1 //===- IntrinsicsAArch64.td - Defines AArch64 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 AArch64-specific intrinsics.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // Advanced SIMD (NEON)
17 let TargetPrefix = "aarch64" in { // All intrinsics start with "llvm.aarch64.".
19 // Vector Absolute Compare (Floating Point)
20 def int_aarch64_neon_vacgeq :
21 Intrinsic<[llvm_v2i64_ty], [llvm_v2f64_ty, llvm_v2f64_ty], [IntrNoMem]>;
22 def int_aarch64_neon_vacgtq :
23 Intrinsic<[llvm_v2i64_ty], [llvm_v2f64_ty, llvm_v2f64_ty], [IntrNoMem]>;
25 // Vector saturating accumulate
26 def int_aarch64_neon_suqadd : Neon_2Arg_Intrinsic;
27 def int_aarch64_neon_usqadd : Neon_2Arg_Intrinsic;
29 // Vector Bitwise reverse
30 def int_aarch64_neon_rbit : Neon_1Arg_Intrinsic;
32 // Vector extract and narrow
33 def int_aarch64_neon_xtn :
34 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
36 // Vector floating-point convert
37 def int_aarch64_neon_frintn : Neon_1Arg_Intrinsic;
38 def int_aarch64_neon_fsqrt : Neon_1Arg_Intrinsic;
39 def int_aarch64_neon_fcvtxn :
40 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
41 def int_aarch64_neon_fcvtns :
42 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
43 def int_aarch64_neon_fcvtnu :
44 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
45 def int_aarch64_neon_fcvtps :
46 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
47 def int_aarch64_neon_fcvtpu :
48 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
49 def int_aarch64_neon_fcvtms :
50 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
51 def int_aarch64_neon_fcvtmu :
52 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
53 def int_aarch64_neon_fcvtas :
54 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
55 def int_aarch64_neon_fcvtau :
56 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
58 // Vector maxNum (Floating Point)
59 def int_aarch64_neon_vmaxnm : Neon_2Arg_Intrinsic;
61 // Vector minNum (Floating Point)
62 def int_aarch64_neon_vminnm : Neon_2Arg_Intrinsic;
64 // Vector Pairwise maxNum (Floating Point)
65 def int_aarch64_neon_vpmaxnm : Neon_2Arg_Intrinsic;
67 // Vector Pairwise minNum (Floating Point)
68 def int_aarch64_neon_vpminnm : Neon_2Arg_Intrinsic;
70 // Vector Multiply Extended and Scalar Multiply Extended (Floating Point)
71 def int_aarch64_neon_vmulx :
72 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>]>;
74 class Neon_N2V_Intrinsic
75 : Intrinsic<[llvm_anyvector_ty], [LLVMMatchType<0>, llvm_i32_ty],
77 class Neon_N3V_Intrinsic
78 : Intrinsic<[llvm_anyvector_ty],
79 [LLVMMatchType<0>, LLVMMatchType<0>, llvm_i32_ty],
81 class Neon_N2V_Narrow_Intrinsic
82 : Intrinsic<[llvm_anyvector_ty],
83 [LLVMExtendedElementVectorType<0>, llvm_i32_ty],
86 // Vector rounding shift right by immediate (Signed)
87 def int_aarch64_neon_vsrshr : Neon_N2V_Intrinsic;
88 def int_aarch64_neon_vurshr : Neon_N2V_Intrinsic;
89 def int_aarch64_neon_vsqshlu : Neon_N2V_Intrinsic;
91 def int_aarch64_neon_vsri : Neon_N3V_Intrinsic;
92 def int_aarch64_neon_vsli : Neon_N3V_Intrinsic;
94 def int_aarch64_neon_vsqshrun : Neon_N2V_Narrow_Intrinsic;
95 def int_aarch64_neon_vrshrn : Neon_N2V_Narrow_Intrinsic;
96 def int_aarch64_neon_vsqrshrun : Neon_N2V_Narrow_Intrinsic;
97 def int_aarch64_neon_vsqshrn : Neon_N2V_Narrow_Intrinsic;
98 def int_aarch64_neon_vuqshrn : Neon_N2V_Narrow_Intrinsic;
99 def int_aarch64_neon_vsqrshrn : Neon_N2V_Narrow_Intrinsic;
100 def int_aarch64_neon_vuqrshrn : Neon_N2V_Narrow_Intrinsic;
103 class Neon_Across_Intrinsic
104 : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
106 class Neon_2Arg_Across_Float_Intrinsic
107 : Intrinsic<[llvm_anyvector_ty], [llvm_v4f32_ty], [IntrNoMem]>;
109 def int_aarch64_neon_saddlv : Neon_Across_Intrinsic;
110 def int_aarch64_neon_uaddlv : Neon_Across_Intrinsic;
111 def int_aarch64_neon_smaxv : Neon_Across_Intrinsic;
112 def int_aarch64_neon_umaxv : Neon_Across_Intrinsic;
113 def int_aarch64_neon_sminv : Neon_Across_Intrinsic;
114 def int_aarch64_neon_uminv : Neon_Across_Intrinsic;
115 def int_aarch64_neon_vaddv : Neon_Across_Intrinsic;
116 def int_aarch64_neon_vmaxv : Neon_Across_Intrinsic;
117 def int_aarch64_neon_vminv : Neon_Across_Intrinsic;
118 def int_aarch64_neon_vmaxnmv : Neon_Across_Intrinsic;
119 def int_aarch64_neon_vminnmv : Neon_Across_Intrinsic;
121 // Vector Table Lookup.
