• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
3  * Copyright (C) 2019-2020 Collabora, Ltd.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include <err.h>
26 #include <fcntl.h>
27 #include <stdint.h>
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <sys/mman.h>
31 #include <sys/stat.h>
32 #include <sys/types.h>
33 
34 #include "compiler/glsl_types.h"
35 #include "compiler/glsl/glsl_to_nir.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "util/half_float.h"
38 #include "util/list.h"
39 #include "util/u_debug.h"
40 #include "util/u_dynarray.h"
41 #include "util/u_math.h"
42 
43 #include "compiler.h"
44 #include "helpers.h"
45 #include "midgard.h"
46 #include "midgard_compile.h"
47 #include "midgard_nir.h"
48 #include "midgard_ops.h"
49 #include "midgard_quirks.h"
50 
51 #include "disassemble.h"
52 
53 static const struct debug_named_value midgard_debug_options[] = {
54    {"shaders", MIDGARD_DBG_SHADERS, "Dump shaders in NIR and MIR"},
55    {"shaderdb", MIDGARD_DBG_SHADERDB, "Prints shader-db statistics"},
56    {"inorder", MIDGARD_DBG_INORDER, "Disables out-of-order scheduling"},
57    {"verbose", MIDGARD_DBG_VERBOSE, "Dump shaders verbosely"},
58    {"internal", MIDGARD_DBG_INTERNAL, "Dump internal shaders"},
59    DEBUG_NAMED_VALUE_END};
60 
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug, "MIDGARD_MESA_DEBUG",
62                             midgard_debug_options, 0)
63 
64 int midgard_debug = 0;
65 
66 static midgard_block *
create_empty_block(compiler_context * ctx)67 create_empty_block(compiler_context *ctx)
68 {
69    midgard_block *blk = rzalloc(ctx, midgard_block);
70 
71    blk->base.predecessors =
72       _mesa_set_create(blk, _mesa_hash_pointer, _mesa_key_pointer_equal);
73 
74    blk->base.name = ctx->block_source_count++;
75 
76    return blk;
77 }
78 
79 static void
schedule_barrier(compiler_context * ctx)80 schedule_barrier(compiler_context *ctx)
81 {
82    midgard_block *temp = ctx->after_block;
83    ctx->after_block = create_empty_block(ctx);
84    ctx->block_count++;
85    list_addtail(&ctx->after_block->base.link, &ctx->blocks);
86    list_inithead(&ctx->after_block->base.instructions);
87    pan_block_add_successor(&ctx->current_block->base, &ctx->after_block->base);
88    ctx->current_block = ctx->after_block;
89    ctx->after_block = temp;
90 }
91 
92 /* Helpers to generate midgard_instruction's using macro magic, since every
93  * driver seems to do it that way */
94 
95 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
96 
97 #define M_LOAD_STORE(name, store, T)                                           \
98    static midgard_instruction m_##name(unsigned ssa, unsigned address)         \
99    {                                                                           \
100       midgard_instruction i = {                                                \
101          .type = TAG_LOAD_STORE_4,                                             \
102          .mask = 0xF,                                                          \
103          .dest = ~0,                                                           \
104          .src = {~0, ~0, ~0, ~0},                                              \
105          .swizzle = SWIZZLE_IDENTITY_4,                                        \
106          .op = midgard_op_##name,                                              \
107          .load_store =                                                         \
108             {                                                                  \
109                .signed_offset = address,                                       \
110             },                                                                 \
111       };                                                                       \
112                                                                                \
113       if (store) {                                                             \
114          i.src[0] = ssa;                                                       \
115          i.src_types[0] = T;                                                   \
116          i.dest_type = T;                                                      \
117       } else {                                                                 \
118          i.dest = ssa;                                                         \
119          i.dest_type = T;                                                      \
120       }                                                                        \
121       return i;                                                                \
122    }
123 
124 #define M_LOAD(name, T)  M_LOAD_STORE(name, false, T)
125 #define M_STORE(name, T) M_LOAD_STORE(name, true, T)
126 
127 M_LOAD(ld_attr_32, nir_type_uint32);
128 M_LOAD(ld_vary_32, nir_type_uint32);
129 M_LOAD(ld_ubo_u8, nir_type_uint32); /* mandatory extension to 32-bit */
130 M_LOAD(ld_ubo_u16, nir_type_uint32);
131 M_LOAD(ld_ubo_32, nir_type_uint32);
132 M_LOAD(ld_ubo_64, nir_type_uint32);
133 M_LOAD(ld_ubo_128, nir_type_uint32);
134 M_LOAD(ld_u8, nir_type_uint8);
135 M_LOAD(ld_u16, nir_type_uint16);
136 M_LOAD(ld_32, nir_type_uint32);
137 M_LOAD(ld_64, nir_type_uint32);
138 M_LOAD(ld_128, nir_type_uint32);
139 M_STORE(st_u8, nir_type_uint8);
140 M_STORE(st_u16, nir_type_uint16);
141 M_STORE(st_32, nir_type_uint32);
142 M_STORE(st_64, nir_type_uint32);
143 M_STORE(st_128, nir_type_uint32);
144 M_LOAD(ld_tilebuffer_raw, nir_type_uint32);
145 M_LOAD(ld_tilebuffer_16f, nir_type_float16);
146 M_LOAD(ld_tilebuffer_32f, nir_type_float32);
147 M_STORE(st_vary_32, nir_type_uint32);
148 M_LOAD(ld_cubemap_coords, nir_type_uint32);
149 M_LOAD(ldst_mov, nir_type_uint32);
150 M_LOAD(ld_image_32f, nir_type_float32);
151 M_LOAD(ld_image_16f, nir_type_float16);
152 M_LOAD(ld_image_32u, nir_type_uint32);
153 M_LOAD(ld_image_32i, nir_type_int32);
154 M_STORE(st_image_32f, nir_type_float32);
155 M_STORE(st_image_16f, nir_type_float16);
156 M_STORE(st_image_32u, nir_type_uint32);
157 M_STORE(st_image_32i, nir_type_int32);
158 M_LOAD(lea_image, nir_type_uint64);
159 
160 #define M_IMAGE(op)                                                            \
161    static midgard_instruction op##_image(nir_alu_type type, unsigned val,      \
162                                          unsigned address)                     \
163    {                                                                           \
164       switch (type) {                                                          \
165       case nir_type_float32:                                                   \
166          return m_##op##_image_32f(val, address);                              \
167       case nir_type_float16:                                                   \
168          return m_##op##_image_16f(val, address);                              \
169       case nir_type_uint32:                                                    \
170          return m_##op##_image_32u(val, address);                              \
171       case nir_type_int32:                                                     \
172          return m_##op##_image_32i(val, address);                              \
173       default:                                                                 \
174          unreachable("Invalid image type");                                    \
175       }                                                                        \
176    }
177 
178 M_IMAGE(ld);
179 M_IMAGE(st);
180 
181 static midgard_instruction
v_branch(bool conditional,bool invert)182 v_branch(bool conditional, bool invert)
183 {
184    midgard_instruction ins = {
185       .type = TAG_ALU_4,
186       .unit = ALU_ENAB_BRANCH,
187       .compact_branch = true,
188       .branch =
189          {
190             .conditional = conditional,
191             .invert_conditional = invert,
192          },
193       .dest = ~0,
194       .src = {~0, ~0, ~0, ~0},
195    };
196 
197    return ins;
198 }
199 
200 static void
attach_constants(compiler_context * ctx,midgard_instruction * ins,void * constants,int name)201 attach_constants(compiler_context *ctx, midgard_instruction *ins,
202                  void *constants, int name)
203 {
204    ins->has_constants = true;
205    memcpy(&ins->constants, constants, 16);
206 }
207 
208 static int
glsl_type_size(const struct glsl_type * type,bool bindless)209 glsl_type_size(const struct glsl_type *type, bool bindless)
210 {
211    return glsl_count_attribute_slots(type, false);
212 }
213 
214 static bool
midgard_nir_lower_global_load_instr(nir_builder * b,nir_intrinsic_instr * intr,void * data)215 midgard_nir_lower_global_load_instr(nir_builder *b, nir_intrinsic_instr *intr,
216                                     void *data)
217 {
218    if (intr->intrinsic != nir_intrinsic_load_global &&
219        intr->intrinsic != nir_intrinsic_load_shared)
220       return false;
221 
222    unsigned compsz = intr->def.bit_size;
223    unsigned totalsz = compsz * intr->def.num_components;
224    /* 8, 16, 32, 64 and 128 bit loads don't need to be lowered */
225    if (util_bitcount(totalsz) < 2 && totalsz <= 128)
226       return false;
227 
228    b->cursor = nir_before_instr(&intr->instr);
229 
230    nir_def *addr = intr->src[0].ssa;
231 
232    nir_def *comps[MIR_VEC_COMPONENTS];
233    unsigned ncomps = 0;
234 
235    while (totalsz) {
236       unsigned loadsz = MIN2(1 << (util_last_bit(totalsz) - 1), 128);
237       unsigned loadncomps = loadsz / compsz;
238 
239       nir_def *load;
240       if (intr->intrinsic == nir_intrinsic_load_global) {
241          load = nir_load_global(b, addr, compsz / 8, loadncomps, compsz);
242       } else {
243          assert(intr->intrinsic == nir_intrinsic_load_shared);
244          nir_intrinsic_instr *shared_load =
245             nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_shared);
246          shared_load->num_components = loadncomps;
247          shared_load->src[0] = nir_src_for_ssa(addr);
248          nir_intrinsic_set_align(shared_load, compsz / 8, 0);
249          nir_intrinsic_set_base(shared_load, nir_intrinsic_base(intr));
250          nir_def_init(&shared_load->instr, &shared_load->def,
251                       shared_load->num_components, compsz);
252          nir_builder_instr_insert(b, &shared_load->instr);
253          load = &shared_load->def;
254       }
255 
256       for (unsigned i = 0; i < loadncomps; i++)
257          comps[ncomps++] = nir_channel(b, load, i);
258 
259       totalsz -= loadsz;
260       addr = nir_iadd_imm(b, addr, loadsz / 8);
261    }
262 
263    assert(ncomps == intr->def.num_components);
264    nir_def_rewrite_uses(&intr->def, nir_vec(b, comps, ncomps));
265 
266    return true;
267 }
268 
269 static bool
midgard_nir_lower_global_load(nir_shader * shader)270 midgard_nir_lower_global_load(nir_shader *shader)
271 {
272    return nir_shader_intrinsics_pass(
273       shader, midgard_nir_lower_global_load_instr,
274       nir_metadata_block_index | nir_metadata_dominance, NULL);
275 }
276 
277 static bool
mdg_should_scalarize(const nir_instr * instr,const void * _unused)278 mdg_should_scalarize(const nir_instr *instr, const void *_unused)
279 {
280    const nir_alu_instr *alu = nir_instr_as_alu(instr);
281 
282    if (nir_src_bit_size(alu->src[0].src) == 64)
283       return true;
284 
285    if (alu->def.bit_size == 64)
286       return true;
287 
288    switch (alu->op) {
289    case nir_op_fdot2:
290    case nir_op_umul_high:
291    case nir_op_imul_high:
292    case nir_op_pack_half_2x16:
293    case nir_op_unpack_half_2x16:
294 
295    /* The LUT unit is scalar */
296    case nir_op_fsqrt:
297    case nir_op_frcp:
298    case nir_op_frsq:
299    case nir_op_fsin_mdg:
300    case nir_op_fcos_mdg:
301    case nir_op_fexp2:
302    case nir_op_flog2:
303       return true;
304    default:
305       return false;
306    }
307 }
308 
309 /* Only vectorize int64 up to vec2 */
310 static uint8_t
midgard_vectorize_filter(const nir_instr * instr,const void * data)311 midgard_vectorize_filter(const nir_instr *instr, const void *data)
312 {
313    if (instr->type != nir_instr_type_alu)
314       return 0;
315 
316    const nir_alu_instr *alu = nir_instr_as_alu(instr);
317    int src_bit_size = nir_src_bit_size(alu->src[0].src);
318    int dst_bit_size = alu->def.bit_size;
319 
320    if (src_bit_size == 64 || dst_bit_size == 64)
321       return 2;
322 
323    return 4;
324 }
325 
326 void
midgard_preprocess_nir(nir_shader * nir,unsigned gpu_id)327 midgard_preprocess_nir(nir_shader *nir, unsigned gpu_id)
328 {
329    unsigned quirks = midgard_get_quirks(gpu_id);
330 
331    /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
332     * (so we don't accidentally duplicate the epilogue since mesa/st has
333     * messed with our I/O quite a bit already).
334     */
335    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
336 
337    if (nir->info.stage == MESA_SHADER_VERTEX) {
338       NIR_PASS_V(nir, nir_lower_viewport_transform);
339       NIR_PASS_V(nir, nir_lower_point_size, 1.0, 0.0);
340    }
341 
342    NIR_PASS_V(nir, nir_lower_var_copies);
343    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
344    NIR_PASS_V(nir, nir_split_var_copies);
345    NIR_PASS_V(nir, nir_lower_var_copies);
346    NIR_PASS_V(nir, nir_lower_global_vars_to_local);
347    NIR_PASS_V(nir, nir_lower_var_copies);
348    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
349 
350    NIR_PASS_V(nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
351               glsl_type_size, 0);
352 
353    if (nir->info.stage == MESA_SHADER_VERTEX) {
354       /* nir_lower[_explicit]_io is lazy and emits mul+add chains even
355        * for offsets it could figure out are constant.  Do some
356        * constant folding before pan_nir_lower_store_component below.
