_SubtargetP9 = 1 << 23,
719       MOF_SubtargetP10 = 1 << 24,
720       MOF_SubtargetSPE = 1 << 25
721     };
722 
723     // The addressing modes for loads and stores.
724     enum AddrMode {
725       AM_None,
726       AM_DForm,
727       AM_DSForm,
728       AM_DQForm,
729       AM_PrefixDForm,
730       AM_XForm,
731       AM_PCRel
732     };
733   } // end namespace PPC
734 
735   class PPCTargetLowering : public TargetLowering {
736     const PPCSubtarget &Subtarget;
737 
738   public:
739     explicit PPCTargetLowering(const PPCTargetMachine &TM,
740                                const PPCSubtarget &STI);
741 
742     /// getTargetNodeName() - This method returns the name of a target specific
743     /// DAG node.
744     const char *getTargetNodeName(unsigned Opcode) const override;
745 
746     bool isSelectSupported(SelectSupportKind Kind) const override {
747       // PowerPC does not support scalar condition selects on vectors.
748       return (Kind != SelectSupportKind::ScalarCondVectorVal);
749     }
750 
751     /// getPreferredVectorAction - The code we generate when vector types are
752     /// legalized by promoting the integer element type is often much worse
753     /// than code we generate if we widen the type for applicable vector types.
754     /// The issue with promoting is that the vector is scalaraized, individual
755     /// elements promoted and then the vector is rebuilt. So say we load a pair
756     /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
757     /// loads, moves back into VSR's (or memory ops if we don't have moves) and
758     /// then the VPERM for the shuffle. All in all a very slow sequence.
759     TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT)
760       const override {
761       // Default handling for scalable and single-element vectors.
762       if (VT.isScalableVector() || VT.getVectorNumElements() == 1)
763         return TargetLoweringBase::getPreferredVectorAction(VT);
764 
765       // Split and promote vNi1 vectors so we don't produce v256i1/v512i1
766       // types as those are only for MMA instructions.
767       if (VT.getScalarSizeInBits() == 1 && VT.getSizeInBits() > 16)
768         return TypeSplitVector;
769       if (VT.getScalarSizeInBits() == 1)
770         return TypePromoteInteger;
771 
772       // Widen vectors that have reasonably sized elements.
773       if (VT.getScalarSizeInBits() % 8 == 0)
774         return TypeWidenVector;
775       return TargetLoweringBase::getPreferredVectorAction(VT);
776     }
777 
778     bool useSoftFloat() const override;
779 
780     bool hasSPE() const;
781 
782     MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
783       return MVT::i32;
784     }
785 
786     bool isCheapToSpeculateCttz(Type *Ty) const override {
787       return true;
788     }
789 
790     bool isCheapToSpeculateCtlz(Type *Ty) const override {
791       return true;
792     }
793 
794     bool isCtlzFast() const override {
795       return true;
796     }
797 
798     bool isEqualityCmpFoldedWithSignedCmp() const override {
799       return false;
800     }
801 
802     bool hasAndNotCompare(SDValue) const override {
803       return true;
804     }
805 
806     bool preferIncOfAddToSubOfNot(EVT VT) const override;
807 
808     bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
809       return VT.isScalarInteger();
810     }
811 
812     SDValue getNegatedExpression(SDValue Op, SelectionDAG &DAG, bool LegalOps,
813                                  bool OptForSize, NegatibleCost &Cost,
814                                  unsigned Depth = 0) const override;
815 
816     /// getSetCCResultType - Return the ISD::SETCC ValueType
817     EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
818                            EVT VT) const override;
819 
820     /// Return true if target always benefits from combining into FMA for a
821     /// given value type. This must typically return false on targets where FMA
822     /// takes more cycles to execute than FADD.
823     bool enableAggressiveFMAFusion(EVT VT) const override;
824 
825     /// getPreIndexedAddressParts - returns true by value, base pointer and
826     /// offset pointer and addressing mode by reference if the node's address
827     /// can be legally represented as pre-indexed load / store address.
