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_vcvtxn :
40 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
41 def int_aarch64_neon_vcvtzs :
42 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
43 def int_aarch64_neon_vcvtzu :
44 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
46 // Vector maxNum (Floating Point)
47 def int_aarch64_neon_vmaxnm : Neon_2Arg_Intrinsic;
49 // Vector minNum (Floating Point)
50 def int_aarch64_neon_vminnm : Neon_2Arg_Intrinsic;
52 // Vector Pairwise maxNum (Floating Point)
53 def int_aarch64_neon_vpmaxnm : Neon_2Arg_Intrinsic;
55 // Vector Pairwise minNum (Floating Point)
56 def int_aarch64_neon_vpminnm : Neon_2Arg_Intrinsic;
58 // Vector Multiply Extended and Scalar Multiply Extended (Floating Point)
59 def int_aarch64_neon_vmulx :
60 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>]>;
62 class Neon_N2V_Intrinsic
63 : Intrinsic<[llvm_anyvector_ty], [LLVMMatchType<0>, llvm_i32_ty],
65 class Neon_N3V_Intrinsic
66 : Intrinsic<[llvm_anyvector_ty],
67 [LLVMMatchType<0>, LLVMMatchType<0>, llvm_i32_ty],
69 class Neon_N2V_Narrow_Intrinsic
70 : Intrinsic<[llvm_anyvector_ty],
71 [LLVMExtendedElementVectorType<0>, llvm_i32_ty],
74 // Vector rounding shift right by immediate (Signed)
75 def int_aarch64_neon_vsrshr : Neon_N2V_Intrinsic;
76 def int_aarch64_neon_vurshr : Neon_N2V_Intrinsic;
77 def int_aarch64_neon_vsqshlu : Neon_N2V_Intrinsic;
79 def int_aarch64_neon_vsri : Neon_N3V_Intrinsic;
80 def int_aarch64_neon_vsli : Neon_N3V_Intrinsic;
82 def int_aarch64_neon_vsqshrun : Neon_N2V_Narrow_Intrinsic;
83 def int_aarch64_neon_vrshrn : Neon_N2V_Narrow_Intrinsic;
84 def int_aarch64_neon_vsqrshrun : Neon_N2V_Narrow_Intrinsic;
85 def int_aarch64_neon_vsqshrn : Neon_N2V_Narrow_Intrinsic;
86 def int_aarch64_neon_vuqshrn : Neon_N2V_Narrow_Intrinsic;
87 def int_aarch64_neon_vsqrshrn : Neon_N2V_Narrow_Intrinsic;
88 def int_aarch64_neon_vuqrshrn : Neon_N2V_Narrow_Intrinsic;
91 class Neon_Across_Intrinsic
92 : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
94 def int_aarch64_neon_saddlv : Neon_Across_Intrinsic;
95 def int_aarch64_neon_uaddlv : Neon_Across_Intrinsic;
96 def int_aarch64_neon_smaxv : Neon_Across_Intrinsic;
97 def int_aarch64_neon_umaxv : Neon_Across_Intrinsic;
98 def int_aarch64_neon_sminv : Neon_Across_Intrinsic;
99 def int_aarch64_neon_uminv : Neon_Across_Intrinsic;
100 def int_aarch64_neon_vaddv : Neon_Across_Intrinsic;
101 def int_aarch64_neon_vmaxv :
102 Intrinsic<[llvm_float_ty], [llvm_v4f32_ty], [IntrNoMem]>;
103 def int_aarch64_neon_vminv :
104 Intrinsic<[llvm_float_ty], [llvm_v4f32_ty], [IntrNoMem]>;
105 def int_aarch64_neon_vmaxnmv :
106 Intrinsic<[llvm_float_ty], [llvm_v4f32_ty], [IntrNoMem]>;
107 def int_aarch64_neon_vminnmv :
108 Intrinsic<[llvm_float_ty], [llvm_v4f32_ty], [IntrNoMem]>;
110 // Vector Table Lookup.
