1 //===- IntrinsicsAMDGPU.td - Defines AMDGPU 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 R600-specific intrinsics.
12 //===----------------------------------------------------------------------===//
14 class AMDGPUReadPreloadRegisterIntrinsic
15 : Intrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
17 class AMDGPUReadPreloadRegisterIntrinsicNamed<string name>
18 : Intrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>, GCCBuiltin<name>;
20 let TargetPrefix = "r600" in {
22 multiclass AMDGPUReadPreloadRegisterIntrinsic_xyz {
23 def _x : AMDGPUReadPreloadRegisterIntrinsic;
24 def _y : AMDGPUReadPreloadRegisterIntrinsic;
25 def _z : AMDGPUReadPreloadRegisterIntrinsic;
28 multiclass AMDGPUReadPreloadRegisterIntrinsic_xyz_named<string prefix> {
29 def _x : AMDGPUReadPreloadRegisterIntrinsicNamed<!strconcat(prefix, "_x")>;
30 def _y : AMDGPUReadPreloadRegisterIntrinsicNamed<!strconcat(prefix, "_y")>;
31 def _z : AMDGPUReadPreloadRegisterIntrinsicNamed<!strconcat(prefix, "_z")>;
34 defm int_r600_read_global_size : AMDGPUReadPreloadRegisterIntrinsic_xyz_named
35 <"__builtin_r600_read_global_size">;
36 defm int_r600_read_ngroups : AMDGPUReadPreloadRegisterIntrinsic_xyz_named
37 <"__builtin_r600_read_ngroups">;
38 defm int_r600_read_tgid : AMDGPUReadPreloadRegisterIntrinsic_xyz_named
39 <"__builtin_r600_read_tgid">;
41 defm int_r600_read_local_size : AMDGPUReadPreloadRegisterIntrinsic_xyz;
42 defm int_r600_read_tidig : AMDGPUReadPreloadRegisterIntrinsic_xyz;
44 def int_r600_group_barrier : GCCBuiltin<"__builtin_r600_group_barrier">,
45 Intrinsic<[], [], [IntrConvergent]>;
47 // AS 7 is PARAM_I_ADDRESS, used for kernel arguments
48 def int_r600_implicitarg_ptr :
49 GCCBuiltin<"__builtin_r600_implicitarg_ptr">,
50 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 7>], [],
51 [IntrNoMem, IntrSpeculatable]>;
53 def int_r600_rat_store_typed :
54 // 1st parameter: Data
55 // 2nd parameter: Index
56 // 3rd parameter: Constant RAT ID
57 Intrinsic<[], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty], []>,
58 GCCBuiltin<"__builtin_r600_rat_store_typed">;
60 def int_r600_recipsqrt_ieee : Intrinsic<
61 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
64 def int_r600_recipsqrt_clamped : Intrinsic<
65 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
68 def int_r600_cube : Intrinsic<
69 [llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem, IntrSpeculatable]
72 } // End TargetPrefix = "r600"
74 let TargetPrefix = "amdgcn" in {
76 //===----------------------------------------------------------------------===//
77 // ABI Special Intrinsics
78 //===----------------------------------------------------------------------===//
80 defm int_amdgcn_workitem_id : AMDGPUReadPreloadRegisterIntrinsic_xyz;
81 defm int_amdgcn_workgroup_id : AMDGPUReadPreloadRegisterIntrinsic_xyz_named
82 <"__builtin_amdgcn_workgroup_id">;
84 def int_amdgcn_dispatch_ptr :
85 GCCBuiltin<"__builtin_amdgcn_dispatch_ptr">,
86 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 2>], [],
87 [IntrNoMem, IntrSpeculatable]>;
89 def int_amdgcn_queue_ptr :
90 GCCBuiltin<"__builtin_amdgcn_queue_ptr">,
91 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 2>], [],
92 [IntrNoMem, IntrSpeculatable]>;
94 def int_amdgcn_kernarg_segment_ptr :
95 GCCBuiltin<"__builtin_amdgcn_kernarg_segment_ptr">,
96 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 2>], [],
97 [IntrNoMem, IntrSpeculatable]>;
99 def int_amdgcn_implicitarg_ptr :
100 GCCBuiltin<"__builtin_amdgcn_implicitarg_ptr">,
101 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 2>], [],
102 [IntrNoMem, IntrSpeculatable]>;
104 def int_amdgcn_groupstaticsize :
105 GCCBuiltin<"__builtin_amdgcn_groupstaticsize">,
106 Intrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
108 def int_amdgcn_dispatch_id :
109 GCCBuiltin<"__builtin_amdgcn_dispatch_id">,
110 Intrinsic<[llvm_i64_ty], [], [IntrNoMem, IntrSpeculatable]>;
112 def int_amdgcn_implicit_buffer_ptr :
113 GCCBuiltin<"__builtin_amdgcn_implicit_buffer_ptr">,
114 Intrinsic<[LLVMQualPointerType<llvm_i8_ty, 2>], [],
115 [IntrNoMem, IntrSpeculatable]>;
117 // Set EXEC to the 64-bit value given.
118 // This is always moved to the beginning of the basic block.
119 def int_amdgcn_init_exec : Intrinsic<[],
120 [llvm_i64_ty], // 64-bit literal constant
123 // Set EXEC according to a thread count packed in an SGPR input:
124 // thread_count = (input >> bitoffset) & 0x7f;
125 // This is always moved to the beginning of the basic block.
