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/CodeGenTypes/MachineValueType.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 between FIRST_MEMORY_OPCODE and LAST_MEMORY_OPCODE |
42 | // are considered memory opcodes and are treated differently than other |
43 | // entries. |
44 | enum NodeType : unsigned { |
45 | // Start the numbering where the builtin ops and target ops leave off. |
46 | FIRST_NUMBER = ISD::BUILTIN_OP_END, |
47 | |
48 | /// FSEL - Traditional three-operand fsel node. |
49 | /// |
50 | FSEL, |
51 | |
52 | /// XSMAXC[DQ]P, XSMINC[DQ]P - C-type min/max instructions. |
53 | XSMAXC, |
54 | XSMINC, |
55 | |
56 | /// FCFID - The FCFID instruction, taking an f64 operand and producing |
57 | /// and f64 value containing the FP representation of the integer that |
58 | /// was temporarily in the f64 operand. |
59 | FCFID, |
60 | |
61 | /// Newer FCFID[US] integer-to-floating-point conversion instructions for |
62 | /// unsigned integers and single-precision outputs. |
63 | FCFIDU, |
64 | FCFIDS, |
65 | FCFIDUS, |
66 | |
67 | /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64 |
68 | /// operand, producing an f64 value containing the integer representation |
69 | /// of that FP value. |
70 | FCTIDZ, |
71 | FCTIWZ, |
72 | |
73 | /// Newer FCTI[D,W]UZ floating-point-to-integer conversion instructions for |
74 | /// unsigned integers with round toward zero. |
75 | FCTIDUZ, |
76 | FCTIWUZ, |
77 | |
78 | /// VEXTS, ByteWidth - takes an input in VSFRC and produces an output in |
79 | /// VSFRC that is sign-extended from ByteWidth to a 64-byte integer. |
80 | VEXTS, |
81 | |
82 | /// Reciprocal estimate instructions (unary FP ops). |
83 | FRE, |
84 | FRSQRTE, |
85 | |
86 | /// Test instruction for software square root. |
87 | FTSQRT, |
88 | |
89 | /// Square root instruction. |
90 | FSQRT, |
91 | |
92 | /// VPERM - The PPC VPERM Instruction. |
93 | /// |
94 | VPERM, |
95 | |
96 | /// XXSPLT - The PPC VSX splat instructions |
97 | /// |
98 | XXSPLT, |
99 | |
100 | /// XXSPLTI_SP_TO_DP - The PPC VSX splat instructions for immediates for |
101 | /// converting immediate single precision numbers to double precision |
102 | /// vector or scalar. |
103 | XXSPLTI_SP_TO_DP, |
104 | |
105 | /// XXSPLTI32DX - The PPC XXSPLTI32DX instruction. |
106 | /// |
107 | XXSPLTI32DX, |
108 | |
109 | /// VECINSERT - The PPC vector insert instruction |
110 | /// |
111 | VECINSERT, |
112 | |
113 | /// VECSHL - The PPC vector shift left instruction |
114 | /// |
115 | VECSHL, |
116 | |
117 | /// XXPERMDI - The PPC XXPERMDI instruction |
118 | /// |
119 | XXPERMDI, |
120 | XXPERM, |
121 | |
122 | /// The CMPB instruction (takes two operands of i32 or i64). |
123 | CMPB, |
124 | |
125 | /// Hi/Lo - These represent the high and low 16-bit parts of a global |
126 | /// address respectively. These nodes have two operands, the first of |
127 | /// which must be a TargetGlobalAddress, and the second of which must be a |
128 | /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C', |
129 | /// though these are usually folded into other nodes. |
130 | Hi, |
131 | Lo, |
132 | |
133 | /// The following two target-specific nodes are used for calls through |
134 | /// function pointers in the 64-bit SVR4 ABI. |
135 | |
136 | /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX) |
137 | /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to |
138 | /// compute an allocation on the stack. |
139 | DYNALLOC, |
140 | |
141 | /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to |
142 | /// compute an offset from native SP to the address of the most recent |
143 | /// dynamic alloca. |
144 | DYNAREAOFFSET, |
145 | |
146 | /// To avoid stack clash, allocation is performed by block and each block is |
147 | /// probed. |
148 | PROBED_ALLOCA, |
149 | |
150 | /// The result of the mflr at function entry, used for PIC code. |
151 | GlobalBaseReg, |
152 | |
153 | /// These nodes represent PPC shifts. |
154 | /// |
155 | /// For scalar types, only the last `n + 1` bits of the shift amounts |
156 | /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc. |
157 | /// for exact behaviors. |
158 | /// |
159 | /// For vector types, only the last n bits are used. See vsld. |
160 | SRL, |
161 | SRA, |
162 | SHL, |
163 | |
164 | /// These nodes represent PPC arithmetic operations with carry. |
165 | ADDC, |
166 | ADDE, |
167 | SUBC, |
168 | SUBE, |
169 | |
170 | /// FNMSUB - Negated multiply-subtract instruction. |
171 | FNMSUB, |
172 | |
173 | /// EXTSWSLI = The PPC extswsli instruction, which does an extend-sign |
174 | /// word and shift left immediate. |
175 | EXTSWSLI, |
176 | |
177 | /// The combination of sra[wd]i and addze used to implemented signed |
178 | /// integer division by a power of 2. The first operand is the dividend, |
179 | /// and the second is the constant shift amount (representing the |
180 | /// divisor). |
181 | SRA_ADDZE, |
182 | |
183 | /// CALL - A direct function call. |
184 | /// CALL_NOP is a call with the special NOP which follows 64-bit |
185 | /// CALL_NOTOC the caller does not use the TOC. |
186 | /// SVR4 calls and 32-bit/64-bit AIX calls. |
187 | CALL, |
188 | CALL_NOP, |
189 | CALL_NOTOC, |
190 | |
191 | /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a |
192 | /// MTCTR instruction. |
193 | MTCTR, |
194 | |
195 | /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a |
196 | /// BCTRL instruction. |
197 | BCTRL, |
198 | |
199 | /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl |
200 | /// instruction and the TOC reload required on 64-bit ELF, 32-bit AIX |
201 | /// and 64-bit AIX. |
202 | BCTRL_LOAD_TOC, |
203 | |
204 | /// The variants that implicitly define rounding mode for calls with |
205 | /// strictfp semantics. |
206 | CALL_RM, |
207 | CALL_NOP_RM, |
208 | CALL_NOTOC_RM, |
209 | BCTRL_RM, |
210 | BCTRL_LOAD_TOC_RM, |
211 | |
212 | /// Return with a glue operand, matched by 'blr' |
213 | RET_GLUE, |
214 | |
215 | /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction. |
216 | /// This copies the bits corresponding to the specified CRREG into the |
217 | /// resultant GPR. Bits corresponding to other CR regs are undefined. |
218 | MFOCRF, |
219 | |
220 | /// Direct move from a VSX register to a GPR |
221 | MFVSR, |
222 | |
223 | /// Direct move from a GPR to a VSX register (algebraic) |
224 | MTVSRA, |
225 | |
226 | /// Direct move from a GPR to a VSX register (zero) |
227 | MTVSRZ, |
228 | |
229 | /// Direct move of 2 consecutive GPR to a VSX register. |
230 | BUILD_FP128, |
231 | |
232 | /// BUILD_SPE64 and EXTRACT_SPE are analogous to BUILD_PAIR and |
233 | /// EXTRACT_ELEMENT but take f64 arguments instead of i64, as i64 is |
234 | /// unsupported for this target. |
235 | /// Merge 2 GPRs to a single SPE register. |
236 | BUILD_SPE64, |
237 | |
238 | /// Extract SPE register component, second argument is high or low. |
239 | , |
240 | |
241 | /// Extract a subvector from signed integer vector and convert to FP. |
242 | /// It is primarily used to convert a (widened) illegal integer vector |
243 | /// type to a legal floating point vector type. |
244 | /// For example v2i32 -> widened to v4i32 -> v2f64 |
245 | SINT_VEC_TO_FP, |
246 | |
247 | /// Extract a subvector from unsigned integer vector and convert to FP. |
248 | /// As with SINT_VEC_TO_FP, used for converting illegal types. |
249 | UINT_VEC_TO_FP, |
250 | |
251 | /// PowerPC instructions that have SCALAR_TO_VECTOR semantics tend to |
252 | /// place the value into the least significant element of the most |
253 | /// significant doubleword in the vector. This is not element zero for |
254 | /// anything smaller than a doubleword on either endianness. This node has |
255 | /// the same semantics as SCALAR_TO_VECTOR except that the value remains in |
256 | /// the aforementioned location in the vector register. |
257 | SCALAR_TO_VECTOR_PERMUTED, |
258 | |
259 | // FIXME: Remove these once the ANDI glue bug is fixed: |
260 | /// i1 = ANDI_rec_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the |
261 | /// eq or gt bit of CR0 after executing andi. x, 1. This is used to |
262 | /// implement truncation of i32 or i64 to i1. |
263 | ANDI_rec_1_EQ_BIT, |
264 | ANDI_rec_1_GT_BIT, |
265 | |
266 | // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit |
267 | // target (returns (Lo, Hi)). It takes a chain operand. |