122 def int_aarch64_neon_vtbl1 :
123 Intrinsic<[llvm_anyvector_ty],
124 [llvm_anyvector_ty, LLVMMatchType<0>], [IntrNoMem]>;
126 def int_aarch64_neon_vtbl2 :
127 Intrinsic<[llvm_anyvector_ty],
128 [llvm_anyvector_ty, LLVMMatchType<1>, LLVMMatchType<0>],
131 def int_aarch64_neon_vtbl3 :
132 Intrinsic<[llvm_anyvector_ty],
133 [llvm_anyvector_ty, LLVMMatchType<1>, LLVMMatchType<1>,
134 LLVMMatchType<0>], [IntrNoMem]>;
136 def int_aarch64_neon_vtbl4 :
137 Intrinsic<[llvm_anyvector_ty],
138 [llvm_anyvector_ty, LLVMMatchType<1>, LLVMMatchType<1>,
139 LLVMMatchType<1>, LLVMMatchType<0>], [IntrNoMem]>;
141 // Vector Table Extension.
142 // Some elements of the destination vector may not be updated, so the original
143 // value of that vector is passed as the first argument. The next 1-4
144 // arguments after that are the table.
145 def int_aarch64_neon_vtbx1 :
146 Intrinsic<[llvm_anyvector_ty],
147 [LLVMMatchType<0>, llvm_anyvector_ty, LLVMMatchType<0>],
150 def int_aarch64_neon_vtbx2 :
151 Intrinsic<[llvm_anyvector_ty],
152 [LLVMMatchType<0>, llvm_anyvector_ty, LLVMMatchType<1>,
153 LLVMMatchType<0>], [IntrNoMem]>;
155 def int_aarch64_neon_vtbx3 :
156 Intrinsic<[llvm_anyvector_ty],
157 [LLVMMatchType<0>, llvm_anyvector_ty, LLVMMatchType<1>,
158 LLVMMatchType<1>, LLVMMatchType<0>], [IntrNoMem]>;
160 def int_aarch64_neon_vtbx4 :
161 Intrinsic<[llvm_anyvector_ty],
162 [LLVMMatchType<0>, llvm_anyvector_ty, LLVMMatchType<1>,
163 LLVMMatchType<1>, LLVMMatchType<1>, LLVMMatchType<0>],
167 def int_aarch64_neon_vld1x2 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>],
168 [llvm_ptr_ty, llvm_i32_ty],
170 def int_aarch64_neon_vld1x3 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
172 [llvm_ptr_ty, llvm_i32_ty],
174 def int_aarch64_neon_vld1x4 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
175 LLVMMatchType<0>, LLVMMatchType<0>],
176 [llvm_ptr_ty, llvm_i32_ty],
179 def int_aarch64_neon_vst1x2 : Intrinsic<[],
180 [llvm_ptr_ty, llvm_anyvector_ty,
181 LLVMMatchType<0>, llvm_i32_ty],
182 [IntrReadWriteArgMem]>;
183 def int_aarch64_neon_vst1x3 : Intrinsic<[],
184 [llvm_ptr_ty, llvm_anyvector_ty,
185 LLVMMatchType<0>, LLVMMatchType<0>,
186 llvm_i32_ty], [IntrReadWriteArgMem]>;
187 def int_aarch64_neon_vst1x4 : Intrinsic<[],
188 [llvm_ptr_ty, llvm_anyvector_ty,
189 LLVMMatchType<0>, LLVMMatchType<0>,
190 LLVMMatchType<0>, llvm_i32_ty],
191 [IntrReadWriteArgMem]>;
194 def int_aarch64_neon_vaddds :
195 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
196 def int_aarch64_neon_vadddu :
197 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
201 def int_aarch64_neon_vsubds :
202 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
203 def int_aarch64_neon_vsubdu :
204 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
209 def int_aarch64_neon_vshlds :
210 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
211 def int_aarch64_neon_vshldu :
212 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
214 // Scalar Saturating Shift Left
215 def int_aarch64_neon_vqshls : Neon_2Arg_Intrinsic;
216 def int_aarch64_neon_vqshlu : Neon_2Arg_Intrinsic;
218 // Scalar Shift Rouding Left
219 def int_aarch64_neon_vrshlds :
220 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
221 def int_aarch64_neon_vrshldu :
222 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
224 // Scalar Saturating Rounding Shift Left
225 def int_aarch64_neon_vqrshls : Neon_2Arg_Intrinsic;
226 def int_aarch64_neon_vqrshlu : Neon_2Arg_Intrinsic;
228 // Scalar Reduce Pairwise Add.
