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
;
94 struct ac_llvm_flow
*flow
;
96 unsigned flow_depth_max
;
98 unsigned range_md_kind
;
99 unsigned invariant_load_md_kind
;
100 unsigned uniform_md_kind
;
101 unsigned fpmath_md_kind
;
102 LLVMValueRef fpmath_md_2p5_ulp
;
103 LLVMValueRef empty_md
;
105 enum chip_class chip_class
;
106 enum radeon_family family
;
112 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
113 enum chip_class chip_class
, enum radeon_family family
);
116 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
);
119 ac_get_llvm_num_components(LLVMValueRef value
);
122 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
);
125 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
129 unsigned ac_get_type_size(LLVMTypeRef type
);
131 LLVMTypeRef
ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
);
132 LLVMValueRef
ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
133 LLVMValueRef
ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
134 LLVMTypeRef
ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
);
135 LLVMValueRef
ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
);
138 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
139 LLVMTypeRef return_type
, LLVMValueRef
*params
,
140 unsigned param_count
, unsigned attrib_mask
);
142 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
);
145 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
146 unsigned count_incoming
, LLVMValueRef
*values
,
147 LLVMBasicBlockRef
*blocks
);
149 void ac_build_s_barrier(struct ac_llvm_context
*ctx
);
150 void ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
151 LLVMValueRef
*pvgpr
);
153 LLVMValueRef
ac_build_shader_clock(struct ac_llvm_context
*ctx
);
155 LLVMValueRef
ac_build_ballot(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
156 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
159 LLVMValueRef
ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
161 LLVMValueRef
ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
163 LLVMValueRef
ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
166 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
167 unsigned value_count
, unsigned component
);
170 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
171 LLVMValueRef
*values
,
172 unsigned value_count
,
173 unsigned value_stride
,
177 ac_build_gather_values(struct ac_llvm_context
*ctx
,
178 LLVMValueRef
*values
,
179 unsigned value_count
);
180 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
182 unsigned num_channels
);
183 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
186 ac_build_fdiv(struct ac_llvm_context
*ctx
,
190 LLVMValueRef
ac_build_fast_udiv(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_nuw(struct ac_llvm_context
*ctx
,
198 LLVMValueRef multiplier
,
199 LLVMValueRef pre_shift
,
200 LLVMValueRef post_shift
,
201 LLVMValueRef increment
);
202 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
204 LLVMValueRef multiplier
,
205 LLVMValueRef post_shift
);
208 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
209 bool is_deriv
, bool is_array
, bool is_lod
,
210 LLVMValueRef
*coords_arg
,
211 LLVMValueRef
*derivs_arg
);
215 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
216 LLVMValueRef llvm_chan
,
217 LLVMValueRef attr_number
,
223 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
224 LLVMValueRef llvm_chan
,
225 LLVMValueRef attr_number
,
231 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
232 LLVMValueRef parameter
,
233 LLVMValueRef llvm_chan
,
234 LLVMValueRef attr_number
,
235 LLVMValueRef params
);
238 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
239 LLVMValueRef base_ptr
,
243 ac_build_gep0(struct ac_llvm_context
*ctx
,
244 LLVMValueRef base_ptr
,
246 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
250 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
251 LLVMValueRef base_ptr
, LLVMValueRef index
,
254 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
256 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
257 LLVMValueRef base_ptr
, LLVMValueRef index
);
258 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
259 LLVMValueRef base_ptr
, LLVMValueRef index
);
260 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
261 LLVMValueRef base_ptr
, LLVMValueRef index
);
264 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
267 unsigned num_channels
,
268 LLVMValueRef voffset
,
269 LLVMValueRef soffset
,
270 unsigned inst_offset
,
273 bool writeonly_memory
,
274 bool swizzle_enable_hint
);
277 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
281 LLVMValueRef voffset
,
282 unsigned num_channels
,
284 bool writeonly_memory
);
287 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
291 LLVMValueRef voffset
,
292 LLVMValueRef soffset
,
293 unsigned inst_offset
,
299 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
302 LLVMValueRef voffset
,
303 unsigned num_channels
,
307 /* load_format that handles the stride & element count better if idxen is
308 * disabled by LLVM. */
309 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
312 LLVMValueRef voffset
,
313 unsigned num_channels
,
318 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
320 LLVMValueRef voffset
,
321 LLVMValueRef soffset
,
322 LLVMValueRef immoffset
,
326 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
328 LLVMValueRef voffset
,
329 LLVMValueRef soffset
,
330 LLVMValueRef immoffset
,
334 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
337 LLVMValueRef voffset
,
338 LLVMValueRef soffset
,
339 LLVMValueRef immoffset
,
340 unsigned num_channels
,
348 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
350 LLVMValueRef voffset
,
351 LLVMValueRef soffset
,
352 LLVMValueRef immoffset
,
353 unsigned num_channels
,
361 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
364 LLVMValueRef voffset
,
365 LLVMValueRef soffset
,
367 bool writeonly_memory
);
370 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
373 LLVMValueRef voffset
,
374 LLVMValueRef soffset
,
376 bool writeonly_memory
);
379 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
383 LLVMValueRef voffset
,
384 LLVMValueRef soffset
,
385 LLVMValueRef immoffset
,
386 unsigned num_channels
,
391 bool writeonly_memory
);
394 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
397 LLVMValueRef voffset
,
398 LLVMValueRef soffset
,
399 LLVMValueRef immoffset
,
400 unsigned num_channels
,
405 bool writeonly_memory
);
408 ac_get_thread_id(struct ac_llvm_context
*ctx
);
410 #define AC_TID_MASK_TOP_LEFT 0xfffffffc
411 #define AC_TID_MASK_TOP 0xfffffffd
412 #define AC_TID_MASK_LEFT 0xfffffffe
415 ac_build_ddxy(struct ac_llvm_context
*ctx
,
420 #define AC_SENDMSG_GS 2
421 #define AC_SENDMSG_GS_DONE 3
423 #define AC_SENDMSG_GS_OP_NOP (0 << 4)
