1
//===- IntrinsicsARM.td - Defines ARM intrinsics -----------*- tablegen -*-===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This file defines all of the ARM-specific intrinsics.
12
//===----------------------------------------------------------------------===//
15
//===----------------------------------------------------------------------===//
18
let TargetPrefix = "arm" in { // All intrinsics start with "llvm.arm.".
19
def int_arm_thread_pointer : GCCBuiltin<"__builtin_thread_pointer">,
20
Intrinsic<[llvm_ptr_ty], [], [IntrNoMem]>;
23
//===----------------------------------------------------------------------===//
24
// Saturating Arithmentic
26
let TargetPrefix = "arm" in { // All intrinsics start with "llvm.arm.".
27
def int_arm_qadd : GCCBuiltin<"__builtin_arm_qadd">,
28
Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty],
29
[IntrNoMem, Commutative]>;
30
def int_arm_qsub : GCCBuiltin<"__builtin_arm_qsub">,
31
Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
32
def int_arm_ssat : GCCBuiltin<"__builtin_arm_ssat">,
33
Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
34
def int_arm_usat : GCCBuiltin<"__builtin_arm_usat">,
35
Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
38
//===----------------------------------------------------------------------===//
41
let TargetPrefix = "arm" in { // All intrinsics start with "llvm.arm.".
42
def int_arm_get_fpscr : GCCBuiltin<"__builtin_arm_get_fpscr">,
43
Intrinsic<[llvm_i32_ty], [], [IntrNoMem]>;
44
def int_arm_set_fpscr : GCCBuiltin<"__builtin_arm_set_fpscr">,
45
Intrinsic<[], [llvm_i32_ty], []>;
46
def int_arm_vcvtr : Intrinsic<[llvm_float_ty], [llvm_anyfloat_ty],
48
def int_arm_vcvtru : Intrinsic<[llvm_float_ty], [llvm_anyfloat_ty],
52
//===----------------------------------------------------------------------===//
53
// Advanced SIMD (NEON)
55
let TargetPrefix = "arm" in { // All intrinsics start with "llvm.arm.".
57
// The following classes do not correspond directly to GCC builtins.
58
class Neon_1Arg_Intrinsic
59
: Intrinsic<[llvm_anyvector_ty], [LLVMMatchType<0>], [IntrNoMem]>;
60
class Neon_1Arg_Narrow_Intrinsic
61
: Intrinsic<[llvm_anyvector_ty],
62
[LLVMExtendedElementVectorType<0>], [IntrNoMem]>;
63
class Neon_2Arg_Intrinsic
64
: Intrinsic<[llvm_anyvector_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
66
class Neon_2Arg_Narrow_Intrinsic
67
: Intrinsic<[llvm_anyvector_ty],
68
[LLVMExtendedElementVectorType<0>,
69
LLVMExtendedElementVectorType<0>],
71
class Neon_2Arg_Long_Intrinsic
72
: Intrinsic<[llvm_anyvector_ty],
73
[LLVMTruncatedElementVectorType<0>,
74
LLVMTruncatedElementVectorType<0>],
76
class Neon_3Arg_Intrinsic
77
: Intrinsic<[llvm_anyvector_ty],
78
[LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
80
class Neon_3Arg_Long_Intrinsic
81
: Intrinsic<[llvm_anyvector_ty],
83
LLVMTruncatedElementVectorType<0>,
84
LLVMTruncatedElementVectorType<0>],
86
class Neon_CvtFxToFP_Intrinsic
87
: Intrinsic<[llvm_anyfloat_ty], [llvm_anyint_ty, llvm_i32_ty], [IntrNoMem]>;
88
class Neon_CvtFPToFx_Intrinsic
89
: Intrinsic<[llvm_anyint_ty], [llvm_anyfloat_ty, llvm_i32_ty], [IntrNoMem]>;
91
// The table operands for VTBL and VTBX consist of 1 to 4 v8i8 vectors.
92
// Besides the table, VTBL has one other v8i8 argument and VTBX has two.