268 | READ_TIME_BASE, |
269 | |
270 | // EH_SJLJ_SETJMP - SjLj exception handling setjmp. |
271 | EH_SJLJ_SETJMP, |
272 | |
273 | // EH_SJLJ_LONGJMP - SjLj exception handling longjmp. |
274 | EH_SJLJ_LONGJMP, |
275 | |
276 | /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP* |
277 | /// instructions. For lack of better number, we use the opcode number |
278 | /// encoding for the OPC field to identify the compare. For example, 838 |
279 | /// is VCMPGTSH. |
280 | VCMP, |
281 | |
282 | /// RESVEC, OUTFLAG = VCMP_rec(LHS, RHS, OPC) - Represents one of the |
283 | /// altivec VCMP*_rec instructions. For lack of better number, we use the |
284 | /// opcode number encoding for the OPC field to identify the compare. For |
285 | /// example, 838 is VCMPGTSH. |
286 | VCMP_rec, |
287 | |
288 | /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This |
289 | /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the |
290 | /// condition register to branch on, OPC is the branch opcode to use (e.g. |
291 | /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is |
292 | /// an optional input flag argument. |
293 | COND_BRANCH, |
294 | |
295 | /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based |
296 | /// loops. |
297 | BDNZ, |
298 | BDZ, |
299 | |
300 | /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding |
301 | /// towards zero. Used only as part of the long double-to-int |
302 | /// conversion sequence. |
303 | FADDRTZ, |
304 | |
305 | /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register. |
306 | MFFS, |
307 | |
308 | /// TC_RETURN - A tail call return. |
309 | /// operand #0 chain |
310 | /// operand #1 callee (register or absolute) |
311 | /// operand #2 stack adjustment |
312 | /// operand #3 optional in flag |
313 | TC_RETURN, |
314 | |
315 | /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls |
316 | CR6SET, |
317 | CR6UNSET, |
318 | |
319 | /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS |
320 | /// for non-position independent code on PPC32. |
321 | PPC32_GOT, |
322 | |
323 | /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and |
324 | /// local dynamic TLS and position indendepent code on PPC32. |
325 | PPC32_PICGOT, |
326 | |
327 | /// G8RC = ADDIS_GOT_TPREL_HA %x2, Symbol - Used by the initial-exec |
328 | /// TLS model, produces an ADDIS8 instruction that adds the GOT |
329 | /// base to sym\@got\@tprel\@ha. |
330 | ADDIS_GOT_TPREL_HA, |
331 | |
332 | /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec |
333 | /// TLS model, produces a LD instruction with base register G8RReg |
334 | /// and offset sym\@got\@tprel\@l. This completes the addition that |
335 | /// finds the offset of "sym" relative to the thread pointer. |
336 | LD_GOT_TPREL_L, |
337 | |
338 | /// G8RC = ADD_TLS G8RReg, Symbol - Can be used by the initial-exec |
339 | /// and local-exec TLS models, produces an ADD instruction that adds |
340 | /// the contents of G8RReg to the thread pointer. Symbol contains a |
341 | /// relocation sym\@tls which is to be replaced by the thread pointer |
342 | /// and identifies to the linker that the instruction is part of a |
343 | /// TLS sequence. |
344 | ADD_TLS, |
345 | |
346 | /// G8RC = ADDIS_TLSGD_HA %x2, Symbol - For the general-dynamic TLS |
347 | /// model, produces an ADDIS8 instruction that adds the GOT base |
348 | /// register to sym\@got\@tlsgd\@ha. |
349 | ADDIS_TLSGD_HA, |
350 | |
351 | /// %x3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS |
352 | /// model, produces an ADDI8 instruction that adds G8RReg to |
353 | /// sym\@got\@tlsgd\@l and stores the result in X3. Hidden by |
354 | /// ADDIS_TLSGD_L_ADDR until after register assignment. |
355 | ADDI_TLSGD_L, |
356 | |
357 | /// %x3 = GET_TLS_ADDR %x3, Symbol - For the general-dynamic TLS |
358 | /// model, produces a call to __tls_get_addr(sym\@tlsgd). Hidden by |
359 | /// ADDIS_TLSGD_L_ADDR until after register assignment. |
360 | GET_TLS_ADDR, |
361 | |
362 | /// %x3 = GET_TPOINTER - Used for the local- and initial-exec TLS model on |
363 | /// 32-bit AIX, produces a call to .__get_tpointer to retrieve the thread |
364 | /// pointer. At the end of the call, the thread pointer is found in R3. |
365 | GET_TPOINTER, |
366 | |
367 | /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that |
368 | /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following |
369 | /// register assignment. |
370 | ADDI_TLSGD_L_ADDR, |
371 | |
372 | /// GPRC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY |
373 | /// G8RC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY |
374 | /// Op that combines two register copies of TOC entries |
375 | /// (region handle into R3 and variable offset into R4) followed by a |
376 | /// GET_TLS_ADDR node which will be expanded to a call to .__tls_get_addr. |
377 | /// This node is used in 64-bit mode as well (in which case the result is |
378 | /// G8RC and inputs are X3/X4). |
379 | TLSGD_AIX, |
380 | |
381 | /// %x3 = GET_TLS_MOD_AIX _$TLSML - For the AIX local-dynamic TLS model, |
382 | /// produces a call to .__tls_get_mod(_$TLSML\@ml). |
383 | GET_TLS_MOD_AIX, |
384 | |
385 | /// [GP|G8]RC = TLSLD_AIX, TOC_ENTRY(module handle) |
386 | /// Op that requires a single input of the module handle TOC entry in R3, |
387 | /// and generates a GET_TLS_MOD_AIX node which will be expanded into a call |
388 | /// to .__tls_get_mod. This node is used in both 32-bit and 64-bit modes. |
389 | /// The only difference is the register class. |
390 | TLSLD_AIX, |
391 | |
392 | /// G8RC = ADDIS_TLSLD_HA %x2, Symbol - For the local-dynamic TLS |
393 | /// model, produces an ADDIS8 instruction that adds the GOT base |
394 | /// register to sym\@got\@tlsld\@ha. |
395 | ADDIS_TLSLD_HA, |
396 | |
397 | /// %x3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS |
398 | /// model, produces an ADDI8 instruction that adds G8RReg to |
399 | /// sym\@got\@tlsld\@l and stores the result in X3. Hidden by |
400 | /// ADDIS_TLSLD_L_ADDR until after register assignment. |
401 | ADDI_TLSLD_L, |
402 | |
403 | /// %x3 = GET_TLSLD_ADDR %x3, Symbol - For the local-dynamic TLS |
404 | /// model, produces a call to __tls_get_addr(sym\@tlsld). Hidden by |
405 | /// ADDIS_TLSLD_L_ADDR until after register assignment. |
406 | GET_TLSLD_ADDR, |
407 | |
408 | /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that |
409 | /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion |
410 | /// following register assignment. |
411 | ADDI_TLSLD_L_ADDR, |
412 | |
413 | /// G8RC = ADDIS_DTPREL_HA %x3, Symbol - For the local-dynamic TLS |
414 | /// model, produces an ADDIS8 instruction that adds X3 to |
415 | /// sym\@dtprel\@ha. |
416 | ADDIS_DTPREL_HA, |
417 | |
418 | /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS |
419 | /// model, produces an ADDI8 instruction that adds G8RReg to |
420 | /// sym\@got\@dtprel\@l. |
421 | ADDI_DTPREL_L, |
422 | |
423 | /// G8RC = PADDI_DTPREL %x3, Symbol - For the pc-rel based local-dynamic TLS |
424 | /// model, produces a PADDI8 instruction that adds X3 to sym\@dtprel. |
425 | PADDI_DTPREL, |
426 | |
427 | /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded |
428 | /// during instruction selection to optimize a BUILD_VECTOR into |
429 | /// operations on splats. This is necessary to avoid losing these |
430 | /// optimizations due to constant folding. |
431 | VADD_SPLAT, |
432 | |
433 | /// CHAIN = SC CHAIN, Imm128 - System call. The 7-bit unsigned |
434 | /// operand identifies the operating system entry point. |
435 | SC, |
436 | |
437 | /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer. |
438 | CLRBHRB, |
439 | |
440 | /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch |
441 | /// history rolling buffer entry. |
442 | MFBHRBE, |
443 | |
444 | /// CHAIN = RFEBB CHAIN, State - Return from event-based branch. |
445 | RFEBB, |
446 | |
447 | /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little |
448 | /// endian. Maps to an xxswapd instruction that corrects an lxvd2x |
449 | /// or stxvd2x instruction. The chain is necessary because the |
450 | /// sequence replaces a load and needs to provide the same number |
451 | /// of outputs. |
452 | XXSWAPD, |
453 | |
454 | /// An SDNode for swaps that are not associated with any loads/stores |
455 | /// and thereby have no chain. |
456 | SWAP_NO_CHAIN, |