828     bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
829                                    SDValue &Offset,
830                                    ISD::MemIndexedMode &AM,
831                                    SelectionDAG &DAG) const override;
832 
833     /// SelectAddressEVXRegReg - Given the specified addressed, check to see if
834     /// it can be more efficiently represented as [r+imm].
835     bool SelectAddressEVXRegReg(SDValue N, SDValue &Base, SDValue &Index,
836                                 SelectionDAG &DAG) const;
837 
838     /// SelectAddressRegReg - Given the specified addressed, check to see if it
839     /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment
840     /// is non-zero, only accept displacement which is not suitable for [r+imm].
841     /// Returns false if it can be represented by [r+imm], which are preferred.
842     bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
843                              SelectionDAG &DAG,
844                              MaybeAlign EncodingAlignment = std::nullopt) const;
845 
846     /// SelectAddressRegImm - Returns true if the address N can be represented
847     /// by a base register plus a signed 16-bit displacement [r+imm], and if it
848     /// is not better represented as reg+reg. If \p EncodingAlignment is
849     /// non-zero, only accept displacements suitable for instruction encoding
850     /// requirement, i.e. multiples of 4 for DS form.
851     bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
852                              SelectionDAG &DAG,
853                              MaybeAlign EncodingAlignment) const;
854     bool SelectAddressRegImm34(SDValue N, SDValue &Disp, SDValue &Base,
855                                SelectionDAG &DAG) const;
856 
857     /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
858     /// represented as an indexed [r+r] operation.
859     bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
860                                  SelectionDAG &DAG) const;
861 
862     /// SelectAddressPCRel - Represent the specified address as pc relative to
863     /// be represented as [pc+imm]
864     bool SelectAddressPCRel(SDValue N, SDValue &Base) const;
865 
866     Sched::Preference getSchedulingPreference(SDNode *N) const override;
867 
868     /// LowerOperation - Provide custom lowering hooks for some operations.
869     ///
870     SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
871 
872     /// ReplaceNodeResults - Replace the results of node with an illegal result
873     /// type with new values built out of custom code.
874     ///
875     void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
876                             SelectionDAG &DAG) const override;
877 
878     SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
879     SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
880 
881     SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
882 
883     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
884                           SmallVectorImpl<SDNode *> &Created) const override;
885 
886     Register getRegisterByName(const char* RegName, LLT VT,
887                                const MachineFunction &MF) const override;
888 
889     void computeKnownBitsForTargetNode(const SDValue Op,
890                                        KnownBits &Known,
891                                        const APInt &DemandedElts,
892                                        const SelectionDAG &DAG,
893                                        unsigned Depth = 0) const override;
894 
895     Align getPrefLoopAlignment(MachineLoop *ML) const override;
896 
897     bool shouldInsertFencesForAtomic(const Instruction *I) const override {
898       return true;
899     }
900 
901     Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
902                                   AtomicOrdering Ord) const override;
903     Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
904                                    AtomicOrdering Ord) const override;
905 
906     bool shouldInlineQuadwordAtomics() const;
907 
908     TargetLowering::AtomicExpansionKind
909     shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
910 
911     TargetLowering::AtomicExpansionKind
912     shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
913 
914     Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder,
915                                         AtomicRMWInst *AI, Value *AlignedAddr,
916                                         Value *Incr, Value *Mask,
917                                         Value *ShiftAmt,
918                                         AtomicOrdering Ord) const override;
919     Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder,
920                                             AtomicCmpXchgInst *CI,
921                                             Value *AlignedAddr, Value *CmpVal,
922                                             Value *NewVal, Value *Mask,
923                                             AtomicOrdering Ord) const override;
924 
925     MachineBasicBlock *
926     EmitInstrWithCustomInserter(MachineInstr &MI,
927                                 MachineBasicBlock *MBB) const override;
928     MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
929                                         MachineBasicBlock *MBB,
930                                         unsigned AtomicSize,
931                                         unsigned BinOpcode,
932                                         unsigned CmpOpcode = 0,
933                                         unsigned CmpPred = 0) const;
934     MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
935                                                 MachineBasicBlock *MBB,
936                                                 bool is8bit,
937                                                 unsigned Opcode,
938                                                 unsigned CmpOpcode = 0,
939                                                 unsigned CmpPred = 0) const;
940 
941     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
942                                         MachineBasicBlock *MBB) const;
943 
944     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
945                                          MachineBasicBlock *MBB) const;
946 
947     MachineBasicBlock *emitProbedAlloca(MachineInstr &MI,
948                                         MachineBasicBlock *MBB) const;
949 
950     bool hasInlineStackProbe(const MachineFunction &MF) const override;
951 
952     unsigned getStackProbeSize(const MachineFunction &MF) const;
953 
954     ConstraintType getConstraintType(StringRef Constraint) const override;
955 
956     /// Examine constraint string and operand type and determine a weight value.