111 def int_aarch64_neon_vtbl1 :
112 Intrinsic<[llvm_anyvector_ty],
113 [llvm_v16i8_ty, LLVMMatchType<0>], [IntrNoMem]>;
115 def int_aarch64_neon_vtbl2 :
116 Intrinsic<[llvm_anyvector_ty],
117 [llvm_v16i8_ty, llvm_v16i8_ty, LLVMMatchType<0>],
120 def int_aarch64_neon_vtbl3 :
121 Intrinsic<[llvm_anyvector_ty],
122 [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty,
123 LLVMMatchType<0>], [IntrNoMem]>;
125 def int_aarch64_neon_vtbl4 :
126 Intrinsic<[llvm_anyvector_ty],
127 [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty,
128 llvm_v16i8_ty, LLVMMatchType<0>], [IntrNoMem]>;
130 // Vector Table Extension.
131 // Some elements of the destination vector may not be updated, so the original
132 // value of that vector is passed as the first argument. The next 1-4
133 // arguments after that are the table.
134 def int_aarch64_neon_vtbx1 :
135 Intrinsic<[llvm_anyvector_ty],
136 [LLVMMatchType<0>, llvm_v16i8_ty, LLVMMatchType<0>],
139 def int_aarch64_neon_vtbx2 :
140 Intrinsic<[llvm_anyvector_ty],
141 [LLVMMatchType<0>, llvm_v16i8_ty, llvm_v16i8_ty,
142 LLVMMatchType<0>], [IntrNoMem]>;
144 def int_aarch64_neon_vtbx3 :
145 Intrinsic<[llvm_anyvector_ty],
146 [LLVMMatchType<0>, llvm_v16i8_ty, llvm_v16i8_ty,
147 llvm_v16i8_ty, LLVMMatchType<0>], [IntrNoMem]>;
149 def int_aarch64_neon_vtbx4 :
150 Intrinsic<[llvm_anyvector_ty],
151 [LLVMMatchType<0>, llvm_v16i8_ty, llvm_v16i8_ty,
152 llvm_v16i8_ty, llvm_v16i8_ty, LLVMMatchType<0>],
156 def int_aarch64_neon_vld1x2 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>],
157 [llvm_ptr_ty, llvm_i32_ty],
159 def int_aarch64_neon_vld1x3 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
161 [llvm_ptr_ty, llvm_i32_ty],
163 def int_aarch64_neon_vld1x4 : Intrinsic<[llvm_anyvector_ty, LLVMMatchType<0>,
164 LLVMMatchType<0>, LLVMMatchType<0>],
165 [llvm_ptr_ty, llvm_i32_ty],
168 def int_aarch64_neon_vst1x2 : Intrinsic<[],
169 [llvm_ptr_ty, llvm_anyvector_ty,
170 LLVMMatchType<0>, llvm_i32_ty],
171 [IntrReadWriteArgMem]>;
172 def int_aarch64_neon_vst1x3 : Intrinsic<[],
173 [llvm_ptr_ty, llvm_anyvector_ty,
174 LLVMMatchType<0>, LLVMMatchType<0>,
175 llvm_i32_ty], [IntrReadWriteArgMem]>;
176 def int_aarch64_neon_vst1x4 : Intrinsic<[],
177 [llvm_ptr_ty, llvm_anyvector_ty,
178 LLVMMatchType<0>, LLVMMatchType<0>,
179 LLVMMatchType<0>, llvm_i32_ty],
180 [IntrReadWriteArgMem]>;
183 def int_aarch64_neon_vaddds :
184 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
185 def int_aarch64_neon_vadddu :
186 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
190 def int_aarch64_neon_vsubds :
191 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
192 def int_aarch64_neon_vsubdu :
193 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
198 def int_aarch64_neon_vshlds :
199 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
200 def int_aarch64_neon_vshldu :
201 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
203 // Scalar Saturating Shift Left
204 def int_aarch64_neon_vqshls : Neon_2Arg_Intrinsic;
205 def int_aarch64_neon_vqshlu : Neon_2Arg_Intrinsic;
207 // Scalar Shift Rouding Left
208 def int_aarch64_neon_vrshlds :
209 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
210 def int_aarch64_neon_vrshldu :
211 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
213 // Scalar Saturating Rounding Shift Left
214 def int_aarch64_neon_vqrshls : Neon_2Arg_Intrinsic;
215 def int_aarch64_neon_vqrshlu : Neon_2Arg_Intrinsic;
217 // Scalar Reduce Pairwise Add.