126 def int_amdgcn_init_exec_from_input : Intrinsic<[],
127 [llvm_i32_ty, // 32-bit SGPR input
128 llvm_i32_ty], // bit offset of the thread count
132 //===----------------------------------------------------------------------===//
133 // Instruction Intrinsics
134 //===----------------------------------------------------------------------===//
136 // The first parameter is s_sendmsg immediate (i16),
137 // the second one is copied to m0
138 def int_amdgcn_s_sendmsg : GCCBuiltin<"__builtin_amdgcn_s_sendmsg">,
139 Intrinsic <[], [llvm_i32_ty, llvm_i32_ty], []>;
140 def int_amdgcn_s_sendmsghalt : GCCBuiltin<"__builtin_amdgcn_s_sendmsghalt">,
141 Intrinsic <[], [llvm_i32_ty, llvm_i32_ty], []>;
143 def int_amdgcn_s_barrier : GCCBuiltin<"__builtin_amdgcn_s_barrier">,
144 Intrinsic<[], [], [IntrConvergent]>;
146 def int_amdgcn_wave_barrier : GCCBuiltin<"__builtin_amdgcn_wave_barrier">,
147 Intrinsic<[], [], [IntrConvergent]>;
149 def int_amdgcn_s_waitcnt : GCCBuiltin<"__builtin_amdgcn_s_waitcnt">,
150 Intrinsic<[], [llvm_i32_ty], []>;
152 def int_amdgcn_div_scale : Intrinsic<
153 // 1st parameter: Numerator
154 // 2nd parameter: Denominator
155 // 3rd parameter: Constant to select select between first and
156 // second. (0 = first, 1 = second).
157 [llvm_anyfloat_ty, llvm_i1_ty],
158 [LLVMMatchType<0>, LLVMMatchType<0>, llvm_i1_ty],
159 [IntrNoMem, IntrSpeculatable]
162 def int_amdgcn_div_fmas : Intrinsic<[llvm_anyfloat_ty],
163 [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>, llvm_i1_ty],
164 [IntrNoMem, IntrSpeculatable]
167 def int_amdgcn_div_fixup : Intrinsic<[llvm_anyfloat_ty],
168 [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
169 [IntrNoMem, IntrSpeculatable]
172 def int_amdgcn_trig_preop : Intrinsic<
173 [llvm_anyfloat_ty], [LLVMMatchType<0>, llvm_i32_ty],
174 [IntrNoMem, IntrSpeculatable]
177 def int_amdgcn_sin : Intrinsic<
178 [llvm_anyfloat_ty], [LLVMMatchType<0>],
179 [IntrNoMem, IntrSpeculatable]
182 def int_amdgcn_cos : Intrinsic<
183 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
186 def int_amdgcn_log_clamp : Intrinsic<
187 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
190 def int_amdgcn_fmul_legacy : GCCBuiltin<"__builtin_amdgcn_fmul_legacy">,
191 Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty],
192 [IntrNoMem, IntrSpeculatable]
195 def int_amdgcn_rcp : Intrinsic<
196 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
199 def int_amdgcn_rcp_legacy : GCCBuiltin<"__builtin_amdgcn_rcp_legacy">,
200 Intrinsic<[llvm_float_ty], [llvm_float_ty],
201 [IntrNoMem, IntrSpeculatable]
204 def int_amdgcn_rsq : Intrinsic<
205 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
208 def int_amdgcn_rsq_legacy : GCCBuiltin<"__builtin_amdgcn_rsq_legacy">,
210 [llvm_float_ty], [llvm_float_ty], [IntrNoMem, IntrSpeculatable]
213 def int_amdgcn_rsq_clamp : Intrinsic<
214 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]>;
216 def int_amdgcn_ldexp : Intrinsic<
217 [llvm_anyfloat_ty], [LLVMMatchType<0>, llvm_i32_ty],
218 [IntrNoMem, IntrSpeculatable]
221 def int_amdgcn_frexp_mant : Intrinsic<
222 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
225 def int_amdgcn_frexp_exp : Intrinsic<
226 [llvm_anyint_ty], [llvm_anyfloat_ty], [IntrNoMem, IntrSpeculatable]
229 // v_fract is buggy on SI/CI. It mishandles infinities, may return 1.0
230 // and always uses rtz, so is not suitable for implementing the OpenCL
231 // fract function. It should be ok on VI.
