blob: 5358b15437ccb9c0ab8805a7df0dde471238aecb [file] [log] [blame]
Andrew Scull5e1ddfa2018-08-14 10:06:54 +01001//===-- llvm/CodeGen/ISDOpcodes.h - CodeGen opcodes -------------*- C++ -*-===//
2//
Andrew Walbran16937d02019-10-22 13:54:20 +01003// 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
Andrew Scull5e1ddfa2018-08-14 10:06:54 +01006//
7//===----------------------------------------------------------------------===//
8//
9// This file declares codegen opcodes and related utilities.
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_CODEGEN_ISDOPCODES_H
14#define LLVM_CODEGEN_ISDOPCODES_H
15
Olivier Deprezf4ef2d02021-04-20 13:36:24 +020016#include "llvm/CodeGen/ValueTypes.h"
17
Andrew Scull5e1ddfa2018-08-14 10:06:54 +010018namespace llvm {
19
20/// ISD namespace - This namespace contains an enum which represents all of the
21/// SelectionDAG node types and value types.
22///
23namespace ISD {
24
Olivier Deprezf4ef2d02021-04-20 13:36:24 +020025//===--------------------------------------------------------------------===//
26/// ISD::NodeType enum - This enum defines the target-independent operators
27/// for a SelectionDAG.
28///
29/// Targets may also define target-dependent operator codes for SDNodes. For
30/// example, on x86, these are the enum values in the X86ISD namespace.
31/// Targets should aim to use target-independent operators to model their
32/// instruction sets as much as possible, and only use target-dependent
33/// operators when they have special requirements.
34///
35/// Finally, during and after selection proper, SNodes may use special
36/// operator codes that correspond directly with MachineInstr opcodes. These
37/// are used to represent selected instructions. See the isMachineOpcode()
38/// and getMachineOpcode() member functions of SDNode.
39///
40enum NodeType {
41
42 /// DELETED_NODE - This is an illegal value that is used to catch
43 /// errors. This opcode is not a legal opcode for any node.
44 DELETED_NODE,
45
46 /// EntryToken - This is the marker used to indicate the start of a region.
47 EntryToken,
48
49 /// TokenFactor - This node takes multiple tokens as input and produces a
50 /// single token result. This is used to represent the fact that the operand
51 /// operators are independent of each other.
52 TokenFactor,
53
54 /// AssertSext, AssertZext - These nodes record if a register contains a
55 /// value that has already been zero or sign extended from a narrower type.
56 /// These nodes take two operands. The first is the node that has already
57 /// been extended, and the second is a value type node indicating the width
58 /// of the extension
59 AssertSext,
60 AssertZext,
61 AssertAlign,
62
63 /// Various leaf nodes.
64 BasicBlock,
65 VALUETYPE,
66 CONDCODE,
67 Register,
68 RegisterMask,
69 Constant,
70 ConstantFP,
71 GlobalAddress,
72 GlobalTLSAddress,
73 FrameIndex,
74 JumpTable,
75 ConstantPool,
76 ExternalSymbol,
77 BlockAddress,
78
79 /// The address of the GOT
80 GLOBAL_OFFSET_TABLE,
81
82 /// FRAMEADDR, RETURNADDR - These nodes represent llvm.frameaddress and
83 /// llvm.returnaddress on the DAG. These nodes take one operand, the index
84 /// of the frame or return address to return. An index of zero corresponds
85 /// to the current function's frame or return address, an index of one to
86 /// the parent's frame or return address, and so on.
87 FRAMEADDR,
88 RETURNADDR,
89
90 /// ADDROFRETURNADDR - Represents the llvm.addressofreturnaddress intrinsic.
91 /// This node takes no operand, returns a target-specific pointer to the
92 /// place in the stack frame where the return address of the current
93 /// function is stored.
94 ADDROFRETURNADDR,
95
96 /// SPONENTRY - Represents the llvm.sponentry intrinsic. Takes no argument
97 /// and returns the stack pointer value at the entry of the current
98 /// function calling this intrinsic.
99 SPONENTRY,
100
101 /// LOCAL_RECOVER - Represents the llvm.localrecover intrinsic.
102 /// Materializes the offset from the local object pointer of another
103 /// function to a particular local object passed to llvm.localescape. The
104 /// operand is the MCSymbol label used to represent this offset, since
105 /// typically the offset is not known until after code generation of the
106 /// parent.
107 LOCAL_RECOVER,
108
109 /// READ_REGISTER, WRITE_REGISTER - This node represents llvm.register on
110 /// the DAG, which implements the named register global variables extension.
111 READ_REGISTER,
112 WRITE_REGISTER,
113
114 /// FRAME_TO_ARGS_OFFSET - This node represents offset from frame pointer to
115 /// first (possible) on-stack argument. This is needed for correct stack
116 /// adjustment during unwind.
117 FRAME_TO_ARGS_OFFSET,
118
119 /// EH_DWARF_CFA - This node represents the pointer to the DWARF Canonical
120 /// Frame Address (CFA), generally the value of the stack pointer at the
121 /// call site in the previous frame.
122 EH_DWARF_CFA,
123
124 /// OUTCHAIN = EH_RETURN(INCHAIN, OFFSET, HANDLER) - This node represents
125 /// 'eh_return' gcc dwarf builtin, which is used to return from
126 /// exception. The general meaning is: adjust stack by OFFSET and pass
127 /// execution to HANDLER. Many platform-related details also :)
128 EH_RETURN,
129
130 /// RESULT, OUTCHAIN = EH_SJLJ_SETJMP(INCHAIN, buffer)
131 /// This corresponds to the eh.sjlj.setjmp intrinsic.
132 /// It takes an input chain and a pointer to the jump buffer as inputs
133 /// and returns an outchain.
134 EH_SJLJ_SETJMP,
135
136 /// OUTCHAIN = EH_SJLJ_LONGJMP(INCHAIN, buffer)
137 /// This corresponds to the eh.sjlj.longjmp intrinsic.
138 /// It takes an input chain and a pointer to the jump buffer as inputs
139 /// and returns an outchain.
140 EH_SJLJ_LONGJMP,
141
142 /// OUTCHAIN = EH_SJLJ_SETUP_DISPATCH(INCHAIN)
143 /// The target initializes the dispatch table here.
144 EH_SJLJ_SETUP_DISPATCH,
145
146 /// TargetConstant* - Like Constant*, but the DAG does not do any folding,
147 /// simplification, or lowering of the constant. They are used for constants
148 /// which are known to fit in the immediate fields of their users, or for
149 /// carrying magic numbers which are not values which need to be
150 /// materialized in registers.
151 TargetConstant,
152 TargetConstantFP,
153
154 /// TargetGlobalAddress - Like GlobalAddress, but the DAG does no folding or
155 /// anything else with this node, and this is valid in the target-specific
156 /// dag, turning into a GlobalAddress operand.
157 TargetGlobalAddress,
158 TargetGlobalTLSAddress,
159 TargetFrameIndex,
160 TargetJumpTable,
161 TargetConstantPool,
162 TargetExternalSymbol,
163 TargetBlockAddress,
164
165 MCSymbol,
166
167 /// TargetIndex - Like a constant pool entry, but with completely
168 /// target-dependent semantics. Holds target flags, a 32-bit index, and a
169 /// 64-bit index. Targets can use this however they like.
170 TargetIndex,
171
172 /// RESULT = INTRINSIC_WO_CHAIN(INTRINSICID, arg1, arg2, ...)
173 /// This node represents a target intrinsic function with no side effects.
174 /// The first operand is the ID number of the intrinsic from the
175 /// llvm::Intrinsic namespace. The operands to the intrinsic follow. The
176 /// node returns the result of the intrinsic.
177 INTRINSIC_WO_CHAIN,
178
179 /// RESULT,OUTCHAIN = INTRINSIC_W_CHAIN(INCHAIN, INTRINSICID, arg1, ...)
180 /// This node represents a target intrinsic function with side effects that
181 /// returns a result. The first operand is a chain pointer. The second is
182 /// the ID number of the intrinsic from the llvm::Intrinsic namespace. The
183 /// operands to the intrinsic follow. The node has two results, the result
184 /// of the intrinsic and an output chain.
