1 /*
2 * Copyright 2016 Bas Nieuwenhuizen
3 *
4 * SPDX-License-Identifier: MIT
5 */
6 #ifndef AC_LLVM_BUILD_H
7 #define AC_LLVM_BUILD_H
8
9 #include "ac_llvm_util.h"
10 #include "ac_shader_abi.h"
11 #include "ac_shader_args.h"
12 #include "ac_shader_util.h"
13 #include "amd_family.h"
14 #include "compiler/nir/nir.h"
15 #include <llvm-c/Core.h>
16
17 #include <stdbool.h>
18
19 #ifdef __cplusplus
20 extern "C" {
21 #endif
22
23 enum
24 {
25 AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
26 AC_ADDR_SPACE_GLOBAL = 1,
27 AC_ADDR_SPACE_GDS = 2,
28 AC_ADDR_SPACE_LDS = 3,
29 AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
30 AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
31 };
32
33 /* Fine-grained wait flags for GFX12, older chips merge them. */
34 #define AC_WAIT_DS (1 << 0) /* s_wait_dscnt (GFX12) or lgkmcnt (GFX6-11): LDS (all gens), GDS (GFX6-11) */
35 #define AC_WAIT_KM (1 << 1) /* s_wait_kmcnt (GFX12) or lgkmcnt (GFX6-11): SMEM, message */
36 #define AC_WAIT_EXP (1 << 2) /* s_wait_expcnt: Exports */
37 #define AC_WAIT_LOAD (1 << 3) /* s_wait_loadcnt (GFX12) or vmcnt (GFX6-11): VMEM loads */
38 #define AC_WAIT_STORE (1 << 4) /* s_wait_storecnt (GFX12) or vscnt (GFX10-11) or vmcnt (GFX6-9): VMEM stores */
39 #define AC_WAIT_SAMPLE (1 << 5) /* s_wait_samplecnt (GFX12) or vmcnt (GFX6-11): VMEM sample/gather */
40 #define AC_WAIT_BVH (1 << 6) /* s_wait_bvhcnt (GFX12) or vmcnt (GFX6-11): BVH */
41
42 struct ac_llvm_flow;
43 struct ac_llvm_compiler;
44 struct radeon_info;
45
46 struct ac_llvm_flow_state {
47 struct ac_llvm_flow *stack;
48 unsigned depth_max;
49 unsigned depth;
50 };
51
52 struct ac_llvm_pointer {
53 union {
54 LLVMValueRef value;
55 LLVMValueRef v;
56 };
57 /* Doesn't support complex types (pointer to pointer to etc...),
58 * but this isn't a problem since there's no place where this
59 * would be required.
60 */
61 union {
62 LLVMTypeRef pointee_type;
63 LLVMTypeRef t;
64 };
65 };
66
67 struct ac_llvm_context {
68 LLVMContextRef context;
69 LLVMModuleRef module;
70 LLVMBuilderRef builder;
71
72 struct ac_llvm_pointer main_function;
73
74 LLVMTypeRef voidt;
75 LLVMTypeRef i1;
76 LLVMTypeRef i8;
77 LLVMTypeRef i16;
78 LLVMTypeRef i32;
79 LLVMTypeRef i64;
80 LLVMTypeRef i128;
81 LLVMTypeRef intptr;
82 LLVMTypeRef f16;
83 LLVMTypeRef f32;
84 LLVMTypeRef f64;
85 LLVMTypeRef v4i8;
86 LLVMTypeRef v2i16;
87 LLVMTypeRef v4i16;
88 LLVMTypeRef v2f16;
89 LLVMTypeRef v4f16;
90 LLVMTypeRef v2i32;
91 LLVMTypeRef v3i32;
92 LLVMTypeRef v4i32;
93 LLVMTypeRef v2f32;
94 LLVMTypeRef v3f32;
95 LLVMTypeRef v4f32;
96 LLVMTypeRef v8i32;
97 LLVMTypeRef iN_wavemask;
98 LLVMTypeRef iN_ballotmask;
99
100 LLVMValueRef i8_0;
101 LLVMValueRef i8_1;
102 LLVMValueRef i16_0;
103 LLVMValueRef i16_1;
104 LLVMValueRef i32_0;
105 LLVMValueRef i32_1;
106 LLVMValueRef i64_0;
107 LLVMValueRef i64_1;
108 LLVMValueRef i128_0;
109 LLVMValueRef i128_1;
110 LLVMValueRef f16_0;
111 LLVMValueRef f16_1;
112 LLVMValueRef f32_0;
113 LLVMValueRef f32_1;
114 LLVMValueRef f64_0;
115 LLVMValueRef f64_1;
116 LLVMValueRef i1true;
117 LLVMValueRef i1false;
118
119 /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
120 * are two ac_llvm_contexts. Declare a pointer here, so that the control
121 * flow stack is shared by both ac_llvm_contexts.
