2 * Copyright 2016 Bas Nieuwenhuizen
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 #ifndef AC_LLVM_BUILD_H
26 #define AC_LLVM_BUILD_H
29 #include <llvm-c/TargetMachine.h>
30 #include "compiler/nir/nir.h"
31 #include "amd_family.h"
38 AC_ADDR_SPACE_FLAT
= 0, /* Slower than global. */
39 AC_ADDR_SPACE_GLOBAL
= 1,
40 AC_ADDR_SPACE_GDS
= 2,
41 AC_ADDR_SPACE_LDS
= 3,
42 AC_ADDR_SPACE_CONST
= 4, /* Global allowing SMEM. */
43 AC_ADDR_SPACE_CONST_32BIT
= 6, /* same as CONST, but the pointer type has 32 bits */
46 /* Combine these with & instead of |. */
47 #define NOOP_WAITCNT 0xcf7f
48 #define LGKM_CNT 0xc07f
49 #define EXP_CNT 0xcf0f
50 #define VM_CNT 0x0f70 /* On GFX9, vmcnt has 6 bits in [0:3] and [14:15] */
54 struct ac_llvm_context
{
55 LLVMContextRef context
;
57 LLVMBuilderRef builder
;
90 struct ac_llvm_flow
*flow
;
92 unsigned flow_depth_max
;
94 unsigned range_md_kind
;
95 unsigned invariant_load_md_kind
;
96 unsigned uniform_md_kind
;
97 unsigned fpmath_md_kind
;
98 LLVMValueRef fpmath_md_2p5_ulp
;
99 LLVMValueRef empty_md
;
101 enum chip_class chip_class
;
102 enum radeon_family family
;
108 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
109 enum chip_class chip_class
, enum radeon_family family
);
112 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
);
115 ac_get_llvm_num_components(LLVMValueRef value
);
118 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
);
121 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
125 unsigned ac_get_type_size(LLVMTypeRef type
);
127 LLVMTypeRef
ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
);
128 LLVMValueRef
ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
129 LLVMValueRef
ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
130 LLVMTypeRef
ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
);
131 LLVMValueRef
ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
134 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
135 LLVMTypeRef return_type
, LLVMValueRef
*params
,
136 unsigned param_count
, unsigned attrib_mask
);
138 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
);
141 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
142 unsigned count_incoming
, LLVMValueRef
*values
,
143 LLVMBasicBlockRef
*blocks
);
145 void ac_build_s_barrier(struct ac_llvm_context
*ctx
);
146 void ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
147 LLVMValueRef
*pvgpr
);
149 LLVMValueRef
ac_build_shader_clock(struct ac_llvm_context
*ctx
);
151 LLVMValueRef
ac_build_ballot(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
153 LLVMValueRef
ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
155 LLVMValueRef
ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
157 LLVMValueRef
ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
160 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
161 unsigned value_count
, unsigned component
);
164 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
165 LLVMValueRef
*values
,
166 unsigned value_count
,
167 unsigned value_stride
,
171 ac_build_gather_values(struct ac_llvm_context
*ctx
,
172 LLVMValueRef
*values
,
173 unsigned value_count
);
174 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
176 unsigned num_channels
);
177 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
180 ac_build_fdiv(struct ac_llvm_context
*ctx
,
184 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
186 LLVMValueRef multiplier
,
187 LLVMValueRef pre_shift
,
188 LLVMValueRef post_shift
,
189 LLVMValueRef increment
);
190 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
192 LLVMValueRef multiplier
,
193 LLVMValueRef pre_shift
,
194 LLVMValueRef post_shift
,
195 LLVMValueRef increment
);
196 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
198 LLVMValueRef multiplier
,
199 LLVMValueRef post_shift
);
202 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
203 bool is_deriv
, bool is_array
, bool is_lod
,
204 LLVMValueRef
*coords_arg
,
205 LLVMValueRef
*derivs_arg
);
209 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
210 LLVMValueRef llvm_chan
,
211 LLVMValueRef attr_number
,
217 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
218 LLVMValueRef llvm_chan
,
219 LLVMValueRef attr_number
,
225 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
226 LLVMValueRef parameter
,
227 LLVMValueRef llvm_chan
,
228 LLVMValueRef attr_number
,
229 LLVMValueRef params
);
232 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
233 LLVMValueRef base_ptr
,
237 ac_build_gep0(struct ac_llvm_context
*ctx
,
238 LLVMValueRef base_ptr
,
240 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
244 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
245 LLVMValueRef base_ptr
, LLVMValueRef index
,
248 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
250 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
251 LLVMValueRef base_ptr
, LLVMValueRef index
);
252 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
253 LLVMValueRef base_ptr
, LLVMValueRef index
);
254 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
255 LLVMValueRef base_ptr
, LLVMValueRef index
);
258 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
261 unsigned num_channels
,
262 LLVMValueRef voffset
,
263 LLVMValueRef soffset
,
264 unsigned inst_offset
,
267 bool writeonly_memory
,
268 bool swizzle_enable_hint
);
271 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
275 LLVMValueRef voffset
,
276 unsigned num_channels
,
278 bool writeonly_memory
);
281 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
285 LLVMValueRef voffset
,
286 LLVMValueRef soffset
,
287 unsigned inst_offset
,
293 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
296 LLVMValueRef voffset
,
297 unsigned num_channels
,
301 /* load_format that handles the stride & element count better if idxen is
302 * disabled by LLVM. */
303 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
306 LLVMValueRef voffset
,
307 unsigned num_channels
,
312 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
314 LLVMValueRef voffset
,
315 LLVMValueRef soffset
,
316 LLVMValueRef immoffset
,
320 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
323 LLVMValueRef voffset
,
324 LLVMValueRef soffset
,
325 LLVMValueRef immoffset
,
326 unsigned num_channels
,
334 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
336 LLVMValueRef voffset
,
337 LLVMValueRef soffset
,
338 LLVMValueRef immoffset
,
339 unsigned num_channels
,
347 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
351 LLVMValueRef voffset
,
352 LLVMValueRef soffset
,
353 LLVMValueRef immoffset
,
354 unsigned num_channels
,
359 bool writeonly_memory
);
362 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
365 LLVMValueRef voffset
,
366 LLVMValueRef soffset
,