185 INTRINSIC_W_CHAIN,
186
187 /// OUTCHAIN = INTRINSIC_VOID(INCHAIN, INTRINSICID, arg1, arg2, ...)
188 /// This node represents a target intrinsic function with side effects that
189 /// does not return a result. The first operand is a chain pointer. The
190 /// second is the ID number of the intrinsic from the llvm::Intrinsic
191 /// namespace. The operands to the intrinsic follow.
192 INTRINSIC_VOID,
193
194 /// CopyToReg - This node has three operands: a chain, a register number to
195 /// set to this value, and a value.
196 CopyToReg,
197
198 /// CopyFromReg - This node indicates that the input value is a virtual or
199 /// physical register that is defined outside of the scope of this
200 /// SelectionDAG. The register is available from the RegisterSDNode object.
201 CopyFromReg,
202
203 /// UNDEF - An undefined node.
204 UNDEF,
205
206 // FREEZE - FREEZE(VAL) returns an arbitrary value if VAL is UNDEF (or
207 // is evaluated to UNDEF), or returns VAL otherwise. Note that each
208 // read of UNDEF can yield different value, but FREEZE(UNDEF) cannot.
209 FREEZE,
210
211 /// EXTRACT_ELEMENT - This is used to get the lower or upper (determined by
212 /// a Constant, which is required to be operand #1) half of the integer or
213 /// float value specified as operand #0. This is only for use before
214 /// legalization, for values that will be broken into multiple registers.
215 EXTRACT_ELEMENT,
216
217 /// BUILD_PAIR - This is the opposite of EXTRACT_ELEMENT in some ways.
218 /// Given two values of the same integer value type, this produces a value
219 /// twice as big. Like EXTRACT_ELEMENT, this can only be used before
220 /// legalization. The lower part of the composite value should be in
221 /// element 0 and the upper part should be in element 1.
222 BUILD_PAIR,
223
224 /// MERGE_VALUES - This node takes multiple discrete operands and returns
225 /// them all as its individual results. This nodes has exactly the same
226 /// number of inputs and outputs. This node is useful for some pieces of the
227 /// code generator that want to think about a single node with multiple
228 /// results, not multiple nodes.
229 MERGE_VALUES,
230
231 /// Simple integer binary arithmetic operators.
232 ADD,
233 SUB,
234 MUL,
235 SDIV,
236 UDIV,
237 SREM,
238 UREM,
239
240 /// SMUL_LOHI/UMUL_LOHI - Multiply two integers of type iN, producing
241 /// a signed/unsigned value of type i[2*N], and return the full value as
242 /// two results, each of type iN.
243 SMUL_LOHI,
244 UMUL_LOHI,
245
246 /// SDIVREM/UDIVREM - Divide two integers and produce both a quotient and
247 /// remainder result.
248 SDIVREM,
249 UDIVREM,
250
251 /// CARRY_FALSE - This node is used when folding other nodes,
252 /// like ADDC/SUBC, which indicate the carry result is always false.
253 CARRY_FALSE,
254
255 /// Carry-setting nodes for multiple precision addition and subtraction.
256 /// These nodes take two operands of the same value type, and produce two
257 /// results. The first result is the normal add or sub result, the second
258 /// result is the carry flag result.
259 /// FIXME: These nodes are deprecated in favor of ADDCARRY and SUBCARRY.
260 /// They are kept around for now to provide a smooth transition path
261 /// toward the use of ADDCARRY/SUBCARRY and will eventually be removed.
262 ADDC,
263 SUBC,
264
265 /// Carry-using nodes for multiple precision addition and subtraction. These
266 /// nodes take three operands: The first two are the normal lhs and rhs to
267 /// the add or sub, and the third is the input carry flag. These nodes
268 /// produce two results; the normal result of the add or sub, and the output
269 /// carry flag. These nodes both read and write a carry flag to allow them
270 /// to them to be chained together for add and sub of arbitrarily large
271 /// values.
272 ADDE,
273 SUBE,
274
275 /// Carry-using nodes for multiple precision addition and subtraction.
276 /// These nodes take three operands: The first two are the normal lhs and
277 /// rhs to the add or sub, and the third is a boolean indicating if there
278 /// is an incoming carry. These nodes produce two results: the normal
279 /// result of the add or sub, and the output carry so they can be chained
280 /// together. The use of this opcode is preferable to adde/sube if the
281 /// target supports it, as the carry is a regular value rather than a
282 /// glue, which allows further optimisation.
283 ADDCARRY,
284 SUBCARRY,
285
286 /// Carry-using overflow-aware nodes for multiple precision addition and
287 /// subtraction. These nodes take three operands: The first two are normal lhs
288 /// and rhs to the add or sub, and the third is a boolean indicating if there
289 /// is an incoming carry. They produce two results: the normal result of the
290 /// add or sub, and a boolean that indicates if an overflow occured (*not*
291 /// flag, because it may be a store to memory, etc.). If the type of the
292 /// boolean is not i1 then the high bits conform to getBooleanContents.
293 SADDO_CARRY,
294 SSUBO_CARRY,
295
296 /// RESULT, BOOL = [SU]ADDO(LHS, RHS) - Overflow-aware nodes for addition.
297 /// These nodes take two operands: the normal LHS and RHS to the add. They
298 /// produce two results: the normal result of the add, and a boolean that
299 /// indicates if an overflow occurred (*not* a flag, because it may be store
300 /// to memory, etc.). If the type of the boolean is not i1 then the high
301 /// bits conform to getBooleanContents.
302 /// These nodes are generated from llvm.[su]add.with.overflow intrinsics.
303 SADDO,
304 UADDO,
305
306 /// Same for subtraction.
307 SSUBO,
308 USUBO,
309
310 /// Same for multiplication.
311 SMULO,
312 UMULO,
313
314 /// RESULT = [US]ADDSAT(LHS, RHS) - Perform saturation addition on 2
315 /// integers with the same bit width (W). If the true value of LHS + RHS
316 /// exceeds the largest value that can be represented by W bits, the
317 /// resulting value is this maximum value. Otherwise, if this value is less
318 /// than the smallest value that can be represented by W bits, the
319 /// resulting value is this minimum value.
320 SADDSAT,
321 UADDSAT,
322
323 /// RESULT = [US]SUBSAT(LHS, RHS) - Perform saturation subtraction on 2
324 /// integers with the same bit width (W). If the true value of LHS - RHS
325 /// exceeds the largest value that can be represented by W bits, the
326 /// resulting value is this maximum value. Otherwise, if this value is less
327 /// than the smallest value that can be represented by W bits, the
328 /// resulting value is this minimum value.
329 SSUBSAT,
330 USUBSAT,
331
332 /// RESULT = [US]SHLSAT(LHS, RHS) - Perform saturation left shift. The first
333 /// operand is the value to be shifted, and the second argument is the amount
334 /// to shift by. Both must be integers of the same bit width (W). If the true
335 /// value of LHS << RHS exceeds the largest value that can be represented by
336 /// W bits, the resulting value is this maximum value, Otherwise, if this
337 /// value is less than the smallest value that can be represented by W bits,
338 /// the resulting value is this minimum value.
339 SSHLSAT,
340 USHLSAT,
341
342 /// RESULT = [US]MULFIX(LHS, RHS, SCALE) - Perform fixed point multiplication
343 /// on
344 /// 2 integers with the same width and scale. SCALE represents the scale of
345 /// both operands as fixed point numbers. This SCALE parameter must be a
346 /// constant integer. A scale of zero is effectively performing
347 /// multiplication on 2 integers.
348 SMULFIX,
349 UMULFIX,
350
351 /// Same as the corresponding unsaturated fixed point instructions, but the
352 /// result is clamped between the min and max values representable by the
353 /// bits of the first 2 operands.
354 SMULFIXSAT,
355 UMULFIXSAT,
356
357 /// RESULT = [US]DIVFIX(LHS, RHS, SCALE) - Perform fixed point division on
358 /// 2 integers with the same width and scale. SCALE represents the scale
359 /// of both operands as fixed point numbers. This SCALE parameter must be a
360 /// constant integer.
361 SDIVFIX,
362 UDIVFIX,
363
364 /// Same as the corresponding unsaturated fixed point instructions, but the
365 /// result is clamped between the min and max values representable by the
366 /// bits of the first 2 operands.