457 | |
458 | /// FP_EXTEND_HALF(VECTOR, IDX) - Custom extend upper (IDX=0) half or |
459 | /// lower (IDX=1) half of v4f32 to v2f64. |
460 | FP_EXTEND_HALF, |
461 | |
462 | /// MAT_PCREL_ADDR = Materialize a PC Relative address. This can be done |
463 | /// either through an add like PADDI or through a PC Relative load like |
464 | /// PLD. |
465 | MAT_PCREL_ADDR, |
466 | |
467 | /// TLS_DYNAMIC_MAT_PCREL_ADDR = Materialize a PC Relative address for |
468 | /// TLS global address when using dynamic access models. This can be done |
469 | /// through an add like PADDI. |
470 | TLS_DYNAMIC_MAT_PCREL_ADDR, |
471 | |
472 | /// TLS_LOCAL_EXEC_MAT_ADDR = Materialize an address for TLS global address |
473 | /// when using local exec access models, and when prefixed instructions are |
474 | /// available. This is used with ADD_TLS to produce an add like PADDI. |
475 | TLS_LOCAL_EXEC_MAT_ADDR, |
476 | |
477 | /// ACC_BUILD = Build an accumulator register from 4 VSX registers. |
478 | ACC_BUILD, |
479 | |
480 | /// PAIR_BUILD = Build a vector pair register from 2 VSX registers. |
481 | PAIR_BUILD, |
482 | |
483 | /// EXTRACT_VSX_REG = Extract one of the underlying vsx registers of |
484 | /// an accumulator or pair register. This node is needed because |
485 | /// EXTRACT_SUBVECTOR expects the input and output vectors to have the same |
486 | /// element type. |
487 | , |
488 | |
489 | /// XXMFACC = This corresponds to the xxmfacc instruction. |
490 | XXMFACC, |
491 | |
492 | // Constrained conversion from floating point to int |
493 | FIRST_STRICTFP_OPCODE, |
494 | STRICT_FCTIDZ = FIRST_STRICTFP_OPCODE, |
495 | STRICT_FCTIWZ, |
496 | STRICT_FCTIDUZ, |
497 | STRICT_FCTIWUZ, |
498 | |
499 | /// Constrained integer-to-floating-point conversion instructions. |
500 | STRICT_FCFID, |
501 | STRICT_FCFIDU, |
502 | STRICT_FCFIDS, |
503 | STRICT_FCFIDUS, |
504 | |
505 | /// Constrained floating point add in round-to-zero mode. |
506 | STRICT_FADDRTZ, |
507 | LAST_STRICTFP_OPCODE = STRICT_FADDRTZ, |
508 | |
509 | /// SETBC - The ISA 3.1 (P10) SETBC instruction. |
510 | SETBC, |
511 | |
512 | /// SETBCR - The ISA 3.1 (P10) SETBCR instruction. |
513 | SETBCR, |
514 | |
515 | // NOTE: The nodes below may require PC-Rel specific patterns if the |
516 | // address could be PC-Relative. When adding new nodes below, consider |
517 | // whether or not the address can be PC-Relative and add the corresponding |
518 | // PC-relative patterns and tests. |
519 | |
520 | /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a |
521 | /// byte-swapping store instruction. It byte-swaps the low "Type" bits of |
522 | /// the GPRC input, then stores it through Ptr. Type can be either i16 or |
523 | /// i32. |
524 | FIRST_MEMORY_OPCODE, |
525 | STBRX = FIRST_MEMORY_OPCODE, |
526 | |
527 | /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a |
528 | /// byte-swapping load instruction. It loads "Type" bits, byte swaps it, |
529 | /// then puts it in the bottom bits of the GPRC. TYPE can be either i16 |
530 | /// or i32. |
531 | LBRX, |
532 | |
533 | /// STFIWX - The STFIWX instruction. The first operand is an input token |
534 | /// chain, then an f64 value to store, then an address to store it to. |
535 | STFIWX, |
536 | |
537 | /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point |
538 | /// load which sign-extends from a 32-bit integer value into the |
539 | /// destination 64-bit register. |
540 | LFIWAX, |
541 | |
542 | /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point |
543 | /// load which zero-extends from a 32-bit integer value into the |
544 | /// destination 64-bit register. |
545 | LFIWZX, |
546 | |
547 | /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an |
548 | /// integer smaller than 64 bits into a VSR. The integer is zero-extended. |
549 | /// This can be used for converting loaded integers to floating point. |
550 | LXSIZX, |
551 | |
552 | /// STXSIX - The STXSI[bh]X instruction. The first operand is an input |
553 | /// chain, then an f64 value to store, then an address to store it to, |
554 | /// followed by a byte-width for the store. |
555 | STXSIX, |
556 | |
557 | /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian. |
558 | /// Maps directly to an lxvd2x instruction that will be followed by |
559 | /// an xxswapd. |
560 | LXVD2X, |
561 | |
562 | /// LXVRZX - Load VSX Vector Rightmost and Zero Extend |
563 | /// This node represents v1i128 BUILD_VECTOR of a zero extending load |
564 | /// instruction from <byte, halfword, word, or doubleword> to i128. |
565 | /// Allows utilization of the Load VSX Vector Rightmost Instructions. |
566 | LXVRZX, |
567 | |
568 | /// VSRC, CHAIN = LOAD_VEC_BE CHAIN, Ptr - Occurs only for little endian. |
569 | /// Maps directly to one of lxvd2x/lxvw4x/lxvh8x/lxvb16x depending on |
570 | /// the vector type to load vector in big-endian element order. |
571 | LOAD_VEC_BE, |
572 | |
573 | /// VSRC, CHAIN = LD_VSX_LH CHAIN, Ptr - This is a floating-point load of a |
574 | /// v2f32 value into the lower half of a VSR register. |
575 | LD_VSX_LH, |
576 | |
577 | /// VSRC, CHAIN = LD_SPLAT, CHAIN, Ptr - a splatting load memory |
578 | /// instructions such as LXVDSX, LXVWSX. |
579 | LD_SPLAT, |
580 | |
581 | /// VSRC, CHAIN = ZEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory |
582 | /// that zero-extends. |
583 | ZEXT_LD_SPLAT, |
584 | |
585 | /// VSRC, CHAIN = SEXT_LD_SPLAT, CHAIN, Ptr - a splatting load memory |
586 | /// that sign-extends. |
587 | SEXT_LD_SPLAT, |
588 | |
589 | /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian. |
590 | /// Maps directly to an stxvd2x instruction that will be preceded by |
591 | /// an xxswapd. |
592 | STXVD2X, |
593 | |
594 | /// CHAIN = STORE_VEC_BE CHAIN, VSRC, Ptr - Occurs only for little endian. |
595 | /// Maps directly to one of stxvd2x/stxvw4x/stxvh8x/stxvb16x depending on |
596 | /// the vector type to store vector in big-endian element order. |
597 | STORE_VEC_BE, |
598 | |
599 | /// Store scalar integers from VSR. |
600 | ST_VSR_SCAL_INT, |
601 | |
602 | /// ATOMIC_CMP_SWAP - the exact same as the target-independent nodes |
603 | /// except they ensure that the compare input is zero-extended for |
604 | /// sub-word versions because the atomic loads zero-extend. |
605 | ATOMIC_CMP_SWAP_8, |
606 | ATOMIC_CMP_SWAP_16, |
607 | |
608 | /// CHAIN,Glue = STORE_COND CHAIN, GPR, Ptr |
609 | /// The store conditional instruction ST[BHWD]ARX that produces a glue |
610 | /// result to attach it to a conditional branch. |
611 | STORE_COND, |
612 | |
613 | /// GPRC = TOC_ENTRY GA, TOC |
614 | /// Loads the entry for GA from the TOC, where the TOC base is given by |
615 | /// the last operand. |
616 | TOC_ENTRY, |
617 | LAST_MEMORY_OPCODE = TOC_ENTRY, |
618 | }; |
619 | |
620 | } // end namespace PPCISD |
621 | |
622 | /// Define some predicates that are used for node matching. |
623 | namespace PPC { |
624 | |
625 | /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a |
626 | /// VPKUHUM instruction. |
627 | bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
628 | SelectionDAG &DAG); |
629 | |
630 | /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a |
631 | /// VPKUWUM instruction. |
632 | bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
633 | SelectionDAG &DAG); |
634 | |
635 | /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a |
636 | /// VPKUDUM instruction. |
637 | bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
638 | SelectionDAG &DAG); |
639 | |
640 | /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for |
641 | /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes). |
642 | bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
643 | unsigned ShuffleKind, SelectionDAG &DAG); |
644 | |
645 | /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for |
646 | /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes). |
647 | bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
648 | unsigned ShuffleKind, SelectionDAG &DAG); |
649 | |
650 | /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for |
651 | /// a VMRGEW or VMRGOW instruction |
652 | bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven, |
653 | unsigned ShuffleKind, SelectionDAG &DAG); |
654 | /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable |
655 | /// for a XXSLDWI instruction. |