357        */
358       NIR_PASS_V(nir, nir_opt_constant_folding);
359       NIR_PASS_V(nir, pan_nir_lower_store_component);
360    }
361 
362    NIR_PASS_V(nir, nir_lower_ssbo);
363    NIR_PASS_V(nir, pan_nir_lower_zs_store);
364 
365    NIR_PASS_V(nir, nir_lower_frexp);
366    NIR_PASS_V(nir, midgard_nir_lower_global_load);
367 
368    nir_lower_idiv_options idiv_options = {
369       .allow_fp16 = true,
370    };
371 
372    NIR_PASS_V(nir, nir_lower_idiv, &idiv_options);
373 
374    nir_lower_tex_options lower_tex_options = {
375       .lower_txs_lod = true,
376       .lower_txp = ~0,
377       .lower_tg4_broadcom_swizzle = true,
378       .lower_txd = true,
379       .lower_invalid_implicit_lod = true,
380    };
381 
382    NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);
383    NIR_PASS_V(nir, nir_lower_image_atomics_to_global);
384 
385    /* TEX_GRAD fails to apply sampler descriptor settings on some
386     * implementations, requiring a lowering.
387     */
388    if (quirks & MIDGARD_BROKEN_LOD)
389       NIR_PASS_V(nir, midgard_nir_lod_errata);
390 
391    /* lower MSAA image operations to 3D load before coordinate lowering */
392    NIR_PASS_V(nir, pan_nir_lower_image_ms);
393 
394    /* Midgard image ops coordinates are 16-bit instead of 32-bit */
395    NIR_PASS_V(nir, midgard_nir_lower_image_bitsize);
396 
397    if (nir->info.stage == MESA_SHADER_FRAGMENT)
398       NIR_PASS_V(nir, nir_lower_helper_writes, true);
399 
400    NIR_PASS_V(nir, pan_lower_helper_invocation);
401    NIR_PASS_V(nir, pan_lower_sample_pos);
402    NIR_PASS_V(nir, midgard_nir_lower_algebraic_early);
403    NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
404    NIR_PASS_V(nir, nir_lower_flrp, 16 | 32 | 64, false /* always_precise */);
405    NIR_PASS_V(nir, nir_lower_var_copies);
406 }
407 
408 static void
optimise_nir(nir_shader * nir,unsigned quirks,bool is_blend)409 optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
410 {
411    bool progress;
412 
413    do {
414       progress = false;
415 
416       NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
417 
418       NIR_PASS(progress, nir, nir_copy_prop);
419       NIR_PASS(progress, nir, nir_opt_remove_phis);
420       NIR_PASS(progress, nir, nir_opt_dce);
421       NIR_PASS(progress, nir, nir_opt_dead_cf);
422       NIR_PASS(progress, nir, nir_opt_cse);
423       NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
424       NIR_PASS(progress, nir, nir_opt_algebraic);
425       NIR_PASS(progress, nir, nir_opt_constant_folding);
426       NIR_PASS(progress, nir, nir_opt_undef);
427       NIR_PASS(progress, nir, nir_lower_undef_to_zero);
428 
429       NIR_PASS(progress, nir, nir_opt_loop_unroll);
430 
431       NIR_PASS(progress, nir, nir_opt_vectorize, midgard_vectorize_filter,
432                NULL);
433    } while (progress);
434 
435    NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
436 
437    /* Run after opts so it can hit more */
438    if (!is_blend)
439       NIR_PASS(progress, nir, nir_fuse_io_16);
440 
441    do {
442       progress = false;
443 
444       NIR_PASS(progress, nir, nir_opt_dce);
445       NIR_PASS(progress, nir, nir_opt_algebraic);
446       NIR_PASS(progress, nir, nir_opt_constant_folding);
447       NIR_PASS(progress, nir, nir_copy_prop);
448    } while (progress);
449 
450    NIR_PASS(progress, nir, nir_opt_algebraic_late);
451    NIR_PASS(progress, nir, nir_opt_algebraic_distribute_src_mods);
452 
453    /* We implement booleans as 32-bit 0/~0 */
454    NIR_PASS(progress, nir, nir_lower_bool_to_int32);
455 
456    /* Now that booleans are lowered, we can run out late opts */
457    NIR_PASS(progress, nir, midgard_nir_lower_algebraic_late);
458    NIR_PASS(progress, nir, midgard_nir_cancel_inot);
459    NIR_PASS_V(nir, midgard_nir_type_csel);
460 
461    /* Clean up after late opts */
462    do {
463       progress = false;
464 
465       NIR_PASS(progress, nir, nir_opt_dce);
466       NIR_PASS(progress, nir, nir_opt_constant_folding);
467       NIR_PASS(progress, nir, nir_copy_prop);
468    } while (progress);
469 
470    /* Backend scheduler is purely local, so do some global optimizations
471     * to reduce register pressure. */
472    nir_move_options move_all = nir_move_const_undef | nir_move_load_ubo |
473                                nir_move_load_input | nir_move_comparisons |
474                                nir_move_copies | nir_move_load_ssbo;
475 
476    NIR_PASS_V(nir, nir_opt_sink, move_all);
477    NIR_PASS_V(nir, nir_opt_move, move_all);
478 
479    /* Take us out of SSA */
480    NIR_PASS(progress, nir, nir_convert_from_ssa, true);
481 
482    /* We are a vector architecture; write combine where possible */
483    NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest, false);
484    NIR_PASS(progress, nir, nir_lower_vec_to_regs, NULL, NULL);
485 
486    NIR_PASS(progress, nir, nir_opt_dce);
487    NIR_PASS_V(nir, nir_trivialize_registers);
488 }
489 
490 /* Do not actually emit a load; instead, cache the constant for inlining */
491 
492 static void
emit_load_const(compiler_context * ctx,nir_load_const_instr * instr)493 emit_load_const(compiler_context *ctx, nir_load_const_instr *instr)
494 {
495    nir_def def = instr->def;
496 
497    midgard_constants *consts = rzalloc(ctx, midgard_constants);
498 
499    assert(instr->def.num_components * instr->def.bit_size <=
500           sizeof(*consts) * 8);
501 
502 #define RAW_CONST_COPY(bits)                                                   \
503    nir_const_value_to_array(consts->u##bits, instr->value,                     \
504                             instr->def.num_components, u##bits)
505 
506    switch (instr->def.bit_size) {
507    case 64:
508       RAW_CONST_COPY(64);
509       break;
510    case 32:
511       RAW_CONST_COPY(32);
512       break;
513    case 16:
514       RAW_CONST_COPY(16);
515       break;
516    case 8:
517       RAW_CONST_COPY(8);
518       break;
519    default:
520       unreachable("Invalid bit_size for load_const instruction\n");
521    }
522 
523    /* Shifted for SSA, +1 for off-by-one */
524    _mesa_hash_table_u64_insert(ctx->ssa_constants, (def.index << 1) + 1,
525                                consts);
526 }
527 
528 /* Normally constants are embedded implicitly, but for I/O and such we have to
529  * explicitly emit a move with the constant source */
530 
531 static void
emit_explicit_constant(compiler_context * ctx,unsigned node)532 emit_explicit_constant(compiler_context *ctx, unsigned node)
533 {
534    void *constant_value =
535       _mesa_hash_table_u64_search(ctx->ssa_constants, node + 1);
536 
537    if (constant_value) {
538       midgard_instruction ins =
539          v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), node);
540       attach_constants(ctx, &ins, constant_value, node + 1);
541       emit_mir_instruction(ctx, ins);
542    }
543 }
544 
545 static bool
nir_is_non_scalar_swizzle(nir_alu_src * src,unsigned nr_components)546 nir_is_non_scalar_swizzle(nir_alu_src *src, unsigned nr_components)
547 {
548    unsigned comp = src->swizzle[0];
549 
550    for (unsigned c = 1; c < nr_components; ++c) {
551       if (src->swizzle[c] != comp)
552          return true;
553    }
554 
555    return false;
556 }
557 
558 #define ALU_CASE(nir, _op)                                                     \
559    case nir_op_##nir:                                                          \
560       op = midgard_alu_op_##_op;                                               \
561       assert(src_bitsize == dst_bitsize);                                      \
562       break;
563 
564 #define ALU_CASE_RTZ(nir, _op)                                                 \
565    case nir_op_##nir:                                                          \
566       op = midgard_alu_op_##_op;                                               \
567       roundmode = MIDGARD_RTZ;                                                 \
568       break;
569 
570 #define ALU_CHECK_CMP()                                                        \
571    assert(src_bitsize == 16 || src_bitsize == 32 || src_bitsize == 64);        \
572    assert(dst_bitsize == 16 || dst_bitsize == 32);
573 
574 #define ALU_CASE_BCAST(nir, _op, count)                                        \
575    case nir_op_##nir:                                                          \
576       op = midgard_alu_op_##_op;                                               \
577       broadcast_swizzle = count;                                               \
578       ALU_CHECK_CMP();                                                         \
579       break;
580 
581 #define ALU_CASE_CMP(nir, _op)                                                 \
582    case nir_op_##nir:                                                          \
583       op = midgard_alu_op_##_op;                                               \
584       ALU_CHECK_CMP();                                                         \
585       break;
586 
587 static void
mir_copy_src(midgard_instruction * ins,nir_alu_instr * instr,unsigned i,unsigned to,bool is_int,unsigned bcast_count)588 mir_copy_src(midgard_instruction *ins, nir_alu_instr *instr, unsigned i,
589              unsigned to, bool is_int, unsigned bcast_count)
590 {
591    nir_alu_src src = instr->src[i];
592    unsigned bits = nir_src_bit_size(src.src);
593 
594    ins->src[to] = nir_src_index(NULL, &src.src);
595    ins->src_types[to] = nir_op_infos[instr->op].input_types[i] | bits;
596 
597    /* Figure out which component we should fill unused channels with. This
598     * doesn't matter too much in the non-broadcast case, but it makes
599     * should that scalar sources are packed with replicated swizzles,
600     * which works around issues seen with the combination of source
601     * expansion and destination shrinking.