957     /// The operand object must already have been set up with the operand type.
958     ConstraintWeight getSingleConstraintMatchWeight(
959       AsmOperandInfo &info, const char *constraint) const override;
960 
961     std::pair<unsigned, const TargetRegisterClass *>
962     getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
963                                  StringRef Constraint, MVT VT) const override;
964 
965     /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
966     /// function arguments in the caller parameter area.  This is the actual
967     /// alignment, not its logarithm.
968     uint64_t getByValTypeAlignment(Type *Ty,
969                                    const DataLayout &DL) const override;
970 
971     /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
972     /// vector.  If it is invalid, don't add anything to Ops.
973     void LowerAsmOperandForConstraint(SDValue Op,
974                                       std::string &Constraint,
975                                       std::vector<SDValue> &Ops,
976                                       SelectionDAG &DAG) const override;
977 
978     unsigned
979     getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
980       if (ConstraintCode == "es")
981         return InlineAsm::Constraint_es;
982       else if (ConstraintCode == "Q")
983         return InlineAsm::Constraint_Q;
984       else if (ConstraintCode == "Z")
985         return InlineAsm::Constraint_Z;
986       else if (ConstraintCode == "Zy")
987         return InlineAsm::Constraint_Zy;
988       return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
989     }
990 
991     void CollectTargetIntrinsicOperands(const CallInst &I,
992                                  SmallVectorImpl<SDValue> &Ops,
993                                  SelectionDAG &DAG) const override;
994 
995     /// isLegalAddressingMode - Return true if the addressing mode represented
996     /// by AM is legal for this target, for a load/store of the specified type.
997     bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
998                                Type *Ty, unsigned AS,
999                                Instruction *I = nullptr) const override;
1000 
1001     /// isLegalICmpImmediate - Return true if the specified immediate is legal
1002     /// icmp immediate, that is the target has icmp instructions which can
1003     /// compare a register against the immediate without having to materialize
1004     /// the immediate into a register.
1005     bool isLegalICmpImmediate(int64_t Imm) const override;
1006 
1007     /// isLegalAddImmediate - Return true if the specified immediate is legal
1008     /// add immediate, that is the target has add instructions which can
1009     /// add a register and the immediate without having to materialize
1010     /// the immediate into a register.
1011     bool isLegalAddImmediate(int64_t Imm) const override;
1012 
1013     /// isTruncateFree - Return true if it's free to truncate a value of
1014     /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
1015     /// register X1 to i32 by referencing its sub-register R1.
1016     bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
1017     bool isTruncateFree(EVT VT1, EVT VT2) const override;
1018 
1019     bool isZExtFree(SDValue Val, EVT VT2) const override;
1020 
1021     bool isFPExtFree(EVT DestVT, EVT SrcVT) const override;
1022 
1023     /// Returns true if it is beneficial to convert a load of a constant
1024     /// to just the constant itself.