656 | bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
657 | bool &Swap, bool IsLE); |
658 | |
659 | /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable |
660 | /// for a XXBRH instruction. |
661 | bool isXXBRHShuffleMask(ShuffleVectorSDNode *N); |
662 | |
663 | /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable |
664 | /// for a XXBRW instruction. |
665 | bool isXXBRWShuffleMask(ShuffleVectorSDNode *N); |
666 | |
667 | /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable |
668 | /// for a XXBRD instruction. |
669 | bool isXXBRDShuffleMask(ShuffleVectorSDNode *N); |
670 | |
671 | /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable |
672 | /// for a XXBRQ instruction. |
673 | bool isXXBRQShuffleMask(ShuffleVectorSDNode *N); |
674 | |
675 | /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable |
676 | /// for a XXPERMDI instruction. |
677 | bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
678 | bool &Swap, bool IsLE); |
679 | |
680 | /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the |
681 | /// shift amount, otherwise return -1. |
682 | int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind, |
683 | SelectionDAG &DAG); |
684 | |
685 | /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand |
686 | /// specifies a splat of a single element that is suitable for input to |
687 | /// VSPLTB/VSPLTH/VSPLTW. |
688 | bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize); |
689 | |
690 | /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by |
691 | /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any |
692 | /// shuffle of v4f32/v4i32 vectors that just inserts one element from one |
693 | /// vector into the other. This function will also set a couple of |
694 | /// output parameters for how much the source vector needs to be shifted and |
695 | /// what byte number needs to be specified for the instruction to put the |
696 | /// element in the desired location of the target vector. |
697 | bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
698 | unsigned &InsertAtByte, bool &Swap, bool IsLE); |
699 | |
700 | /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is |
701 | /// appropriate for PPC mnemonics (which have a big endian bias - namely |
702 | /// elements are counted from the left of the vector register). |
703 | unsigned getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize, |
704 | SelectionDAG &DAG); |
705 | |
706 | /// get_VSPLTI_elt - If this is a build_vector of constants which can be |
707 | /// formed by using a vspltis[bhw] instruction of the specified element |
708 | /// size, return the constant being splatted. The ByteSize field indicates |
709 | /// the number of bytes of each element [124] -> [bhw]. |
710 | SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG); |
711 | |
712 | // Flags for computing the optimal addressing mode for loads and stores. |
713 | enum MemOpFlags { |
714 | MOF_None = 0, |
715 | |
716 | // Extension mode for integer loads. |
717 | MOF_SExt = 1, |
718 | MOF_ZExt = 1 << 1, |
719 | MOF_NoExt = 1 << 2, |
720 | |
721 | // Address computation flags. |
722 | MOF_NotAddNorCst = 1 << 5, // Not const. or sum of ptr and scalar. |
723 | MOF_RPlusSImm16 = 1 << 6, // Reg plus signed 16-bit constant. |
724 | MOF_RPlusLo = 1 << 7, // Reg plus signed 16-bit relocation |
725 | MOF_RPlusSImm16Mult4 = 1 << 8, // Reg plus 16-bit signed multiple of 4. |
726 | MOF_RPlusSImm16Mult16 = 1 << 9, // Reg plus 16-bit signed multiple of 16. |
727 | MOF_RPlusSImm34 = 1 << 10, // Reg plus 34-bit signed constant. |
728 | MOF_RPlusR = 1 << 11, // Sum of two variables. |
729 | MOF_PCRel = 1 << 12, // PC-Relative relocation. |
730 | MOF_AddrIsSImm32 = 1 << 13, // A simple 32-bit constant. |
731 | |
732 | // The in-memory type. |
733 | MOF_SubWordInt = 1 << 15, |
734 | MOF_WordInt = 1 << 16, |
735 | MOF_DoubleWordInt = 1 << 17, |
736 | MOF_ScalarFloat = 1 << 18, // Scalar single or double precision. |
737 | MOF_Vector = 1 << 19, // Vector types and quad precision scalars. |
738 | MOF_Vector256 = 1 << 20, |
739 | |
740 | // Subtarget features. |
741 | MOF_SubtargetBeforeP9 = 1 << 22, |
742 | MOF_SubtargetP9 = 1 << 23, |
743 | MOF_SubtargetP10 = 1 << 24, |
744 | MOF_SubtargetSPE = 1 << 25 |
745 | }; |
746 | |
747 | // The addressing modes for loads and stores. |
748 | enum AddrMode { |
749 | AM_None, |
750 | AM_DForm, |
751 | AM_DSForm, |
752 | AM_DQForm, |
753 | AM_PrefixDForm, |
754 | AM_XForm, |
755 | AM_PCRel |
756 | }; |
757 | } // end namespace PPC |
758 | |
759 | class PPCTargetLowering : public TargetLowering { |
760 | const PPCSubtarget &Subtarget; |
761 | |
762 | public: |
763 | explicit PPCTargetLowering(const PPCTargetMachine &TM, |
764 | const PPCSubtarget &STI); |
765 | |
766 | /// getTargetNodeName() - This method returns the name of a target specific |
767 | /// DAG node. |
768 | const char *getTargetNodeName(unsigned Opcode) const override; |
769 | |
770 | bool isSelectSupported(SelectSupportKind Kind) const override { |
771 | // PowerPC does not support scalar condition selects on vectors. |
772 | return (Kind != SelectSupportKind::ScalarCondVectorVal); |
773 | } |
774 | |
775 | /// getPreferredVectorAction - The code we generate when vector types are |
776 | /// legalized by promoting the integer element type is often much worse |
777 | /// than code we generate if we widen the type for applicable vector types. |
778 | /// The issue with promoting is that the vector is scalaraized, individual |
779 | /// elements promoted and then the vector is rebuilt. So say we load a pair |
780 | /// of v4i8's and shuffle them. This will turn into a mess of 8 extending |
781 | /// loads, moves back into VSR's (or memory ops if we don't have moves) and |
782 | /// then the VPERM for the shuffle. All in all a very slow sequence. |
783 | TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT) |
784 | const override { |
785 | // Default handling for scalable and single-element vectors. |
786 | if (VT.isScalableVector() || VT.getVectorNumElements() == 1) |
787 | return TargetLoweringBase::getPreferredVectorAction(VT); |
788 | |
789 | // Split and promote vNi1 vectors so we don't produce v256i1/v512i1 |
790 | // types as those are only for MMA instructions. |
791 | if (VT.getScalarSizeInBits() == 1 && VT.getSizeInBits() > 16) |
792 | return TypeSplitVector; |
793 | if (VT.getScalarSizeInBits() == 1) |
794 | return TypePromoteInteger; |
795 | |
796 | // Widen vectors that have reasonably sized elements. |
797 | if (VT.getScalarSizeInBits() % 8 == 0) |
798 | return TypeWidenVector; |
799 | return TargetLoweringBase::getPreferredVectorAction(VT); |
800 | } |
801 | |
802 | bool useSoftFloat() const override; |
803 | |
804 | bool hasSPE() const; |
805 | |
806 | MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override { |
807 | return MVT::i32; |
808 | } |
809 | |
810 | bool isCheapToSpeculateCttz(Type *Ty) const override { |
811 | return true; |
812 | } |
813 | |
814 | bool isCheapToSpeculateCtlz(Type *Ty) const override { |
815 | return true; |
816 | } |
817 | |
818 | bool |
819 | (Type *VectorTy, |
820 | unsigned ElemSizeInBits, |
821 | unsigned &Index) const override; |
822 | |
823 | bool isCtlzFast() const override { |
824 | return true; |
825 | } |
826 | |
827 | bool isEqualityCmpFoldedWithSignedCmp() const override { |
828 | return false; |
829 | } |
830 | |
831 | bool hasAndNotCompare(SDValue) const override { |
832 | return true; |
833 | } |
834 | |
835 | bool preferIncOfAddToSubOfNot(EVT VT) const override; |
836 | |
837 | bool convertSetCCLogicToBitwiseLogic(EVT VT) const override { |
838 | return VT.isScalarInteger(); |
839 | } |
840 | |
841 | SDValue getNegatedExpression(SDValue Op, SelectionDAG &DAG, bool LegalOps, |
842 | bool OptForSize, NegatibleCost &Cost, |
843 | unsigned Depth = 0) const override; |
844 | |
845 | /// getSetCCResultType - Return the ISD::SETCC ValueType |
846 | EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, |
847 | EVT VT) const override; |
848 | |
849 | /// Return true if target always benefits from combining into FMA for a |
850 | /// given value type. This must typically return false on targets where FMA |
851 | /// takes more cycles to execute than FADD. |
852 | bool enableAggressiveFMAFusion(EVT VT) const override; |
853 | |