232 def int_amdgcn_fract : Intrinsic<
233 [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
236 def int_amdgcn_cvt_pkrtz : Intrinsic<
237 [llvm_v2f16_ty], [llvm_float_ty, llvm_float_ty],
238 [IntrNoMem, IntrSpeculatable]
241 def int_amdgcn_cvt_pknorm_i16 : Intrinsic<
242 [llvm_v2i16_ty], [llvm_float_ty, llvm_float_ty],
243 [IntrNoMem, IntrSpeculatable]
246 def int_amdgcn_cvt_pknorm_u16 : Intrinsic<
247 [llvm_v2i16_ty], [llvm_float_ty, llvm_float_ty],
248 [IntrNoMem, IntrSpeculatable]
251 def int_amdgcn_cvt_pk_i16 : Intrinsic<
252 [llvm_v2i16_ty], [llvm_i32_ty, llvm_i32_ty],
253 [IntrNoMem, IntrSpeculatable]
256 def int_amdgcn_cvt_pk_u16 : Intrinsic<
257 [llvm_v2i16_ty], [llvm_i32_ty, llvm_i32_ty],
258 [IntrNoMem, IntrSpeculatable]
261 def int_amdgcn_class : Intrinsic<
262 [llvm_i1_ty], [llvm_anyfloat_ty, llvm_i32_ty],
263 [IntrNoMem, IntrSpeculatable]
266 def int_amdgcn_fmed3 : GCCBuiltin<"__builtin_amdgcn_fmed3">,
267 Intrinsic<[llvm_anyfloat_ty],
268 [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
269 [IntrNoMem, IntrSpeculatable]
272 def int_amdgcn_cubeid : GCCBuiltin<"__builtin_amdgcn_cubeid">,
273 Intrinsic<[llvm_float_ty],
274 [llvm_float_ty, llvm_float_ty, llvm_float_ty],
275 [IntrNoMem, IntrSpeculatable]
278 def int_amdgcn_cubema : GCCBuiltin<"__builtin_amdgcn_cubema">,
279 Intrinsic<[llvm_float_ty],
280 [llvm_float_ty, llvm_float_ty, llvm_float_ty],
281 [IntrNoMem, IntrSpeculatable]
284 def int_amdgcn_cubesc : GCCBuiltin<"__builtin_amdgcn_cubesc">,
285 Intrinsic<[llvm_float_ty],
286 [llvm_float_ty, llvm_float_ty, llvm_float_ty],
287 [IntrNoMem, IntrSpeculatable]
290 def int_amdgcn_cubetc : GCCBuiltin<"__builtin_amdgcn_cubetc">,
291 Intrinsic<[llvm_float_ty],
292 [llvm_float_ty, llvm_float_ty, llvm_float_ty],
293 [IntrNoMem, IntrSpeculatable]
296 // v_ffbh_i32, as opposed to v_ffbh_u32. For v_ffbh_u32, llvm.ctlz
298 def int_amdgcn_sffbh :
299 Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>],
300 [IntrNoMem, IntrSpeculatable]
304 // Fields should mirror atomicrmw
305 class AMDGPUAtomicIncIntrin : Intrinsic<[llvm_anyint_ty],
308 llvm_i32_ty, // ordering
309 llvm_i32_ty, // scope
310 llvm_i1_ty], // isVolatile
311 [IntrArgMemOnly, NoCapture<0>], "",
315 def int_amdgcn_atomic_inc : AMDGPUAtomicIncIntrin;
316 def int_amdgcn_atomic_dec : AMDGPUAtomicIncIntrin;
318 class AMDGPUImageLoad<bit NoMem = 0> : Intrinsic <
319 [llvm_anyfloat_ty], // vdata(VGPR)
320 [llvm_anyint_ty, // vaddr(VGPR)
321 llvm_anyint_ty, // rsrc(SGPR)
322 llvm_i32_ty, // dmask(imm)
323 llvm_i1_ty, // glc(imm)
324 llvm_i1_ty, // slc(imm)
325 llvm_i1_ty, // lwe(imm)
326 llvm_i1_ty], // da(imm)
327 !if(NoMem, [IntrNoMem], [IntrReadMem]), "",
328 !if(NoMem, [], [SDNPMemOperand])>;
330 def int_amdgcn_image_load : AMDGPUImageLoad;
331 def int_amdgcn_image_load_mip : AMDGPUImageLoad;
332 def int_amdgcn_image_getresinfo : AMDGPUImageLoad<1>;
334 class AMDGPUImageStore : Intrinsic <
336 [llvm_anyfloat_ty, // vdata(VGPR)
337 llvm_anyint_ty, // vaddr(VGPR)
338 llvm_anyint_ty, // rsrc(SGPR)
339 llvm_i32_ty, // dmask(imm)
340 llvm_i1_ty, // glc(imm)
341 llvm_i1_ty, // slc(imm)
342 llvm_i1_ty, // lwe(imm)
343 llvm_i1_ty], // da(imm)
344 [IntrWriteMem], "", [SDNPMemOperand]>;
346 def int_amdgcn_image_store : AMDGPUImageStore;
347 def int_amdgcn_image_store_mip : AMDGPUImageStore;
349 class AMDGPUImageSample<bit NoMem = 0> : Intrinsic <
350 [llvm_anyfloat_ty], // vdata(VGPR)
351 [llvm_anyfloat_ty, // vaddr(VGPR)
352 llvm_anyint_ty, // rsrc(SGPR)
353 llvm_v4i32_ty, // sampler(SGPR)
354 llvm_i32_ty, // dmask(imm)
355 llvm_i1_ty, // unorm(imm)
356 llvm_i1_ty, // glc(imm)
357 llvm_i1_ty, // slc(imm)
358 llvm_i1_ty, // lwe(imm)
359 llvm_i1_ty], // da(imm)
360 !if(NoMem, [IntrNoMem], [IntrReadMem]), "",