1025     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1026                                            Type *Ty) const override;
1027 
1028     bool convertSelectOfConstantsToMath(EVT VT) const override {
1029       return true;
1030     }
1031 
1032     bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
1033                                 SDValue C) const override;
1034 
1035     bool isDesirableToTransformToIntegerOp(unsigned Opc,
1036                                            EVT VT) const override {
1037       // Only handle float load/store pair because float(fpr) load/store
1038       // instruction has more cycles than integer(gpr) load/store in PPC.
1039       if (Opc != ISD::LOAD && Opc != ISD::STORE)
1040         return false;
1041       if (VT != MVT::f32 && VT != MVT::f64)
1042         return false;
1043 
1044       return true;
1045     }
1046 
1047     // Returns true if the address of the global is stored in TOC entry.
1048     bool isAccessedAsGotIndirect(SDValue N) const;
1049 
1050     bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
1051 
1052     bool getTgtMemIntrinsic(IntrinsicInfo &Info,
1053                             const CallInst &I,
1054                             MachineFunction &MF,
1055                             unsigned Intrinsic) const override;
1056 
1057     /// It returns EVT::Other if the type should be determined using generic
1058     /// target-independent logic.
1059     EVT getOptimalMemOpType(const MemOp &Op,
1060                             const AttributeList &FuncAttributes) const override;
1061 
1062     /// Is unaligned memory access allowed for the given type, and is it fast
1063     /// relative to software emulation.
1064     bool allowsMisalignedMemoryAccesses(
1065         EVT VT, unsigned AddrSpace, Align Alignment = Align(1),
1066         MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
1067         unsigned *Fast = nullptr) const override;
1068 
1069     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
1070     /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
1071     /// expanded to FMAs when this method returns true, otherwise fmuladd is
1072     /// expanded to fmul + fadd.
1073     bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
1074                                     EVT VT) const override;
1075 
1076     bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override;
1077 
1078     /// isProfitableToHoist - Check if it is profitable to hoist instruction
1079     /// \p I to its dominator block.
1080     /// For example, it is not profitable if \p I and it's only user can form a
1081     /// FMA instruction, because Powerpc prefers FMADD.
1082     bool isProfitableToHoist(Instruction *I) const override;
1083 
1084     const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1085 
1086     // Should we expand the build vector with shuffles?
1087     bool
1088     shouldExpandBuildVectorWithShuffles(EVT VT,
1089                                         unsigned DefinedValues) const override;
1090 
1091     // Keep the zero-extensions for arguments to libcalls.
1092     bool shouldKeepZExtForFP16Conv() const override { return true; }
1093 
1094     /// createFastISel - This method returns a target-specific FastISel object,
1095     /// or null if the target does not support "fast" instruction selection.
1096     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1097                              const TargetLibraryInfo *LibInfo) const override;
1098 
1099     /// Returns true if an argument of type Ty needs to be passed in a
1100     /// contiguous block of registers in calling convention CallConv.
1101     bool functionArgumentNeedsConsecutiveRegisters(
1102         Type *Ty, CallingConv::ID CallConv, bool isVarArg,
1103         const DataLayout &DL) const override {
1104       // We support any array type as "consecutive" block in the parameter
1105       // save area.  The element type defines the alignment requirement and
1106       // whether the argument should go in GPRs, FPRs, or VRs if available.
1107       //
1108       // Note that clang uses this capability both to implement the ELFv2
1109       // homogeneous float/vector aggregate ABI, and to avoid having to use
1110       // "byval" when passing aggregates that might fully fit in registers.
1111       return Ty->isArrayTy();
1112     }
1113 
1114     /// If a physical register, this returns the register that receives the
1115     /// exception address on entry to an EH pad.
1116     Register
1117     getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1118 
1119     /// If a physical register, this returns the register that receives the
1120     /// exception typeid on entry to a landing pad.
1121     Register
1122     getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1123 
1124     /// Override to support customized stack guard loading.