602     */
603    unsigned replicate_c = 0;
604    if (bcast_count) {
605       replicate_c = bcast_count - 1;
606    } else {
607       for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c) {
608          if (nir_alu_instr_channel_used(instr, i, c))
609             replicate_c = c;
610       }
611    }
612 
613    for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c) {
614       ins->swizzle[to][c] =
615          src.swizzle[((!bcast_count || c < bcast_count) &&
616                       nir_alu_instr_channel_used(instr, i, c))
617                         ? c
618                         : replicate_c];
619    }
620 }
621 
622 static void
emit_alu(compiler_context * ctx,nir_alu_instr * instr)623 emit_alu(compiler_context *ctx, nir_alu_instr *instr)
624 {
625    /* Derivatives end up emitted on the texture pipe, not the ALUs. This
626     * is handled elsewhere */
627 
628    if (instr->op == nir_op_fddx || instr->op == nir_op_fddy) {
629       midgard_emit_derivatives(ctx, instr);
630       return;
631    }
632 
633    unsigned nr_components = instr->def.num_components;
634    unsigned nr_inputs = nir_op_infos[instr->op].num_inputs;
635    unsigned op = 0;
636 
637    /* Number of components valid to check for the instruction (the rest
638     * will be forced to the last), or 0 to use as-is. Relevant as
639     * ball-type instructions have a channel count in NIR but are all vec4
640     * in Midgard */
641 
642    unsigned broadcast_swizzle = 0;
643 
644    /* Should we swap arguments? */
645    bool flip_src12 = false;
646 
647    ASSERTED unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
648    unsigned dst_bitsize = instr->def.bit_size;
649 
650    enum midgard_roundmode roundmode = MIDGARD_RTE;
651 
652    switch (instr->op) {
653       ALU_CASE(fadd, fadd);
654       ALU_CASE(fmul, fmul);
655       ALU_CASE(fmin, fmin);
656       ALU_CASE(fmax, fmax);
657       ALU_CASE(imin, imin);
658       ALU_CASE(imax, imax);
659       ALU_CASE(umin, umin);
660       ALU_CASE(umax, umax);
661       ALU_CASE(ffloor, ffloor);
662       ALU_CASE(fround_even, froundeven);
663       ALU_CASE(ftrunc, ftrunc);
664       ALU_CASE(fceil, fceil);
665       ALU_CASE(fdot3, fdot3);
666       ALU_CASE(fdot4, fdot4);
667       ALU_CASE(iadd, iadd);
668       ALU_CASE(isub, isub);
669       ALU_CASE(iadd_sat, iaddsat);
670       ALU_CASE(isub_sat, isubsat);
671       ALU_CASE(uadd_sat, uaddsat);
672       ALU_CASE(usub_sat, usubsat);
673       ALU_CASE(imul, imul);
674       ALU_CASE(imul_high, imul);
675       ALU_CASE(umul_high, imul);
676       ALU_CASE(uclz, iclz);
677 
678       /* Zero shoved as second-arg */
679       ALU_CASE(iabs, iabsdiff);
680 
681       ALU_CASE(uabs_isub, iabsdiff);
682       ALU_CASE(uabs_usub, uabsdiff);
683 
684       ALU_CASE(mov, imov);
685 
686       ALU_CASE_CMP(feq32, feq);
687       ALU_CASE_CMP(fneu32, fne);
688       ALU_CASE_CMP(flt32, flt);
689       ALU_CASE_CMP(ieq32, ieq);
690       ALU_CASE_CMP(ine32, ine);
691       ALU_CASE_CMP(ilt32, ilt);
692       ALU_CASE_CMP(ult32, ult);
693 
694       /* We don't have a native b2f32 instruction. Instead, like many
695        * GPUs, we exploit booleans as 0/~0 for false/true, and
696        * correspondingly AND
697        * by 1.0 to do the type conversion. For the moment, prime us
698        * to emit:
699        *
700        * iand [whatever], #0
701        *
702        * At the end of emit_alu (as MIR), we'll fix-up the constant
703        */
704 
705       ALU_CASE_CMP(b2f32, iand);
706       ALU_CASE_CMP(b2f16, iand);
707       ALU_CASE_CMP(b2i32, iand);
708 
709       ALU_CASE(frcp, frcp);
710       ALU_CASE(frsq, frsqrt);
711       ALU_CASE(fsqrt, fsqrt);
712       ALU_CASE(fexp2, fexp2);
713       ALU_CASE(flog2, flog2);
714 
715       ALU_CASE_RTZ(f2i64, f2i_rte);
716       ALU_CASE_RTZ(f2u64, f2u_rte);
717       ALU_CASE_RTZ(i2f64, i2f_rte);
718       ALU_CASE_RTZ(u2f64, u2f_rte);
719 
720       ALU_CASE_RTZ(f2i32, f2i_rte);
721       ALU_CASE_RTZ(f2u32, f2u_rte);
722       ALU_CASE_RTZ(i2f32, i2f_rte);
723       ALU_CASE_RTZ(u2f32, u2f_rte);
724 
725       ALU_CASE_RTZ(f2i8, f2i_rte);
726       ALU_CASE_RTZ(f2u8, f2u_rte);
727 
728       ALU_CASE_RTZ(f2i16, f2i_rte);
729       ALU_CASE_RTZ(f2u16, f2u_rte);
730       ALU_CASE_RTZ(i2f16, i2f_rte);
731       ALU_CASE_RTZ(u2f16, u2f_rte);
732 
733       ALU_CASE(fsin_mdg, fsinpi);
734       ALU_CASE(fcos_mdg, fcospi);
735 
736       /* We'll get 0 in the second arg, so:
737        * ~a = ~(a | 0) = nor(a, 0) */
738       ALU_CASE(inot, inor);
739       ALU_CASE(iand, iand);
740       ALU_CASE(ior, ior);
741       ALU_CASE(ixor, ixor);
742       ALU_CASE(ishl, ishl);
743       ALU_CASE(ishr, iasr);
744       ALU_CASE(ushr, ilsr);
745 
746       ALU_CASE_BCAST(b32all_fequal2, fball_eq, 2);
747       ALU_CASE_BCAST(b32all_fequal3, fball_eq, 3);
748       ALU_CASE_CMP(b32all_fequal4, fball_eq);
749 
750       ALU_CASE_BCAST(b32any_fnequal2, fbany_neq, 2);
751       ALU_CASE_BCAST(b32any_fnequal3, fbany_neq, 3);
752       ALU_CASE_CMP(b32any_fnequal4, fbany_neq);
753 
754       ALU_CASE_BCAST(b32all_iequal2, iball_eq, 2);
755       ALU_CASE_BCAST(b32all_iequal3, iball_eq, 3);
756       ALU_CASE_CMP(b32all_iequal4, iball_eq);
757 
758       ALU_CASE_BCAST(b32any_inequal2, ibany_neq, 2);
759       ALU_CASE_BCAST(b32any_inequal3, ibany_neq, 3);
760       ALU_CASE_CMP(b32any_inequal4, ibany_neq);
761 
762       /* Source mods will be shoved in later */
763       ALU_CASE(fabs, fmov);
764       ALU_CASE(fneg, fmov);
765       ALU_CASE(fsat, fmov);
766       ALU_CASE(fsat_signed_mali, fmov);
767       ALU_CASE(fclamp_pos_mali, fmov);
768 
769       /* For size conversion, we use a move. Ideally though we would squash
770        * these ops together; maybe that has to happen after in NIR as part of
771        * propagation...? An earlier algebraic pass ensured we step down by
772        * only / exactly one size. If stepping down, we use a dest override to
773        * reduce the size; if stepping up, we use a larger-sized move with a
774        * half source and a sign/zero-extension modifier */
775 
776    case nir_op_i2i8:
777    case nir_op_i2i16:
778    case nir_op_i2i32:
779    case nir_op_i2i64:
780    case nir_op_u2u8:
781    case nir_op_u2u16:
782    case nir_op_u2u32:
783    case nir_op_u2u64:
784    case nir_op_f2f16:
785    case nir_op_f2f32:
786    case nir_op_f2f64: {
787       if (instr->op == nir_op_f2f16 || instr->op == nir_op_f2f32 ||
788           instr->op == nir_op_f2f64)
789          op = midgard_alu_op_fmov;
790       else
791          op = midgard_alu_op_imov;
792 
793       break;
794    }
795 
796       /* For greater-or-equal, we lower to less-or-equal and flip the
797        * arguments */
798 
799    case nir_op_fge:
800    case nir_op_fge32:
801    case nir_op_ige32:
802    case nir_op_uge32: {
803       op = instr->op == nir_op_fge     ? midgard_alu_op_fle
804            : instr->op == nir_op_fge32 ? midgard_alu_op_fle
805            : instr->op == nir_op_ige32 ? midgard_alu_op_ile
806            : instr->op == nir_op_uge32 ? midgard_alu_op_ule
807                                        : 0;
808 
809       flip_src12 = true;
810       ALU_CHECK_CMP();
811       break;
812    }
813 
814    case nir_op_b32csel:
815    case nir_op_b32fcsel_mdg: {
816       bool mixed = nir_is_non_scalar_swizzle(&instr->src[0], nr_components);
817       bool is_float = instr->op == nir_op_b32fcsel_mdg;
818       op = is_float ? (mixed ? midgard_alu_op_fcsel_v : midgard_alu_op_fcsel)
819                     : (mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel);
820 
821       int index = nir_src_index(ctx, &instr->src[0].src);
822       emit_explicit_constant(ctx, index);
823 
824       break;
825    }
826 
827    case nir_op_unpack_32_2x16:
828    case nir_op_unpack_32_4x8:
829    case nir_op_pack_32_2x16:
830    case nir_op_pack_32_4x8: {
831       op = midgard_alu_op_imov;
832       break;
833    }
834 
835    default:
836       mesa_loge("Unhandled ALU op %s\n", nir_op_infos[instr->op].name);
837       assert(0);
838       return;
839    }
840 
841    /* Promote imov to fmov if it might help inline a constant */
842    if (op == midgard_alu_op_imov && nir_src_is_const(instr->src[0].src) &&
843        nir_src_bit_size(instr->src[0].src) == 32 &&
844        nir_is_same_comp_swizzle(instr->src[0].swizzle,
845                                 nir_src_num_components(instr->src[0].src))) {
846       op = midgard_alu_op_fmov;
847    }
848 
849    /* Midgard can perform certain modifiers on output of an ALU op */
850 
851    unsigned outmod = 0;
852    bool is_int = midgard_is_integer_op(op);
853 
854    if (instr->op == nir_op_umul_high || instr->op == nir_op_imul_high) {
855       outmod = midgard_outmod_keephi;
856    } else if (midgard_is_integer_out_op(op)) {
857       outmod = midgard_outmod_keeplo;
858    } else if (instr->op == nir_op_fsat) {
859       outmod = midgard_outmod_clamp_0_1;
860    } else if (instr->op == nir_op_fsat_signed_mali) {
861       outmod = midgard_outmod_clamp_m1_1;
862    } else if (instr->op == nir_op_fclamp_pos_mali) {
863       outmod = midgard_outmod_clamp_0_inf;
864    }
865 
866    /* Fetch unit, quirks, etc information */
867    unsigned opcode_props = alu_opcode_props[op].props;
868    bool quirk_flipped_r24 = opcode_props & QUIRK_FLIPPED_R24;
869 
870    midgard_instruction ins = {
871       .type = TAG_ALU_4,
872       .dest_type = nir_op_infos[instr->op].output_type | dst_bitsize,
873       .roundmode = roundmode,
874    };
875 
876    ins.dest = nir_def_index_with_mask(&instr->def, &ins.mask);
877 
878    for (unsigned i = nr_inputs; i < ARRAY_SIZE(ins.src); ++i)
879       ins.src[i] = ~0;
880 
881    if (quirk_flipped_r24) {
882       ins.src[0] = ~0;
883       mir_copy_src(&ins, instr, 0, 1, is_int, broadcast_swizzle);
884    } else {
885       for (unsigned i = 0; i < nr_inputs; ++i) {
886          unsigned to = i;
887 
888          if (instr->op == nir_op_b32csel || instr->op == nir_op_b32fcsel_mdg) {
889             /* The condition is the first argument; move
890              * the other arguments up one to be a binary
891              * instruction for Midgard with the condition
892              * last */
893 
894             if (i == 0)
895                to = 2;
896             else if (flip_src12)
897                to = 2 - i;
898             else
899                to = i - 1;
900          } else if (flip_src12) {
901             to = 1 - to;
902          }
903 
904          mir_copy_src(&ins, instr, i, to, is_int, broadcast_swizzle);
905       }
906    }
907 
908    if (instr->op == nir_op_fneg || instr->op == nir_op_fabs) {
909       /* Lowered to move */
910       if (instr->op == nir_op_fneg)
911          ins.src_neg[1] ^= true;
912 
913       if (instr->op == nir_op_fabs)
914          ins.src_abs[1] = true;
915    }
916 
917    ins.op = op;
918    ins.outmod = outmod;
919 
920    /* Late fixup for emulated instructions */
921 
922    if (instr->op == nir_op_b2f32 || instr->op == nir_op_b2i32) {
923       /* Presently, our second argument is an inline #0 constant.