367 SDIVFIXSAT,
368 UDIVFIXSAT,
369
370 /// Simple binary floating point operators.
371 FADD,
372 FSUB,
373 FMUL,
374 FDIV,
375 FREM,
376
377 /// Constrained versions of the binary floating point operators.
378 /// These will be lowered to the simple operators before final selection.
379 /// They are used to limit optimizations while the DAG is being
380 /// optimized.
381 STRICT_FADD,
382 STRICT_FSUB,
383 STRICT_FMUL,
384 STRICT_FDIV,
385 STRICT_FREM,
386 STRICT_FMA,
387
388 /// Constrained versions of libm-equivalent floating point intrinsics.
389 /// These will be lowered to the equivalent non-constrained pseudo-op
390 /// (or expanded to the equivalent library call) before final selection.
391 /// They are used to limit optimizations while the DAG is being optimized.
392 STRICT_FSQRT,
393 STRICT_FPOW,
394 STRICT_FPOWI,
395 STRICT_FSIN,
396 STRICT_FCOS,
397 STRICT_FEXP,
398 STRICT_FEXP2,
399 STRICT_FLOG,
400 STRICT_FLOG10,
401 STRICT_FLOG2,
402 STRICT_FRINT,
403 STRICT_FNEARBYINT,
404 STRICT_FMAXNUM,
405 STRICT_FMINNUM,
406 STRICT_FCEIL,
407 STRICT_FFLOOR,
408 STRICT_FROUND,
409 STRICT_FROUNDEVEN,
410 STRICT_FTRUNC,
411 STRICT_LROUND,
412 STRICT_LLROUND,
413 STRICT_LRINT,
414 STRICT_LLRINT,
415 STRICT_FMAXIMUM,
416 STRICT_FMINIMUM,
417
418 /// STRICT_FP_TO_[US]INT - Convert a floating point value to a signed or
419 /// unsigned integer. These have the same semantics as fptosi and fptoui
420 /// in IR.
421 /// They are used to limit optimizations while the DAG is being optimized.
422 STRICT_FP_TO_SINT,
423 STRICT_FP_TO_UINT,
424
425 /// STRICT_[US]INT_TO_FP - Convert a signed or unsigned integer to
426 /// a floating point value. These have the same semantics as sitofp and
427 /// uitofp in IR.
428 /// They are used to limit optimizations while the DAG is being optimized.
429 STRICT_SINT_TO_FP,
430 STRICT_UINT_TO_FP,
431
432 /// X = STRICT_FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating
433 /// point type down to the precision of the destination VT. TRUNC is a
434 /// flag, which is always an integer that is zero or one. If TRUNC is 0,
435 /// this is a normal rounding, if it is 1, this FP_ROUND is known to not
436 /// change the value of Y.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100437 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200438 /// The TRUNC = 1 case is used in cases where we know that the value will
439 /// not be modified by the node, because Y is not using any of the extra
440 /// precision of source type. This allows certain transformations like
441 /// STRICT_FP_EXTEND(STRICT_FP_ROUND(X,1)) -> X which are not safe for
442 /// STRICT_FP_EXTEND(STRICT_FP_ROUND(X,0)) because the extra bits aren't
443 /// removed.
444 /// It is used to limit optimizations while the DAG is being optimized.
445 STRICT_FP_ROUND,
446
447 /// X = STRICT_FP_EXTEND(Y) - Extend a smaller FP type into a larger FP
448 /// type.
449 /// It is used to limit optimizations while the DAG is being optimized.
450 STRICT_FP_EXTEND,
451
452 /// STRICT_FSETCC/STRICT_FSETCCS - Constrained versions of SETCC, used
453 /// for floating-point operands only. STRICT_FSETCC performs a quiet
454 /// comparison operation, while STRICT_FSETCCS performs a signaling
455 /// comparison operation.
456 STRICT_FSETCC,
457 STRICT_FSETCCS,
458
459 /// FMA - Perform a * b + c with no intermediate rounding step.
460 FMA,
461
462 /// FMAD - Perform a * b + c, while getting the same result as the
463 /// separately rounded operations.
464 FMAD,
465
466 /// FCOPYSIGN(X, Y) - Return the value of X with the sign of Y. NOTE: This
467 /// DAG node does not require that X and Y have the same type, just that
468 /// they are both floating point. X and the result must have the same type.
469 /// FCOPYSIGN(f32, f64) is allowed.
470 FCOPYSIGN,
471
472 /// INT = FGETSIGN(FP) - Return the sign bit of the specified floating point
473 /// value as an integer 0/1 value.
474 FGETSIGN,
475
476 /// Returns platform specific canonical encoding of a floating point number.
477 FCANONICALIZE,
478
479 /// BUILD_VECTOR(ELT0, ELT1, ELT2, ELT3,...) - Return a fixed-width vector
480 /// with the specified, possibly variable, elements. The types of the
481 /// operands must match the vector element type, except that integer types
482 /// are allowed to be larger than the element type, in which case the
483 /// operands are implicitly truncated. The types of the operands must all
484 /// be the same.
485 BUILD_VECTOR,
486
487 /// INSERT_VECTOR_ELT(VECTOR, VAL, IDX) - Returns VECTOR with the element
488 /// at IDX replaced with VAL. If the type of VAL is larger than the vector
489 /// element type then VAL is truncated before replacement.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100490 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200491 /// If VECTOR is a scalable vector, then IDX may be larger than the minimum
492 /// vector width. IDX is not first scaled by the runtime scaling factor of
493 /// VECTOR.
494 INSERT_VECTOR_ELT,
495
496 /// EXTRACT_VECTOR_ELT(VECTOR, IDX) - Returns a single element from VECTOR
497 /// identified by the (potentially variable) element number IDX. If the return
498 /// type is an integer type larger than the element type of the vector, the
499 /// result is extended to the width of the return type. In that case, the high
500 /// bits are undefined.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100501 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200502 /// If VECTOR is a scalable vector, then IDX may be larger than the minimum
503 /// vector width. IDX is not first scaled by the runtime scaling factor of
504 /// VECTOR.
505 EXTRACT_VECTOR_ELT,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100506
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200507 /// CONCAT_VECTORS(VECTOR0, VECTOR1, ...) - Given a number of values of
508 /// vector type with the same length and element type, this produces a
509 /// concatenated vector result value, with length equal to the sum of the
510 /// lengths of the input vectors. If VECTOR0 is a fixed-width vector, then
511 /// VECTOR1..VECTORN must all be fixed-width vectors. Similarly, if VECTOR0
512 /// is a scalable vector, then VECTOR1..VECTORN must all be scalable vectors.
513 CONCAT_VECTORS,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100514
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200515 /// INSERT_SUBVECTOR(VECTOR1, VECTOR2, IDX) - Returns a vector with VECTOR2
516 /// inserted into VECTOR1. IDX represents the starting element number at which
517 /// VECTOR2 will be inserted. IDX must be a constant multiple of T's known
518 /// minimum vector length. Let the type of VECTOR2 be T, then if T is a
519 /// scalable vector, IDX is first scaled by the runtime scaling factor of T.
520 /// The elements of VECTOR1 starting at IDX are overwritten with VECTOR2.
521 /// Elements IDX through (IDX + num_elements(T) - 1) must be valid VECTOR1
522 /// indices. If this condition cannot be determined statically but is false at
523 /// runtime, then the result vector is undefined.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100524 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200525 /// This operation supports inserting a fixed-width vector into a scalable
526 /// vector, but not the other way around.
527 INSERT_SUBVECTOR,
528
529 /// EXTRACT_SUBVECTOR(VECTOR, IDX) - Returns a subvector from VECTOR.
530 /// Let the result type be T, then IDX represents the starting element number
531 /// from which a subvector of type T is extracted. IDX must be a constant
532 /// multiple of T's known minimum vector length. If T is a scalable vector,
533 /// IDX is first scaled by the runtime scaling factor of T. Elements IDX
534 /// through (IDX + num_elements(T) - 1) must be valid VECTOR indices. If this
535 /// condition cannot be determined statically but is false at runtime, then
536 /// the result vector is undefined. The IDX parameter must be a vector index
537 /// constant type, which for most targets will be an integer pointer type.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100538 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200539 /// This operation supports extracting a fixed-width vector from a scalable
540 /// vector, but not the other way around.