93
// Overall, the classes range from 2 to 6 v8i8 arguments.
94
class Neon_Tbl2Arg_Intrinsic
95
: Intrinsic<[llvm_v8i8_ty],
96
[llvm_v8i8_ty, llvm_v8i8_ty], [IntrNoMem]>;
97
class Neon_Tbl3Arg_Intrinsic
98
: Intrinsic<[llvm_v8i8_ty],
99
[llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty], [IntrNoMem]>;
100
class Neon_Tbl4Arg_Intrinsic
101
: Intrinsic<[llvm_v8i8_ty],
102
[llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty],
104
class Neon_Tbl5Arg_Intrinsic
105
: Intrinsic<[llvm_v8i8_ty],
106
[llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty,
107
llvm_v8i8_ty], [IntrNoMem]>;
108
class Neon_Tbl6Arg_Intrinsic
109
: Intrinsic<[llvm_v8i8_ty],
110
[llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty, llvm_v8i8_ty,
111
llvm_v8i8_ty, llvm_v8i8_ty], [IntrNoMem]>;
116
let Properties = [IntrNoMem, Commutative] in {
119
def int_arm_neon_vhadds : Neon_2Arg_Intrinsic;
120
def int_arm_neon_vhaddu : Neon_2Arg_Intrinsic;
121
def int_arm_neon_vrhadds : Neon_2Arg_Intrinsic;
122
def int_arm_neon_vrhaddu : Neon_2Arg_Intrinsic;
123
def int_arm_neon_vqadds : Neon_2Arg_Intrinsic;
124
def int_arm_neon_vqaddu : Neon_2Arg_Intrinsic;
125
def int_arm_neon_vaddhn : Neon_2Arg_Narrow_Intrinsic;
126
def int_arm_neon_vraddhn : Neon_2Arg_Narrow_Intrinsic;
129
def int_arm_neon_vmulp : Neon_2Arg_Intrinsic;
130
def int_arm_neon_vqdmulh : Neon_2Arg_Intrinsic;
131
def int_arm_neon_vqrdmulh : Neon_2Arg_Intrinsic;
132
def int_arm_neon_vmullp : Neon_2Arg_Long_Intrinsic;
133
def int_arm_neon_vqdmull : Neon_2Arg_Long_Intrinsic;
134
def int_arm_neon_vqdmlal : Neon_3Arg_Long_Intrinsic;
135
def int_arm_neon_vqdmlsl : Neon_3Arg_Long_Intrinsic;
138
def int_arm_neon_vmaxs : Neon_2Arg_Intrinsic;
139
def int_arm_neon_vmaxu : Neon_2Arg_Intrinsic;
142
def int_arm_neon_vmins : Neon_2Arg_Intrinsic;
143
def int_arm_neon_vminu : Neon_2Arg_Intrinsic;
145
// Vector Reciprocal Step.
146
def int_arm_neon_vrecps : Neon_2Arg_Intrinsic;
148
// Vector Reciprocal Square Root Step.
149
def int_arm_neon_vrsqrts : Neon_2Arg_Intrinsic;
153
def int_arm_neon_vhsubs : Neon_2Arg_Intrinsic;
154
def int_arm_neon_vhsubu : Neon_2Arg_Intrinsic;
155
def int_arm_neon_vqsubs : Neon_2Arg_Intrinsic;
156
def int_arm_neon_vqsubu : Neon_2Arg_Intrinsic;
157
def int_arm_neon_vsubhn : Neon_2Arg_Narrow_Intrinsic;
158
def int_arm_neon_vrsubhn : Neon_2Arg_Narrow_Intrinsic;
160
// Vector Absolute Compare.