122 */
123 struct ac_llvm_flow_state *flow;
124
125 unsigned range_md_kind;
126 unsigned invariant_load_md_kind;
127 unsigned uniform_md_kind;
128 unsigned fpmath_md_kind;
129 LLVMValueRef empty_md;
130 LLVMValueRef three_md;
131
132 const struct radeon_info *info;
133 enum amd_gfx_level gfx_level;
134
135 unsigned wave_size;
136 unsigned ballot_mask_bits;
137
138 unsigned float_mode;
139
140 bool exports_color_null;
141 bool exports_mrtz;
142
143 struct ac_llvm_pointer lds;
144
145 LLVMValueRef ring_offsets;
146 int ring_offsets_index;
147 };
148
149 void ac_llvm_context_init(struct ac_llvm_context *ctx, struct ac_llvm_compiler *compiler,
150 const struct radeon_info *info, enum ac_float_mode float_mode,
151 unsigned wave_size, unsigned ballot_mask_bits, bool exports_color_null,
152 bool exports_mrtz);
153
154 void ac_llvm_context_dispose(struct ac_llvm_context *ctx);
155
156 int ac_get_llvm_num_components(LLVMValueRef value);
157
158 int ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
159
160 LLVMValueRef ac_llvm_extract_elem(struct ac_llvm_context *ac, LLVMValueRef value, int index);
161
162 unsigned ac_get_type_size(LLVMTypeRef type);
163
164 LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
165 LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
166 LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
167 LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
168 LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
169
170 LLVMValueRef ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
171 LLVMTypeRef return_type, LLVMValueRef *params, unsigned param_count,
172 unsigned attrib_mask);
173
174 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
175
176 LLVMValueRef ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type, unsigned count_incoming,
177 LLVMValueRef *values, LLVMBasicBlockRef *blocks);
178
179 void ac_build_s_barrier(struct ac_llvm_context *ctx, gl_shader_stage stage);
180 void ac_build_optimization_barrier(struct ac_llvm_context *ctx, LLVMValueRef *pgpr, bool sgpr);
181
182 LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx, mesa_scope scope);
183
184 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
185
186 LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
187
188 LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
189
190 LLVMValueRef ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
191 unsigned value_count, unsigned component);
192
193 LLVMValueRef ac_build_gather_values_extended(struct ac_llvm_context *ctx, LLVMValueRef *values,
194 unsigned value_count, unsigned value_stride,
195 bool always_vector);
196 LLVMValueRef ac_build_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
197 unsigned value_count);
198
199 LLVMValueRef ac_build_concat(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
200
201 LLVMValueRef ac_extract_components(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned start,
202 unsigned channels);
203
204 LLVMValueRef ac_build_expand(struct ac_llvm_context *ctx, LLVMValueRef value,
205 unsigned src_channels, unsigned dst_channels);
206
207 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx, LLVMValueRef value,
208 unsigned num_channels);
209 LLVMValueRef ac_build_fdiv(struct ac_llvm_context *ctx, LLVMValueRef num, LLVMValueRef den);
210
211 LLVMValueRef ac_build_fs_interp(struct ac_llvm_context *ctx, LLVMValueRef llvm_chan,
212 LLVMValueRef attr_number, LLVMValueRef params, LLVMValueRef i,
213 LLVMValueRef j);
214
215 LLVMValueRef ac_build_fs_interp_f16(struct ac_llvm_context *ctx, LLVMValueRef llvm_chan,
216 LLVMValueRef attr_number, LLVMValueRef params, LLVMValueRef i,
217 LLVMValueRef j, bool high_16bits);
218
219 LLVMValueRef ac_build_fs_interp_mov(struct ac_llvm_context *ctx, unsigned parameter,
220 LLVMValueRef llvm_chan, LLVMValueRef attr_number,
221 LLVMValueRef params);
222
223 LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index);
224
225 void ac_build_indexed_store(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index,