218 def int_aarch64_neon_vpadd :
219 Intrinsic<[llvm_v1i64_ty], [llvm_v2i64_ty],[IntrNoMem]>;
220 def int_aarch64_neon_vpfadd :
221 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
223 // Scalar Reduce Pairwise Floating Point Max/Min.
224 def int_aarch64_neon_vpmax :
225 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
226 def int_aarch64_neon_vpmin :
227 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
229 // Scalar Reduce Pairwise Floating Point Maxnm/Minnm.
230 def int_aarch64_neon_vpfmaxnm :
231 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
232 def int_aarch64_neon_vpfminnm :
233 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
235 // Scalar Signed Integer Convert To Floating-point
236 def int_aarch64_neon_vcvtint2fps :
237 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
239 // Scalar Unsigned Integer Convert To Floating-point
240 def int_aarch64_neon_vcvtint2fpu :
241 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty], [IntrNoMem]>;
243 // Scalar Floating-point Convert
244 def int_aarch64_neon_fcvtxn :
245 Intrinsic<[llvm_float_ty], [llvm_double_ty], [IntrNoMem]>;
246 def int_aarch64_neon_fcvtns :
247 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
248 def int_aarch64_neon_fcvtnu :
249 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
250 def int_aarch64_neon_fcvtps :
251 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
252 def int_aarch64_neon_fcvtpu :
253 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
254 def int_aarch64_neon_fcvtms :
255 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
256 def int_aarch64_neon_fcvtmu :
257 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
258 def int_aarch64_neon_fcvtas :
259 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
260 def int_aarch64_neon_fcvtau :
261 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
262 def int_aarch64_neon_fcvtzs :
263 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
264 def int_aarch64_neon_fcvtzu :
265 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
267 // Scalar Floating-point Reciprocal Estimate.
268 def int_aarch64_neon_vrecpe :
269 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
271 // Scalar Floating-point Reciprocal Exponent
272 def int_aarch64_neon_vrecpx :
273 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
275 // Scalar Floating-point Reciprocal Square Root Estimate
276 def int_aarch64_neon_vrsqrte :
277 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
279 // Scalar Floating-point Reciprocal Step
280 def int_aarch64_neon_vrecps :
281 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
284 // Scalar Floating-point Reciprocal Square Root Step
285 def int_aarch64_neon_vrsqrts :
286 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
289 // Compare with vector operands.
290 class Neon_Cmp_Intrinsic :
291 Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_anyvector_ty],
294 // Floating-point compare with scalar operands.
295 class Neon_Float_Cmp_Intrinsic :
296 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty, llvm_anyfloat_ty],
299 // Scalar Compare Equal
300 def int_aarch64_neon_vceq : Neon_Cmp_Intrinsic;
301 def int_aarch64_neon_fceq : Neon_Float_Cmp_Intrinsic;
303 // Scalar Compare Greater-Than or Equal
304 def int_aarch64_neon_vcge : Neon_Cmp_Intrinsic;
305 def int_aarch64_neon_vchs : Neon_Cmp_Intrinsic;
306 def int_aarch64_neon_fcge : Neon_Float_Cmp_Intrinsic;
307 def int_aarch64_neon_fchs : Neon_Float_Cmp_Intrinsic;
309 // Scalar Compare Less-Than or Equal
310 def int_aarch64_neon_vclez : Neon_Cmp_Intrinsic;
311 def int_aarch64_neon_fclez : Neon_Float_Cmp_Intrinsic;
313 // Scalar Compare Less-Than
314 def int_aarch64_neon_vcltz : Neon_Cmp_Intrinsic;