1125     bool useLoadStackGuardNode() const override;
1126     void insertSSPDeclarations(Module &M) const override;
1127     Value *getSDagStackGuard(const Module &M) const override;
1128 
1129     bool isFPImmLegal(const APFloat &Imm, EVT VT,
1130                       bool ForCodeSize) const override;
1131 
1132     unsigned getJumpTableEncoding() const override;
1133     bool isJumpTableRelative() const override;
1134     SDValue getPICJumpTableRelocBase(SDValue Table,
1135                                      SelectionDAG &DAG) const override;
1136     const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1137                                                unsigned JTI,
1138                                                MCContext &Ctx) const override;
1139 
1140     /// SelectOptimalAddrMode - Based on a node N and it's Parent (a MemSDNode),
1141     /// compute the address flags of the node, get the optimal address mode
1142     /// based on the flags, and set the Base and Disp based on the address mode.
1143     PPC::AddrMode SelectOptimalAddrMode(const SDNode *Parent, SDValue N,
1144                                         SDValue &Disp, SDValue &Base,
1145                                         SelectionDAG &DAG,
1146                                         MaybeAlign Align) const;
1147     /// SelectForceXFormMode - Given the specified address, force it to be
1148     /// represented as an indexed [r+r] operation (an XForm instruction).
1149     PPC::AddrMode SelectForceXFormMode(SDValue N, SDValue &Disp, SDValue &Base,
1150                                        SelectionDAG &DAG) const;
1151 
1152     bool splitValueIntoRegisterParts(
1153         SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
1154         unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC)
1155         const override;
1156     /// Structure that collects some common arguments that get passed around
1157     /// between the functions for call lowering.
1158     struct CallFlags {
1159       const CallingConv::ID CallConv;
1160       const bool IsTailCall : 1;
1161       const bool IsVarArg : 1;
1162       const bool IsPatchPoint : 1;
1163       const bool IsIndirect : 1;
1164       const bool HasNest : 1;
1165       const bool NoMerge : 1;
1166 
1167       CallFlags(CallingConv::ID CC, bool IsTailCall, bool IsVarArg,
1168                 bool IsPatchPoint, bool IsIndirect, bool HasNest, bool NoMerge)
1169           : CallConv(CC), IsTailCall(IsTailCall), IsVarArg(IsVarArg),
1170             IsPatchPoint(IsPatchPoint), IsIndirect(IsIndirect),
1171             HasNest(HasNest), NoMerge(NoMerge) {}
1172     };
1173 
1174     CCAssignFn *ccAssignFnForCall(CallingConv::ID CC, bool Return,
1175                                   bool IsVarArg) const;
1176     bool supportsTailCallFor(const CallBase *CB) const;
1177 
1178   private:
1179     struct ReuseLoadInfo {
1180       SDValue Ptr;
1181       SDValue Chain;
1182       SDValue ResChain;
1183       MachinePointerInfo MPI;
1184       bool IsDereferenceable = false;
1185       bool IsInvariant = false;
1186       Align Alignment;
1187       AAMDNodes AAInfo;
1188       const MDNode *Ranges = nullptr;
1189 
1190       ReuseLoadInfo() = default;
1191 
1192       MachineMemOperand::Flags MMOFlags() const {
1193         MachineMemOperand::Flags F = MachineMemOperand::MONone;
1194         if (IsDereferenceable)
1195           F |= MachineMemOperand::MODereferenceable;
1196         if (IsInvariant)
1197           F |= MachineMemOperand::MOInvariant;
1198         return F;
1199       }
1200     };
1201 
1202     // Map that relates a set of common address flags to PPC addressing modes.