161
let TargetPrefix = "arm" in {
162
def int_arm_neon_vacged : Intrinsic<[llvm_v2i32_ty],
163
[llvm_v2f32_ty, llvm_v2f32_ty],
165
def int_arm_neon_vacgeq : Intrinsic<[llvm_v4i32_ty],
166
[llvm_v4f32_ty, llvm_v4f32_ty],
168
def int_arm_neon_vacgtd : Intrinsic<[llvm_v2i32_ty],
169
[llvm_v2f32_ty, llvm_v2f32_ty],
171
def int_arm_neon_vacgtq : Intrinsic<[llvm_v4i32_ty],
172
[llvm_v4f32_ty, llvm_v4f32_ty],
176
// Vector Absolute Differences.
177
def int_arm_neon_vabds : Neon_2Arg_Intrinsic;
178
def int_arm_neon_vabdu : Neon_2Arg_Intrinsic;
180
// Vector Pairwise Add.
181
def int_arm_neon_vpadd : Neon_2Arg_Intrinsic;
183
// Vector Pairwise Add Long.
184
// Note: This is different than the other "long" NEON intrinsics because
185
// the result vector has half as many elements as the source vector.
186
// The source and destination vector types must be specified separately.
187
let TargetPrefix = "arm" in {
188
def int_arm_neon_vpaddls : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty],
190
def int_arm_neon_vpaddlu : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty],
194
// Vector Pairwise Add and Accumulate Long.
195
// Note: This is similar to vpaddl but the destination vector also appears
196
// as the first argument.
197
let TargetPrefix = "arm" in {
198
def int_arm_neon_vpadals : Intrinsic<[llvm_anyvector_ty],
199
[LLVMMatchType<0>, llvm_anyvector_ty],
201
def int_arm_neon_vpadalu : Intrinsic<[llvm_anyvector_ty],
202
[LLVMMatchType<0>, llvm_anyvector_ty],
206
// Vector Pairwise Maximum and Minimum.
207
def int_arm_neon_vpmaxs : Neon_2Arg_Intrinsic;
208
def int_arm_neon_vpmaxu : Neon_2Arg_Intrinsic;
209
def int_arm_neon_vpmins : Neon_2Arg_Intrinsic;
210
def int_arm_neon_vpminu : Neon_2Arg_Intrinsic;
214
// The various saturating and rounding vector shift operations need to be
215
// represented by intrinsics in LLVM, and even the basic VSHL variable shift
216
// operation cannot be safely translated to LLVM's shift operators. VSHL can
217
// be used for both left and right shifts, or even combinations of the two,
218
// depending on the signs of the shift amounts. It also has well-defined
219
// behavior for shift amounts that LLVM leaves undefined. Only basic shifts
220
// by constants can be represented with LLVM's shift operators.
222
// The shift counts for these intrinsics are always vectors, even for constant
223
// shifts, where the constant is replicated. For consistency with VSHL (and
224
// other variable shift instructions), left shifts have positive shift counts
225
// and right shifts have negative shift counts. This convention is also used
226
// for constant right shift intrinsics, and to help preserve sanity, the
227
// intrinsic names use "shift" instead of either "shl" or "shr". Where
228
// applicable, signed and unsigned versions of the intrinsics are
229
// distinguished with "s" and "u" suffixes. A few NEON shift instructions,
230
// such as VQSHLU, take signed operands but produce unsigned results; these
231
// use a "su" suffix.
234
def int_arm_neon_vshifts : Neon_2Arg_Intrinsic;
235
def int_arm_neon_vshiftu : Neon_2Arg_Intrinsic;
236
def int_arm_neon_vshiftls : Neon_2Arg_Long_Intrinsic;
237
def int_arm_neon_vshiftlu : Neon_2Arg_Long_Intrinsic;
238
def int_arm_neon_vshiftn : Neon_2Arg_Narrow_Intrinsic;
240
// Vector Rounding Shift.
241
def int_arm_neon_vrshifts : Neon_2Arg_Intrinsic;
242
def int_arm_neon_vrshiftu : Neon_2Arg_Intrinsic;
243
def int_arm_neon_vrshiftn : Neon_2Arg_Narrow_Intrinsic;
245
// Vector Saturating Shift.