229 def int_aarch64_neon_vpadd :
230 Intrinsic<[llvm_v1i64_ty], [llvm_v2i64_ty],[IntrNoMem]>;
231 def int_aarch64_neon_vpfadd :
232 Intrinsic<[llvm_v1f32_ty], [llvm_v2f32_ty], [IntrNoMem]>;
233 def int_aarch64_neon_vpfaddq :
234 Intrinsic<[llvm_v1f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
236 // Scalar Reduce Pairwise Floating Point Max/Min.
237 def int_aarch64_neon_vpmax :
238 Intrinsic<[llvm_v1f32_ty], [llvm_v2f32_ty], [IntrNoMem]>;
239 def int_aarch64_neon_vpmaxq :
240 Intrinsic<[llvm_v1f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
241 def int_aarch64_neon_vpmin :
242 Intrinsic<[llvm_v1f32_ty], [llvm_v2f32_ty], [IntrNoMem]>;
243 def int_aarch64_neon_vpminq :
244 Intrinsic<[llvm_v1f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
246 // Scalar Reduce Pairwise Floating Point Maxnm/Minnm.
247 def int_aarch64_neon_vpfmaxnm :
248 Intrinsic<[llvm_v1f32_ty], [llvm_v2f32_ty], [IntrNoMem]>;
249 def int_aarch64_neon_vpfmaxnmq :
250 Intrinsic<[llvm_v1f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
251 def int_aarch64_neon_vpfminnm :
252 Intrinsic<[llvm_v1f32_ty], [llvm_v2f32_ty], [IntrNoMem]>;
253 def int_aarch64_neon_vpfminnmq :
254 Intrinsic<[llvm_v1f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
256 // Scalar Signed Integer Convert To Floating-point
257 def int_aarch64_neon_vcvtf32_s32 :
258 Intrinsic<[llvm_float_ty], [llvm_v1i32_ty], [IntrNoMem]>;
259 def int_aarch64_neon_vcvtf64_s64 :
260 Intrinsic<[llvm_double_ty], [llvm_v1i64_ty], [IntrNoMem]>;
262 // Scalar Unsigned Integer Convert To Floating-point
263 def int_aarch64_neon_vcvtf32_u32 :
264 Intrinsic<[llvm_float_ty], [llvm_v1i32_ty], [IntrNoMem]>;
265 def int_aarch64_neon_vcvtf64_u64 :
266 Intrinsic<[llvm_double_ty], [llvm_v1i64_ty], [IntrNoMem]>;
268 // Scalar Floating-point Reciprocal Exponent
269 def int_aarch64_neon_vrecpx : Neon_1Arg_Intrinsic;
271 class Neon_Cmp_Intrinsic
272 : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_anyvector_ty],
275 // Scalar Compare Equal
276 def int_aarch64_neon_vceq : Neon_Cmp_Intrinsic;
278 // Scalar Compare Greater-Than or Equal
279 def int_aarch64_neon_vcge : Neon_Cmp_Intrinsic;
280 def int_aarch64_neon_vchs : Neon_Cmp_Intrinsic;
282 // Scalar Compare Less-Than or Equal
283 def int_aarch64_neon_vclez : Neon_Cmp_Intrinsic;
285 // Scalar Compare Less-Than
286 def int_aarch64_neon_vcltz : Neon_Cmp_Intrinsic;
288 // Scalar Compare Greater-Than
289 def int_aarch64_neon_vcgt : Neon_Cmp_Intrinsic;
290 def int_aarch64_neon_vchi : Neon_Cmp_Intrinsic;
292 // Scalar Compare Bitwise Test Bits
293 def int_aarch64_neon_vtstd : Neon_Cmp_Intrinsic;
295 // Scalar Floating-point Absolute Compare Greater Than Or Equal
296 def int_aarch64_neon_vcage : Neon_Cmp_Intrinsic;
298 // Scalar Floating-point Absolute Compare Greater Than
299 def int_aarch64_neon_vcagt : Neon_Cmp_Intrinsic;
301 // Scalar Signed Saturating Accumulated of Unsigned Value
302 def int_aarch64_neon_vuqadd : Neon_2Arg_Intrinsic;
304 // Scalar Unsigned Saturating Accumulated of Signed Value
305 def int_aarch64_neon_vsqadd : Neon_2Arg_Intrinsic;
307 // Scalar Absolute Value