1203     std::map<PPC::AddrMode, SmallVector<unsigned, 16>> AddrModesMap;
1204     void initializeAddrModeMap();
1205 
1206     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
1207                              SelectionDAG &DAG,
1208                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
1209     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
1210                          SelectionDAG &DAG) const;
1211 
1212     void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
1213                                 SelectionDAG &DAG, const SDLoc &dl) const;
1214     SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
1215                                      const SDLoc &dl) const;
1216 
1217     bool directMoveIsProfitable(const SDValue &Op) const;
1218     SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
1219                                      const SDLoc &dl) const;
1220 
1221     SDValue LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG,
1222                                  const SDLoc &dl) const;
1223 
1224     SDValue LowerTRUNCATEVector(SDValue Op, SelectionDAG &DAG) const;
1225 
1226     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
1227     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
1228 
1229     bool IsEligibleForTailCallOptimization(
1230         const GlobalValue *CalleeGV, CallingConv::ID CalleeCC,
1231         CallingConv::ID CallerCC, bool isVarArg,
1232         const SmallVectorImpl<ISD::InputArg> &Ins) const;
1233 
1234     bool IsEligibleForTailCallOptimization_64SVR4(
1235         const GlobalValue *CalleeGV, CallingConv::ID CalleeCC,
1236         CallingConv::ID CallerCC, const CallBase *CB, bool isVarArg,
1237         const SmallVectorImpl<ISD::OutputArg> &Outs,
1238         const SmallVectorImpl<ISD::InputArg> &Ins, const Function *CallerFunc,
1239         bool isCalleeExternalSymbol) const;
1240 
1241     bool isEligibleForTCO(const GlobalValue *CalleeGV, CallingConv::ID CalleeCC,
1242                           CallingConv::ID CallerCC, const CallBase *CB,
1243                           bool isVarArg,
1244                           const SmallVectorImpl<ISD::OutputArg> &Outs,
1245                           const SmallVectorImpl<ISD::InputArg> &Ins,
1246                           const Function *CallerFunc,
1247                           bool isCalleeExternalSymbol) const;
1248 
1249     SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
1250                                          SDValue Chain, SDValue &LROpOut,
1251                                          SDValue &FPOpOut,
1252                                          const SDLoc &dl) const;
1253 
1254     SDValue getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, SDValue GA) const;
1255 
1256     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1257     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1258     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1259     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1260     SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1261     SDValue LowerGlobalTLSAddressAIX(SDValue Op, SelectionDAG &DAG) const;
1262     SDValue LowerGlobalTLSAddressLinux(SDValue Op, SelectionDAG &DAG) const;
1263     SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1264     SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1265     SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1266     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1267     SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1268     SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const;
1269     SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1270     SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1271     SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
1272     SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
1273     SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1274     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1275     SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
1276     SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
1277     SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
1278     SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1279     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
1280     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
1281                            const SDLoc &dl) const;
1282     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1283     SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
1284     SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1285     SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1286     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
1287     SDValue LowerFunnelShift(SDValue Op, SelectionDAG &DAG) const;
1288     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1289     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
1290     SDValue LowerVPERM(SDValue Op, SelectionDAG &DAG, ArrayRef<int> PermMask,
1291                        EVT VT, SDValue V1, SDValue V2) const;
1292     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1293     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1294     SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
1295     SDValue LowerBSWAP(SDValue Op, SelectionDAG &DAG) const;
1296     SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
1297     SDValue LowerIS_FPCLASS(SDValue Op, SelectionDAG &DAG) const;
1298     SDValue lowerToLibCall(const char *LibCallName, SDValue Op,
1299                            SelectionDAG &DAG) const;
1300     SDValue lowerLibCallBasedOnType(const char *LibCallFloatName,
1301                                     const char *LibCallDoubleName, SDValue Op,
1302                                     SelectionDAG &DAG) const;
1303     bool isLowringToMASSFiniteSafe(SDValue Op) const;
1304     bool isLowringToMASSSafe(SDValue Op) const;
1305     bool isScalarMASSConversionEnabled() const;
1306     SDValue lowerLibCallBase(const char *LibCallDoubleName,
1307                              const char *LibCallFloatName,
1308                              const char *LibCallDoubleNameFinite,
1309                              const char *LibCallFloatNameFinite, SDValue Op,
1310                              SelectionDAG &DAG) const;
1311     SDValue lowerPow(SDValue Op, SelectionDAG &DAG) const;
1312     SDValue lowerSin(SDValue Op, SelectionDAG &DAG) const;
1313     SDValue lowerCos(SDValue Op, SelectionDAG &DAG) const;
1314     SDValue lowerLog(SDValue Op, SelectionDAG &DAG) const;
1315     SDValue lowerLog10(SDValue Op, SelectionDAG &DAG) const;
1316     SDValue lowerExp(SDValue Op, SelectionDAG &DAG) const;