226 LLVMValueRef value);
227
228 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
229 LLVMValueRef index);
230 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
231 LLVMValueRef index);
232
233 void ac_build_buffer_store_dword(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
234 LLVMValueRef vindex, LLVMValueRef voffset, LLVMValueRef soffset,
235 enum gl_access_qualifier access);
236
237 void ac_build_buffer_store_format(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef data,
238 LLVMValueRef vindex, LLVMValueRef voffset, enum gl_access_qualifier access);
239
240 LLVMValueRef ac_build_buffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc, int num_channels,
241 LLVMValueRef vindex, LLVMValueRef voffset, LLVMValueRef soffset,
242 LLVMTypeRef channel_type, enum gl_access_qualifier access,
243 bool can_speculate, bool allow_smem);
244
245 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
246 LLVMValueRef vindex, LLVMValueRef voffset,
247 unsigned num_channels, enum gl_access_qualifier access,
248 bool can_speculate, bool d16, bool tfe);
249
250 LLVMValueRef ac_build_buffer_load_short(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
251 LLVMValueRef voffset, LLVMValueRef soffset,
252 enum gl_access_qualifier access);
253
254 LLVMValueRef ac_build_buffer_load_byte(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
255 LLVMValueRef voffset, LLVMValueRef soffset,
256 enum gl_access_qualifier access);
257
258 LLVMValueRef ac_build_safe_tbuffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
259 LLVMValueRef vindex, LLVMValueRef voffset,
260 LLVMValueRef soffset,
261 const enum pipe_format format,
262 unsigned channel_bit_size,
263 unsigned const_offset,
264 unsigned align_offset,
265 unsigned align_mul,
266 unsigned num_channels,
267 enum gl_access_qualifier access,
268 bool can_speculate);
269
270 void ac_build_buffer_store_short(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
271 LLVMValueRef vdata, LLVMValueRef voffset, LLVMValueRef soffset,
272 enum gl_access_qualifier access);
273
274 void ac_build_buffer_store_byte(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
275 LLVMValueRef voffset, LLVMValueRef soffset, enum gl_access_qualifier access);
276
277 void ac_set_range_metadata(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned lo,
278 unsigned hi);
279 LLVMValueRef ac_get_thread_id(struct ac_llvm_context *ctx);
280
281 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
282 #define AC_TID_MASK_TOP 0xfffffffd
283 #define AC_TID_MASK_LEFT 0xfffffffe
284
285 LLVMValueRef ac_build_ddxy(struct ac_llvm_context *ctx, uint32_t mask, int idx, LLVMValueRef val);
286
287 void ac_build_sendmsg(struct ac_llvm_context *ctx, uint32_t imm, LLVMValueRef m0_content);
288
289 LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx, LLVMValueRef arg, LLVMTypeRef dst_type);
290
291 LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx, LLVMValueRef arg, LLVMTypeRef dst_type,
292 bool rev);
293 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
294 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
295 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
296 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
297 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
298 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
299 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
300
301 struct ac_export_args {
302 LLVMValueRef out[4];
303 unsigned target;
304 unsigned enabled_channels;
305 bool compr;
306 bool done;
307 bool valid_mask;
308 };
309
310 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
311
312 void ac_build_export_null(struct ac_llvm_context *ctx, bool uses_discard);
313
314 enum ac_image_opcode
315 {
316 ac_image_sample,
317 ac_image_gather4,
318 ac_image_load,
319 ac_image_load_mip,
320 ac_image_store,
321 ac_image_store_mip,
322 ac_image_get_lod,
323 ac_image_get_resinfo,
324 ac_image_atomic,