315 def int_aarch64_neon_fcltz : Neon_Float_Cmp_Intrinsic;
317 // Scalar Compare Greater-Than
318 def int_aarch64_neon_vcgt : Neon_Cmp_Intrinsic;
319 def int_aarch64_neon_vchi : Neon_Cmp_Intrinsic;
320 def int_aarch64_neon_fcgt : Neon_Float_Cmp_Intrinsic;
321 def int_aarch64_neon_fchi : Neon_Float_Cmp_Intrinsic;
323 // Scalar Compare Bitwise Test Bits
324 def int_aarch64_neon_vtstd : Neon_Cmp_Intrinsic;
326 // Scalar Floating-point Absolute Compare Greater Than Or Equal
327 def int_aarch64_neon_vcage : Neon_Cmp_Intrinsic;
328 def int_aarch64_neon_fcage : Neon_Float_Cmp_Intrinsic;
330 // Scalar Floating-point Absolute Compare Greater Than
331 def int_aarch64_neon_vcagt : Neon_Cmp_Intrinsic;
332 def int_aarch64_neon_fcagt : Neon_Float_Cmp_Intrinsic;
334 // Scalar Signed Saturating Accumulated of Unsigned Value
335 def int_aarch64_neon_vuqadd : Neon_2Arg_Intrinsic;
337 // Scalar Unsigned Saturating Accumulated of Signed Value
338 def int_aarch64_neon_vsqadd : Neon_2Arg_Intrinsic;
340 // Scalar Absolute Value
341 def int_aarch64_neon_vabs :
342 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty], [IntrNoMem]>;
344 // Scalar Absolute Difference
345 def int_aarch64_neon_vabd :
346 Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
349 // Scalar Negate Value
350 def int_aarch64_neon_vneg :
351 Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty], [IntrNoMem]>;
353 // Signed Saturating Doubling Multiply-Add Long
354 def int_aarch64_neon_vqdmlal : Neon_3Arg_Long_Intrinsic;
356 // Signed Saturating Doubling Multiply-Subtract Long
357 def int_aarch64_neon_vqdmlsl : Neon_3Arg_Long_Intrinsic;
359 def int_aarch64_neon_vmull_p64 :
360 Intrinsic<[llvm_v16i8_ty], [llvm_v1i64_ty, llvm_v1i64_ty], [IntrNoMem]>;
362 class Neon_2Arg_ShiftImm_Intrinsic
363 : Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_i32_ty], [IntrNoMem]>;
365 class Neon_3Arg_ShiftImm_Intrinsic
366 : Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty, llvm_v1i64_ty, llvm_i32_ty],
369 // Scalar Shift Right (Immediate)
370 def int_aarch64_neon_vshrds_n : Neon_2Arg_ShiftImm_Intrinsic;
371 def int_aarch64_neon_vshrdu_n : Neon_2Arg_ShiftImm_Intrinsic;
373 // Scalar Shift Right and Accumulate (Immediate)
374 def int_aarch64_neon_vsrads_n : Neon_3Arg_ShiftImm_Intrinsic;
375 def int_aarch64_neon_vsradu_n : Neon_3Arg_ShiftImm_Intrinsic;
377 // Scalar Rounding Shift Right and Accumulate (Immediate)
378 def int_aarch64_neon_vrsrads_n : Neon_3Arg_ShiftImm_Intrinsic;
379 def int_aarch64_neon_vrsradu_n : Neon_3Arg_ShiftImm_Intrinsic;
381 // Scalar Shift Left (Immediate)
382 def int_aarch64_neon_vshld_n : Neon_2Arg_ShiftImm_Intrinsic;
384 // Scalar Saturating Shift Left (Immediate)
385 def int_aarch64_neon_vqshls_n : Neon_N2V_Intrinsic;
386 def int_aarch64_neon_vqshlu_n : Neon_N2V_Intrinsic;
388 // Scalar Signed Saturating Shift Left Unsigned (Immediate)
389 def int_aarch64_neon_vqshlus_n : Neon_N2V_Intrinsic;
391 // Scalar Signed Fixed-point Convert To Floating-Point (Immediate)
392 def int_aarch64_neon_vcvtfxs2fp_n :
393 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty, llvm_i32_ty], [IntrNoMem]>;
395 // Scalar Unsigned Fixed-point Convert To Floating-Point (Immediate)
396 def int_aarch64_neon_vcvtfxu2fp_n :
397 Intrinsic<[llvm_anyfloat_ty], [llvm_anyvector_ty, llvm_i32_ty], [IntrNoMem]>;
399 // Scalar Floating-point Convert To Signed Fixed-point (Immediate)
400 def int_aarch64_neon_vcvtfp2fxs_n :
401 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty, llvm_i32_ty], [IntrNoMem]>;
403 // Scalar Floating-point Convert To Unsigned Fixed-point (Immediate)
404 def int_aarch64_neon_vcvtfp2fxu_n :
405 Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty, llvm_i32_ty], [IntrNoMem]>;