361 !if(NoMem, [], [SDNPMemOperand])>;
364 def int_amdgcn_image_sample : AMDGPUImageSample;
365 def int_amdgcn_image_sample_cl : AMDGPUImageSample;
366 def int_amdgcn_image_sample_d : AMDGPUImageSample;
367 def int_amdgcn_image_sample_d_cl : AMDGPUImageSample;
368 def int_amdgcn_image_sample_l : AMDGPUImageSample;
369 def int_amdgcn_image_sample_b : AMDGPUImageSample;
370 def int_amdgcn_image_sample_b_cl : AMDGPUImageSample;
371 def int_amdgcn_image_sample_lz : AMDGPUImageSample;
372 def int_amdgcn_image_sample_cd : AMDGPUImageSample;
373 def int_amdgcn_image_sample_cd_cl : AMDGPUImageSample;
375 // Sample with comparison
376 def int_amdgcn_image_sample_c : AMDGPUImageSample;
377 def int_amdgcn_image_sample_c_cl : AMDGPUImageSample;
378 def int_amdgcn_image_sample_c_d : AMDGPUImageSample;
379 def int_amdgcn_image_sample_c_d_cl : AMDGPUImageSample;
380 def int_amdgcn_image_sample_c_l : AMDGPUImageSample;
381 def int_amdgcn_image_sample_c_b : AMDGPUImageSample;
382 def int_amdgcn_image_sample_c_b_cl : AMDGPUImageSample;
383 def int_amdgcn_image_sample_c_lz : AMDGPUImageSample;
384 def int_amdgcn_image_sample_c_cd : AMDGPUImageSample;
385 def int_amdgcn_image_sample_c_cd_cl : AMDGPUImageSample;
387 // Sample with offsets
388 def int_amdgcn_image_sample_o : AMDGPUImageSample;
389 def int_amdgcn_image_sample_cl_o : AMDGPUImageSample;
390 def int_amdgcn_image_sample_d_o : AMDGPUImageSample;
391 def int_amdgcn_image_sample_d_cl_o : AMDGPUImageSample;
392 def int_amdgcn_image_sample_l_o : AMDGPUImageSample;
393 def int_amdgcn_image_sample_b_o : AMDGPUImageSample;
394 def int_amdgcn_image_sample_b_cl_o : AMDGPUImageSample;
395 def int_amdgcn_image_sample_lz_o : AMDGPUImageSample;
396 def int_amdgcn_image_sample_cd_o : AMDGPUImageSample;
397 def int_amdgcn_image_sample_cd_cl_o : AMDGPUImageSample;
399 // Sample with comparison and offsets
400 def int_amdgcn_image_sample_c_o : AMDGPUImageSample;
401 def int_amdgcn_image_sample_c_cl_o : AMDGPUImageSample;
402 def int_amdgcn_image_sample_c_d_o : AMDGPUImageSample;
403 def int_amdgcn_image_sample_c_d_cl_o : AMDGPUImageSample;
404 def int_amdgcn_image_sample_c_l_o : AMDGPUImageSample;
405 def int_amdgcn_image_sample_c_b_o : AMDGPUImageSample;
406 def int_amdgcn_image_sample_c_b_cl_o : AMDGPUImageSample;
407 def int_amdgcn_image_sample_c_lz_o : AMDGPUImageSample;
408 def int_amdgcn_image_sample_c_cd_o : AMDGPUImageSample;
409 def int_amdgcn_image_sample_c_cd_cl_o : AMDGPUImageSample;
412 def int_amdgcn_image_gather4 : AMDGPUImageSample;
413 def int_amdgcn_image_gather4_cl : AMDGPUImageSample;
414 def int_amdgcn_image_gather4_l : AMDGPUImageSample;
415 def int_amdgcn_image_gather4_b : AMDGPUImageSample;
416 def int_amdgcn_image_gather4_b_cl : AMDGPUImageSample;
417 def int_amdgcn_image_gather4_lz : AMDGPUImageSample;
419 // Gather4 with comparison
420 def int_amdgcn_image_gather4_c : AMDGPUImageSample;
421 def int_amdgcn_image_gather4_c_cl : AMDGPUImageSample;
422 def int_amdgcn_image_gather4_c_l : AMDGPUImageSample;
423 def int_amdgcn_image_gather4_c_b : AMDGPUImageSample;
424 def int_amdgcn_image_gather4_c_b_cl : AMDGPUImageSample;
425 def int_amdgcn_image_gather4_c_lz : AMDGPUImageSample;
427 // Gather4 with offsets
428 def int_amdgcn_image_gather4_o : AMDGPUImageSample;
429 def int_amdgcn_image_gather4_cl_o : AMDGPUImageSample;
430 def int_amdgcn_image_gather4_l_o : AMDGPUImageSample;
431 def int_amdgcn_image_gather4_b_o : AMDGPUImageSample;
432 def int_amdgcn_image_gather4_b_cl_o : AMDGPUImageSample;
433 def int_amdgcn_image_gather4_lz_o : AMDGPUImageSample;
435 // Gather4 with comparison and offsets
436 def int_amdgcn_image_gather4_c_o : AMDGPUImageSample;
437 def int_amdgcn_image_gather4_c_cl_o : AMDGPUImageSample;