924        * Switch over to an embedded 1.0 constant (that can't fit
925        * inline, since we're 32-bit, not 16-bit like the inline
926        * constants) */
927 
928       ins.has_inline_constant = false;
929       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
930       ins.src_types[1] = nir_type_float32;
931       ins.has_constants = true;
932 
933       if (instr->op == nir_op_b2f32)
934          ins.constants.f32[0] = 1.0f;
935       else
936          ins.constants.i32[0] = 1;
937 
938       for (unsigned c = 0; c < 16; ++c)
939          ins.swizzle[1][c] = 0;
940    } else if (instr->op == nir_op_b2f16) {
941       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
942       ins.src_types[1] = nir_type_float16;
943       ins.has_constants = true;
944       ins.constants.i16[0] = _mesa_float_to_half(1.0);
945 
946       for (unsigned c = 0; c < 16; ++c)
947          ins.swizzle[1][c] = 0;
948    } else if (nr_inputs == 1 && !quirk_flipped_r24) {
949       /* Lots of instructions need a 0 plonked in */
950       ins.has_inline_constant = false;
951       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
952       ins.src_types[1] = ins.src_types[0];
953       ins.has_constants = true;
954       ins.constants.u32[0] = 0;
955 
956       for (unsigned c = 0; c < 16; ++c)
957          ins.swizzle[1][c] = 0;
958    } else if (instr->op == nir_op_pack_32_2x16) {
959       ins.dest_type = nir_type_uint16;
960       ins.mask = mask_of(nr_components * 2);
961       ins.is_pack = true;
962    } else if (instr->op == nir_op_pack_32_4x8) {
963       ins.dest_type = nir_type_uint8;
964       ins.mask = mask_of(nr_components * 4);
965       ins.is_pack = true;
966    } else if (instr->op == nir_op_unpack_32_2x16) {
967       ins.dest_type = nir_type_uint32;
968       ins.mask = mask_of(nr_components >> 1);
969       ins.is_pack = true;
970    } else if (instr->op == nir_op_unpack_32_4x8) {
971       ins.dest_type = nir_type_uint32;
972       ins.mask = mask_of(nr_components >> 2);
973       ins.is_pack = true;
974    }
975 
976    emit_mir_instruction(ctx, ins);
977 }
978 
979 #undef ALU_CASE
980 
981 static void
mir_set_intr_mask(nir_instr * instr,midgard_instruction * ins,bool is_read)982 mir_set_intr_mask(nir_instr *instr, midgard_instruction *ins, bool is_read)
983 {
984    nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
985    unsigned nir_mask = 0;
986    unsigned dsize = 0;
987 
988    if (is_read) {
989       nir_mask = mask_of(nir_intrinsic_dest_components(intr));
990 
991       /* Extension is mandatory for 8/16-bit loads */
992       dsize = intr->def.bit_size == 64 ? 64 : 32;
993    } else {
994       nir_mask = nir_intrinsic_write_mask(intr);
995       dsize = OP_IS_COMMON_STORE(ins->op) ? nir_src_bit_size(intr->src[0]) : 32;
996    }
997 
998    /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
999    unsigned bytemask = pan_to_bytemask(dsize, nir_mask);
1000    ins->dest_type = nir_type_uint | dsize;
1001    mir_set_bytemask(ins, bytemask);
1002 }
1003 
1004 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1005  * optimized) versions of UBO #0 */
1006 
1007 static midgard_instruction *
emit_ubo_read(compiler_context * ctx,nir_instr * instr,unsigned dest,unsigned offset,nir_src * indirect_offset,unsigned indirect_shift,unsigned index,unsigned nr_comps)1008 emit_ubo_read(compiler_context *ctx, nir_instr *instr, unsigned dest,
1009               unsigned offset, nir_src *indirect_offset,
1010               unsigned indirect_shift, unsigned index, unsigned nr_comps)
1011 {
1012    midgard_instruction ins;
1013 
1014    unsigned dest_size = (instr->type == nir_instr_type_intrinsic)
1015                            ? nir_instr_as_intrinsic(instr)->def.bit_size
1016                            : 32;
1017 
1018    unsigned bitsize = dest_size * nr_comps;
1019 
1020    /* Pick the smallest intrinsic to avoid out-of-bounds reads */
1021    if (bitsize <= 8)
1022       ins = m_ld_ubo_u8(dest, 0);
1023    else if (bitsize <= 16)
1024       ins = m_ld_ubo_u16(dest, 0);
1025    else if (bitsize <= 32)
1026       ins = m_ld_ubo_32(dest, 0);
1027    else if (bitsize <= 64)
1028       ins = m_ld_ubo_64(dest, 0);
1029    else if (bitsize <= 128)
1030       ins = m_ld_ubo_128(dest, 0);
1031    else
1032       unreachable("Invalid UBO read size");
1033 
1034    ins.constants.u32[0] = offset;
1035 
1036    if (instr->type == nir_instr_type_intrinsic)
1037       mir_set_intr_mask(instr, &ins, true);
1038 
1039    if (indirect_offset) {
1040       ins.src[2] = nir_src_index(ctx, indirect_offset);
1041       ins.src_types[2] = nir_type_uint32;
1042       ins.load_store.index_shift = indirect_shift;
1043 
1044       /* X component for the whole swizzle to prevent register
1045        * pressure from ballooning from the extra components */
1046       for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[2]); ++i)
1047          ins.swizzle[2][i] = 0;
1048    } else {
1049       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1050    }
1051 
1052    if (indirect_offset && !indirect_shift)
1053       mir_set_ubo_offset(&ins, indirect_offset, offset);
1054 
1055    midgard_pack_ubo_index_imm(&ins.load_store, index);
1056 
1057    return emit_mir_instruction(ctx, ins);
1058 }
1059 
1060 /* Globals are like UBOs if you squint. And shared memory is like globals if
1061  * you squint even harder */
1062 
1063 static void
emit_global(compiler_context * ctx,nir_instr * instr,bool is_read,unsigned srcdest,nir_src * offset,unsigned seg)1064 emit_global(compiler_context *ctx, nir_instr *instr, bool is_read,
1065             unsigned srcdest, nir_src *offset, unsigned seg)
1066 {
1067    midgard_instruction ins;
1068 
1069    nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1070    if (is_read) {
1071       unsigned bitsize = intr->def.bit_size * intr->def.num_components;
1072 
1073       switch (bitsize) {
1074       case 8:
1075          ins = m_ld_u8(srcdest, 0);
1076          break;
1077       case 16:
1078          ins = m_ld_u16(srcdest, 0);
1079          break;
1080       case 32:
1081          ins = m_ld_32(srcdest, 0);
1082          break;
1083       case 64:
1084          ins = m_ld_64(srcdest, 0);
1085          break;
1086       case 128:
1087          ins = m_ld_128(srcdest, 0);
1088          break;
1089       default:
1090          unreachable("Invalid global read size");
1091       }
1092 
1093       mir_set_intr_mask(instr, &ins, is_read);
1094 
1095       /* For anything not aligned on 32bit, make sure we write full
1096        * 32 bits registers. */
1097       if (bitsize & 31) {
1098          unsigned comps_per_32b = 32 / intr->def.bit_size;
1099 
1100          for (unsigned c = 0; c < 4 * comps_per_32b; c += comps_per_32b) {
1101             if (!(ins.mask & BITFIELD_RANGE(c, comps_per_32b)))
1102                continue;
1103 
1104             unsigned base = ~0;
1105             for (unsigned i = 0; i < comps_per_32b; i++) {
1106                if (ins.mask & BITFIELD_BIT(c + i)) {
1107                   base = ins.swizzle[0][c + i];
1108                   break;
1109                }
1110             }
1111 
1112             assert(base != ~0);
1113 
1114             for (unsigned i = 0; i < comps_per_32b; i++) {
1115                if (!(ins.mask & BITFIELD_BIT(c + i))) {
1116                   ins.swizzle[0][c + i] = base + i;
1117                   ins.mask |= BITFIELD_BIT(c + i);
1118                }
1119                assert(ins.swizzle[0][c + i] == base + i);
1120             }
1121          }
1122       }
1123    } else {
1124       unsigned bitsize =
1125          nir_src_bit_size(intr->src[0]) * nir_src_num_components(intr->src[0]);
1126 
1127       if (bitsize == 8)
1128          ins = m_st_u8(srcdest, 0);
1129       else if (bitsize == 16)
1130          ins = m_st_u16(srcdest, 0);
1131       else if (bitsize <= 32)
1132          ins = m_st_32(srcdest, 0);
1133       else if (bitsize <= 64)
1134          ins = m_st_64(srcdest, 0);
1135       else if (bitsize <= 128)
1136          ins = m_st_128(srcdest, 0);
1137       else
1138          unreachable("Invalid global store size");
1139 
1140       mir_set_intr_mask(instr, &ins, is_read);
1141    }
1142 
1143    mir_set_offset(ctx, &ins, offset, seg);
1144 
1145    /* Set a valid swizzle for masked out components */
1146    assert(ins.mask);
1147    unsigned first_component = __builtin_ffs(ins.mask) - 1;
1148 
1149    for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i) {
1150       if (!(ins.mask & (1 << i)))
1151          ins.swizzle[0][i] = first_component;
1152    }
1153 
1154    emit_mir_instruction(ctx, ins);
1155 }
1156 
1157 static midgard_load_store_op
translate_atomic_op(nir_atomic_op op)1158 translate_atomic_op(nir_atomic_op op)
1159 {
1160    /* clang-format off */
1161    switch (op) {
1162    case nir_atomic_op_xchg:    return midgard_op_atomic_xchg;
1163    case nir_atomic_op_cmpxchg: return midgard_op_atomic_cmpxchg;
1164    case nir_atomic_op_iadd:    return midgard_op_atomic_add;
1165    case nir_atomic_op_iand:    return midgard_op_atomic_and;
1166    case nir_atomic_op_imax:    return midgard_op_atomic_imax;
1167    case nir_atomic_op_imin:    return midgard_op_atomic_imin;
1168    case nir_atomic_op_ior:     return midgard_op_atomic_or;
1169    case nir_atomic_op_umax:    return midgard_op_atomic_umax;
1170    case nir_atomic_op_umin:    return midgard_op_atomic_umin;
1171    case nir_atomic_op_ixor:    return midgard_op_atomic_xor;
1172    default: unreachable("Unexpected atomic");
1173    }
1174    /* clang-format on */
1175 }
1176 
1177 /* Emit an atomic to shared memory or global memory. */
1178 static void
emit_atomic(compiler_context * ctx,nir_intrinsic_instr * instr)1179 emit_atomic(compiler_context *ctx, nir_intrinsic_instr *instr)
1180 {
1181    midgard_load_store_op op =
1182       translate_atomic_op(nir_intrinsic_atomic_op(instr));
1183 
1184    nir_alu_type type =
1185       (op == midgard_op_atomic_imin || op == midgard_op_atomic_imax)
1186          ? nir_type_int
1187          : nir_type_uint;
1188 
1189    bool is_shared = (instr->intrinsic == nir_intrinsic_shared_atomic) ||
1190                     (instr->intrinsic == nir_intrinsic_shared_atomic_swap);
1191 
1192    unsigned dest = nir_def_index(&instr->def);
1193    unsigned val = nir_src_index(ctx, &instr->src[1]);
1194    unsigned bitsize = nir_src_bit_size(instr->src[1]);
1195    emit_explicit_constant(ctx, val);
1196 
1197    midgard_instruction ins = {
1198       .type = TAG_LOAD_STORE_4,
1199       .mask = 0xF,
1200       .dest = dest,
1201       .src = {~0, ~0, ~0, val},
1202       .src_types = {0, 0, 0, type | bitsize},
1203       .op = op,
1204    };
1205 
1206    nir_src *src_offset = nir_get_io_offset_src(instr);
1207 
1208    if (op == midgard_op_atomic_cmpxchg) {
1209       unsigned xchg_val = nir_src_index(ctx, &instr->src[2]);
1210       emit_explicit_constant(ctx, xchg_val);
1211 
1212       ins.src[2] = val;
1213       ins.src_types[2] = type | bitsize;
1214       ins.src[3] = xchg_val;
1215 
1216       if (is_shared) {
1217          ins.load_store.arg_reg = REGISTER_LDST_LOCAL_STORAGE_PTR;
1218          ins.load_store.arg_comp = COMPONENT_Z;
1219          ins.load_store.bitsize_toggle = true;
1220       } else {
1221          for (unsigned i = 0; i < 2; ++i)
1222             ins.swizzle[1][i] = i;
1223 
1224          ins.src[1] = nir_src_index(ctx, src_offset);
1225          ins.src_types[1] = nir_type_uint64;
1226       }
1227    } else
1228       mir_set_offset(ctx, &ins, src_offset,
1229                      is_shared ? LDST_SHARED : LDST_GLOBAL);
1230 
1231    mir_set_intr_mask(&instr->instr, &ins, true);
1232 
1233    emit_mir_instruction(ctx, ins);
1234 }
1235 
1236 static void
emit_varying_read(compiler_context * ctx,unsigned dest,unsigned offset,unsigned nr_comp,unsigned component,nir_src * indirect_offset,nir_alu_type type,bool flat)1237 emit_varying_read(compiler_context *ctx, unsigned dest, unsigned offset,
1238                   unsigned nr_comp, unsigned component,
1239                   nir_src *indirect_offset, nir_alu_type type, bool flat)
1240 {
1241    midgard_instruction ins = m_ld_vary_32(dest, PACK_LDST_ATTRIB_OFS(offset));
1242    ins.mask = mask_of(nr_comp);
1243    ins.dest_type = type;
1244 
1245    if (type == nir_type_float16) {
1246       /* Ensure we are aligned so we can pack it later */
1247       ins.mask = mask_of(ALIGN_POT(nr_comp, 2));
1248    }
1249 
1250    for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i)
1251       ins.swizzle[0][i] = MIN2(i + component, COMPONENT_W);
1252 
1253    midgard_varying_params p = {
1254       .flat_shading = flat,
1255       .perspective_correction = 1,
1256       .interpolate_sample = true,
1257    };
1258    midgard_pack_varying_params(&ins.load_store, p);
1259 
1260    if (indirect_offset) {
1261       ins.src[2] = nir_src_index(ctx, indirect_offset);
1262       ins.src_types[2] = nir_type_uint32;
1263    } else
1264       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1265 
1266    ins.load_store.arg_reg = REGISTER_LDST_ZERO;
1267    ins.load_store.index_format = midgard_index_address_u32;
1268 
1269    /* For flat shading, we always use .u32 and require 32-bit mode. For
1270     * smooth shading, we use the appropriate floating-point type.
1271     *
1272     * This could be optimized, but it makes it easy to check correctness.