246
def int_arm_neon_vqshifts : Neon_2Arg_Intrinsic;
247
def int_arm_neon_vqshiftu : Neon_2Arg_Intrinsic;
248
def int_arm_neon_vqshiftsu : Neon_2Arg_Intrinsic;
249
def int_arm_neon_vqshiftns : Neon_2Arg_Narrow_Intrinsic;
250
def int_arm_neon_vqshiftnu : Neon_2Arg_Narrow_Intrinsic;
251
def int_arm_neon_vqshiftnsu : Neon_2Arg_Narrow_Intrinsic;
253
// Vector Saturating Rounding Shift.
254
def int_arm_neon_vqrshifts : Neon_2Arg_Intrinsic;
255
def int_arm_neon_vqrshiftu : Neon_2Arg_Intrinsic;
256
def int_arm_neon_vqrshiftns : Neon_2Arg_Narrow_Intrinsic;
257
def int_arm_neon_vqrshiftnu : Neon_2Arg_Narrow_Intrinsic;
258
def int_arm_neon_vqrshiftnsu : Neon_2Arg_Narrow_Intrinsic;
260
// Vector Shift and Insert.
261
def int_arm_neon_vshiftins : Neon_3Arg_Intrinsic;
263
// Vector Absolute Value and Saturating Absolute Value.
264
def int_arm_neon_vabs : Neon_1Arg_Intrinsic;
265
def int_arm_neon_vqabs : Neon_1Arg_Intrinsic;
267
// Vector Saturating Negate.
268
def int_arm_neon_vqneg : Neon_1Arg_Intrinsic;
270
// Vector Count Leading Sign/Zero Bits.
271
def int_arm_neon_vcls : Neon_1Arg_Intrinsic;
272
def int_arm_neon_vclz : Neon_1Arg_Intrinsic;
274
// Vector Count One Bits.
275
def int_arm_neon_vcnt : Neon_1Arg_Intrinsic;
277
// Vector Reciprocal Estimate.
278
def int_arm_neon_vrecpe : Neon_1Arg_Intrinsic;
280
// Vector Reciprocal Square Root Estimate.
281
def int_arm_neon_vrsqrte : Neon_1Arg_Intrinsic;
283
// Vector Conversions Between Floating-point and Fixed-point.
284
def int_arm_neon_vcvtfp2fxs : Neon_CvtFPToFx_Intrinsic;
285
def int_arm_neon_vcvtfp2fxu : Neon_CvtFPToFx_Intrinsic;
286
def int_arm_neon_vcvtfxs2fp : Neon_CvtFxToFP_Intrinsic;
287
def int_arm_neon_vcvtfxu2fp : Neon_CvtFxToFP_Intrinsic;
289
// Narrowing Saturating Vector Moves.
290
def int_arm_neon_vqmovns : Neon_1Arg_Narrow_Intrinsic;
291
def int_arm_neon_vqmovnu : Neon_1Arg_Narrow_Intrinsic;
292
def int_arm_neon_vqmovnsu : Neon_1Arg_Narrow_Intrinsic;
294
// Vector Table Lookup.
295
// The first 1-4 arguments are the table.
296
def int_arm_neon_vtbl1 : Neon_Tbl2Arg_Intrinsic;
297
def int_arm_neon_vtbl2 : Neon_Tbl3Arg_Intrinsic;
298
def int_arm_neon_vtbl3 : Neon_Tbl4Arg_Intrinsic;
299
def int_arm_neon_vtbl4 : Neon_Tbl5Arg_Intrinsic;
301
// Vector Table Extension.
302
// Some elements of the destination vector may not be updated, so the original
303
// value of that vector is passed as the first argument. The next 1-4
304
// arguments after that are the table.
305
def int_arm_neon_vtbx1 : Neon_Tbl3Arg_Intrinsic;
306
def int_arm_neon_vtbx2 : Neon_Tbl4Arg_Intrinsic;
307
def int_arm_neon_vtbx3 : Neon_Tbl5Arg_Intrinsic;
308
def int_arm_neon_vtbx4 : Neon_Tbl6Arg_Intrinsic;
310
let TargetPrefix = "arm" in {
312
// De-interleaving vector loads from N-element structures.