438 def int_amdgcn_image_gather4_c_l_o : AMDGPUImageSample;
439 def int_amdgcn_image_gather4_c_b_o : AMDGPUImageSample;
440 def int_amdgcn_image_gather4_c_b_cl_o : AMDGPUImageSample;
441 def int_amdgcn_image_gather4_c_lz_o : AMDGPUImageSample;
443 def int_amdgcn_image_getlod : AMDGPUImageSample<1>;
445 class AMDGPUImageAtomic : Intrinsic <
447 [llvm_i32_ty, // vdata(VGPR)
448 llvm_anyint_ty, // vaddr(VGPR)
449 llvm_v8i32_ty, // rsrc(SGPR)
450 llvm_i1_ty, // r128(imm)
451 llvm_i1_ty, // da(imm)
452 llvm_i1_ty], // slc(imm)
453 [], "", [SDNPMemOperand]>;
455 def int_amdgcn_image_atomic_swap : AMDGPUImageAtomic;
456 def int_amdgcn_image_atomic_add : AMDGPUImageAtomic;
457 def int_amdgcn_image_atomic_sub : AMDGPUImageAtomic;
458 def int_amdgcn_image_atomic_smin : AMDGPUImageAtomic;
459 def int_amdgcn_image_atomic_umin : AMDGPUImageAtomic;
460 def int_amdgcn_image_atomic_smax : AMDGPUImageAtomic;
461 def int_amdgcn_image_atomic_umax : AMDGPUImageAtomic;
462 def int_amdgcn_image_atomic_and : AMDGPUImageAtomic;
463 def int_amdgcn_image_atomic_or : AMDGPUImageAtomic;
464 def int_amdgcn_image_atomic_xor : AMDGPUImageAtomic;
465 def int_amdgcn_image_atomic_inc : AMDGPUImageAtomic;
466 def int_amdgcn_image_atomic_dec : AMDGPUImageAtomic;
467 def int_amdgcn_image_atomic_cmpswap : Intrinsic <
469 [llvm_i32_ty, // src(VGPR)
470 llvm_i32_ty, // cmp(VGPR)
471 llvm_anyint_ty, // vaddr(VGPR)
472 llvm_v8i32_ty, // rsrc(SGPR)
473 llvm_i1_ty, // r128(imm)
474 llvm_i1_ty, // da(imm)
475 llvm_i1_ty], // slc(imm)
476 [], "", [SDNPMemOperand]>;
478 class AMDGPUBufferLoad : Intrinsic <
480 [llvm_v4i32_ty, // rsrc(SGPR)
481 llvm_i32_ty, // vindex(VGPR)
482 llvm_i32_ty, // offset(SGPR/VGPR/imm)
483 llvm_i1_ty, // glc(imm)
484 llvm_i1_ty], // slc(imm)
485 [IntrReadMem], "", [SDNPMemOperand]>;
486 def int_amdgcn_buffer_load_format : AMDGPUBufferLoad;
487 def int_amdgcn_buffer_load : AMDGPUBufferLoad;
489 class AMDGPUBufferStore : Intrinsic <
491 [llvm_anyfloat_ty, // vdata(VGPR) -- can currently only select f32, v2f32, v4f32
492 llvm_v4i32_ty, // rsrc(SGPR)
493 llvm_i32_ty, // vindex(VGPR)
494 llvm_i32_ty, // offset(SGPR/VGPR/imm)
495 llvm_i1_ty, // glc(imm)
496 llvm_i1_ty], // slc(imm)
497 [IntrWriteMem], "", [SDNPMemOperand]>;
498 def int_amdgcn_buffer_store_format : AMDGPUBufferStore;
499 def int_amdgcn_buffer_store : AMDGPUBufferStore;
501 def int_amdgcn_tbuffer_load : Intrinsic <
502 [llvm_any_ty], // overloaded for types f32/i32, v2f32/v2i32, v4f32/v4i32
503 [llvm_v4i32_ty, // rsrc(SGPR)
504 llvm_i32_ty, // vindex(VGPR)
505 llvm_i32_ty, // voffset(VGPR)
506 llvm_i32_ty, // soffset(SGPR)
507 llvm_i32_ty, // offset(imm)
508 llvm_i32_ty, // dfmt(imm)
509 llvm_i32_ty, // nfmt(imm)
510 llvm_i1_ty, // glc(imm)
511 llvm_i1_ty], // slc(imm)
512 [IntrReadMem], "", [SDNPMemOperand]>;
514 def int_amdgcn_tbuffer_store : Intrinsic <
516 [llvm_any_ty, // vdata(VGPR), overloaded for types f32/i32, v2f32/v2i32, v4f32/v4i32
517 llvm_v4i32_ty, // rsrc(SGPR)
518 llvm_i32_ty, // vindex(VGPR)
519 llvm_i32_ty, // voffset(VGPR)
520 llvm_i32_ty, // soffset(SGPR)
521 llvm_i32_ty, // offset(imm)
522 llvm_i32_ty, // dfmt(imm)
523 llvm_i32_ty, // nfmt(imm)
524 llvm_i1_ty, // glc(imm)
525 llvm_i1_ty], // slc(imm)
526 [IntrWriteMem], "", [SDNPMemOperand]>;
528 class AMDGPUBufferAtomic : Intrinsic <
530 [llvm_i32_ty, // vdata(VGPR)
531 llvm_v4i32_ty, // rsrc(SGPR)
532 llvm_i32_ty, // vindex(VGPR)
533 llvm_i32_ty, // offset(SGPR/VGPR/imm)
534 llvm_i1_ty], // slc(imm)
535 [], "", [SDNPMemOperand]>;
536 def int_amdgcn_buffer_atomic_swap : AMDGPUBufferAtomic;
537 def int_amdgcn_buffer_atomic_add : AMDGPUBufferAtomic;
538 def int_amdgcn_buffer_atomic_sub : AMDGPUBufferAtomic;