1273     */
1274    if (flat) {
1275       assert(nir_alu_type_get_type_size(type) == 32);
1276       ins.op = midgard_op_ld_vary_32u;
1277    } else {
1278       assert(nir_alu_type_get_base_type(type) == nir_type_float);
1279 
1280       ins.op = (nir_alu_type_get_type_size(type) == 32) ? midgard_op_ld_vary_32
1281                                                         : midgard_op_ld_vary_16;
1282    }
1283 
1284    emit_mir_instruction(ctx, ins);
1285 }
1286 
1287 static midgard_instruction
emit_image_op(compiler_context * ctx,nir_intrinsic_instr * instr)1288 emit_image_op(compiler_context *ctx, nir_intrinsic_instr *instr)
1289 {
1290    enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
1291    unsigned nr_attr = ctx->stage == MESA_SHADER_VERTEX
1292                          ? util_bitcount64(ctx->nir->info.inputs_read)
1293                          : 0;
1294    unsigned nr_dim = glsl_get_sampler_dim_coordinate_components(dim);
1295    bool is_array = nir_intrinsic_image_array(instr);
1296    bool is_store = instr->intrinsic == nir_intrinsic_image_store;
1297 
1298    assert(dim != GLSL_SAMPLER_DIM_MS && "MSAA'd image not lowered");
1299 
1300    unsigned coord_reg = nir_src_index(ctx, &instr->src[1]);
1301    emit_explicit_constant(ctx, coord_reg);
1302 
1303    nir_src *index = &instr->src[0];
1304    bool is_direct = nir_src_is_const(*index);
1305 
1306    /* For image opcodes, address is used as an index into the attribute
1307     * descriptor */
1308    unsigned address = nr_attr;
1309    if (is_direct)
1310       address += nir_src_as_uint(*index);
1311 
1312    midgard_instruction ins;
1313    if (is_store) { /* emit st_image_* */
1314       unsigned val = nir_src_index(ctx, &instr->src[3]);
1315       emit_explicit_constant(ctx, val);
1316 
1317       nir_alu_type type = nir_intrinsic_src_type(instr);
1318       ins = st_image(type, val, PACK_LDST_ATTRIB_OFS(address));
1319       nir_alu_type base_type = nir_alu_type_get_base_type(type);
1320       ins.src_types[0] = base_type | nir_src_bit_size(instr->src[3]);
1321    } else if (instr->intrinsic == nir_intrinsic_image_texel_address) {
1322       ins =
1323          m_lea_image(nir_def_index(&instr->def), PACK_LDST_ATTRIB_OFS(address));
1324       ins.mask = mask_of(2); /* 64-bit memory address */
1325    } else {                  /* emit ld_image_* */
1326       nir_alu_type type = nir_intrinsic_dest_type(instr);
1327       ins = ld_image(type, nir_def_index(&instr->def),
1328                      PACK_LDST_ATTRIB_OFS(address));
1329       ins.mask = mask_of(nir_intrinsic_dest_components(instr));
1330       ins.dest_type = type;
1331    }
1332 
1333    /* Coord reg */
1334    ins.src[1] = coord_reg;
1335    ins.src_types[1] = nir_type_uint16;
1336    if (nr_dim == 3 || is_array) {
1337       ins.load_store.bitsize_toggle = true;
1338    }
1339 
1340    /* Image index reg */
1341    if (!is_direct) {
1342       ins.src[2] = nir_src_index(ctx, index);
1343       ins.src_types[2] = nir_type_uint32;
1344    } else
1345       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1346 
1347    emit_mir_instruction(ctx, ins);
1348 
1349    return ins;
1350 }
1351 
1352 static void
emit_attr_read(compiler_context * ctx,unsigned dest,unsigned offset,unsigned nr_comp,nir_alu_type t)1353 emit_attr_read(compiler_context *ctx, unsigned dest, unsigned offset,
1354                unsigned nr_comp, nir_alu_type t)
1355 {
1356    midgard_instruction ins = m_ld_attr_32(dest, PACK_LDST_ATTRIB_OFS(offset));
1357    ins.load_store.arg_reg = REGISTER_LDST_ZERO;
1358    ins.load_store.index_reg = REGISTER_LDST_ZERO;
1359    ins.mask = mask_of(nr_comp);
1360 
1361    /* Use the type appropriate load */
1362    switch (t) {
1363    case nir_type_uint:
1364    case nir_type_bool:
1365       ins.op = midgard_op_ld_attr_32u;
1366       break;
1367    case nir_type_int:
1368       ins.op = midgard_op_ld_attr_32i;
1369       break;
1370    case nir_type_float:
1371       ins.op = midgard_op_ld_attr_32;
1372       break;
1373    default:
1374       unreachable("Attempted to load unknown type");
1375       break;
1376    }
1377 
1378    emit_mir_instruction(ctx, ins);
1379 }
1380 
1381 static unsigned
compute_builtin_arg(nir_intrinsic_op op)1382 compute_builtin_arg(nir_intrinsic_op op)
1383 {
1384    switch (op) {
1385    case nir_intrinsic_load_workgroup_id:
1386       return REGISTER_LDST_GROUP_ID;
1387    case nir_intrinsic_load_local_invocation_id:
1388       return REGISTER_LDST_LOCAL_THREAD_ID;
1389    case nir_intrinsic_load_global_invocation_id:
1390    case nir_intrinsic_load_global_invocation_id_zero_base:
1391       return REGISTER_LDST_GLOBAL_THREAD_ID;
1392    default:
1393       unreachable("Invalid compute paramater loaded");
1394    }
1395 }
1396 
1397 static void
emit_fragment_store(compiler_context * ctx,unsigned src,unsigned src_z,unsigned src_s,enum midgard_rt_id rt,unsigned sample_iter)1398 emit_fragment_store(compiler_context *ctx, unsigned src, unsigned src_z,
1399                     unsigned src_s, enum midgard_rt_id rt, unsigned sample_iter)
1400 {
1401    assert(rt < ARRAY_SIZE(ctx->writeout_branch));
1402    assert(sample_iter < ARRAY_SIZE(ctx->writeout_branch[0]));
1403 
1404    midgard_instruction *br = ctx->writeout_branch[rt][sample_iter];
1405 
1406    assert(!br);
1407 
1408    emit_explicit_constant(ctx, src);
1409 
1410    struct midgard_instruction ins = v_branch(false, false);
1411 
1412    bool depth_only = (rt == MIDGARD_ZS_RT);
1413 
1414    ins.writeout = depth_only ? 0 : PAN_WRITEOUT_C;
1415 
1416    /* Add dependencies */
1417    ins.src[0] = src;
1418    ins.src_types[0] = nir_type_uint32;
1419 
1420    if (depth_only)
1421       ins.constants.u32[0] = 0xFF;
1422    else
1423       ins.constants.u32[0] = ((rt - MIDGARD_COLOR_RT0) << 8) | sample_iter;
1424 
1425    for (int i = 0; i < 4; ++i)
1426       ins.swizzle[0][i] = i;
1427 
1428    if (~src_z) {
1429       emit_explicit_constant(ctx, src_z);
1430       ins.src[2] = src_z;
1431       ins.src_types[2] = nir_type_uint32;
1432       ins.writeout |= PAN_WRITEOUT_Z;
1433    }
1434    if (~src_s) {
1435       emit_explicit_constant(ctx, src_s);
1436       ins.src[3] = src_s;
1437       ins.src_types[3] = nir_type_uint32;
1438       ins.writeout |= PAN_WRITEOUT_S;
1439    }
1440 
1441    /* Emit the branch */
1442    br = emit_mir_instruction(ctx, ins);
1443    schedule_barrier(ctx);
1444    ctx->writeout_branch[rt][sample_iter] = br;
1445 
1446    /* Push our current location = current block count - 1 = where we'll
1447     * jump to. Maybe a bit too clever for my own good */
1448 
1449    br->branch.target_block = ctx->block_count - 1;
1450 }
1451 
1452 static void
emit_compute_builtin(compiler_context * ctx,nir_intrinsic_instr * instr)1453 emit_compute_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1454 {
1455    unsigned reg = nir_def_index(&instr->def);
1456    midgard_instruction ins = m_ldst_mov(reg, 0);
1457    ins.mask = mask_of(3);
1458    ins.swizzle[0][3] = COMPONENT_X; /* xyzx */
1459    ins.load_store.arg_reg = compute_builtin_arg(instr->intrinsic);
1460    emit_mir_instruction(ctx, ins);
1461 }
1462 
1463 static unsigned
vertex_builtin_arg(nir_intrinsic_op op)1464 vertex_builtin_arg(nir_intrinsic_op op)
1465 {
1466    switch (op) {
1467    case nir_intrinsic_load_vertex_id_zero_base:
1468       return PAN_VERTEX_ID;
1469    case nir_intrinsic_load_instance_id:
1470       return PAN_INSTANCE_ID;
1471    default:
1472       unreachable("Invalid vertex builtin");
1473    }
1474 }
1475 
1476 static void
emit_vertex_builtin(compiler_context * ctx,nir_intrinsic_instr * instr)1477 emit_vertex_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1478 {
1479    unsigned reg = nir_def_index(&instr->def);
1480    emit_attr_read(ctx, reg, vertex_builtin_arg(instr->intrinsic), 1,
1481                   nir_type_int);
1482 }
1483 
1484 static void
emit_special(compiler_context * ctx,nir_intrinsic_instr * instr,unsigned idx)1485 emit_special(compiler_context *ctx, nir_intrinsic_instr *instr, unsigned idx)
1486 {
1487    unsigned reg = nir_def_index(&instr->def);
1488 
1489    midgard_instruction ld = m_ld_tilebuffer_raw(reg, 0);
1490    ld.op = midgard_op_ld_special_32u;
1491    ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(idx);
1492    ld.load_store.index_reg = REGISTER_LDST_ZERO;
1493 
1494    for (int i = 0; i < 4; ++i)
1495       ld.swizzle[0][i] = COMPONENT_X;
1496 
1497    emit_mir_instruction(ctx, ld);
1498 }
1499 
1500 static void
emit_control_barrier(compiler_context * ctx)1501 emit_control_barrier(compiler_context *ctx)
1502 {
1503    midgard_instruction ins = {
1504       .type = TAG_TEXTURE_4,
1505       .dest = ~0,
1506       .src = {~0, ~0, ~0, ~0},
1507       .op = midgard_tex_op_barrier,
1508    };
1509 
1510    emit_mir_instruction(ctx, ins);
1511 }
1512 
1513 static uint8_t
output_load_rt_addr(compiler_context * ctx,nir_intrinsic_instr * instr)1514 output_load_rt_addr(compiler_context *ctx, nir_intrinsic_instr *instr)
1515 {
1516    unsigned loc = nir_intrinsic_io_semantics(instr).location;
1517 
1518    if (loc >= FRAG_RESULT_DATA0)
1519       return loc - FRAG_RESULT_DATA0;
1520 
1521    if (loc == FRAG_RESULT_DEPTH)
1522       return 0x1F;
1523    if (loc == FRAG_RESULT_STENCIL)
1524       return 0x1E;
1525 
1526    unreachable("Invalid RT to load from");
1527 }
1528 
1529 static void
emit_intrinsic(compiler_context * ctx,nir_intrinsic_instr * instr)1530 emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
1531 {
1532    unsigned offset = 0, reg;
1533 
1534    switch (instr->intrinsic) {
1535    case nir_intrinsic_decl_reg:
1536    case nir_intrinsic_store_reg:
1537       /* Always fully consumed */
1538       break;
1539 
1540    case nir_intrinsic_load_reg: {
1541       /* NIR guarantees that, for typical isel, this will always be fully
1542        * consumed. However, we also do our own nir_scalar chasing for
1543        * address arithmetic, bypassing the source chasing helpers. So we can end
1544        * up with unconsumed load_register instructions. Translate them here. 99%
1545        * of the time, these moves will be DCE'd away.