541 EXTRACT_SUBVECTOR,
542
543 /// VECTOR_SHUFFLE(VEC1, VEC2) - Returns a vector, of the same type as
544 /// VEC1/VEC2. A VECTOR_SHUFFLE node also contains an array of constant int
545 /// values that indicate which value (or undef) each result element will
546 /// get. These constant ints are accessible through the
547 /// ShuffleVectorSDNode class. This is quite similar to the Altivec
548 /// 'vperm' instruction, except that the indices must be constants and are
549 /// in terms of the element size of VEC1/VEC2, not in terms of bytes.
550 VECTOR_SHUFFLE,
551
552 /// SCALAR_TO_VECTOR(VAL) - This represents the operation of loading a
553 /// scalar value into element 0 of the resultant vector type. The top
554 /// elements 1 to N-1 of the N-element vector are undefined. The type
555 /// of the operand must match the vector element type, except when they
556 /// are integer types. In this case the operand is allowed to be wider
557 /// than the vector element type, and is implicitly truncated to it.
558 SCALAR_TO_VECTOR,
559
560 /// SPLAT_VECTOR(VAL) - Returns a vector with the scalar value VAL
561 /// duplicated in all lanes. The type of the operand must match the vector
562 /// element type, except when they are integer types. In this case the
563 /// operand is allowed to be wider than the vector element type, and is
564 /// implicitly truncated to it.
565 SPLAT_VECTOR,
566
567 /// MULHU/MULHS - Multiply high - Multiply two integers of type iN,
568 /// producing an unsigned/signed value of type i[2*N], then return the top
569 /// part.
570 MULHU,
571 MULHS,
572
573 /// [US]{MIN/MAX} - Binary minimum or maximum or signed or unsigned
574 /// integers.
575 SMIN,
576 SMAX,
577 UMIN,
578 UMAX,
579
580 /// Bitwise operators - logical and, logical or, logical xor.
581 AND,
582 OR,
583 XOR,
584
585 /// ABS - Determine the unsigned absolute value of a signed integer value of
586 /// the same bitwidth.
587 /// Note: A value of INT_MIN will return INT_MIN, no saturation or overflow
588 /// is performed.
589 ABS,
590
591 /// Shift and rotation operations. After legalization, the type of the
592 /// shift amount is known to be TLI.getShiftAmountTy(). Before legalization
593 /// the shift amount can be any type, but care must be taken to ensure it is
594 /// large enough. TLI.getShiftAmountTy() is i8 on some targets, but before
595 /// legalization, types like i1024 can occur and i8 doesn't have enough bits
596 /// to represent the shift amount.
597 /// When the 1st operand is a vector, the shift amount must be in the same
598 /// type. (TLI.getShiftAmountTy() will return the same type when the input
599 /// type is a vector.)
600 /// For rotates and funnel shifts, the shift amount is treated as an unsigned
601 /// amount modulo the element size of the first operand.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100602 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200603 /// Funnel 'double' shifts take 3 operands, 2 inputs and the shift amount.
604 /// fshl(X,Y,Z): (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
605 /// fshr(X,Y,Z): (X << (BW - (Z % BW))) | (Y >> (Z % BW))
606 SHL,
607 SRA,
608 SRL,
609 ROTL,
610 ROTR,
611 FSHL,
612 FSHR,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100613
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200614 /// Byte Swap and Counting operators.
615 BSWAP,
616 CTTZ,
617 CTLZ,
618 CTPOP,
619 BITREVERSE,
620 PARITY,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100621
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200622 /// Bit counting operators with an undefined result for zero inputs.
623 CTTZ_ZERO_UNDEF,
624 CTLZ_ZERO_UNDEF,
625
626 /// Select(COND, TRUEVAL, FALSEVAL). If the type of the boolean COND is not
627 /// i1 then the high bits must conform to getBooleanContents.
628 SELECT,
629
630 /// Select with a vector condition (op #0) and two vector operands (ops #1
631 /// and #2), returning a vector result. All vectors have the same length.
632 /// Much like the scalar select and setcc, each bit in the condition selects
633 /// whether the corresponding result element is taken from op #1 or op #2.
634 /// At first, the VSELECT condition is of vXi1 type. Later, targets may
635 /// change the condition type in order to match the VSELECT node using a
636 /// pattern. The condition follows the BooleanContent format of the target.
637 VSELECT,
638
639 /// Select with condition operator - This selects between a true value and
640 /// a false value (ops #2 and #3) based on the boolean result of comparing
641 /// the lhs and rhs (ops #0 and #1) of a conditional expression with the
642 /// condition code in op #4, a CondCodeSDNode.
643 SELECT_CC,
644
645 /// SetCC operator - This evaluates to a true value iff the condition is
646 /// true. If the result value type is not i1 then the high bits conform
647 /// to getBooleanContents. The operands to this are the left and right
648 /// operands to compare (ops #0, and #1) and the condition code to compare
649 /// them with (op #2) as a CondCodeSDNode. If the operands are vector types
650 /// then the result type must also be a vector type.
651 SETCC,
652
653 /// Like SetCC, ops #0 and #1 are the LHS and RHS operands to compare, but
654 /// op #2 is a boolean indicating if there is an incoming carry. This
655 /// operator checks the result of "LHS - RHS - Carry", and can be used to
656 /// compare two wide integers:
657 /// (setcccarry lhshi rhshi (subcarry lhslo rhslo) cc).
658 /// Only valid for integers.
659 SETCCCARRY,
660
661 /// SHL_PARTS/SRA_PARTS/SRL_PARTS - These operators are used for expanded
662 /// integer shift operations. The operation ordering is:
663 /// [Lo,Hi] = op [LoLHS,HiLHS], Amt
664 SHL_PARTS,
665 SRA_PARTS,
666 SRL_PARTS,
667
668 /// Conversion operators. These are all single input single output
669 /// operations. For all of these, the result type must be strictly
670 /// wider or narrower (depending on the operation) than the source
671 /// type.
672
673 /// SIGN_EXTEND - Used for integer types, replicating the sign bit
674 /// into new bits.
675 SIGN_EXTEND,
676
677 /// ZERO_EXTEND - Used for integer types, zeroing the new bits.
678 ZERO_EXTEND,
679
680 /// ANY_EXTEND - Used for integer types. The high bits are undefined.
681 ANY_EXTEND,
682
683 /// TRUNCATE - Completely drop the high bits.
684 TRUNCATE,
685
686 /// [SU]INT_TO_FP - These operators convert integers (whose interpreted sign
687 /// depends on the first letter) to floating point.
688 SINT_TO_FP,
689 UINT_TO_FP,
690
691 /// SIGN_EXTEND_INREG - This operator atomically performs a SHL/SRA pair to
692 /// sign extend a small value in a large integer register (e.g. sign
693 /// extending the low 8 bits of a 32-bit register to fill the top 24 bits
694 /// with the 7th bit). The size of the smaller type is indicated by the 1th
695 /// operand, a ValueType node.
696 SIGN_EXTEND_INREG,
697
698 /// ANY_EXTEND_VECTOR_INREG(Vector) - This operator represents an
699 /// in-register any-extension of the low lanes of an integer vector. The
700 /// result type must have fewer elements than the operand type, and those
701 /// elements must be larger integer types such that the total size of the
702 /// operand type is less than or equal to the size of the result type. Each
703 /// of the low operand elements is any-extended into the corresponding,
704 /// wider result elements with the high bits becoming undef.
705 /// NOTE: The type legalizer prefers to make the operand and result size
706 /// the same to allow expansion to shuffle vector during op legalization.
707 ANY_EXTEND_VECTOR_INREG,
708
709 /// SIGN_EXTEND_VECTOR_INREG(Vector) - This operator represents an
710 /// in-register sign-extension of the low lanes of an integer vector. The
711 /// result type must have fewer elements than the operand type, and those
712 /// elements must be larger integer types such that the total size of the
713 /// operand type is less than or equal to the size of the result type. Each
714 /// of the low operand elements is sign-extended into the corresponding,
715 /// wider result elements.
716 /// NOTE: The type legalizer prefers to make the operand and result size
717 /// the same to allow expansion to shuffle vector during op legalization.