539 def int_amdgcn_buffer_atomic_smin : AMDGPUBufferAtomic;
540 def int_amdgcn_buffer_atomic_umin : AMDGPUBufferAtomic;
541 def int_amdgcn_buffer_atomic_smax : AMDGPUBufferAtomic;
542 def int_amdgcn_buffer_atomic_umax : AMDGPUBufferAtomic;
543 def int_amdgcn_buffer_atomic_and : AMDGPUBufferAtomic;
544 def int_amdgcn_buffer_atomic_or : AMDGPUBufferAtomic;
545 def int_amdgcn_buffer_atomic_xor : AMDGPUBufferAtomic;
546 def int_amdgcn_buffer_atomic_cmpswap : Intrinsic<
548 [llvm_i32_ty, // src(VGPR)
549 llvm_i32_ty, // cmp(VGPR)
550 llvm_v4i32_ty, // rsrc(SGPR)
551 llvm_i32_ty, // vindex(VGPR)
552 llvm_i32_ty, // offset(SGPR/VGPR/imm)
553 llvm_i1_ty], // slc(imm)
554 [], "", [SDNPMemOperand]>;
556 // Uses that do not set the done bit should set IntrWriteMem on the
558 def int_amdgcn_exp : Intrinsic <[], [
561 llvm_any_ty, // src0 (f32 or i32)
562 LLVMMatchType<0>, // src1
563 LLVMMatchType<0>, // src2
564 LLVMMatchType<0>, // src3
571 // exp with compr bit set.
572 def int_amdgcn_exp_compr : Intrinsic <[], [
575 llvm_anyvector_ty, // src0 (v2f16 or v2i16)
576 LLVMMatchType<0>, // src1
582 def int_amdgcn_buffer_wbinvl1_sc :
583 GCCBuiltin<"__builtin_amdgcn_buffer_wbinvl1_sc">,
584 Intrinsic<[], [], []>;
586 def int_amdgcn_buffer_wbinvl1 :
587 GCCBuiltin<"__builtin_amdgcn_buffer_wbinvl1">,
588 Intrinsic<[], [], []>;
590 def int_amdgcn_s_dcache_inv :
591 GCCBuiltin<"__builtin_amdgcn_s_dcache_inv">,
592 Intrinsic<[], [], []>;
594 def int_amdgcn_s_memtime :
595 GCCBuiltin<"__builtin_amdgcn_s_memtime">,
596 Intrinsic<[llvm_i64_ty], [], [IntrReadMem]>;
598 def int_amdgcn_s_sleep :
599 GCCBuiltin<"__builtin_amdgcn_s_sleep">,
600 Intrinsic<[], [llvm_i32_ty], []> {
603 def int_amdgcn_s_incperflevel :
604 GCCBuiltin<"__builtin_amdgcn_s_incperflevel">,
605 Intrinsic<[], [llvm_i32_ty], []> {
608 def int_amdgcn_s_decperflevel :
609 GCCBuiltin<"__builtin_amdgcn_s_decperflevel">,
610 Intrinsic<[], [llvm_i32_ty], []> {
613 def int_amdgcn_s_getreg :
614 GCCBuiltin<"__builtin_amdgcn_s_getreg">,
615 Intrinsic<[llvm_i32_ty], [llvm_i32_ty],
616 [IntrReadMem, IntrSpeculatable]
619 // int_amdgcn_s_getpc is provided to allow a specific style of position
620 // independent code to determine the high part of its address when it is
621 // known (through convention) that the code and any data of interest does
622 // not cross a 4Gb address boundary. Use for any other purpose may not
623 // produce the desired results as optimizations may cause code movement,
624 // especially as we explicitly use IntrNoMem to allow optimizations.
625 def int_amdgcn_s_getpc :
626 GCCBuiltin<"__builtin_amdgcn_s_getpc">,
627 Intrinsic<[llvm_i64_ty], [], [IntrNoMem, IntrSpeculatable]>;
629 // __builtin_amdgcn_interp_mov <param>, <attr_chan>, <attr>, <m0>
630 // param values: 0 = P10, 1 = P20, 2 = P0
631 def int_amdgcn_interp_mov :
632 GCCBuiltin<"__builtin_amdgcn_interp_mov">,
633 Intrinsic<[llvm_float_ty],
634 [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
635 [IntrNoMem, IntrSpeculatable]>;
637 // __builtin_amdgcn_interp_p1 <i>, <attr_chan>, <attr>, <m0>
638 // This intrinsic reads from lds, but the memory values are constant,
639 // so it behaves like IntrNoMem.
640 def int_amdgcn_interp_p1 :
641 GCCBuiltin<"__builtin_amdgcn_interp_p1">,
642 Intrinsic<[llvm_float_ty],
643 [llvm_float_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
644 [IntrNoMem, IntrSpeculatable]>;
646 // __builtin_amdgcn_interp_p2 <p1>, <j>, <attr_chan>, <attr>, <m0>
647 def int_amdgcn_interp_p2 :
648 GCCBuiltin<"__builtin_amdgcn_interp_p2">,
649 Intrinsic<[llvm_float_ty],
650 [llvm_float_ty, llvm_float_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
651 [IntrNoMem, IntrSpeculatable]>;
652 // See int_amdgcn_v_interp_p1 for why this is IntrNoMem.