308 def int_aarch64_neon_vabs :
309 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty], [IntrNoMem]>;
311 // Scalar Negate Value
312 def int_aarch64_neon_vneg :
313 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty], [IntrNoMem]>;
315 // Signed Saturating Doubling Multiply-Add Long
316 def int_aarch64_neon_vqdmlal : Neon_3Arg_Long_Intrinsic;
318 // Signed Saturating Doubling Multiply-Subtract Long
319 def int_aarch64_neon_vqdmlsl : Neon_3Arg_Long_Intrinsic;
321 class Neon_2Arg_ShiftImm_Intrinsic
322 : Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_i32_ty], [IntrNoMem]>;
324 class Neon_3Arg_ShiftImm_Intrinsic
325 : Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty, llvm_i32_ty],
328 // Scalar Shift Right (Immediate)
329 def int_aarch64_neon_vshrds_n : Neon_2Arg_ShiftImm_Intrinsic;
330 def int_aarch64_neon_vshrdu_n : Neon_2Arg_ShiftImm_Intrinsic;
332 // Scalar Shift Right and Accumulate (Immediate)
333 def int_aarch64_neon_vsrads_n : Neon_3Arg_ShiftImm_Intrinsic;
334 def int_aarch64_neon_vsradu_n : Neon_3Arg_ShiftImm_Intrinsic;
336 // Scalar Rounding Shift Right and Accumulate (Immediate)
337 def int_aarch64_neon_vrsrads_n : Neon_3Arg_ShiftImm_Intrinsic;
338 def int_aarch64_neon_vrsradu_n : Neon_3Arg_ShiftImm_Intrinsic;
340 // Scalar Shift Left (Immediate)
341 def int_aarch64_neon_vshld_n : Neon_2Arg_ShiftImm_Intrinsic;
343 // Scalar Saturating Shift Left (Immediate)
344 def int_aarch64_neon_vqshls_n : Neon_N2V_Intrinsic;
345 def int_aarch64_neon_vqshlu_n : Neon_N2V_Intrinsic;
347 // Scalar Signed Saturating Shift Left Unsigned (Immediate)
348 def int_aarch64_neon_vqshlus_n : Neon_N2V_Intrinsic;
350 // Scalar Signed Fixed-point Convert To Floating-Point (Immediate)
351 def int_aarch64_neon_vcvtf32_n_s32 :
352 Intrinsic<[llvm_float_ty], [llvm_v1i32_ty, llvm_i32_ty], [IntrNoMem]>;
353 def int_aarch64_neon_vcvtf64_n_s64 :
354 Intrinsic<[llvm_double_ty], [llvm_v1i64_ty, llvm_i32_ty], [IntrNoMem]>;
356 // Scalar Unsigned Fixed-point Convert To Floating-Point (Immediate)
357 def int_aarch64_neon_vcvtf32_n_u32 :
358 Intrinsic<[llvm_float_ty], [llvm_v1i32_ty, llvm_i32_ty], [IntrNoMem]>;
359 def int_aarch64_neon_vcvtf64_n_u64 :
360 Intrinsic<[llvm_double_ty], [llvm_v1i64_ty, llvm_i32_ty], [IntrNoMem]>;
362 // Scalar Floating-point Convert To Signed Fixed-point (Immediate)
363 def int_aarch64_neon_vcvts_n_s32_f32 :
364 Intrinsic<[llvm_v1i32_ty], [llvm_v1f32_ty, llvm_i32_ty], [IntrNoMem]>;
365 def int_aarch64_neon_vcvtd_n_s64_f64 :
366 Intrinsic<[llvm_v1i64_ty], [llvm_v1f64_ty, llvm_i32_ty], [IntrNoMem]>;
368 // Scalar Floating-point Convert To Unsigned Fixed-point (Immediate)
369 def int_aarch64_neon_vcvts_n_u32_f32 :
370 Intrinsic<[llvm_v1i32_ty], [llvm_v1f32_ty, llvm_i32_ty], [IntrNoMem]>;
371 def int_aarch64_neon_vcvtd_n_u64_f64 :
372 Intrinsic<[llvm_v1i64_ty], [llvm_v1f64_ty, llvm_i32_ty], [IntrNoMem]>;
374 class Neon_SHA_Intrinsic
375 : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v1i32_ty, llvm_v4i32_ty],
378 def int_aarch64_neon_sha1c : Neon_SHA_Intrinsic;
379 def int_aarch64_neon_sha1m : Neon_SHA_Intrinsic;
380 def int_aarch64_neon_sha1p : Neon_SHA_Intrinsic;