718 SIGN_EXTEND_VECTOR_INREG,
719
720 /// ZERO_EXTEND_VECTOR_INREG(Vector) - This operator represents an
721 /// in-register zero-extension of the low lanes of an integer vector. The
722 /// result type must have fewer elements than the operand type, and those
723 /// elements must be larger integer types such that the total size of the
724 /// operand type is less than or equal to the size of the result type. Each
725 /// of the low operand elements is zero-extended into the corresponding,
726 /// wider result elements.
727 /// NOTE: The type legalizer prefers to make the operand and result size
728 /// the same to allow expansion to shuffle vector during op legalization.
729 ZERO_EXTEND_VECTOR_INREG,
730
731 /// FP_TO_[US]INT - Convert a floating point value to a signed or unsigned
732 /// integer. These have the same semantics as fptosi and fptoui in IR. If
733 /// the FP value cannot fit in the integer type, the results are undefined.
734 FP_TO_SINT,
735 FP_TO_UINT,
736
737 /// FP_TO_[US]INT_SAT - Convert floating point value in operand 0 to a
738 /// signed or unsigned integer type with the bit width given in operand 1 with
739 /// the following semantics:
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100740 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200741 /// * If the value is NaN, zero is returned.
742 /// * If the value is larger/smaller than the largest/smallest integer,
743 /// the largest/smallest integer is returned (saturation).
744 /// * Otherwise the result of rounding the value towards zero is returned.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100745 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200746 /// The width given in operand 1 must be equal to, or smaller than, the scalar
747 /// result type width. It may end up being smaller than the result witdh as a
748 /// result of integer type legalization.
749 FP_TO_SINT_SAT,
750 FP_TO_UINT_SAT,
751
752 /// X = FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating point type
753 /// down to the precision of the destination VT. TRUNC is a flag, which is
754 /// always an integer that is zero or one. If TRUNC is 0, this is a
755 /// normal rounding, if it is 1, this FP_ROUND is known to not change the
756 /// value of Y.
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100757 ///
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200758 /// The TRUNC = 1 case is used in cases where we know that the value will
759 /// not be modified by the node, because Y is not using any of the extra
760 /// precision of source type. This allows certain transformations like
761 /// FP_EXTEND(FP_ROUND(X,1)) -> X which are not safe for
762 /// FP_EXTEND(FP_ROUND(X,0)) because the extra bits aren't removed.
763 FP_ROUND,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100764
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200765 /// FLT_ROUNDS_ - Returns current rounding mode:
766 /// -1 Undefined
767 /// 0 Round to 0
768 /// 1 Round to nearest
769 /// 2 Round to +inf
770 /// 3 Round to -inf
771 /// Result is rounding mode and chain. Input is a chain.
772 FLT_ROUNDS_,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100773
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200774 /// X = FP_EXTEND(Y) - Extend a smaller FP type into a larger FP type.
775 FP_EXTEND,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100776
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200777 /// BITCAST - This operator converts between integer, vector and FP
778 /// values, as if the value was stored to memory with one type and loaded
779 /// from the same address with the other type (or equivalently for vector
780 /// format conversions, etc). The source and result are required to have
781 /// the same bit size (e.g. f32 <-> i32). This can also be used for
782 /// int-to-int or fp-to-fp conversions, but that is a noop, deleted by
783 /// getNode().
784 ///
785 /// This operator is subtly different from the bitcast instruction from
786 /// LLVM-IR since this node may change the bits in the register. For
787 /// example, this occurs on big-endian NEON and big-endian MSA where the
788 /// layout of the bits in the register depends on the vector type and this
789 /// operator acts as a shuffle operation for some vector type combinations.
790 BITCAST,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100791
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200792 /// ADDRSPACECAST - This operator converts between pointers of different
793 /// address spaces.
794 ADDRSPACECAST,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100795
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200796 /// FP16_TO_FP, FP_TO_FP16 - These operators are used to perform promotions
797 /// and truncation for half-precision (16 bit) floating numbers. These nodes
798 /// form a semi-softened interface for dealing with f16 (as an i16), which
799 /// is often a storage-only type but has native conversions.
800 FP16_TO_FP,
801 FP_TO_FP16,
802 STRICT_FP16_TO_FP,
803 STRICT_FP_TO_FP16,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100804
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200805 /// Perform various unary floating-point operations inspired by libm. For
806 /// FPOWI, the result is undefined if if the integer operand doesn't fit
807 /// into 32 bits.
808 FNEG,
809 FABS,
810 FSQRT,
811 FCBRT,
812 FSIN,
813 FCOS,
814 FPOWI,
815 FPOW,
816 FLOG,
817 FLOG2,
818 FLOG10,
819 FEXP,
820 FEXP2,
821 FCEIL,
822 FTRUNC,
823 FRINT,
824 FNEARBYINT,
825 FROUND,
826 FROUNDEVEN,
827 FFLOOR,
828 LROUND,
829 LLROUND,
830 LRINT,
831 LLRINT,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100832
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200833 /// FMINNUM/FMAXNUM - Perform floating-point minimum or maximum on two
834 /// values.
835 //
836 /// In the case where a single input is a NaN (either signaling or quiet),
837 /// the non-NaN input is returned.
838 ///
839 /// The return value of (FMINNUM 0.0, -0.0) could be either 0.0 or -0.0.
840 FMINNUM,
841 FMAXNUM,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100842
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200843 /// FMINNUM_IEEE/FMAXNUM_IEEE - Perform floating-point minimum or maximum on
844 /// two values, following the IEEE-754 2008 definition. This differs from
845 /// FMINNUM/FMAXNUM in the handling of signaling NaNs. If one input is a
846 /// signaling NaN, returns a quiet NaN.
847 FMINNUM_IEEE,
848 FMAXNUM_IEEE,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100849
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200850 /// FMINIMUM/FMAXIMUM - NaN-propagating minimum/maximum that also treat -0.0
851 /// as less than 0.0. While FMINNUM_IEEE/FMAXNUM_IEEE follow IEEE 754-2008
852 /// semantics, FMINIMUM/FMAXIMUM follow IEEE 754-2018 draft semantics.
853 FMINIMUM,
854 FMAXIMUM,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100855
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200856 /// FSINCOS - Compute both fsin and fcos as a single operation.
857 FSINCOS,
Andrew Scull5e1ddfa2018-08-14 10:06:54 +0100858
Olivier Deprezf4ef2d02021-04-20 13:36:24 +0200859 /// LOAD and STORE have token chains as their first operand, then the same
860 /// operands as an LLVM load/store instruction, then an offset node that
861 /// is added / subtracted from the base pointer to form the address (for
862 /// indexed memory ops).
863 LOAD,
864 STORE,
865
866 /// DYNAMIC_STACKALLOC - Allocate some number of bytes on the stack aligned
867 /// to a specified boundary. This node always has two return values: a new
868 /// stack pointer value and a chain. The first operand is the token chain,
869 /// the second is the number of bytes to allocate, and the third is the
870 /// alignment boundary. The size is guaranteed to be a multiple of the
871 /// stack alignment, and the alignment is guaranteed to be bigger than the
872 /// stack alignment (if required) or 0 to get standard stack alignment.
873 DYNAMIC_STACKALLOC,
874
875 /// Control flow instructions. These all have token chains.
876
877 /// BR - Unconditional branch. The first operand is the chain
878 /// operand, the second is the MBB to branch to.
879 BR,
880
881 /// BRIND - Indirect branch. The first operand is the chain, the second
882 /// is the value to branch to, which must be of the same type as the
883 /// target's pointer type.
884 BRIND,
885
886 /// BR_JT - Jumptable branch. The first operand is the chain, the second
887 /// is the jumptable index, the last one is the jumptable entry index.
888 BR_JT,
889
890 /// BRCOND - Conditional branch. The first operand is the chain, the
891 /// second is the condition, the third is the block to branch to if the
892 /// condition is true. If the type of the condition is not i1, then the
893 /// high bits must conform to getBooleanContents. If the condition is undef,
894 /// it nondeterministically jumps to the block.