313
// Source operands are the address and alignment.
314
def int_arm_neon_vld1 : Intrinsic<[llvm_anyvector_ty],
315
[llvm_ptr_ty, llvm_i32_ty],
317
def int_arm_neon_vld2 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>],
318
[llvm_ptr_ty, llvm_i32_ty],
320
def int_arm_neon_vld3 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
322
[llvm_ptr_ty, llvm_i32_ty],
324
def int_arm_neon_vld4 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
325
LLVMMatchType<0>, LLVMMatchType<0>],
326
[llvm_ptr_ty, llvm_i32_ty],
329
// Vector load N-element structure to one lane.
330
// Source operands are: the address, the N input vectors (since only one
331
// lane is assigned), the lane number, and the alignment.
332
def int_arm_neon_vld2lane : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>],
333
[llvm_ptr_ty, LLVMMatchType<0>,
334
LLVMMatchType<0>, llvm_i32_ty,
335
llvm_i32_ty], [IntrReadArgMem]>;
336
def int_arm_neon_vld3lane : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
338
[llvm_ptr_ty, LLVMMatchType<0>,
339
LLVMMatchType<0>, LLVMMatchType<0>,
340
llvm_i32_ty, llvm_i32_ty],
342
def int_arm_neon_vld4lane : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
343
LLVMMatchType<0>, LLVMMatchType<0>],
344
[llvm_ptr_ty, LLVMMatchType<0>,
345
LLVMMatchType<0>, LLVMMatchType<0>,
346
LLVMMatchType<0>, llvm_i32_ty,
347
llvm_i32_ty], [IntrReadArgMem]>;
349
// Interleaving vector stores from N-element structures.
350
// Source operands are: the address, the N vectors, and the alignment.
351
def int_arm_neon_vst1 : Intrinsic<[],
352
[llvm_ptr_ty, llvm_anyvector_ty,
353
llvm_i32_ty], [IntrReadWriteArgMem]>;
354
def int_arm_neon_vst2 : Intrinsic<[],
355
[llvm_ptr_ty, llvm_anyvector_ty,
356
LLVMMatchType<0>, llvm_i32_ty],
357
[IntrReadWriteArgMem]>;
358
def int_arm_neon_vst3 : Intrinsic<[],
359
[llvm_ptr_ty, llvm_anyvector_ty,
360
LLVMMatchType<0>, LLVMMatchType<0>,
361
llvm_i32_ty], [IntrReadWriteArgMem]>;
362
def int_arm_neon_vst4 : Intrinsic<[],
363
[llvm_ptr_ty, llvm_anyvector_ty,
364
LLVMMatchType<0>, LLVMMatchType<0>,
365
LLVMMatchType<0>, llvm_i32_ty],
366
[IntrReadWriteArgMem]>;
368
// Vector store N-element structure from one lane.
369
// Source operands are: the address, the N vectors, the lane number, and
371
def int_arm_neon_vst2lane : Intrinsic<[],
372
[llvm_ptr_ty, llvm_anyvector_ty,
373
LLVMMatchType<0>, llvm_i32_ty,
374
llvm_i32_ty], [IntrReadWriteArgMem]>;
375
def int_arm_neon_vst3lane : Intrinsic<[],
376
[llvm_ptr_ty, llvm_anyvector_ty,
377
LLVMMatchType<0>, LLVMMatchType<0>,
378
llvm_i32_ty, llvm_i32_ty],
379
[IntrReadWriteArgMem]>;
380
def int_arm_neon_vst4lane : Intrinsic<[],
381
[llvm_ptr_ty, llvm_anyvector_ty,
382
LLVMMatchType<0>, LLVMMatchType<0>,
383
LLVMMatchType<0>, llvm_i32_ty,
384
llvm_i32_ty], [IntrReadWriteArgMem]>;
added
removed
libclamav/c++/llvm/include/llvm/IntrinsicsARM.td