654 // Pixel shaders only: whether the current pixel is live (i.e. not a helper
655 // invocation for derivative computation).
656 def int_amdgcn_ps_live : Intrinsic <
661 def int_amdgcn_mbcnt_lo :
662 GCCBuiltin<"__builtin_amdgcn_mbcnt_lo">,
663 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
665 def int_amdgcn_mbcnt_hi :
666 GCCBuiltin<"__builtin_amdgcn_mbcnt_hi">,
667 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
669 // llvm.amdgcn.ds.swizzle src offset
670 def int_amdgcn_ds_swizzle :
671 GCCBuiltin<"__builtin_amdgcn_ds_swizzle">,
672 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
674 def int_amdgcn_ubfe : Intrinsic<[llvm_anyint_ty],
675 [LLVMMatchType<0>, llvm_i32_ty, llvm_i32_ty],
676 [IntrNoMem, IntrSpeculatable]
679 def int_amdgcn_sbfe : Intrinsic<[llvm_anyint_ty],
680 [LLVMMatchType<0>, llvm_i32_ty, llvm_i32_ty],
681 [IntrNoMem, IntrSpeculatable]
684 def int_amdgcn_lerp :
685 GCCBuiltin<"__builtin_amdgcn_lerp">,
686 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
687 [IntrNoMem, IntrSpeculatable]
690 def int_amdgcn_sad_u8 :
691 GCCBuiltin<"__builtin_amdgcn_sad_u8">,
692 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
693 [IntrNoMem, IntrSpeculatable]
696 def int_amdgcn_msad_u8 :
697 GCCBuiltin<"__builtin_amdgcn_msad_u8">,
698 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
699 [IntrNoMem, IntrSpeculatable]
702 def int_amdgcn_sad_hi_u8 :
703 GCCBuiltin<"__builtin_amdgcn_sad_hi_u8">,
704 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
705 [IntrNoMem, IntrSpeculatable]
708 def int_amdgcn_sad_u16 :
709 GCCBuiltin<"__builtin_amdgcn_sad_u16">,
710 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
711 [IntrNoMem, IntrSpeculatable]
714 def int_amdgcn_qsad_pk_u16_u8 :
715 GCCBuiltin<"__builtin_amdgcn_qsad_pk_u16_u8">,
716 Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i32_ty, llvm_i64_ty],
717 [IntrNoMem, IntrSpeculatable]
720 def int_amdgcn_mqsad_pk_u16_u8 :
721 GCCBuiltin<"__builtin_amdgcn_mqsad_pk_u16_u8">,
722 Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i32_ty, llvm_i64_ty],
723 [IntrNoMem, IntrSpeculatable]
726 def int_amdgcn_mqsad_u32_u8 :
727 GCCBuiltin<"__builtin_amdgcn_mqsad_u32_u8">,
728 Intrinsic<[llvm_v4i32_ty], [llvm_i64_ty, llvm_i32_ty, llvm_v4i32_ty],
729 [IntrNoMem, IntrSpeculatable]
732 def int_amdgcn_cvt_pk_u8_f32 :
733 GCCBuiltin<"__builtin_amdgcn_cvt_pk_u8_f32">,
734 Intrinsic<[llvm_i32_ty], [llvm_float_ty, llvm_i32_ty, llvm_i32_ty],
735 [IntrNoMem, IntrSpeculatable]
738 def int_amdgcn_icmp :
739 Intrinsic<[llvm_i64_ty], [llvm_anyint_ty, LLVMMatchType<0>, llvm_i32_ty],
740 [IntrNoMem, IntrConvergent]>;
742 def int_amdgcn_fcmp :
743 Intrinsic<[llvm_i64_ty], [llvm_anyfloat_ty, LLVMMatchType<0>, llvm_i32_ty],
744 [IntrNoMem, IntrConvergent]>;
746 def int_amdgcn_readfirstlane :
747 GCCBuiltin<"__builtin_amdgcn_readfirstlane">,
748 Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
750 // The lane argument must be uniform across the currently active threads of the
751 // current wave. Otherwise, the result is undefined.
752 def int_amdgcn_readlane :
753 GCCBuiltin<"__builtin_amdgcn_readlane">,
754 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
756 def int_amdgcn_alignbit : Intrinsic<[llvm_i32_ty],
757 [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
758 [IntrNoMem, IntrSpeculatable]
761 def int_amdgcn_alignbyte : Intrinsic<[llvm_i32_ty],
762 [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
763 [IntrNoMem, IntrSpeculatable]
767 // Copies the source value to the destination value, with the guarantee that
768 // the source value is computed as if the entire program were executed in WQM.
769 def int_amdgcn_wqm : Intrinsic<[llvm_any_ty],
770 [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
773 // Return true if at least one thread within the pixel quad passes true into
775 def int_amdgcn_wqm_vote : Intrinsic<[llvm_i1_ty],
776 [llvm_i1_ty], [IntrNoMem, IntrConvergent]
779 // If false, set EXEC=0 for the current thread until the end of program.