895 /// TODO: Its semantics w.r.t undef requires further discussion; we need to
896 /// make it sure that it is consistent with optimizations in MIR & the
897 /// meaning of IMPLICIT_DEF. See https://reviews.llvm.org/D92015
898 BRCOND,
899
900 /// BR_CC - Conditional branch. The behavior is like that of SELECT_CC, in
901 /// that the condition is represented as condition code, and two nodes to
902 /// compare, rather than as a combined SetCC node. The operands in order
903 /// are chain, cc, lhs, rhs, block to branch to if condition is true. If
904 /// condition is undef, it nondeterministically jumps to the block.
905 BR_CC,
906
907 /// INLINEASM - Represents an inline asm block. This node always has two
908 /// return values: a chain and a flag result. The inputs are as follows:
909 /// Operand #0 : Input chain.
910 /// Operand #1 : a ExternalSymbolSDNode with a pointer to the asm string.
911 /// Operand #2 : a MDNodeSDNode with the !srcloc metadata.
912 /// Operand #3 : HasSideEffect, IsAlignStack bits.
913 /// After this, it is followed by a list of operands with this format:
914 /// ConstantSDNode: Flags that encode whether it is a mem or not, the
915 /// of operands that follow, etc. See InlineAsm.h.
916 /// ... however many operands ...
917 /// Operand #last: Optional, an incoming flag.
918 ///
919 /// The variable width operands are required to represent target addressing
920 /// modes as a single "operand", even though they may have multiple
921 /// SDOperands.
922 INLINEASM,
923
924 /// INLINEASM_BR - Branching version of inline asm. Used by asm-goto.
925 INLINEASM_BR,
926
927 /// EH_LABEL - Represents a label in mid basic block used to track
928 /// locations needed for debug and exception handling tables. These nodes
929 /// take a chain as input and return a chain.
930 EH_LABEL,
931
932 /// ANNOTATION_LABEL - Represents a mid basic block label used by
933 /// annotations. This should remain within the basic block and be ordered
934 /// with respect to other call instructions, but loads and stores may float
935 /// past it.
936 ANNOTATION_LABEL,
937
938 /// CATCHRET - Represents a return from a catch block funclet. Used for
939 /// MSVC compatible exception handling. Takes a chain operand and a
940 /// destination basic block operand.
941 CATCHRET,
942
943 /// CLEANUPRET - Represents a return from a cleanup block funclet. Used for
944 /// MSVC compatible exception handling. Takes only a chain operand.
945 CLEANUPRET,
946
947 /// STACKSAVE - STACKSAVE has one operand, an input chain. It produces a
948 /// value, the same type as the pointer type for the system, and an output
949 /// chain.
950 STACKSAVE,
951
952 /// STACKRESTORE has two operands, an input chain and a pointer to restore
953 /// to it returns an output chain.
954 STACKRESTORE,
955
956 /// CALLSEQ_START/CALLSEQ_END - These operators mark the beginning and end
957 /// of a call sequence, and carry arbitrary information that target might
958 /// want to know. The first operand is a chain, the rest are specified by
959 /// the target and not touched by the DAG optimizers.
960 /// Targets that may use stack to pass call arguments define additional
961 /// operands:
962 /// - size of the call frame part that must be set up within the
963 /// CALLSEQ_START..CALLSEQ_END pair,
964 /// - part of the call frame prepared prior to CALLSEQ_START.
965 /// Both these parameters must be constants, their sum is the total call
966 /// frame size.
967 /// CALLSEQ_START..CALLSEQ_END pairs may not be nested.
968 CALLSEQ_START, // Beginning of a call sequence
969 CALLSEQ_END, // End of a call sequence
970
971 /// VAARG - VAARG has four operands: an input chain, a pointer, a SRCVALUE,
972 /// and the alignment. It returns a pair of values: the vaarg value and a
973 /// new chain.
974 VAARG,
975
976 /// VACOPY - VACOPY has 5 operands: an input chain, a destination pointer,
977 /// a source pointer, a SRCVALUE for the destination, and a SRCVALUE for the
978 /// source.
979 VACOPY,
980
981 /// VAEND, VASTART - VAEND and VASTART have three operands: an input chain,
982 /// pointer, and a SRCVALUE.
983 VAEND,
984 VASTART,
985
986 // PREALLOCATED_SETUP - This has 2 operands: an input chain and a SRCVALUE
987 // with the preallocated call Value.
988 PREALLOCATED_SETUP,
989 // PREALLOCATED_ARG - This has 3 operands: an input chain, a SRCVALUE
990 // with the preallocated call Value, and a constant int.
991 PREALLOCATED_ARG,
992
993 /// SRCVALUE - This is a node type that holds a Value* that is used to
994 /// make reference to a value in the LLVM IR.
995 SRCVALUE,
996
997 /// MDNODE_SDNODE - This is a node that holdes an MDNode*, which is used to
998 /// reference metadata in the IR.
999 MDNODE_SDNODE,
1000
1001 /// PCMARKER - This corresponds to the pcmarker intrinsic.
1002 PCMARKER,
1003
1004 /// READCYCLECOUNTER - This corresponds to the readcyclecounter intrinsic.
1005 /// It produces a chain and one i64 value. The only operand is a chain.
1006 /// If i64 is not legal, the result will be expanded into smaller values.
1007 /// Still, it returns an i64, so targets should set legality for i64.
1008 /// The result is the content of the architecture-specific cycle
1009 /// counter-like register (or other high accuracy low latency clock source).
1010 READCYCLECOUNTER,
1011
1012 /// HANDLENODE node - Used as a handle for various purposes.
1013 HANDLENODE,
1014
1015 /// INIT_TRAMPOLINE - This corresponds to the init_trampoline intrinsic. It
1016 /// takes as input a token chain, the pointer to the trampoline, the pointer
1017 /// to the nested function, the pointer to pass for the 'nest' parameter, a
1018 /// SRCVALUE for the trampoline and another for the nested function
1019 /// (allowing targets to access the original Function*).
1020 /// It produces a token chain as output.
1021 INIT_TRAMPOLINE,
1022
1023 /// ADJUST_TRAMPOLINE - This corresponds to the adjust_trampoline intrinsic.
1024 /// It takes a pointer to the trampoline and produces a (possibly) new
1025 /// pointer to the same trampoline with platform-specific adjustments
1026 /// applied. The pointer it returns points to an executable block of code.
1027 ADJUST_TRAMPOLINE,
1028
1029 /// TRAP - Trapping instruction
1030 TRAP,
1031
1032 /// DEBUGTRAP - Trap intended to get the attention of a debugger.
1033 DEBUGTRAP,
1034
1035 /// UBSANTRAP - Trap with an immediate describing the kind of sanitizer failure.
1036 UBSANTRAP,
1037
1038 /// PREFETCH - This corresponds to a prefetch intrinsic. The first operand
1039 /// is the chain. The other operands are the address to prefetch,
1040 /// read / write specifier, locality specifier and instruction / data cache
1041 /// specifier.
1042 PREFETCH,
1043
1044 /// OUTCHAIN = ATOMIC_FENCE(INCHAIN, ordering, scope)
1045 /// This corresponds to the fence instruction. It takes an input chain, and
1046 /// two integer constants: an AtomicOrdering and a SynchronizationScope.
1047 ATOMIC_FENCE,
1048
1049 /// Val, OUTCHAIN = ATOMIC_LOAD(INCHAIN, ptr)
1050 /// This corresponds to "load atomic" instruction.
1051 ATOMIC_LOAD,
1052
1053 /// OUTCHAIN = ATOMIC_STORE(INCHAIN, ptr, val)
1054 /// This corresponds to "store atomic" instruction.
1055 ATOMIC_STORE,
1056
1057 /// Val, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap)
1058 /// For double-word atomic operations:
1059 /// ValLo, ValHi, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmpLo, cmpHi,
1060 /// swapLo, swapHi)
1061 /// This corresponds to the cmpxchg instruction.
1062 ATOMIC_CMP_SWAP,
1063
1064 /// Val, Success, OUTCHAIN
1065 /// = ATOMIC_CMP_SWAP_WITH_SUCCESS(INCHAIN, ptr, cmp, swap)
1066 /// N.b. this is still a strong cmpxchg operation, so
1067 /// Success == "Val == cmp".