854 | /// getPreIndexedAddressParts - returns true by value, base pointer and |
855 | /// offset pointer and addressing mode by reference if the node's address |
856 | /// can be legally represented as pre-indexed load / store address. |
857 | bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, |
858 | SDValue &Offset, |
859 | ISD::MemIndexedMode &AM, |
860 | SelectionDAG &DAG) const override; |
861 | |
862 | /// SelectAddressEVXRegReg - Given the specified addressed, check to see if |
863 | /// it can be more efficiently represented as [r+imm]. |
864 | bool SelectAddressEVXRegReg(SDValue N, SDValue &Base, SDValue &Index, |
865 | SelectionDAG &DAG) const; |
866 | |
867 | /// SelectAddressRegReg - Given the specified addressed, check to see if it |
868 | /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment |
869 | /// is non-zero, only accept displacement which is not suitable for [r+imm]. |
870 | /// Returns false if it can be represented by [r+imm], which are preferred. |
871 | bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index, |
872 | SelectionDAG &DAG, |
873 | MaybeAlign EncodingAlignment = std::nullopt) const; |
874 | |
875 | /// SelectAddressRegImm - Returns true if the address N can be represented |
876 | /// by a base register plus a signed 16-bit displacement [r+imm], and if it |
877 | /// is not better represented as reg+reg. If \p EncodingAlignment is |
878 | /// non-zero, only accept displacements suitable for instruction encoding |
879 | /// requirement, i.e. multiples of 4 for DS form. |
880 | bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base, |
881 | SelectionDAG &DAG, |
882 | MaybeAlign EncodingAlignment) const; |
883 | bool SelectAddressRegImm34(SDValue N, SDValue &Disp, SDValue &Base, |
884 | SelectionDAG &DAG) const; |
885 | |
886 | /// SelectAddressRegRegOnly - Given the specified addressed, force it to be |
887 | /// represented as an indexed [r+r] operation. |
888 | bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index, |
889 | SelectionDAG &DAG) const; |
890 | |
891 | /// SelectAddressPCRel - Represent the specified address as pc relative to |
892 | /// be represented as [pc+imm] |
893 | bool SelectAddressPCRel(SDValue N, SDValue &Base) const; |
894 | |
895 | Sched::Preference getSchedulingPreference(SDNode *N) const override; |
896 | |
897 | /// LowerOperation - Provide custom lowering hooks for some operations. |
898 | /// |
899 | SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; |
900 | |
901 | /// ReplaceNodeResults - Replace the results of node with an illegal result |
902 | /// type with new values built out of custom code. |
903 | /// |
904 | void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, |
905 | SelectionDAG &DAG) const override; |
906 | |
907 | SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const; |
908 | SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const; |
909 | |
910 | SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; |
911 | |
912 | SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, |
913 | SmallVectorImpl<SDNode *> &Created) const override; |
914 | |
915 | Register getRegisterByName(const char* RegName, LLT VT, |
916 | const MachineFunction &MF) const override; |
917 | |
918 | void computeKnownBitsForTargetNode(const SDValue Op, |
919 | KnownBits &Known, |
920 | const APInt &DemandedElts, |
921 | const SelectionDAG &DAG, |
922 | unsigned Depth = 0) const override; |
923 | |
924 | Align getPrefLoopAlignment(MachineLoop *ML) const override; |
925 | |
926 | bool shouldInsertFencesForAtomic(const Instruction *I) const override { |
927 | return true; |
928 | } |
929 | |
930 | Value *emitLoadLinked(IRBuilderBase &Builder, Type *ValueTy, Value *Addr, |
931 | AtomicOrdering Ord) const override; |
932 | |
933 | Value *emitStoreConditional(IRBuilderBase &Builder, Value *Val, Value *Addr, |
934 | AtomicOrdering Ord) const override; |
935 | |
936 | Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst, |
937 | AtomicOrdering Ord) const override; |
938 | Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst, |
939 | AtomicOrdering Ord) const override; |
940 | |
941 | bool shouldInlineQuadwordAtomics() const; |
942 | |
943 | TargetLowering::AtomicExpansionKind |
944 | shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override; |
945 | |
946 | TargetLowering::AtomicExpansionKind |
947 | shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override; |
948 | |
949 | Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder, |
950 | AtomicRMWInst *AI, Value *AlignedAddr, |
951 | Value *Incr, Value *Mask, |
952 | Value *ShiftAmt, |
953 | AtomicOrdering Ord) const override; |
954 | Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder, |
955 | AtomicCmpXchgInst *CI, |
956 | Value *AlignedAddr, Value *CmpVal, |
957 | Value *NewVal, Value *Mask, |
958 | AtomicOrdering Ord) const override; |
959 | |
960 | MachineBasicBlock * |
961 | EmitInstrWithCustomInserter(MachineInstr &MI, |
962 | MachineBasicBlock *MBB) const override; |
963 | MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI, |
964 | MachineBasicBlock *MBB, |
965 | unsigned AtomicSize, |
966 | unsigned BinOpcode, |
967 | unsigned CmpOpcode = 0, |
968 | unsigned CmpPred = 0) const; |
969 | MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI, |
970 | MachineBasicBlock *MBB, |
971 | bool is8bit, |
972 | unsigned Opcode, |
973 | unsigned CmpOpcode = 0, |
974 | unsigned CmpPred = 0) const; |
975 | |
976 | MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI, |
977 | MachineBasicBlock *MBB) const; |
978 | |
979 | MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI, |
980 | MachineBasicBlock *MBB) const; |
981 | |
982 | MachineBasicBlock *emitProbedAlloca(MachineInstr &MI, |
983 | MachineBasicBlock *MBB) const; |
984 | |
985 | bool hasInlineStackProbe(const MachineFunction &MF) const override; |
986 | |
987 | unsigned getStackProbeSize(const MachineFunction &MF) const; |
988 | |
989 | ConstraintType getConstraintType(StringRef Constraint) const override; |
990 | |
991 | /// Examine constraint string and operand type and determine a weight value. |
992 | /// The operand object must already have been set up with the operand type. |
993 | ConstraintWeight getSingleConstraintMatchWeight( |
994 | AsmOperandInfo &info, const char *constraint) const override; |
995 | |
996 | std::pair<unsigned, const TargetRegisterClass *> |
997 | getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
998 | StringRef Constraint, MVT VT) const override; |
999 | |
1000 | /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate |
1001 | /// function arguments in the caller parameter area. |
1002 | Align getByValTypeAlignment(Type *Ty, const DataLayout &DL) const override; |
1003 | |
1004 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
1005 | /// vector. If it is invalid, don't add anything to Ops. |
1006 | void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint, |
1007 | std::vector<SDValue> &Ops, |
1008 | SelectionDAG &DAG) const override; |
1009 | |
1010 | InlineAsm::ConstraintCode |
1011 | getInlineAsmMemConstraint(StringRef ConstraintCode) const override { |
1012 | if (ConstraintCode == "es" ) |
1013 | return InlineAsm::ConstraintCode::es; |
1014 | else if (ConstraintCode == "Q" ) |
1015 | return InlineAsm::ConstraintCode::Q; |
1016 | else if (ConstraintCode == "Z" ) |
1017 | return InlineAsm::ConstraintCode::Z; |
1018 | else if (ConstraintCode == "Zy" ) |
1019 | return InlineAsm::ConstraintCode::Zy; |
1020 | return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); |
1021 | } |
1022 | |
1023 | void CollectTargetIntrinsicOperands(const CallInst &I, |
1024 | SmallVectorImpl<SDValue> &Ops, |
1025 | SelectionDAG &DAG) const override; |
1026 | |
1027 | /// isLegalAddressingMode - Return true if the addressing mode represented |
1028 | /// by AM is legal for this target, for a load/store of the specified type. |
1029 | bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, |
1030 | Type *Ty, unsigned AS, |
1031 | Instruction *I = nullptr) const override; |
1032 | |
1033 | /// isLegalICmpImmediate - Return true if the specified immediate is legal |
1034 | /// icmp immediate, that is the target has icmp instructions which can |
1035 | /// compare a register against the immediate without having to materialize |
1036 | /// the immediate into a register. |