325 ac_image_atomic_cmpswap,
326 };
327
328 enum ac_atomic_op
329 {
330 ac_atomic_swap,
331 ac_atomic_add,
332 ac_atomic_sub,
333 ac_atomic_smin,
334 ac_atomic_umin,
335 ac_atomic_smax,
336 ac_atomic_umax,
337 ac_atomic_and,
338 ac_atomic_or,
339 ac_atomic_xor,
340 ac_atomic_inc_wrap,
341 ac_atomic_dec_wrap,
342 ac_atomic_fmin,
343 ac_atomic_fmax,
344 };
345
346 struct ac_image_args {
347 enum ac_image_opcode opcode;
348 enum ac_atomic_op atomic; /* for the ac_image_atomic opcode */
349 enum ac_image_dim dim;
350 enum gl_access_qualifier access;
351 unsigned dmask : 4;
352 bool unorm : 1;
353 bool level_zero : 1;
354 bool d16 : 1; /* GFX8+: data and return values are 16-bit */
355 bool a16 : 1; /* GFX9+: address components except compare, offset and bias are 16-bit */
356 bool g16 : 1; /* GFX10+: derivatives are 16-bit; GFX<=9: must be equal to a16 */
357 bool tfe : 1;
358 unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
359
360 LLVMValueRef resource;
361 LLVMValueRef sampler;
362 LLVMValueRef data[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
363 LLVMValueRef offset;
364 LLVMValueRef bias;
365 LLVMValueRef compare;
366 LLVMValueRef derivs[6];
367 LLVMValueRef coords[4];
368 LLVMValueRef lod; // also used by ac_image_get_resinfo
369 LLVMValueRef min_lod;
370 };
371
372 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx, struct ac_image_args *a);
373 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
374 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
375 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
376 LLVMValueRef ac_build_cvt_pknorm_i16_f16(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
377 LLVMValueRef ac_build_cvt_pknorm_u16_f16(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
378 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx, LLVMValueRef args[2], unsigned bits,
379 bool hi);
380 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx, LLVMValueRef args[2], unsigned bits,
381 bool hi);
382 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
383 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
384 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input, LLVMValueRef offset,
385 LLVMValueRef width, bool is_signed);
386 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef s1,
387 LLVMValueRef s2);
388 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef s1,
389 LLVMValueRef s2);
390
391 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
392
393 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize);
394 LLVMValueRef ac_const_uint_vec(struct ac_llvm_context *ctx, LLVMTypeRef type, uint64_t value);
395 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0);
396 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src);
397 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
398
399 LLVMValueRef ac_build_fsat(struct ac_llvm_context *ctx, LLVMValueRef src,
400 LLVMTypeRef type);
401
402 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx, LLVMValueRef src0);
403
404 LLVMValueRef ac_build_sudot_4x8(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef s1,
405 LLVMValueRef s2, bool clamp, unsigned neg_lo);
406
407 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
408
409 void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
410
411 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx, LLVMTypeRef dst_type, LLVMValueRef src0);
412
413 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
414 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
415
416 void ac_build_bgnloop(struct ac_llvm_context *ctx, int lable_id);
417 void ac_build_break(struct ac_llvm_context *ctx);
418 void ac_build_continue(struct ac_llvm_context *ctx);
419 void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
420 void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
421 void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
422 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
423
424 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type, const char *name);