367 LLVMValueRef immoffset
,
368 unsigned num_channels
,
373 bool writeonly_memory
);
376 ac_get_thread_id(struct ac_llvm_context
*ctx
);
378 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
379 #define AC_TID_MASK_TOP 0xfffffffd
380 #define AC_TID_MASK_LEFT 0xfffffffe
383 ac_build_ddxy(struct ac_llvm_context
*ctx
,
388 #define AC_SENDMSG_GS 2
389 #define AC_SENDMSG_GS_DONE 3
391 #define AC_SENDMSG_GS_OP_NOP (0 << 4)
392 #define AC_SENDMSG_GS_OP_CUT (1 << 4)
393 #define AC_SENDMSG_GS_OP_EMIT (2 << 4)
394 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
396 void ac_build_sendmsg(struct ac_llvm_context
*ctx
,
398 LLVMValueRef wave_id
);
400 LLVMValueRef
ac_build_imsb(struct ac_llvm_context
*ctx
,
402 LLVMTypeRef dst_type
);
404 LLVMValueRef
ac_build_umsb(struct ac_llvm_context
*ctx
,
406 LLVMTypeRef dst_type
);
407 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
409 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
411 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
413 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
415 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
, LLVMValueRef b
);
416 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
418 struct ac_export_args
{
421 unsigned enabled_channels
;
427 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
);
429 void ac_build_export_null(struct ac_llvm_context
*ctx
);
431 enum ac_image_opcode
{
439 ac_image_get_resinfo
,
441 ac_image_atomic_cmpswap
,
461 ac_image_cube
, // includes cube arrays
465 ac_image_2darraymsaa
,
468 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
469 enum ac_image_cache_policy
{
474 struct ac_image_args
{
475 enum ac_image_opcode opcode
: 4;
476 enum ac_atomic_op atomic
: 4; /* for the ac_image_atomic opcode */
477 enum ac_image_dim dim
: 3;
479 unsigned cache_policy
: 2;
482 unsigned attributes
; /* additional call-site specific AC_FUNC_ATTRs */
484 LLVMValueRef resource
;
485 LLVMValueRef sampler
;
486 LLVMValueRef data
[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
489 LLVMValueRef compare
;
490 LLVMValueRef derivs
[6];
491 LLVMValueRef coords
[4];
492 LLVMValueRef lod
; // also used by ac_image_get_resinfo
495 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
496 struct ac_image_args
*a
);
497 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
498 LLVMValueRef args
[2]);
499 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
500 LLVMValueRef args
[2]);
501 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
502 LLVMValueRef args
[2]);
503 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
504 LLVMValueRef args
[2], unsigned bits
, bool hi
);
505 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
506 LLVMValueRef args
[2], unsigned bits
, bool hi
);
507 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
);
508 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
);
509 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
510 LLVMValueRef offset
, LLVMValueRef width
,
512 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
513 LLVMValueRef s1
, LLVMValueRef s2
);
514 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
515 LLVMValueRef s1
, LLVMValueRef s2
);
517 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
);
519 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
522 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
525 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
528 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
);
530 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
533 void ac_optimize_vs_outputs(struct ac_llvm_context
*ac
,
534 LLVMValueRef main_fn
,
535 uint8_t *vs_output_param_offset
,
536 uint32_t num_outputs
,
537 uint8_t *num_param_exports
);
538 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
);
540 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ac
);
541 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
542 LLVMValueRef dw_addr
);
543 void ac_lds_store(struct ac_llvm_context
*ctx
,
544 LLVMValueRef dw_addr
, LLVMValueRef value
);
546 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
547 LLVMTypeRef dst_type
,
550 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
);
551 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
);
553 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int lable_id
);
554 void ac_build_break(struct ac_llvm_context
*ctx
);
555 void ac_build_continue(struct ac_llvm_context
*ctx
);
556 void ac_build_else(struct ac_llvm_context
*ctx
, int lable_id
);
557 void ac_build_endif(struct ac_llvm_context
*ctx
, int lable_id
);
558 void ac_build_endloop(struct ac_llvm_context
*ctx
, int lable_id
);
559 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
);
560 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
562 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
565 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
567 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
570 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
573 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
576 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
577 unsigned rshift
, unsigned bitwidth
);
579 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
580 LLVMValueRef
*addr
, bool is_array_tex
);
583 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
);
586 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
);
589 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
);
592 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
);
595 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
);
598 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
);
601 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
);
604 * Common arguments for a scan/reduce operation that accumulates per-wave
605 * values across an entire workgroup, while respecting the order of waves.
609 bool enable_exclusive
;
610 bool enable_inclusive
;
612 LLVMValueRef src
; /* clobbered! */
613 LLVMValueRef result_reduce
;
614 LLVMValueRef result_exclusive
;
615 LLVMValueRef result_inclusive
;
617 LLVMValueRef waveidx
;
618 LLVMValueRef numwaves
; /* only needed for "reduce" operations */
620 /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
621 LLVMValueRef scratch
;
626 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
628 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
630 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
633 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
635 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
637 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
640 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
641 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
);
644 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
);