1037 | bool isLegalICmpImmediate(int64_t Imm) const override; |
1038 | |
1039 | /// isLegalAddImmediate - Return true if the specified immediate is legal |
1040 | /// add immediate, that is the target has add instructions which can |
1041 | /// add a register and the immediate without having to materialize |
1042 | /// the immediate into a register. |
1043 | bool isLegalAddImmediate(int64_t Imm) const override; |
1044 | |
1045 | /// isTruncateFree - Return true if it's free to truncate a value of |
1046 | /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in |
1047 | /// register X1 to i32 by referencing its sub-register R1. |
1048 | bool isTruncateFree(Type *Ty1, Type *Ty2) const override; |
1049 | bool isTruncateFree(EVT VT1, EVT VT2) const override; |
1050 | |
1051 | bool isZExtFree(SDValue Val, EVT VT2) const override; |
1052 | |
1053 | bool isFPExtFree(EVT DestVT, EVT SrcVT) const override; |
1054 | |
1055 | /// Returns true if it is beneficial to convert a load of a constant |
1056 | /// to just the constant itself. |
1057 | bool shouldConvertConstantLoadToIntImm(const APInt &Imm, |
1058 | Type *Ty) const override; |
1059 | |
1060 | bool convertSelectOfConstantsToMath(EVT VT) const override { |
1061 | return true; |
1062 | } |
1063 | |
1064 | bool decomposeMulByConstant(LLVMContext &Context, EVT VT, |
1065 | SDValue C) const override; |
1066 | |
1067 | bool isDesirableToTransformToIntegerOp(unsigned Opc, |
1068 | EVT VT) const override { |
1069 | // Only handle float load/store pair because float(fpr) load/store |
1070 | // instruction has more cycles than integer(gpr) load/store in PPC. |
1071 | if (Opc != ISD::LOAD && Opc != ISD::STORE) |
1072 | return false; |
1073 | if (VT != MVT::f32 && VT != MVT::f64) |
1074 | return false; |
1075 | |
1076 | return true; |
1077 | } |
1078 | |
1079 | // Returns true if the address of the global is stored in TOC entry. |
1080 | bool isAccessedAsGotIndirect(SDValue N) const; |
1081 | |
1082 | bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; |
1083 | |
1084 | bool getTgtMemIntrinsic(IntrinsicInfo &Info, |
1085 | const CallInst &I, |
1086 | MachineFunction &MF, |
1087 | unsigned Intrinsic) const override; |
1088 | |
1089 | /// It returns EVT::Other if the type should be determined using generic |
1090 | /// target-independent logic. |
1091 | EVT getOptimalMemOpType(const MemOp &Op, |
1092 | const AttributeList &FuncAttributes) const override; |
1093 | |
1094 | /// Is unaligned memory access allowed for the given type, and is it fast |
1095 | /// relative to software emulation. |
1096 | bool allowsMisalignedMemoryAccesses( |
1097 | EVT VT, unsigned AddrSpace, Align Alignment = Align(1), |
1098 | MachineMemOperand::Flags Flags = MachineMemOperand::MONone, |
1099 | unsigned *Fast = nullptr) const override; |
1100 | |
1101 | /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster |
1102 | /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be |
1103 | /// expanded to FMAs when this method returns true, otherwise fmuladd is |
1104 | /// expanded to fmul + fadd. |
1105 | bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, |
1106 | EVT VT) const override; |
1107 | |
1108 | bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override; |
1109 | |
1110 | /// isProfitableToHoist - Check if it is profitable to hoist instruction |
1111 | /// \p I to its dominator block. |
1112 | /// For example, it is not profitable if \p I and it's only user can form a |
1113 | /// FMA instruction, because Powerpc prefers FMADD. |
1114 | bool isProfitableToHoist(Instruction *I) const override; |
1115 | |
1116 | const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override; |
1117 | |
1118 | // Should we expand the build vector with shuffles? |
1119 | bool |
1120 | shouldExpandBuildVectorWithShuffles(EVT VT, |
1121 | unsigned DefinedValues) const override; |
1122 | |
1123 | // Keep the zero-extensions for arguments to libcalls. |
1124 | bool shouldKeepZExtForFP16Conv() const override { return true; } |
1125 | |
1126 | /// createFastISel - This method returns a target-specific FastISel object, |
1127 | /// or null if the target does not support "fast" instruction selection. |
1128 | FastISel *createFastISel(FunctionLoweringInfo &FuncInfo, |
1129 | const TargetLibraryInfo *LibInfo) const override; |
1130 | |
1131 | /// Returns true if an argument of type Ty needs to be passed in a |
1132 | /// contiguous block of registers in calling convention CallConv. |
1133 | bool functionArgumentNeedsConsecutiveRegisters( |
1134 | Type *Ty, CallingConv::ID CallConv, bool isVarArg, |
1135 | const DataLayout &DL) const override { |
1136 | // We support any array type as "consecutive" block in the parameter |
1137 | // save area. The element type defines the alignment requirement and |
1138 | // whether the argument should go in GPRs, FPRs, or VRs if available. |
1139 | // |
1140 | // Note that clang uses this capability both to implement the ELFv2 |
1141 | // homogeneous float/vector aggregate ABI, and to avoid having to use |
1142 | // "byval" when passing aggregates that might fully fit in registers. |
1143 | return Ty->isArrayTy(); |
1144 | } |
1145 | |
1146 | /// If a physical register, this returns the register that receives the |
1147 | /// exception address on entry to an EH pad. |
1148 | Register |
1149 | getExceptionPointerRegister(const Constant *PersonalityFn) const override; |
1150 | |
1151 | /// If a physical register, this returns the register that receives the |
1152 | /// exception typeid on entry to a landing pad. |
1153 | Register |
1154 | getExceptionSelectorRegister(const Constant *PersonalityFn) const override; |
1155 | |
1156 | /// Override to support customized stack guard loading. |
1157 | bool useLoadStackGuardNode(const Module &M) const override; |
1158 | void insertSSPDeclarations(Module &M) const override; |
1159 | Value *getSDagStackGuard(const Module &M) const override; |
1160 | |
1161 | bool isFPImmLegal(const APFloat &Imm, EVT VT, |
1162 | bool ForCodeSize) const override; |
1163 | |
1164 | unsigned getJumpTableEncoding() const override; |
1165 | bool isJumpTableRelative() const override; |
1166 | SDValue getPICJumpTableRelocBase(SDValue Table, |
1167 | SelectionDAG &DAG) const override; |
1168 | const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF, |
1169 | unsigned JTI, |
1170 | MCContext &Ctx) const override; |
1171 | |
1172 | /// SelectOptimalAddrMode - Based on a node N and it's Parent (a MemSDNode), |
1173 | /// compute the address flags of the node, get the optimal address mode |
1174 | /// based on the flags, and set the Base and Disp based on the address mode. |
1175 | PPC::AddrMode SelectOptimalAddrMode(const SDNode *Parent, SDValue N, |
1176 | SDValue &Disp, SDValue &Base, |
1177 | SelectionDAG &DAG, |
1178 | MaybeAlign Align) const; |
1179 | /// SelectForceXFormMode - Given the specified address, force it to be |
1180 | /// represented as an indexed [r+r] operation (an XForm instruction). |
1181 | PPC::AddrMode SelectForceXFormMode(SDValue N, SDValue &Disp, SDValue &Base, |
1182 | SelectionDAG &DAG) const; |
1183 | |
1184 | bool splitValueIntoRegisterParts( |
1185 | SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts, |
1186 | unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) |
1187 | const override; |
1188 | /// Structure that collects some common arguments that get passed around |
1189 | /// between the functions for call lowering. |
1190 | struct CallFlags { |
1191 | const CallingConv::ID CallConv; |
1192 | const bool IsTailCall : 1; |
1193 | const bool IsVarArg : 1; |
1194 | const bool IsPatchPoint : 1; |
1195 | const bool IsIndirect : 1; |
1196 | const bool HasNest : 1; |
1197 | const bool NoMerge : 1; |
1198 | |
1199 | CallFlags(CallingConv::ID CC, bool IsTailCall, bool IsVarArg, |
1200 | bool IsPatchPoint, bool IsIndirect, bool HasNest, bool NoMerge) |
1201 | : CallConv(CC), IsTailCall(IsTailCall), IsVarArg(IsVarArg), |
1202 | IsPatchPoint(IsPatchPoint), IsIndirect(IsIndirect), |
1203 | HasNest(HasNest), NoMerge(NoMerge) {} |
1204 | }; |
1205 | |
1206 | CCAssignFn *ccAssignFnForCall(CallingConv::ID CC, bool Return, |
1207 | bool IsVarArg) const; |
1208 | bool supportsTailCallFor(const CallBase *CB) const; |
1209 | |
1210 | private: |
1211 | struct ReuseLoadInfo { |
1212 | SDValue Ptr; |
1213 | SDValue Chain; |
1214 | SDValue ResChain; |