425
426 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned count);
427
428 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param, unsigned rshift,
429 unsigned bitwidth);
430
431 LLVMValueRef ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
432
433 LLVMValueRef ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
434
435 LLVMValueRef ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value,
436 LLVMValueRef lane);
437
438 LLVMValueRef ac_build_mbcnt_add(struct ac_llvm_context *ctx, LLVMValueRef mask, LLVMValueRef add_src);
439 LLVMValueRef ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask);
440
441 LLVMValueRef ac_build_wqm(struct ac_llvm_context *ctx, LLVMValueRef src);
442
443 LLVMValueRef ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
444
445 LLVMValueRef ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op);
446
447 LLVMValueRef ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op,
448 unsigned cluster_size);
449
450 LLVMValueRef ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned lane0,
451 unsigned lane1, unsigned lane2, unsigned lane3,
452 bool use_wqm);
453
454 LLVMValueRef ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
455
456 LLVMValueRef ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize);
457
458 LLVMValueRef ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize);
459
460 LLVMValueRef ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
461 unsigned bitsize);
462
463 LLVMValueRef ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
464
465 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMTypeRef fn_type, LLVMValueRef func,
466 LLVMValueRef *args, unsigned num_args);
467
468 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
469 LLVMValueRef ptr, LLVMValueRef val, const char *sync_scope);
470
471 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
472 LLVMValueRef cmp, LLVMValueRef val, const char *sync_scope);
473
474 void ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth, LLVMValueRef stencil,
475 LLVMValueRef samplemask, LLVMValueRef mrt0_alpha, bool is_last,
476 struct ac_export_args *args);
477
478 struct ac_ngg_prim {
479 unsigned num_vertices;
480 LLVMValueRef isnull;
481 LLVMValueRef index[3];
482 LLVMValueRef edgeflags;
483 LLVMValueRef passthrough;
484 };
485
486 LLVMTypeRef ac_arg_type_to_pointee_type(struct ac_llvm_context *ctx, enum ac_arg_type type);
487
ac_get_arg(struct ac_llvm_context * ctx,struct ac_arg arg)488 static inline LLVMValueRef ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
489 {
490 assert(arg.used);
491 if (arg.arg_index == ctx->ring_offsets_index)
492 return ctx->ring_offsets;
493 int offset = arg.arg_index > ctx->ring_offsets_index ? -1 : 0;
494 return LLVMGetParam(ctx->main_function.value, arg.arg_index + offset);
495 }
496
497 static inline struct ac_llvm_pointer
ac_get_ptr_arg(struct ac_llvm_context * ctx,const struct ac_shader_args * args,struct ac_arg arg)498 ac_get_ptr_arg(struct ac_llvm_context *ctx, const struct ac_shader_args *args, struct ac_arg arg)
499 {
500 struct ac_llvm_pointer ptr;
501 ptr.pointee_type = ac_arg_type_to_pointee_type(ctx, args->args[arg.arg_index].type);
502 ptr.value = ac_get_arg(ctx, arg);
503 return ptr;
504 }
505
506 enum ac_llvm_calling_convention
507 {
508 AC_LLVM_AMDGPU_VS = 87,
509 AC_LLVM_AMDGPU_GS = 88,
510 AC_LLVM_AMDGPU_PS = 89,
511 AC_LLVM_AMDGPU_CS = 90,
512 AC_LLVM_AMDGPU_HS = 93,
513 };
514
515 struct ac_llvm_pointer ac_build_main(const struct ac_shader_args *args, struct ac_llvm_context *ctx,
516 enum ac_llvm_calling_convention convention, const char *name,
517 LLVMTypeRef ret_type, LLVMModuleRef module);
518
519 LLVMValueRef ac_build_is_inf_or_nan(struct ac_llvm_context *ctx, LLVMValueRef a);
520
521 void ac_build_dual_src_blend_swizzle(struct ac_llvm_context *ctx,
522 struct ac_export_args *mrt0,
523 struct ac_export_args *mrt1);
524
525 #ifdef __cplusplus
526 }
527 #endif
528
529 #endif
530