1317     SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) const;
1318     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1319     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
1320     SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
1321     SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
1322     SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const;
1323 
1324     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
1325     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
1326 
1327     SDValue LowerCallResult(SDValue Chain, SDValue InGlue,
1328                             CallingConv::ID CallConv, bool isVarArg,
1329                             const SmallVectorImpl<ISD::InputArg> &Ins,
1330                             const SDLoc &dl, SelectionDAG &DAG,
1331                             SmallVectorImpl<SDValue> &InVals) const;
1332 
1333     SDValue FinishCall(CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG,
1334                        SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
1335                        SDValue InGlue, SDValue Chain, SDValue CallSeqStart,
1336                        SDValue &Callee, int SPDiff, unsigned NumBytes,
1337                        const SmallVectorImpl<ISD::InputArg> &Ins,
1338                        SmallVectorImpl<SDValue> &InVals,
1339                        const CallBase *CB) const;
1340 
1341     SDValue
1342     LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1343                          const SmallVectorImpl<ISD::InputArg> &Ins,
1344                          const SDLoc &dl, SelectionDAG &DAG,
1345                          SmallVectorImpl<SDValue> &InVals) const override;
1346 
1347     SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
1348                       SmallVectorImpl<SDValue> &InVals) const override;
1349 
1350     bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1351                         bool isVarArg,
1352                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1353                         LLVMContext &Context) const override;
1354 
1355     SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1356                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1357                         const SmallVectorImpl<SDValue> &OutVals,
1358                         const SDLoc &dl, SelectionDAG &DAG) const override;
1359 
1360     SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1361                               SelectionDAG &DAG, SDValue ArgVal,
1362                               const SDLoc &dl) const;
1363 
1364     SDValue LowerFormalArguments_AIX(
1365         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1366         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1367         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1368     SDValue LowerFormalArguments_64SVR4(
1369         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1370         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1371         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1372     SDValue LowerFormalArguments_32SVR4(
1373         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1374         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1375         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1376 
1377     SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
1378                                        SDValue CallSeqStart,
1379                                        ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1380                                        const SDLoc &dl) const;
1381 
1382     SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1383                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1384                              const SmallVectorImpl<SDValue> &OutVals,
1385                              const SmallVectorImpl<ISD::InputArg> &Ins,
1386                              const SDLoc &dl, SelectionDAG &DAG,
1387                              SmallVectorImpl<SDValue> &InVals,
1388                              const CallBase *CB) const;
1389     SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1390                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1391                              const SmallVectorImpl<SDValue> &OutVals,
1392                              const SmallVectorImpl<ISD::InputArg> &Ins,
1393                              const SDLoc &dl, SelectionDAG &DAG,
1394                              SmallVectorImpl<SDValue> &InVals,
1395                              const CallBase *CB) const;
1396     SDValue LowerCall_AIX(SDValue Chain, SDValue Callee, CallFlags CFlags,
1397                           const SmallVectorImpl<ISD::OutputArg> &Outs,
1398                           const SmallVectorImpl<SDValue> &OutVals,
1399                           const SmallVectorImpl<ISD::InputArg> &Ins,
1400                           const SDLoc &dl, SelectionDAG &DAG,
1401                           SmallVectorImpl<SDValue> &InVals,
1402                           const CallBase *CB) const;
1403 
1404     SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1405     SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1406     SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
1407 
1408     SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1409     SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1410     SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1411     SDValue combineStoreFPToInt(SDNode *N, DAGCombinerInfo &DCI) const;
1412     SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1413     SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1414     SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1415     SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1416     SDValue combineMUL(SDNode *N, DAGCombinerInfo &DCI) const;
1417     SDValue combineADD(SDNode *N, DAGCombinerInfo &DCI) const;
1418     SDValue combineFMALike(SDNode *N, DAGCombinerInfo &DCI) const;
1419     SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const;
1420     SDValue combineSetCC(SDNode *N, DAGCombinerInfo &DCI) const;
1421     SDValue combineVectorShuffle(ShuffleVectorSDNode *SVN,
1422                                  SelectionDAG &DAG) const;
1423     SDValue combineVReverseMemOP(ShuffleVectorSDNode *SVN, LSBaseSDNode *LSBase,
1424                                  DAGCombinerInfo &DCI) const;
1425 
1426     /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1427     /// SETCC with integer subtraction when (1) there is a legal way of doing it
1428     /// (2) keeping the result of comparison in GPR has performance benefit.