1215 | MachinePointerInfo MPI; |
1216 | bool IsDereferenceable = false; |
1217 | bool IsInvariant = false; |
1218 | Align Alignment; |
1219 | AAMDNodes AAInfo; |
1220 | const MDNode *Ranges = nullptr; |
1221 | |
1222 | ReuseLoadInfo() = default; |
1223 | |
1224 | MachineMemOperand::Flags MMOFlags() const { |
1225 | MachineMemOperand::Flags F = MachineMemOperand::MONone; |
1226 | if (IsDereferenceable) |
1227 | F |= MachineMemOperand::MODereferenceable; |
1228 | if (IsInvariant) |
1229 | F |= MachineMemOperand::MOInvariant; |
1230 | return F; |
1231 | } |
1232 | }; |
1233 | |
1234 | // Map that relates a set of common address flags to PPC addressing modes. |
1235 | std::map<PPC::AddrMode, SmallVector<unsigned, 16>> AddrModesMap; |
1236 | void initializeAddrModeMap(); |
1237 | |
1238 | bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI, |
1239 | SelectionDAG &DAG, |
1240 | ISD::LoadExtType ET = ISD::NON_EXTLOAD) const; |
1241 | |
1242 | void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI, |
1243 | SelectionDAG &DAG, const SDLoc &dl) const; |
1244 | SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG, |
1245 | const SDLoc &dl) const; |
1246 | |
1247 | bool directMoveIsProfitable(const SDValue &Op) const; |
1248 | SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG, |
1249 | const SDLoc &dl) const; |
1250 | |
1251 | SDValue LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG, |
1252 | const SDLoc &dl) const; |
1253 | |
1254 | SDValue LowerTRUNCATEVector(SDValue Op, SelectionDAG &DAG) const; |
1255 | |
1256 | SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const; |
1257 | SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const; |
1258 | |
1259 | bool IsEligibleForTailCallOptimization( |
1260 | const GlobalValue *CalleeGV, CallingConv::ID CalleeCC, |
1261 | CallingConv::ID CallerCC, bool isVarArg, |
1262 | const SmallVectorImpl<ISD::InputArg> &Ins) const; |
1263 | |
1264 | bool IsEligibleForTailCallOptimization_64SVR4( |
1265 | const GlobalValue *CalleeGV, CallingConv::ID CalleeCC, |
1266 | CallingConv::ID CallerCC, const CallBase *CB, bool isVarArg, |
1267 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1268 | const SmallVectorImpl<ISD::InputArg> &Ins, const Function *CallerFunc, |
1269 | bool isCalleeExternalSymbol) const; |
1270 | |
1271 | bool isEligibleForTCO(const GlobalValue *CalleeGV, CallingConv::ID CalleeCC, |
1272 | CallingConv::ID CallerCC, const CallBase *CB, |
1273 | bool isVarArg, |
1274 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1275 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1276 | const Function *CallerFunc, |
1277 | bool isCalleeExternalSymbol) const; |
1278 | |
1279 | SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff, |
1280 | SDValue Chain, SDValue &LROpOut, |
1281 | SDValue &FPOpOut, |
1282 | const SDLoc &dl) const; |
1283 | |
1284 | SDValue getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, SDValue GA) const; |
1285 | |
1286 | SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; |
1287 | SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; |
1288 | SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; |
1289 | SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; |
1290 | SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; |
1291 | SDValue LowerGlobalTLSAddressAIX(SDValue Op, SelectionDAG &DAG) const; |
1292 | SDValue LowerGlobalTLSAddressLinux(SDValue Op, SelectionDAG &DAG) const; |
1293 | SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; |
1294 | SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const; |
1295 | SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const; |
1296 | SDValue (SDValue Op, SelectionDAG &DAG) const; |
1297 | SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const; |
1298 | SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const; |
1299 | SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const; |
1300 | SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const; |
1301 | SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const; |
1302 | SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const; |
1303 | SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const; |
1304 | SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const; |
1305 | SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; |
1306 | SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const; |
1307 | SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const; |
1308 | SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; |
1309 | SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const; |
1310 | SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; |
1311 | SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, |
1312 | const SDLoc &dl) const; |
1313 | SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; |
1314 | SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; |
1315 | SDValue LowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; |
1316 | SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const; |
1317 | SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const; |
1318 | SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const; |
1319 | SDValue LowerFunnelShift(SDValue Op, SelectionDAG &DAG) const; |
1320 | SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; |
1321 | SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; |
1322 | SDValue LowerVPERM(SDValue Op, SelectionDAG &DAG, ArrayRef<int> PermMask, |
1323 | EVT VT, SDValue V1, SDValue V2) const; |
1324 | SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; |
1325 | SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; |
1326 | SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const; |
1327 | SDValue LowerBSWAP(SDValue Op, SelectionDAG &DAG) const; |
1328 | SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const; |
1329 | SDValue LowerIS_FPCLASS(SDValue Op, SelectionDAG &DAG) const; |
1330 | SDValue LowerADDSUBO_CARRY(SDValue Op, SelectionDAG &DAG) const; |
1331 | SDValue LowerADDSUBO(SDValue Op, SelectionDAG &DAG) const; |
1332 | SDValue lowerToLibCall(const char *LibCallName, SDValue Op, |
1333 | SelectionDAG &DAG) const; |
1334 | SDValue lowerLibCallBasedOnType(const char *LibCallFloatName, |
1335 | const char *LibCallDoubleName, SDValue Op, |
1336 | SelectionDAG &DAG) const; |
1337 | bool isLowringToMASSFiniteSafe(SDValue Op) const; |
1338 | bool isLowringToMASSSafe(SDValue Op) const; |
1339 | bool isScalarMASSConversionEnabled() const; |
1340 | SDValue lowerLibCallBase(const char *LibCallDoubleName, |
1341 | const char *LibCallFloatName, |
1342 | const char *LibCallDoubleNameFinite, |
1343 | const char *LibCallFloatNameFinite, SDValue Op, |
1344 | SelectionDAG &DAG) const; |
1345 | SDValue lowerPow(SDValue Op, SelectionDAG &DAG) const; |
1346 | SDValue lowerSin(SDValue Op, SelectionDAG &DAG) const; |
1347 | SDValue lowerCos(SDValue Op, SelectionDAG &DAG) const; |
1348 | SDValue lowerLog(SDValue Op, SelectionDAG &DAG) const; |
1349 | SDValue lowerLog10(SDValue Op, SelectionDAG &DAG) const; |
1350 | SDValue lowerExp(SDValue Op, SelectionDAG &DAG) const; |
1351 | SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) const; |
1352 | SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const; |
1353 | SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const; |
1354 | SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const; |
1355 | SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const; |
1356 | SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const; |
1357 | |
1358 | SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const; |
1359 | SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const; |
1360 | SDValue LowerDMFVectorLoad(SDValue Op, SelectionDAG &DAG) const; |
1361 | SDValue LowerDMFVectorStore(SDValue Op, SelectionDAG &DAG) const; |
1362 | |
1363 | SDValue LowerCallResult(SDValue Chain, SDValue InGlue, |
1364 | CallingConv::ID CallConv, bool isVarArg, |
1365 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1366 | const SDLoc &dl, SelectionDAG &DAG, |
1367 | SmallVectorImpl<SDValue> &InVals) const; |
1368 | |
1369 | SDValue FinishCall(CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG, |
1370 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, |
1371 | SDValue InGlue, SDValue Chain, SDValue CallSeqStart, |