1546        */
1547       nir_def *handle = instr->src[0].ssa;
1548 
1549       midgard_instruction ins =
1550          v_mov(nir_reg_index(handle), nir_def_index(&instr->def));
1551 
1552       ins.dest_type = ins.src_types[1] = nir_type_uint | instr->def.bit_size;
1553 
1554       ins.mask = BITFIELD_MASK(instr->def.num_components);
1555       emit_mir_instruction(ctx, ins);
1556       break;
1557    }
1558 
1559    case nir_intrinsic_discard_if:
1560    case nir_intrinsic_discard: {
1561       bool conditional = instr->intrinsic == nir_intrinsic_discard_if;
1562       struct midgard_instruction discard = v_branch(conditional, false);
1563       discard.branch.target_type = TARGET_DISCARD;
1564 
1565       if (conditional) {
1566          discard.src[0] = nir_src_index(ctx, &instr->src[0]);
1567          discard.src_types[0] = nir_type_uint32;
1568       }
1569 
1570       emit_mir_instruction(ctx, discard);
1571       schedule_barrier(ctx);
1572 
1573       break;
1574    }
1575 
1576    case nir_intrinsic_image_load:
1577    case nir_intrinsic_image_store:
1578    case nir_intrinsic_image_texel_address:
1579       emit_image_op(ctx, instr);
1580       break;
1581 
1582    case nir_intrinsic_load_ubo:
1583    case nir_intrinsic_load_global:
1584    case nir_intrinsic_load_global_constant:
1585    case nir_intrinsic_load_shared:
1586    case nir_intrinsic_load_scratch:
1587    case nir_intrinsic_load_input:
1588    case nir_intrinsic_load_interpolated_input: {
1589       bool is_ubo = instr->intrinsic == nir_intrinsic_load_ubo;
1590       bool is_global = instr->intrinsic == nir_intrinsic_load_global ||
1591                        instr->intrinsic == nir_intrinsic_load_global_constant;
1592       bool is_shared = instr->intrinsic == nir_intrinsic_load_shared;
1593       bool is_scratch = instr->intrinsic == nir_intrinsic_load_scratch;
1594       bool is_flat = instr->intrinsic == nir_intrinsic_load_input;
1595       bool is_interp =
1596          instr->intrinsic == nir_intrinsic_load_interpolated_input;
1597 
1598       /* Get the base type of the intrinsic */
1599       /* TODO: Infer type? Does it matter? */
1600       nir_alu_type t = (is_interp) ? nir_type_float
1601                        : (is_flat) ? nir_intrinsic_dest_type(instr)
1602                                    : nir_type_uint;
1603 
1604       t = nir_alu_type_get_base_type(t);
1605 
1606       if (!(is_ubo || is_global || is_scratch)) {
1607          offset = nir_intrinsic_base(instr);
1608       }
1609 
1610       unsigned nr_comp = nir_intrinsic_dest_components(instr);
1611 
1612       nir_src *src_offset = nir_get_io_offset_src(instr);
1613 
1614       bool direct = nir_src_is_const(*src_offset);
1615       nir_src *indirect_offset = direct ? NULL : src_offset;
1616 
1617       if (direct)
1618          offset += nir_src_as_uint(*src_offset);
1619 
1620       /* We may need to apply a fractional offset */
1621       int component =
1622          (is_flat || is_interp) ? nir_intrinsic_component(instr) : 0;
1623       reg = nir_def_index(&instr->def);
1624 
1625       if (is_ubo) {
1626          nir_src index = instr->src[0];
1627 
1628          /* TODO: Is indirect block number possible? */
1629          assert(nir_src_is_const(index));
1630 
1631          uint32_t uindex = nir_src_as_uint(index);
1632          emit_ubo_read(ctx, &instr->instr, reg, offset, indirect_offset, 0,
1633                        uindex, nr_comp);
1634       } else if (is_global || is_shared || is_scratch) {
1635          unsigned seg =
1636             is_global ? LDST_GLOBAL : (is_shared ? LDST_SHARED : LDST_SCRATCH);
1637          emit_global(ctx, &instr->instr, true, reg, src_offset, seg);
1638       } else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->inputs->is_blend) {
1639          emit_varying_read(ctx, reg, offset, nr_comp, component,
1640                            indirect_offset, t | instr->def.bit_size, is_flat);
1641       } else if (ctx->inputs->is_blend) {
1642          /* ctx->blend_input will be precoloured to r0/r2, where
1643           * the input is preloaded */
1644 
1645          unsigned *input = offset ? &ctx->blend_src1 : &ctx->blend_input;
1646 
1647          if (*input == ~0)
1648             *input = reg;
1649          else
1650             emit_mir_instruction(ctx, v_mov(*input, reg));
1651       } else if (ctx->stage == MESA_SHADER_VERTEX) {
1652          emit_attr_read(ctx, reg, offset, nr_comp, t);
1653       } else {
1654          unreachable("Unknown load");
1655       }
1656 
1657       break;
1658    }
1659 
1660    /* Handled together with load_interpolated_input */
1661    case nir_intrinsic_load_barycentric_pixel:
1662    case nir_intrinsic_load_barycentric_centroid:
1663    case nir_intrinsic_load_barycentric_sample:
1664       break;
1665 
1666       /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1667 
1668    case nir_intrinsic_load_raw_output_pan: {
1669       reg = nir_def_index(&instr->def);
1670 
1671       /* T720 and below use different blend opcodes with slightly
1672        * different semantics than T760 and up */
1673 
1674       midgard_instruction ld = m_ld_tilebuffer_raw(reg, 0);
1675 
1676       unsigned target = output_load_rt_addr(ctx, instr);
1677       ld.load_store.index_comp = target & 0x3;
1678       ld.load_store.index_reg = target >> 2;
1679 
1680       if (nir_src_is_const(instr->src[0])) {
1681          unsigned sample = nir_src_as_uint(instr->src[0]);
1682          ld.load_store.arg_comp = sample & 0x3;
1683          ld.load_store.arg_reg = sample >> 2;
1684       } else {
1685          /* Enable sample index via register. */
1686          ld.load_store.signed_offset |= 1;
1687          ld.src[1] = nir_src_index(ctx, &instr->src[0]);
1688          ld.src_types[1] = nir_type_int32;
1689       }
1690 
1691       if (ctx->quirks & MIDGARD_OLD_BLEND) {
1692          ld.op = midgard_op_ld_special_32u;
1693          ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(16);
1694          ld.load_store.index_reg = REGISTER_LDST_ZERO;
1695       }
1696 
1697       emit_mir_instruction(ctx, ld);
1698       break;
1699    }
1700 
1701    case nir_intrinsic_load_output: {
1702       reg = nir_def_index(&instr->def);
1703 
1704       unsigned bits = instr->def.bit_size;
1705 
1706       midgard_instruction ld;
1707       if (bits == 16)
1708          ld = m_ld_tilebuffer_16f(reg, 0);
1709       else
1710          ld = m_ld_tilebuffer_32f(reg, 0);
1711 
1712       unsigned index = output_load_rt_addr(ctx, instr);
1713       ld.load_store.index_comp = index & 0x3;
1714       ld.load_store.index_reg = index >> 2;
1715 
1716       for (unsigned c = 4; c < 16; ++c)
1717          ld.swizzle[0][c] = 0;
1718 
1719       if (ctx->quirks & MIDGARD_OLD_BLEND) {
1720          if (bits == 16)
1721             ld.op = midgard_op_ld_special_16f;
1722          else
1723             ld.op = midgard_op_ld_special_32f;
1724          ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(1);
1725          ld.load_store.index_reg = REGISTER_LDST_ZERO;
1726       }
1727 
1728       emit_mir_instruction(ctx, ld);
1729       break;
1730    }
1731 
1732    case nir_intrinsic_store_output:
1733    case nir_intrinsic_store_combined_output_pan:
1734       assert(nir_src_is_const(instr->src[1]) && "no indirect outputs");
1735 
1736       reg = nir_src_index(ctx, &instr->src[0]);
1737 
1738       if (ctx->stage == MESA_SHADER_FRAGMENT) {
1739          bool combined =
1740             instr->intrinsic == nir_intrinsic_store_combined_output_pan;
1741 
1742          enum midgard_rt_id rt;
1743 
1744          unsigned reg_z = ~0, reg_s = ~0, reg_2 = ~0;
1745          unsigned writeout = PAN_WRITEOUT_C;
1746          if (combined) {
1747             writeout = nir_intrinsic_component(instr);
1748             if (writeout & PAN_WRITEOUT_Z)
1749                reg_z = nir_src_index(ctx, &instr->src[2]);
1750             if (writeout & PAN_WRITEOUT_S)
1751                reg_s = nir_src_index(ctx, &instr->src[3]);
1752             if (writeout & PAN_WRITEOUT_2)
1753                reg_2 = nir_src_index(ctx, &instr->src[4]);
1754          }
1755 
1756          if (writeout & PAN_WRITEOUT_C) {
1757             nir_io_semantics sem = nir_intrinsic_io_semantics(instr);
1758 
1759             rt = MIDGARD_COLOR_RT0 + (sem.location - FRAG_RESULT_DATA0);
1760          } else {
1761             rt = MIDGARD_ZS_RT;
1762          }
1763 
1764          /* Dual-source blend writeout is done by leaving the
1765           * value in r2 for the blend shader to use. */
1766          if (~reg_2) {
1767             emit_explicit_constant(ctx, reg_2);
1768 
1769             unsigned out = make_compiler_temp(ctx);
1770 
1771             midgard_instruction ins = v_mov(reg_2, out);
1772             emit_mir_instruction(ctx, ins);
1773 
1774             ctx->blend_src1 = out;
1775          }
1776 
1777          emit_fragment_store(ctx, reg, reg_z, reg_s, rt, 0);
1778       } else if (ctx->stage == MESA_SHADER_VERTEX) {
1779          assert(instr->intrinsic == nir_intrinsic_store_output);
1780 
1781          /* We should have been vectorized, though we don't
1782           * currently check that st_vary is emitted only once
1783           * per slot (this is relevant, since there's not a mask
1784           * parameter available on the store [set to 0 by the
1785           * blob]). We do respect the component by adjusting the
1786           * swizzle. If this is a constant source, we'll need to
1787           * emit that explicitly. */
1788 
1789          emit_explicit_constant(ctx, reg);
1790 
1791          offset = nir_intrinsic_base(instr) + nir_src_as_uint(instr->src[1]);
1792 
1793          unsigned dst_component = nir_intrinsic_component(instr);
1794          unsigned nr_comp = nir_src_num_components(instr->src[0]);
1795 
1796          /* ABI: Format controlled by the attribute descriptor.
1797           * This simplifies flat shading, although it prevents
1798           * certain (unimplemented) 16-bit optimizations.
1799           *
1800           * In particular, it lets the driver handle internal
1801           * TGSI shaders that set flat in the VS but smooth in
1802           * the FS. This matches our handling on Bifrost.
1803           */
1804          bool auto32 = true;
1805          assert(nir_alu_type_get_type_size(nir_intrinsic_src_type(instr)) ==
1806                 32);
1807 
1808          /* ABI: varyings in the secondary attribute table */
1809          bool secondary_table = true;
1810 
1811          midgard_instruction st =
1812             m_st_vary_32(reg, PACK_LDST_ATTRIB_OFS(offset));
1813          st.load_store.arg_reg = REGISTER_LDST_ZERO;
1814          st.load_store.index_reg = REGISTER_LDST_ZERO;
1815 
1816          /* Attribute instruction uses these 2-bits for the
1817           * a32 and table bits, pack this specially.
1818           */
1819          st.load_store.index_format =
1820             (auto32 ? (1 << 0) : 0) | (secondary_table ? (1 << 1) : 0);
1821 
1822          /* nir_intrinsic_component(store_intr) encodes the
1823           * destination component start. Source component offset
1824           * adjustment is taken care of in
1825           * install_registers_instr(), when offset_swizzle() is
1826           * called.
1827           */
1828          unsigned src_component = COMPONENT_X;
1829 
1830          assert(nr_comp > 0);
1831          for (unsigned i = 0; i < ARRAY_SIZE(st.swizzle); ++i) {
1832             st.swizzle[0][i] = src_component;
1833             if (i >= dst_component && i < dst_component + nr_comp - 1)
1834                src_component++;
1835          }
1836 
1837          emit_mir_instruction(ctx, st);
1838       } else {
1839          unreachable("Unknown store");
1840       }
1841 
1842       break;
1843 
1844    /* Special case of store_output for lowered blend shaders */
1845    case nir_intrinsic_store_raw_output_pan: {
1846       assert(ctx->stage == MESA_SHADER_FRAGMENT);
1847       reg = nir_src_index(ctx, &instr->src[0]);
1848 
1849       nir_io_semantics sem = nir_intrinsic_io_semantics(instr);
1850       assert(sem.location >= FRAG_RESULT_DATA0);
1851       unsigned rt = sem.location - FRAG_RESULT_DATA0;
1852 
1853       emit_fragment_store(ctx, reg, ~0, ~0, rt + MIDGARD_COLOR_RT0,
1854                           nir_intrinsic_base(instr));
1855       break;
1856    }
1857 
1858    case nir_intrinsic_store_global:
1859    case nir_intrinsic_store_shared:
1860    case nir_intrinsic_store_scratch:
1861       reg = nir_src_index(ctx, &instr->src[0]);
1862       emit_explicit_constant(ctx, reg);
1863 
1864       unsigned seg;
1865       if (instr->intrinsic == nir_intrinsic_store_global)
1866          seg = LDST_GLOBAL;
1867       else if (instr->intrinsic == nir_intrinsic_store_shared)
1868          seg = LDST_SHARED;
1869       else
1870          seg = LDST_SCRATCH;
1871 
1872       emit_global(ctx, &instr->instr, false, reg, &instr->src[1], seg);
1873       break;
1874 
1875    case nir_intrinsic_load_workgroup_id:
1876    case nir_intrinsic_load_local_invocation_id:
1877    case nir_intrinsic_load_global_invocation_id:
1878    case nir_intrinsic_load_global_invocation_id_zero_base:
1879       emit_compute_builtin(ctx, instr);
1880       break;
1881 
1882    case nir_intrinsic_load_vertex_id_zero_base:
1883    case nir_intrinsic_load_instance_id:
1884       emit_vertex_builtin(ctx, instr);
1885       break;
1886 
1887    case nir_intrinsic_load_sample_mask_in:
1888       emit_special(ctx, instr, 96);
1889       break;
1890 
1891    case nir_intrinsic_load_sample_id:
1892       emit_special(ctx, instr, 97);
1893       break;
1894 
1895    case nir_intrinsic_barrier:
1896       if (nir_intrinsic_execution_scope(instr) != SCOPE_NONE) {
1897          schedule_barrier(ctx);
1898          emit_control_barrier(ctx);
1899          schedule_barrier(ctx);
1900       } else if (nir_intrinsic_memory_scope(instr) != SCOPE_NONE) {
1901          /* Midgard doesn't seem to want special handling, though we do need to
1902           * take care when scheduling to avoid incorrect reordering.
1903           *
1904           * Note this is an "else if" since the handling for the execution scope
1905           * case already covers the case when both scopes are present.