424 #define AC_SENDMSG_GS_OP_CUT (1 << 4)
425 #define AC_SENDMSG_GS_OP_EMIT (2 << 4)
426 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
428 void ac_build_sendmsg(struct ac_llvm_context
*ctx
,
430 LLVMValueRef wave_id
);
432 LLVMValueRef
ac_build_imsb(struct ac_llvm_context
*ctx
,
434 LLVMTypeRef dst_type
);
436 LLVMValueRef
ac_build_umsb(struct ac_llvm_context
*ctx
,
438 LLVMTypeRef dst_type
);
439 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
441 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
443 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
445 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
447 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
, LLVMValueRef b
);
448 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
, LLVMValueRef b
);
449 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
);
451 struct ac_export_args
{
454 unsigned enabled_channels
;
460 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
);
462 void ac_build_export_null(struct ac_llvm_context
*ctx
);
464 enum ac_image_opcode
{
472 ac_image_get_resinfo
,
474 ac_image_atomic_cmpswap
,
494 ac_image_cube
, // includes cube arrays
498 ac_image_2darraymsaa
,
501 /* These cache policy bits match the definitions used by the LLVM intrinsics. */
502 enum ac_image_cache_policy
{
507 struct ac_image_args
{
508 enum ac_image_opcode opcode
: 4;
509 enum ac_atomic_op atomic
: 4; /* for the ac_image_atomic opcode */
510 enum ac_image_dim dim
: 3;
512 unsigned cache_policy
: 2;
515 unsigned attributes
; /* additional call-site specific AC_FUNC_ATTRs */
517 LLVMValueRef resource
;
518 LLVMValueRef sampler
;
519 LLVMValueRef data
[2]; /* data[0] is source data (vector); data[1] is cmp for cmpswap */
522 LLVMValueRef compare
;
523 LLVMValueRef derivs
[6];
524 LLVMValueRef coords
[4];
525 LLVMValueRef lod
; // also used by ac_image_get_resinfo
528 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
529 struct ac_image_args
*a
);
530 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
531 LLVMValueRef args
[2]);
532 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
533 LLVMValueRef args
[2]);
534 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
535 LLVMValueRef args
[2]);
536 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
537 LLVMValueRef args
[2], unsigned bits
, bool hi
);
538 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
539 LLVMValueRef args
[2], unsigned bits
, bool hi
);
540 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
);
541 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
);
542 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
543 LLVMValueRef offset
, LLVMValueRef width
,
545 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
546 LLVMValueRef s1
, LLVMValueRef s2
);
547 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
548 LLVMValueRef s1
, LLVMValueRef s2
);
550 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
);
552 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
555 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
556 LLVMValueRef src1
, LLVMValueRef src2
,
559 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
562 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
565 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
);
567 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
570 void ac_optimize_vs_outputs(struct ac_llvm_context
*ac
,
571 LLVMValueRef main_fn
,
572 uint8_t *vs_output_param_offset
,
573 uint32_t num_outputs
,
574 uint8_t *num_param_exports
);
575 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
);
577 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ac
);
578 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
579 LLVMValueRef dw_addr
);
580 void ac_lds_store(struct ac_llvm_context
*ctx
,
581 LLVMValueRef dw_addr
, LLVMValueRef value
);
583 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
584 LLVMTypeRef dst_type
,
587 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
);
588 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
);
590 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int lable_id
);
591 void ac_build_break(struct ac_llvm_context
*ctx
);
592 void ac_build_continue(struct ac_llvm_context
*ctx
);
593 void ac_build_else(struct ac_llvm_context
*ctx
, int lable_id
);
594 void ac_build_endif(struct ac_llvm_context
*ctx
, int lable_id
);
595 void ac_build_endloop(struct ac_llvm_context
*ctx
, int lable_id
);
596 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
);
597 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
599 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
602 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
604 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
607 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
610 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
613 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
614 unsigned rshift
, unsigned bitwidth
);
616 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
617 LLVMValueRef
*addr
, bool is_array_tex
);
620 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
);
623 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
);
626 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
);
629 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
);
632 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
);
635 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
);
638 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
);
641 * Common arguments for a scan/reduce operation that accumulates per-wave
642 * values across an entire workgroup, while respecting the order of waves.
646 bool enable_exclusive
;
647 bool enable_inclusive
;
649 LLVMValueRef src
; /* clobbered! */
650 LLVMValueRef result_reduce
;
651 LLVMValueRef result_exclusive
;
652 LLVMValueRef result_inclusive
;
654 LLVMValueRef waveidx
;
655 LLVMValueRef numwaves
; /* only needed for "reduce" operations */
657 /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
658 LLVMValueRef scratch
;
663 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
665 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
667 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
670 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
672 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
674 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
);
677 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
678 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
);
681 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
);
684 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
688 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
692 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
);
695 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
);