1372 | SDValue &Callee, int SPDiff, unsigned NumBytes, |
1373 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1374 | SmallVectorImpl<SDValue> &InVals, |
1375 | const CallBase *CB) const; |
1376 | |
1377 | SDValue |
1378 | LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1379 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1380 | const SDLoc &dl, SelectionDAG &DAG, |
1381 | SmallVectorImpl<SDValue> &InVals) const override; |
1382 | |
1383 | SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI, |
1384 | SmallVectorImpl<SDValue> &InVals) const override; |
1385 | |
1386 | bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, |
1387 | bool isVarArg, |
1388 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1389 | LLVMContext &Context, const Type *RetTy) const override; |
1390 | |
1391 | SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1392 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1393 | const SmallVectorImpl<SDValue> &OutVals, |
1394 | const SDLoc &dl, SelectionDAG &DAG) const override; |
1395 | |
1396 | SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT, |
1397 | SelectionDAG &DAG, SDValue ArgVal, |
1398 | const SDLoc &dl) const; |
1399 | |
1400 | SDValue LowerFormalArguments_AIX( |
1401 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1402 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1403 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const; |
1404 | SDValue LowerFormalArguments_64SVR4( |
1405 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1406 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1407 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const; |
1408 | SDValue LowerFormalArguments_32SVR4( |
1409 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1410 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1411 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const; |
1412 | |
1413 | SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff, |
1414 | SDValue CallSeqStart, |
1415 | ISD::ArgFlagsTy Flags, SelectionDAG &DAG, |
1416 | const SDLoc &dl) const; |
1417 | |
1418 | SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags, |
1419 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1420 | const SmallVectorImpl<SDValue> &OutVals, |
1421 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1422 | const SDLoc &dl, SelectionDAG &DAG, |
1423 | SmallVectorImpl<SDValue> &InVals, |
1424 | const CallBase *CB) const; |
1425 | SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags, |
1426 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1427 | const SmallVectorImpl<SDValue> &OutVals, |
1428 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1429 | const SDLoc &dl, SelectionDAG &DAG, |
1430 | SmallVectorImpl<SDValue> &InVals, |
1431 | const CallBase *CB) const; |
1432 | SDValue LowerCall_AIX(SDValue Chain, SDValue Callee, CallFlags CFlags, |
1433 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1434 | const SmallVectorImpl<SDValue> &OutVals, |
1435 | const SmallVectorImpl<ISD::InputArg> &Ins, |
1436 | const SDLoc &dl, SelectionDAG &DAG, |
1437 | SmallVectorImpl<SDValue> &InVals, |
1438 | const CallBase *CB) const; |
1439 | |
1440 | SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const; |
1441 | SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const; |
1442 | SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const; |
1443 | |
1444 | SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const; |
1445 | SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const; |
1446 | SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const; |
1447 | SDValue combineStoreFPToInt(SDNode *N, DAGCombinerInfo &DCI) const; |
1448 | SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const; |
1449 | SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const; |
1450 | SDValue combineVectorShift(SDNode *N, DAGCombinerInfo &DCI) const; |
1451 | SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const; |
1452 | SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const; |
1453 | SDValue combineMUL(SDNode *N, DAGCombinerInfo &DCI) const; |
1454 | SDValue combineADD(SDNode *N, DAGCombinerInfo &DCI) const; |
1455 | SDValue combineFMALike(SDNode *N, DAGCombinerInfo &DCI) const; |
1456 | SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const; |
1457 | SDValue combineSetCC(SDNode *N, DAGCombinerInfo &DCI) const; |
1458 | SDValue combineVectorShuffle(ShuffleVectorSDNode *SVN, |
1459 | SelectionDAG &DAG) const; |
1460 | SDValue combineVReverseMemOP(ShuffleVectorSDNode *SVN, LSBaseSDNode *LSBase, |
1461 | DAGCombinerInfo &DCI) const; |
1462 | |
1463 | /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces |
1464 | /// SETCC with integer subtraction when (1) there is a legal way of doing it |
1465 | /// (2) keeping the result of comparison in GPR has performance benefit. |
1466 | SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const; |
1467 | |
1468 | SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled, |
1469 | int &RefinementSteps, bool &UseOneConstNR, |
1470 | bool Reciprocal) const override; |
1471 | SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled, |
1472 | int &RefinementSteps) const override; |
1473 | SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG, |
1474 | const DenormalMode &Mode) const override; |
1475 | SDValue getSqrtResultForDenormInput(SDValue Operand, |
1476 | SelectionDAG &DAG) const override; |
1477 | unsigned combineRepeatedFPDivisors() const override; |
1478 | |
1479 | SDValue |
1480 | combineElementTruncationToVectorTruncation(SDNode *N, |
1481 | DAGCombinerInfo &DCI) const; |
1482 | |
1483 | /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be |
1484 | /// handled by the VINSERTH instruction introduced in ISA 3.0. This is |
1485 | /// essentially any shuffle of v8i16 vectors that just inserts one element |
1486 | /// from one vector into the other. |
1487 | SDValue lowerToVINSERTH(ShuffleVectorSDNode *N, SelectionDAG &DAG) const; |
1488 | |
1489 | /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be |
1490 | /// handled by the VINSERTB instruction introduced in ISA 3.0. This is |
1491 | /// essentially v16i8 vector version of VINSERTH. |
1492 | SDValue lowerToVINSERTB(ShuffleVectorSDNode *N, SelectionDAG &DAG) const; |
1493 | |
1494 | /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be |
1495 | /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1. |
1496 | SDValue lowerToXXSPLTI32DX(ShuffleVectorSDNode *N, SelectionDAG &DAG) const; |
1497 | |
1498 | // Return whether the call instruction can potentially be optimized to a |
1499 | // tail call. This will cause the optimizers to attempt to move, or |
1500 | // duplicate return instructions to help enable tail call optimizations. |
1501 | bool mayBeEmittedAsTailCall(const CallInst *CI) const override; |
1502 | bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override; |
1503 | |
1504 | /// getAddrModeForFlags - Based on the set of address flags, select the most |
1505 | /// optimal instruction format to match by. |
1506 | PPC::AddrMode getAddrModeForFlags(unsigned Flags) const; |
1507 | |
1508 | /// computeMOFlags - Given a node N and it's Parent (a MemSDNode), compute |
1509 | /// the address flags of the load/store instruction that is to be matched. |
1510 | /// The address flags are stored in a map, which is then searched |
1511 | /// through to determine the optimal load/store instruction format. |
1512 | unsigned computeMOFlags(const SDNode *Parent, SDValue N, |
1513 | SelectionDAG &DAG) const; |
1514 | }; // end class PPCTargetLowering |
1515 | |
1516 | namespace PPC { |
1517 | |
1518 | FastISel *createFastISel(FunctionLoweringInfo &FuncInfo, |
1519 | const TargetLibraryInfo *LibInfo); |
1520 | |
1521 | } // end namespace PPC |
1522 | |
1523 | bool isIntS16Immediate(SDNode *N, int16_t &Imm); |
1524 | bool isIntS16Immediate(SDValue Op, int16_t &Imm); |
1525 | bool isIntS34Immediate(SDNode *N, int64_t &Imm); |
1526 | bool isIntS34Immediate(SDValue Op, int64_t &Imm); |
1527 | |
1528 | bool convertToNonDenormSingle(APInt &ArgAPInt); |
1529 | bool convertToNonDenormSingle(APFloat &ArgAPFloat); |
1530 | bool checkConvertToNonDenormSingle(APFloat &ArgAPFloat); |
1531 | |
1532 | } // end namespace llvm |
1533 | |
1534 | #endif // LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H |
1535 | |