1906           */
1907          schedule_barrier(ctx);
1908       }
1909       break;
1910 
1911    case nir_intrinsic_shared_atomic:
1912    case nir_intrinsic_shared_atomic_swap:
1913    case nir_intrinsic_global_atomic:
1914    case nir_intrinsic_global_atomic_swap:
1915       emit_atomic(ctx, instr);
1916       break;
1917 
1918    default:
1919       fprintf(stderr, "Unhandled intrinsic %s\n",
1920               nir_intrinsic_infos[instr->intrinsic].name);
1921       assert(0);
1922       break;
1923    }
1924 }
1925 
1926 /* Returns dimension with 0 special casing cubemaps */
1927 static unsigned
midgard_tex_format(enum glsl_sampler_dim dim)1928 midgard_tex_format(enum glsl_sampler_dim dim)
1929 {
1930    switch (dim) {
1931    case GLSL_SAMPLER_DIM_1D:
1932    case GLSL_SAMPLER_DIM_BUF:
1933       return 1;
1934 
1935    case GLSL_SAMPLER_DIM_2D:
1936    case GLSL_SAMPLER_DIM_MS:
1937    case GLSL_SAMPLER_DIM_EXTERNAL:
1938    case GLSL_SAMPLER_DIM_RECT:
1939       return 2;
1940 
1941    case GLSL_SAMPLER_DIM_3D:
1942       return 3;
1943 
1944    case GLSL_SAMPLER_DIM_CUBE:
1945       return 0;
1946 
1947    default:
1948       unreachable("Unknown sampler dim type");
1949    }
1950 }
1951 
1952 /* Tries to attach an explicit LOD or bias as a constant. Returns whether this
1953  * was successful */
1954 
1955 static bool
pan_attach_constant_bias(compiler_context * ctx,nir_src lod,midgard_texture_word * word)1956 pan_attach_constant_bias(compiler_context *ctx, nir_src lod,
1957                          midgard_texture_word *word)
1958 {
1959    /* To attach as constant, it has to *be* constant */
1960 
1961    if (!nir_src_is_const(lod))
1962       return false;
1963 
1964    float f = nir_src_as_float(lod);
1965 
1966    /* Break into fixed-point */
1967    signed lod_int = f;
1968    float lod_frac = f - lod_int;
1969 
1970    /* Carry over negative fractions */
1971    if (lod_frac < 0.0) {
1972       lod_int--;
1973       lod_frac += 1.0;
1974    }
1975 
1976    /* Encode */
1977    word->bias = float_to_ubyte(lod_frac);
1978    word->bias_int = lod_int;
1979 
1980    return true;
1981 }
1982 
1983 static enum mali_texture_mode
mdg_texture_mode(nir_tex_instr * instr)1984 mdg_texture_mode(nir_tex_instr *instr)
1985 {
1986    if (instr->op == nir_texop_tg4 && instr->is_shadow)
1987       return TEXTURE_GATHER_SHADOW;
1988    else if (instr->op == nir_texop_tg4)
1989       return TEXTURE_GATHER_X + instr->component;
1990    else if (instr->is_shadow)
1991       return TEXTURE_SHADOW;
1992    else
1993       return TEXTURE_NORMAL;
1994 }
1995 
1996 static void
set_tex_coord(compiler_context * ctx,nir_tex_instr * instr,midgard_instruction * ins)1997 set_tex_coord(compiler_context *ctx, nir_tex_instr *instr,
1998               midgard_instruction *ins)
1999 {
2000    int coord_idx = nir_tex_instr_src_index(instr, nir_tex_src_coord);
2001 
2002    assert(coord_idx >= 0);
2003 
2004    int comparator_idx = nir_tex_instr_src_index(instr, nir_tex_src_comparator);
2005    int ms_idx = nir_tex_instr_src_index(instr, nir_tex_src_ms_index);
2006    assert(comparator_idx < 0 || ms_idx < 0);
2007    int ms_or_comparator_idx = ms_idx >= 0 ? ms_idx : comparator_idx;
2008 
2009    unsigned coords = nir_src_index(ctx, &instr->src[coord_idx].src);
2010 
2011    emit_explicit_constant(ctx, coords);
2012 
2013    ins->src_types[1] = nir_tex_instr_src_type(instr, coord_idx) |
2014                        nir_src_bit_size(instr->src[coord_idx].src);
2015 
2016    unsigned nr_comps = instr->coord_components;
2017    unsigned written_mask = 0, write_mask = 0;
2018 
2019    /* Initialize all components to coord.x which is expected to always be
2020     * present. Swizzle is updated below based on the texture dimension
2021     * and extra attributes that are packed in the coordinate argument.
2022     */
2023    for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++)
2024       ins->swizzle[1][c] = COMPONENT_X;
2025 
2026    /* Shadow ref value is part of the coordinates if there's no comparator
2027     * source, in that case it's always placed in the last component.
2028     * Midgard wants the ref value in coord.z.
2029     */
2030    if (instr->is_shadow && comparator_idx < 0) {
2031       ins->swizzle[1][COMPONENT_Z] = --nr_comps;
2032       write_mask |= 1 << COMPONENT_Z;
2033    }
2034 
2035    /* The array index is the last component if there's no shadow ref value
2036     * or second last if there's one. We already decremented the number of
2037     * components to account for the shadow ref value above.
2038     * Midgard wants the array index in coord.w.
2039     */
2040    if (instr->is_array) {
2041       ins->swizzle[1][COMPONENT_W] = --nr_comps;
2042       write_mask |= 1 << COMPONENT_W;
2043    }
2044 
2045    if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
2046       /* texelFetch is undefined on samplerCube */
2047       assert(ins->op != midgard_tex_op_fetch);
2048 
2049       ins->src[1] = make_compiler_temp_reg(ctx);
2050 
2051       /* For cubemaps, we use a special ld/st op to select the face
2052        * and copy the xy into the texture register
2053        */
2054       midgard_instruction ld = m_ld_cubemap_coords(ins->src[1], 0);
2055       ld.src[1] = coords;
2056       ld.src_types[1] = ins->src_types[1];
2057       ld.mask = 0x3; /* xy */
2058       ld.load_store.bitsize_toggle = true;
2059       ld.swizzle[1][3] = COMPONENT_X;
2060       emit_mir_instruction(ctx, ld);
2061 
2062       /* We packed cube coordiates (X,Y,Z) into (X,Y), update the
2063        * written mask accordingly and decrement the number of
2064        * components
2065        */
2066       nr_comps--;
2067       written_mask |= 3;
2068    }
2069 
2070    /* Now flag tex coord components that have not been written yet */
2071    write_mask |= mask_of(nr_comps) & ~written_mask;
2072    for (unsigned c = 0; c < nr_comps; c++)
2073       ins->swizzle[1][c] = c;
2074 
2075    /* Sample index and shadow ref are expected in coord.z */
2076    if (ms_or_comparator_idx >= 0) {
2077       assert(!((write_mask | written_mask) & (1 << COMPONENT_Z)));
2078 
2079       unsigned sample_or_ref =
2080          nir_src_index(ctx, &instr->src[ms_or_comparator_idx].src);
2081 
2082       emit_explicit_constant(ctx, sample_or_ref);
2083 
2084       if (ins->src[1] == ~0)
2085          ins->src[1] = make_compiler_temp_reg(ctx);
2086 
2087       midgard_instruction mov = v_mov(sample_or_ref, ins->src[1]);
2088 
2089       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++)
2090          mov.swizzle[1][c] = COMPONENT_X;
2091 
2092       mov.mask = 1 << COMPONENT_Z;
2093       written_mask |= 1 << COMPONENT_Z;
2094       ins->swizzle[1][COMPONENT_Z] = COMPONENT_Z;
2095       emit_mir_instruction(ctx, mov);
2096    }
2097 
2098    /* Texelfetch coordinates uses all four elements (xyz/index) regardless
2099     * of texture dimensionality, which means it's necessary to zero the
2100     * unused components to keep everything happy.
2101     */
2102    if (ins->op == midgard_tex_op_fetch && (written_mask | write_mask) != 0xF) {
2103       if (ins->src[1] == ~0)
2104          ins->src[1] = make_compiler_temp_reg(ctx);
2105 
2106       /* mov index.zw, #0, or generalized */
2107       midgard_instruction mov =
2108          v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), ins->src[1]);
2109       mov.has_constants = true;
2110       mov.mask = (written_mask | write_mask) ^ 0xF;
2111       emit_mir_instruction(ctx, mov);
2112       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++) {
2113          if (mov.mask & (1 << c))
2114             ins->swizzle[1][c] = c;
2115       }
2116    }
2117 
2118    if (ins->src[1] == ~0) {
2119       /* No temporary reg created, use the src coords directly */
2120       ins->src[1] = coords;
2121    } else if (write_mask) {
2122       /* Move the remaining coordinates to the temporary reg */
2123       midgard_instruction mov = v_mov(coords, ins->src[1]);
2124 
2125       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++) {
2126          if ((1 << c) & write_mask) {
2127             mov.swizzle[1][c] = ins->swizzle[1][c];
2128             ins->swizzle[1][c] = c;
2129          } else {
2130             mov.swizzle[1][c] = COMPONENT_X;
2131          }
2132       }
2133 
2134       mov.mask = write_mask;
2135       emit_mir_instruction(ctx, mov);
2136    }
2137 }
2138 
2139 static void
emit_texop_native(compiler_context * ctx,nir_tex_instr * instr,unsigned midgard_texop)2140 emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
2141                   unsigned midgard_texop)
2142 {
2143    int texture_index = instr->texture_index;
2144    int sampler_index = instr->sampler_index;
2145 
2146    /* If txf is used, we assume there is a valid sampler bound at index 0. Use
2147     * it for txf operations, since there may be no other valid samplers. This is
2148     * a workaround: txf does not require a sampler in NIR (so sampler_index is
2149     * undefined) but we need one in the hardware. This is ABI with the driver.
2150     */
2151    if (!nir_tex_instr_need_sampler(instr))
2152       sampler_index = 0;
2153 
2154    midgard_instruction ins = {
2155       .type = TAG_TEXTURE_4,
2156       .mask = 0xF,
2157       .dest = nir_def_index(&instr->def),
2158       .src = {~0, ~0, ~0, ~0},
2159       .dest_type = instr->dest_type,
2160       .swizzle = SWIZZLE_IDENTITY_4,
2161       .outmod = midgard_outmod_none,
2162       .op = midgard_texop,
2163       .texture = {
2164          .format = midgard_tex_format(instr->sampler_dim),
2165          .texture_handle = texture_index,
2166          .sampler_handle = sampler_index,
2167          .mode = mdg_texture_mode(instr),
2168       }};
2169 
2170    if (instr->is_shadow && !instr->is_new_style_shadow &&
2171        instr->op != nir_texop_tg4)
2172       for (int i = 0; i < 4; ++i)
2173          ins.swizzle[0][i] = COMPONENT_X;
2174 
2175    for (unsigned i = 0; i < instr->num_srcs; ++i) {
2176       int index = nir_src_index(ctx, &instr->src[i].src);
2177       unsigned sz = nir_src_bit_size(instr->src[i].src);
2178       nir_alu_type T = nir_tex_instr_src_type(instr, i) | sz;
2179 
2180       switch (instr->src[i].src_type) {
2181       case nir_tex_src_coord:
2182          set_tex_coord(ctx, instr, &ins);
2183          break;
2184 
2185       case nir_tex_src_bias:
2186       case nir_tex_src_lod: {
2187          /* Try as a constant if we can */
2188 
2189          bool is_txf = midgard_texop == midgard_tex_op_fetch;
2190          if (!is_txf &&
2191              pan_attach_constant_bias(ctx, instr->src[i].src, &ins.texture))
2192             break;
2193 
2194          ins.texture.lod_register = true;
2195          ins.src[2] = index;
2196          ins.src_types[2] = T;
2197 
2198          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
2199             ins.swizzle[2][c] = COMPONENT_X;
2200 
2201          emit_explicit_constant(ctx, index);
2202 
2203          break;
2204       };
2205 
2206       case nir_tex_src_offset: {
2207          ins.texture.offset_register = true;
2208          ins.src[3] = index;
2209          ins.src_types[3] = T;
2210 
2211          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
2212             ins.swizzle[3][c] = (c > COMPONENT_Z) ? 0 : c;
2213 
2214          emit_explicit_constant(ctx, index);
2215          break;
2216       };
2217 
2218       case nir_tex_src_comparator:
2219       case nir_tex_src_ms_index:
2220          /* Nothing to do, handled in set_tex_coord() */
2221          break;
2222 
2223       default: {
2224          fprintf(stderr, "Unknown texture source type: %d\n",
2225                  instr->src[i].src_type);
2226          assert(0);
2227       }
2228       }
2229    }
2230 
2231    emit_mir_instruction(ctx, ins);
2232 }
2233 
2234 static void
emit_tex(compiler_context * ctx,nir_tex_instr * instr)2235 emit_tex(compiler_context *ctx, nir_tex_instr *instr)
2236 {
2237    switch (instr->op) {
2238    case nir_texop_tex:
2239    case nir_texop_txb:
2240       emit_texop_native(ctx, instr, midgard_tex_op_normal);
2241       break;
2242    case nir_texop_txl:
2243    case nir_texop_tg4:
2244       emit_texop_native(ctx, instr, midgard_tex_op_gradient);
2245       break;
2246    case nir_texop_txf:
2247    case nir_texop_txf_ms:
2248       emit_texop_native(ctx, instr, midgard_tex_op_fetch);
2249       break;
2250    default: {
2251       fprintf(stderr, "Unhandled texture op: %d\n", instr->op);
2252       assert(0);
2253    }
2254    }
2255 }
2256 
2257 static void
emit_jump(compiler_context * ctx,nir_jump_instr * instr)2258 emit_jump(compiler_context *ctx, nir_jump_instr *instr)
2259 {
2260    switch (instr->type) {
2261    case nir_jump_break: {
2262       /* Emit a branch out of the loop */
2263       struct midgard_instruction br = v_branch(false, false);
2264       br.branch.target_type = TARGET_BREAK;
2265       br.branch.target_break = ctx->current_loop_depth;
2266       emit_mir_instruction(ctx, br);
2267       break;
2268    }
2269 
2270    default:
2271       unreachable("Unhandled jump");
2272    }
2273 }
2274 
2275 static void
emit_instr(compiler_context * ctx,struct nir_instr * instr)2276 emit_instr(compiler_context *ctx, struct nir_instr *instr)
2277 {
2278    switch (instr->type) {
2279    case nir_instr_type_load_const:
2280       emit_load_const(ctx, nir_instr_as_load_const(instr));
2281       break;
2282 
2283    case nir_instr_type_intrinsic:
2284       emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
2285       break;
2286 
2287    case nir_instr_type_alu:
2288       emit_alu(ctx, nir_instr_as_alu(instr));
2289       break;
2290 
2291    case nir_instr_type_tex:
2292       emit_tex(ctx, nir_instr_as_tex(instr));
2293       break;
2294 
2295    case nir_instr_type_jump:
2296       emit_jump(ctx, nir_instr_as_jump(instr));
2297       break;
2298 
2299    case nir_instr_type_undef:
2300       /* Spurious */
2301       break;
2302 
2303    default:
2304       unreachable("Unhandled instruction type");
2305    }
2306 }
2307 
2308 /* ALU instructions can inline or embed constants, which decreases register
2309  * pressure and saves space. */
2310 
2311 #define CONDITIONAL_ATTACH(idx)                                                \
2312    {                                                                           \
2313       void *entry =                                                            \
2314          _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1);   \
2315                                                                                \
2316       if (entry) {                                                             \
2317          attach_constants(ctx, alu, entry, alu->src[idx] + 1);                 \
2318          alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);                \
2319       }                                                                        \
2320    }
2321 
2322 static void
inline_alu_constants(compiler_context * ctx,midgard_block * block)2323 inline_alu_constants(compiler_context *ctx, midgard_block *block)
2324 {
2325    mir_foreach_instr_in_block(block, alu) {
2326       /* Other instructions cannot inline constants */
2327       if (alu->type != TAG_ALU_4)
2328          continue;
2329       if (alu->compact_branch)
2330          continue;
2331 
2332       /* If there is already a constant here, we can do nothing */
2333       if (alu->has_constants)
2334          continue;
2335 
2336       CONDITIONAL_ATTACH(0);
2337 
2338       if (!alu->has_constants) {
2339          CONDITIONAL_ATTACH(1)
2340       } else if (!alu->inline_constant) {
2341          /* Corner case: _two_ vec4 constants, for instance with a
2342           * csel. For this case, we can only use a constant
2343           * register for one, we'll have to emit a move for the
2344           * other. */
2345 
2346          void *entry =
2347             _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[1] + 1);
2348          unsigned scratch = make_compiler_temp(ctx);
2349 
2350          if (entry) {
2351             midgard_instruction ins =
2352                v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), scratch);
2353             attach_constants(ctx, &ins, entry, alu->src[1] + 1);
2354 
2355             /* Set the source */
2356             alu->src[1] = scratch;
2357 
2358             /* Inject us -before- the last instruction which set r31, if
2359              * possible.