780 def int_amdgcn_kill : Intrinsic<[], [llvm_i1_ty], []>;
782 // Copies the active channels of the source value to the destination value,
783 // with the guarantee that the source value is computed as if the entire
784 // program were executed in Whole Wavefront Mode, i.e. with all channels
785 // enabled, with a few exceptions: - Phi nodes with require WWM return an
787 def int_amdgcn_wwm : Intrinsic<[llvm_any_ty],
788 [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
791 // Given a value, copies it while setting all the inactive lanes to a given
792 // value. Note that OpenGL helper lanes are considered active, so if the
793 // program ever uses WQM, then the instruction and the first source will be
795 def int_amdgcn_set_inactive :
796 Intrinsic<[llvm_anyint_ty],
797 [LLVMMatchType<0>, // value to be copied
798 LLVMMatchType<0>], // value for the inactive lanes to take
799 [IntrNoMem, IntrConvergent]>;
801 //===----------------------------------------------------------------------===//
803 //===----------------------------------------------------------------------===//
805 def int_amdgcn_s_dcache_inv_vol :
806 GCCBuiltin<"__builtin_amdgcn_s_dcache_inv_vol">,
807 Intrinsic<[], [], []>;
809 def int_amdgcn_buffer_wbinvl1_vol :
810 GCCBuiltin<"__builtin_amdgcn_buffer_wbinvl1_vol">,
811 Intrinsic<[], [], []>;
813 //===----------------------------------------------------------------------===//
815 //===----------------------------------------------------------------------===//
817 // llvm.amdgcn.mov.dpp.i32 <src> <dpp_ctrl> <row_mask> <bank_mask> <bound_ctrl>
818 def int_amdgcn_mov_dpp :
819 Intrinsic<[llvm_anyint_ty],
820 [LLVMMatchType<0>, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty,
821 llvm_i1_ty], [IntrNoMem, IntrConvergent]>;
823 // llvm.amdgcn.update.dpp.i32 <old> <src> <dpp_ctrl> <row_mask> <bank_mask> <bound_ctrl>
824 // Should be equivalent to:
825 // v_mov_b32 <dest> <old>
826 // v_mov_b32 <dest> <src> <dpp_ctrl> <row_mask> <bank_mask> <bound_ctrl>
827 def int_amdgcn_update_dpp :
828 Intrinsic<[llvm_anyint_ty],
829 [LLVMMatchType<0>, LLVMMatchType<0>, llvm_i32_ty, llvm_i32_ty,
830 llvm_i32_ty, llvm_i1_ty], [IntrNoMem, IntrConvergent]>;
832 def int_amdgcn_s_dcache_wb :
833 GCCBuiltin<"__builtin_amdgcn_s_dcache_wb">,
834 Intrinsic<[], [], []>;
836 def int_amdgcn_s_dcache_wb_vol :
837 GCCBuiltin<"__builtin_amdgcn_s_dcache_wb_vol">,
838 Intrinsic<[], [], []>;
840 def int_amdgcn_s_memrealtime :
841 GCCBuiltin<"__builtin_amdgcn_s_memrealtime">,
842 Intrinsic<[llvm_i64_ty], [], [IntrReadMem]>;
844 // llvm.amdgcn.ds.permute <index> <src>
845 def int_amdgcn_ds_permute :
846 GCCBuiltin<"__builtin_amdgcn_ds_permute">,
847 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
849 // llvm.amdgcn.ds.bpermute <index> <src>
850 def int_amdgcn_ds_bpermute :
851 GCCBuiltin<"__builtin_amdgcn_ds_bpermute">,
852 Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
855 //===----------------------------------------------------------------------===//
856 // Special Intrinsics for backend internal use only. No frontend
857 // should emit calls to these.
858 // ===----------------------------------------------------------------------===//
859 def int_amdgcn_if : Intrinsic<[llvm_i1_ty, llvm_i64_ty],
860 [llvm_i1_ty], [IntrConvergent]
863 def int_amdgcn_else : Intrinsic<[llvm_i1_ty, llvm_i64_ty],
864 [llvm_i64_ty], [IntrConvergent]
867 def int_amdgcn_break : Intrinsic<[llvm_i64_ty],
868 [llvm_i64_ty], [IntrNoMem, IntrConvergent]
871 def int_amdgcn_if_break : Intrinsic<[llvm_i64_ty],
872 [llvm_i1_ty, llvm_i64_ty], [IntrNoMem, IntrConvergent]
875 def int_amdgcn_else_break : Intrinsic<[llvm_i64_ty],
876 [llvm_i64_ty, llvm_i64_ty], [IntrNoMem, IntrConvergent]
879 def int_amdgcn_loop : Intrinsic<[llvm_i1_ty],
880 [llvm_i64_ty], [IntrConvergent]
883 def int_amdgcn_end_cf : Intrinsic<[], [llvm_i64_ty], [IntrConvergent]>;
885 // Represent unreachable in a divergent region.
886 def int_amdgcn_unreachable : Intrinsic<[], [], [IntrConvergent]>;
888 // Emit 2.5 ulp, no denormal division. Should only be inserted by
889 // pass based on !fpmath metadata.
890 def int_amdgcn_fdiv_fast : Intrinsic<
891 [llvm_float_ty], [llvm_float_ty, llvm_float_ty],
892 [IntrNoMem, IntrSpeculatable]