1429     SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1430 
1431     SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1432                             int &RefinementSteps, bool &UseOneConstNR,
1433                             bool Reciprocal) const override;
1434     SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1435                              int &RefinementSteps) const override;
1436     SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG,
1437                              const DenormalMode &Mode) const override;
1438     SDValue getSqrtResultForDenormInput(SDValue Operand,
1439                                         SelectionDAG &DAG) const override;
1440     unsigned combineRepeatedFPDivisors() const override;
1441 
1442     SDValue
1443     combineElementTruncationToVectorTruncation(SDNode *N,
1444                                                DAGCombinerInfo &DCI) const;
1445 
1446     /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be
1447     /// handled by the VINSERTH instruction introduced in ISA 3.0. This is
1448     /// essentially any shuffle of v8i16 vectors that just inserts one element
1449     /// from one vector into the other.
1450     SDValue lowerToVINSERTH(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1451 
1452     /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be
1453     /// handled by the VINSERTB instruction introduced in ISA 3.0. This is
1454     /// essentially v16i8 vector version of VINSERTH.
1455     SDValue lowerToVINSERTB(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1456 
1457     /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be
1458     /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1.
1459     SDValue lowerToXXSPLTI32DX(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1460 
1461     // Return whether the call instruction can potentially be optimized to a
1462     // tail call. This will cause the optimizers to attempt to move, or
1463     // duplicate return instructions to help enable tail call optimizations.
1464     bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1465     bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
1466 
1467     /// getAddrModeForFlags - Based on the set of address flags, select the most
1468     /// optimal instruction format to match by.
1469     PPC::AddrMode getAddrModeForFlags(unsigned Flags) const;
1470 
1471     /// computeMOFlags - Given a node N and it's Parent (a MemSDNode), compute
1472     /// the address flags of the load/store instruction that is to be matched.
1473     /// The address flags are stored in a map, which is then searched
1474     /// through to determine the optimal load/store instruction format.
1475     unsigned computeMOFlags(const SDNode *Parent, SDValue N,
1476                             SelectionDAG &DAG) const;
1477   }; // end class PPCTargetLowering
1478 
1479   namespace PPC {
1480 
1481     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1482                              const TargetLibraryInfo *LibInfo);
1483 
1484   } // end namespace PPC
1485 
1486   bool isIntS16Immediate(SDNode *N, int16_t &Imm);
1487   bool isIntS16Immediate(SDValue Op, int16_t &Imm);
1488   bool isIntS34Immediate(SDNode *N, int64_t &Imm);
1489   bool isIntS34Immediate(SDValue Op, int64_t &Imm);
1490 
1491   bool convertToNonDenormSingle(APInt &ArgAPInt);
1492   bool convertToNonDenormSingle(APFloat &ArgAPFloat);
1493   bool checkConvertToNonDenormSingle(APFloat &ArgAPFloat);
1494 
1495 } // end namespace llvm
1496 
1497 #endif // LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
1498