1068 ATOMIC_CMP_SWAP_WITH_SUCCESS,
1069
1070 /// Val, OUTCHAIN = ATOMIC_SWAP(INCHAIN, ptr, amt)
1071 /// Val, OUTCHAIN = ATOMIC_LOAD_[OpName](INCHAIN, ptr, amt)
1072 /// For double-word atomic operations:
1073 /// ValLo, ValHi, OUTCHAIN = ATOMIC_SWAP(INCHAIN, ptr, amtLo, amtHi)
1074 /// ValLo, ValHi, OUTCHAIN = ATOMIC_LOAD_[OpName](INCHAIN, ptr, amtLo, amtHi)
1075 /// These correspond to the atomicrmw instruction.
1076 ATOMIC_SWAP,
1077 ATOMIC_LOAD_ADD,
1078 ATOMIC_LOAD_SUB,
1079 ATOMIC_LOAD_AND,
1080 ATOMIC_LOAD_CLR,
1081 ATOMIC_LOAD_OR,
1082 ATOMIC_LOAD_XOR,
1083 ATOMIC_LOAD_NAND,
1084 ATOMIC_LOAD_MIN,
1085 ATOMIC_LOAD_MAX,
1086 ATOMIC_LOAD_UMIN,
1087 ATOMIC_LOAD_UMAX,
1088 ATOMIC_LOAD_FADD,
1089 ATOMIC_LOAD_FSUB,
1090
1091 // Masked load and store - consecutive vector load and store operations
1092 // with additional mask operand that prevents memory accesses to the
1093 // masked-off lanes.
1094 //
1095 // Val, OutChain = MLOAD(BasePtr, Mask, PassThru)
1096 // OutChain = MSTORE(Value, BasePtr, Mask)
1097 MLOAD,
1098 MSTORE,
1099
1100 // Masked gather and scatter - load and store operations for a vector of
1101 // random addresses with additional mask operand that prevents memory
1102 // accesses to the masked-off lanes.
1103 //
1104 // Val, OutChain = GATHER(InChain, PassThru, Mask, BasePtr, Index, Scale)
1105 // OutChain = SCATTER(InChain, Value, Mask, BasePtr, Index, Scale)
1106 //
1107 // The Index operand can have more vector elements than the other operands
1108 // due to type legalization. The extra elements are ignored.
1109 MGATHER,
1110 MSCATTER,
1111
1112 /// This corresponds to the llvm.lifetime.* intrinsics. The first operand
1113 /// is the chain and the second operand is the alloca pointer.
1114 LIFETIME_START,
1115 LIFETIME_END,
1116
1117 /// GC_TRANSITION_START/GC_TRANSITION_END - These operators mark the
1118 /// beginning and end of GC transition sequence, and carry arbitrary
1119 /// information that target might need for lowering. The first operand is
1120 /// a chain, the rest are specified by the target and not touched by the DAG
1121 /// optimizers. GC_TRANSITION_START..GC_TRANSITION_END pairs may not be
1122 /// nested.
1123 GC_TRANSITION_START,
1124 GC_TRANSITION_END,
1125
1126 /// GET_DYNAMIC_AREA_OFFSET - get offset from native SP to the address of
1127 /// the most recent dynamic alloca. For most targets that would be 0, but
1128 /// for some others (e.g. PowerPC, PowerPC64) that would be compile-time
1129 /// known nonzero constant. The only operand here is the chain.
1130 GET_DYNAMIC_AREA_OFFSET,
1131
1132 /// Pseudo probe for AutoFDO, as a place holder in a basic block to improve
1133 /// the sample counts quality.
1134 PSEUDO_PROBE,
1135
1136 /// VSCALE(IMM) - Returns the runtime scaling factor used to calculate the
1137 /// number of elements within a scalable vector. IMM is a constant integer
1138 /// multiplier that is applied to the runtime value.
1139 VSCALE,
1140
1141 /// Generic reduction nodes. These nodes represent horizontal vector
1142 /// reduction operations, producing a scalar result.
1143 /// The SEQ variants perform reductions in sequential order. The first
1144 /// operand is an initial scalar accumulator value, and the second operand
1145 /// is the vector to reduce.
1146 /// E.g. RES = VECREDUCE_SEQ_FADD f32 ACC, <4 x f32> SRC_VEC
1147 /// ... is equivalent to
1148 /// RES = (((ACC + SRC_VEC[0]) + SRC_VEC[1]) + SRC_VEC[2]) + SRC_VEC[3]
1149 VECREDUCE_SEQ_FADD,
1150 VECREDUCE_SEQ_FMUL,
1151
1152 /// These reductions have relaxed evaluation order semantics, and have a
1153 /// single vector operand. The order of evaluation is unspecified. For
1154 /// pow-of-2 vectors, one valid legalizer expansion is to use a tree
1155 /// reduction, i.e.:
1156 /// For RES = VECREDUCE_FADD <8 x f16> SRC_VEC
1157 /// PART_RDX = FADD SRC_VEC[0:3], SRC_VEC[4:7]
1158 /// PART_RDX2 = FADD PART_RDX[0:1], PART_RDX[2:3]
1159 /// RES = FADD PART_RDX2[0], PART_RDX2[1]
1160 /// For non-pow-2 vectors, this can be computed by extracting each element
1161 /// and performing the operation as if it were scalarized.
1162 VECREDUCE_FADD,
1163 VECREDUCE_FMUL,
1164 /// FMIN/FMAX nodes can have flags, for NaN/NoNaN variants.
1165 VECREDUCE_FMAX,
1166 VECREDUCE_FMIN,
1167 /// Integer reductions may have a result type larger than the vector element
1168 /// type. However, the reduction is performed using the vector element type
1169 /// and the value in the top bits is unspecified.
1170 VECREDUCE_ADD,
1171 VECREDUCE_MUL,
1172 VECREDUCE_AND,
1173 VECREDUCE_OR,
1174 VECREDUCE_XOR,
1175 VECREDUCE_SMAX,
1176 VECREDUCE_SMIN,
1177 VECREDUCE_UMAX,
1178 VECREDUCE_UMIN,
1179
1180// Vector Predication
1181#define BEGIN_REGISTER_VP_SDNODE(VPSDID, ...) VPSDID,
1182#include "llvm/IR/VPIntrinsics.def"
1183
1184 /// BUILTIN_OP_END - This must be the last enum value in this list.
1185 /// The target-specific pre-isel opcode values start here.
1186 BUILTIN_OP_END
1187};
1188
1189/// FIRST_TARGET_STRICTFP_OPCODE - Target-specific pre-isel operations
1190/// which cannot raise FP exceptions should be less than this value.
1191/// Those that do must not be less than this value.
1192static const int FIRST_TARGET_STRICTFP_OPCODE = BUILTIN_OP_END + 400;
1193
1194/// FIRST_TARGET_MEMORY_OPCODE - Target-specific pre-isel operations
1195/// which do not reference a specific memory location should be less than
1196/// this value. Those that do must not be less than this value, and can
1197/// be used with SelectionDAG::getMemIntrinsicNode.
1198static const int FIRST_TARGET_MEMORY_OPCODE = BUILTIN_OP_END + 500;
1199
1200/// Get underlying scalar opcode for VECREDUCE opcode.
1201/// For example ISD::AND for ISD::VECREDUCE_AND.
1202NodeType getVecReduceBaseOpcode(unsigned VecReduceOpcode);
1203
1204/// Whether this is a vector-predicated Opcode.
1205bool isVPOpcode(unsigned Opcode);
1206
1207/// The operand position of the vector mask.
1208Optional<unsigned> getVPMaskIdx(unsigned Opcode);
1209
1210/// The operand position of the explicit vector length parameter.
1211Optional<unsigned> getVPExplicitVectorLengthIdx(unsigned Opcode);
1212
1213//===--------------------------------------------------------------------===//
1214/// MemIndexedMode enum - This enum defines the load / store indexed
1215/// addressing modes.
1216///
1217/// UNINDEXED "Normal" load / store. The effective address is already
1218/// computed and is available in the base pointer. The offset
1219/// operand is always undefined. In addition to producing a
1220/// chain, an unindexed load produces one value (result of the
1221/// load); an unindexed store does not produce a value.
1222///
1223/// PRE_INC Similar to the unindexed mode where the effective address is
1224/// PRE_DEC the value of the base pointer add / subtract the offset.
1225/// It considers the computation as being folded into the load /
1226/// store operation (i.e. the load / store does the address
1227/// computation as well as performing the memory transaction).
1228/// The base operand is always undefined. In addition to
1229/// producing a chain, pre-indexed load produces two values
1230/// (result of the load and the result of the address
1231/// computation); a pre-indexed store produces one value (result
1232/// of the address computation).
1233///
1234/// POST_INC The effective address is the value of the base pointer. The
1235/// POST_DEC value of the offset operand is then added to / subtracted
1236/// from the base after memory transaction. In addition to
1237/// producing a chain, post-indexed load produces two values
1238/// (the result of the load and the result of the base +/- offset
1239/// computation); a post-indexed store produces one value (the
1240/// the result of the base +/- offset computation).
1241enum MemIndexedMode { UNINDEXED = 0, PRE_INC, PRE_DEC, POST_INC, POST_DEC };
1242
1243static const int LAST_INDEXED_MODE = POST_DEC + 1;
1244
1245//===--------------------------------------------------------------------===//
1246/// MemIndexType enum - This enum defines how to interpret MGATHER/SCATTER's
1247/// index parameter when calculating addresses.
1248///
1249/// SIGNED_SCALED Addr = Base + ((signed)Index * sizeof(element))
1250/// SIGNED_UNSCALED Addr = Base + (signed)Index
1251/// UNSIGNED_SCALED Addr = Base + ((unsigned)Index * sizeof(element))
1252/// UNSIGNED_UNSCALED Addr = Base + (unsigned)Index
1253enum MemIndexType {
1254 SIGNED_SCALED = 0,
1255 SIGNED_UNSCALED,
1256 UNSIGNED_SCALED,
1257 UNSIGNED_UNSCALED
1258};
1259
1260static const int LAST_MEM_INDEX_TYPE = UNSIGNED_UNSCALED + 1;
1261
1262//===--------------------------------------------------------------------===//
1263/// LoadExtType enum - This enum defines the three variants of LOADEXT
1264/// (load with extension).
1265///
1266/// SEXTLOAD loads the integer operand and sign extends it to a larger
1267/// integer result type.
1268/// ZEXTLOAD loads the integer operand and zero extends it to a larger
1269/// integer result type.
1270/// EXTLOAD is used for two things: floating point extending loads and
1271/// integer extending loads [the top bits are undefined].
1272enum LoadExtType { NON_EXTLOAD = 0, EXTLOAD, SEXTLOAD, ZEXTLOAD };
1273
1274static const int LAST_LOADEXT_TYPE = ZEXTLOAD + 1;
1275
1276NodeType getExtForLoadExtType(bool IsFP, LoadExtType);
1277
1278//===--------------------------------------------------------------------===//
1279/// ISD::CondCode enum - These are ordered carefully to make the bitfields
1280/// below work out, when considering SETFALSE (something that never exists
1281/// dynamically) as 0. "U" -> Unsigned (for integer operands) or Unordered
1282/// (for floating point), "L" -> Less than, "G" -> Greater than, "E" -> Equal
1283/// to. If the "N" column is 1, the result of the comparison is undefined if
1284/// the input is a NAN.
1285///
1286/// All of these (except for the 'always folded ops') should be handled for
1287/// floating point. For integer, only the SETEQ,SETNE,SETLT,SETLE,SETGT,
1288/// SETGE,SETULT,SETULE,SETUGT, and SETUGE opcodes are used.
1289///
1290/// Note that these are laid out in a specific order to allow bit-twiddling
1291/// to transform conditions.
1292enum CondCode {
1293 // Opcode N U L G E Intuitive operation
1294 SETFALSE, // 0 0 0 0 Always false (always folded)
1295 SETOEQ, // 0 0 0 1 True if ordered and equal
1296 SETOGT, // 0 0 1 0 True if ordered and greater than
1297 SETOGE, // 0 0 1 1 True if ordered and greater than or equal
1298 SETOLT, // 0 1 0 0 True if ordered and less than
1299 SETOLE, // 0 1 0 1 True if ordered and less than or equal
1300 SETONE, // 0 1 1 0 True if ordered and operands are unequal
1301 SETO, // 0 1 1 1 True if ordered (no nans)
1302 SETUO, // 1 0 0 0 True if unordered: isnan(X) | isnan(Y)
1303 SETUEQ, // 1 0 0 1 True if unordered or equal
1304 SETUGT, // 1 0 1 0 True if unordered or greater than
1305 SETUGE, // 1 0 1 1 True if unordered, greater than, or equal
1306 SETULT, // 1 1 0 0 True if unordered or less than
1307 SETULE, // 1 1 0 1 True if unordered, less than, or equal
1308 SETUNE, // 1 1 1 0 True if unordered or not equal
1309 SETTRUE, // 1 1 1 1 Always true (always folded)
1310 // Don't care operations: undefined if the input is a nan.
1311 SETFALSE2, // 1 X 0 0 0 Always false (always folded)
1312 SETEQ, // 1 X 0 0 1 True if equal
1313 SETGT, // 1 X 0 1 0 True if greater than
1314 SETGE, // 1 X 0 1 1 True if greater than or equal
1315 SETLT, // 1 X 1 0 0 True if less than
1316 SETLE, // 1 X 1 0 1 True if less than or equal
1317 SETNE, // 1 X 1 1 0 True if not equal
1318 SETTRUE2, // 1 X 1 1 1 Always true (always folded)
1319
1320 SETCC_INVALID // Marker value.
1321};
1322
1323/// Return true if this is a setcc instruction that performs a signed
1324/// comparison when used with integer operands.
1325inline bool isSignedIntSetCC(CondCode Code) {
1326 return Code == SETGT || Code == SETGE || Code == SETLT || Code == SETLE;
1327}
1328
1329/// Return true if this is a setcc instruction that performs an unsigned
1330/// comparison when used with integer operands.
1331inline bool isUnsignedIntSetCC(CondCode Code) {
1332 return Code == SETUGT || Code == SETUGE || Code == SETULT || Code == SETULE;
1333}
1334
1335/// Return true if the specified condition returns true if the two operands to
1336/// the condition are equal. Note that if one of the two operands is a NaN,
1337/// this value is meaningless.
1338inline bool isTrueWhenEqual(CondCode Cond) { return ((int)Cond & 1) != 0; }
1339
1340/// This function returns 0 if the condition is always false if an operand is
1341/// a NaN, 1 if the condition is always true if the operand is a NaN, and 2 if
1342/// the condition is undefined if the operand is a NaN.
1343inline unsigned getUnorderedFlavor(CondCode Cond) {
1344 return ((int)Cond >> 3) & 3;
1345}
1346
1347/// Return the operation corresponding to !(X op Y), where 'op' is a valid
1348/// SetCC operation.
1349CondCode getSetCCInverse(CondCode Operation, EVT Type);
1350
1351namespace GlobalISel {
1352/// Return the operation corresponding to !(X op Y), where 'op' is a valid
1353/// SetCC operation. The U bit of the condition code has different meanings
1354/// between floating point and integer comparisons and LLT's don't provide
1355/// this distinction. As such we need to be told whether the comparison is
1356/// floating point or integer-like. Pointers should use integer-like
1357/// comparisons.
1358CondCode getSetCCInverse(CondCode Operation, bool isIntegerLike);
1359} // end namespace GlobalISel
1360
1361/// Return the operation corresponding to (Y op X) when given the operation
1362/// for (X op Y).
1363CondCode getSetCCSwappedOperands(CondCode Operation);
1364
1365/// Return the result of a logical OR between different comparisons of
1366/// identical values: ((X op1 Y) | (X op2 Y)). This function returns
1367/// SETCC_INVALID if it is not possible to represent the resultant comparison.
1368CondCode getSetCCOrOperation(CondCode Op1, CondCode Op2, EVT Type);
1369
1370/// Return the result of a logical AND between different comparisons of
1371/// identical values: ((X op1 Y) & (X op2 Y)). This function returns
1372/// SETCC_INVALID if it is not possible to represent the resultant comparison.
1373CondCode getSetCCAndOperation(CondCode Op1, CondCode Op2, EVT Type);
1374
1375} // namespace ISD
1376
1377} // namespace llvm
Andrew Scull5e1ddfa2018-08-14 10:06:54 +01001378
1379#endif