2360              */
2361             midgard_instruction *first = list_first_entry(
2362                &block->base.instructions, midgard_instruction, link);
2363 
2364             if (alu == first) {
2365                mir_insert_instruction_before(ctx, alu, ins);
2366             } else {
2367                mir_insert_instruction_before(ctx, mir_prev_op(alu), ins);
2368             }
2369          }
2370       }
2371    }
2372 }
2373 
2374 unsigned
max_bitsize_for_alu(midgard_instruction * ins)2375 max_bitsize_for_alu(midgard_instruction *ins)
2376 {
2377    unsigned max_bitsize = 0;
2378    for (int i = 0; i < MIR_SRC_COUNT; i++) {
2379       if (ins->src[i] == ~0)
2380          continue;
2381       unsigned src_bitsize = nir_alu_type_get_type_size(ins->src_types[i]);
2382       max_bitsize = MAX2(src_bitsize, max_bitsize);
2383    }
2384    unsigned dst_bitsize = nir_alu_type_get_type_size(ins->dest_type);
2385    max_bitsize = MAX2(dst_bitsize, max_bitsize);
2386 
2387    /* We emulate 8-bit as 16-bit for simplicity of packing */
2388    max_bitsize = MAX2(max_bitsize, 16);
2389 
2390    /* We don't have fp16 LUTs, so we'll want to emit code like:
2391     *
2392     *      vlut.fsinr hr0, hr0
2393     *
2394     * where both input and output are 16-bit but the operation is carried
2395     * out in 32-bit
2396     */
2397 
2398    switch (ins->op) {
2399    case midgard_alu_op_fsqrt:
2400    case midgard_alu_op_frcp:
2401    case midgard_alu_op_frsqrt:
2402    case midgard_alu_op_fsinpi:
2403    case midgard_alu_op_fcospi:
2404    case midgard_alu_op_fexp2:
2405    case midgard_alu_op_flog2:
2406       max_bitsize = MAX2(max_bitsize, 32);
2407       break;
2408 
2409    default:
2410       break;
2411    }
2412 
2413    /* High implies computing at a higher bitsize, e.g umul_high of 32-bit
2414     * requires computing at 64-bit */
2415    if (midgard_is_integer_out_op(ins->op) &&
2416        ins->outmod == midgard_outmod_keephi) {
2417       max_bitsize *= 2;
2418       assert(max_bitsize <= 64);
2419    }
2420 
2421    return max_bitsize;
2422 }
2423 
2424 midgard_reg_mode
reg_mode_for_bitsize(unsigned bitsize)2425 reg_mode_for_bitsize(unsigned bitsize)
2426 {
2427    switch (bitsize) {
2428       /* use 16 pipe for 8 since we don't support vec16 yet */
2429    case 8:
2430    case 16:
2431       return midgard_reg_mode_16;
2432    case 32:
2433       return midgard_reg_mode_32;
2434    case 64:
2435       return midgard_reg_mode_64;
2436    default:
2437       unreachable("invalid bit size");
2438    }
2439 }
2440 
2441 /* Midgard supports two types of constants, embedded constants (128-bit) and
2442  * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2443  * constants can be demoted to inline constants, for space savings and
2444  * sometimes a performance boost */
2445 
2446 static void
embedded_to_inline_constant(compiler_context * ctx,midgard_block * block)2447 embedded_to_inline_constant(compiler_context *ctx, midgard_block *block)
2448 {
2449    mir_foreach_instr_in_block(block, ins) {
2450       if (!ins->has_constants)
2451          continue;
2452       if (ins->has_inline_constant)
2453          continue;
2454 
2455       unsigned max_bitsize = max_bitsize_for_alu(ins);
2456 
2457       /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2458       bool is_16 = max_bitsize == 16;
2459       bool is_32 = max_bitsize == 32;
2460 
2461       if (!(is_16 || is_32))
2462          continue;
2463 
2464       /* src1 cannot be an inline constant due to encoding
2465        * restrictions. So, if possible we try to flip the arguments
2466        * in that case */
2467 
2468       int op = ins->op;
2469 
2470       if (ins->src[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT) &&
2471           alu_opcode_props[op].props & OP_COMMUTES) {
2472          mir_flip(ins);
2473       }
2474 
2475       if (ins->src[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
2476          /* Component is from the swizzle. Take a nonzero component */
2477          assert(ins->mask);
2478          unsigned first_comp = ffs(ins->mask) - 1;
2479          unsigned component = ins->swizzle[1][first_comp];
2480 
2481          /* Scale constant appropriately, if we can legally */
2482          int16_t scaled_constant = 0;
2483 
2484          if (is_16) {
2485             scaled_constant = ins->constants.u16[component];
2486          } else if (midgard_is_integer_op(op)) {
2487             scaled_constant = ins->constants.u32[component];
2488 
2489             /* Constant overflow after resize */
2490             if (scaled_constant != ins->constants.u32[component])
2491                continue;
2492          } else {
2493             float original = ins->constants.f32[component];
2494             scaled_constant = _mesa_float_to_half(original);
2495 
2496             /* Check for loss of precision. If this is
2497              * mediump, we don't care, but for a highp
2498              * shader, we need to pay attention. NIR
2499              * doesn't yet tell us which mode we're in!
2500              * Practically this prevents most constants
2501              * from being inlined, sadly. */
2502 
2503             float fp32 = _mesa_half_to_float(scaled_constant);
2504 
2505             if (fp32 != original)
2506                continue;
2507          }
2508 
2509          /* Should've been const folded */
2510          if (ins->src_abs[1] || ins->src_neg[1])
2511             continue;
2512 
2513          /* Make sure that the constant is not itself a vector
2514           * by checking if all accessed values are the same. */
2515 
2516          const midgard_constants *cons = &ins->constants;
2517          uint32_t value = is_16 ? cons->u16[component] : cons->u32[component];
2518 
2519          bool is_vector = false;
2520          unsigned mask = effective_writemask(ins->op, ins->mask);
2521 
2522          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c) {
2523             /* We only care if this component is actually used */
2524             if (!(mask & (1 << c)))
2525                continue;
2526 
2527             uint32_t test = is_16 ? cons->u16[ins->swizzle[1][c]]
2528                                   : cons->u32[ins->swizzle[1][c]];
2529 
2530             if (test != value) {
2531                is_vector = true;
2532                break;
2533             }
2534          }
2535 
2536          if (is_vector)
2537             continue;
2538 
2539          /* Get rid of the embedded constant */
2540          ins->has_constants = false;
2541          ins->src[1] = ~0;
2542          ins->has_inline_constant = true;
2543          ins->inline_constant = scaled_constant;
2544       }
2545    }
2546 }
2547 
2548 /* Dead code elimination for branches at the end of a block - only one branch
2549  * per block is legal semantically */
2550 
2551 static void
midgard_cull_dead_branch(compiler_context * ctx,midgard_block * block)2552 midgard_cull_dead_branch(compiler_context *ctx, midgard_block *block)
2553 {
2554    bool branched = false;
2555 
2556    mir_foreach_instr_in_block_safe(block, ins) {
2557       if (!midgard_is_branch_unit(ins->unit))
2558          continue;
2559 
2560       if (branched)
2561          mir_remove_instruction(ins);
2562 
2563       branched = true;
2564    }
2565 }
2566 
2567 /* We want to force the invert on AND/OR to the second slot to legalize into
2568  * iandnot/iornot. The relevant patterns are for AND (and OR respectively)
2569  *
2570  *   ~a & #b = ~a & ~(#~b)
2571  *   ~a & b = b & ~a
2572  */
2573 
2574 static void
midgard_legalize_invert(compiler_context * ctx,midgard_block * block)2575 midgard_legalize_invert(compiler_context *ctx, midgard_block *block)
2576 {
2577    mir_foreach_instr_in_block(block, ins) {
2578       if (ins->type != TAG_ALU_4)
2579          continue;
2580 
2581       if (ins->op != midgard_alu_op_iand && ins->op != midgard_alu_op_ior)
2582          continue;
2583 
2584       if (ins->src_invert[1] || !ins->src_invert[0])
2585          continue;
2586 
2587       if (ins->has_inline_constant) {
2588          /* ~(#~a) = ~(~#a) = a, so valid, and forces both
2589           * inverts on */
2590          ins->inline_constant = ~ins->inline_constant;
2591          ins->src_invert[1] = true;
2592       } else {
2593          /* Flip to the right invert order. Note
2594           * has_inline_constant false by assumption on the
2595           * branch, so flipping makes sense. */
2596          mir_flip(ins);
2597       }
2598    }
2599 }
2600 
2601 static unsigned
emit_fragment_epilogue(compiler_context * ctx,unsigned rt,unsigned sample_iter)2602 emit_fragment_epilogue(compiler_context *ctx, unsigned rt, unsigned sample_iter)
2603 {
2604    /* Loop to ourselves */
2605    midgard_instruction *br = ctx->writeout_branch[rt][sample_iter];
2606    struct midgard_instruction ins = v_branch(false, false);
2607    ins.writeout = br->writeout;
2608    ins.branch.target_block = ctx->block_count - 1;
2609    ins.constants.u32[0] = br->constants.u32[0];
2610    memcpy(&ins.src_types, &br->src_types, sizeof(ins.src_types));
2611    emit_mir_instruction(ctx, ins);
2612 
2613    ctx->current_block->epilogue = true;
2614    schedule_barrier(ctx);
2615    return ins.branch.target_block;
2616 }
2617 
2618 static midgard_block *
emit_block_init(compiler_context * ctx)2619 emit_block_init(compiler_context *ctx)
2620 {
2621    midgard_block *this_block = ctx->after_block;
2622    ctx->after_block = NULL;
2623 
2624    if (!this_block)
2625       this_block = create_empty_block(ctx);
2626 
2627    list_addtail(&this_block->base.link, &ctx->blocks);
2628 
2629    this_block->scheduled = false;
2630    ++ctx->block_count;
2631 
2632    /* Set up current block */
2633    list_inithead(&this_block->base.instructions);
2634    ctx->current_block = this_block;
2635 
2636    return this_block;
2637 }
2638 
2639 static midgard_block *
emit_block(compiler_context * ctx,nir_block * block)2640 emit_block(compiler_context *ctx, nir_block *block)
2641 {
2642    midgard_block *this_block = emit_block_init(ctx);
2643 
2644    nir_foreach_instr(instr, block) {
2645       emit_instr(ctx, instr);
2646       ++ctx->instruction_count;
2647    }
2648 
2649    return this_block;
2650 }
2651 
2652 static midgard_block *emit_cf_list(struct compiler_context *ctx,
2653                                    struct exec_list *list);
2654 
2655 static void
emit_if(struct compiler_context * ctx,nir_if * nif)2656 emit_if(struct compiler_context *ctx, nir_if *nif)
2657 {
2658    midgard_block *before_block = ctx->current_block;
2659 
2660    /* Speculatively emit the branch, but we can't fill it in until later */
2661    EMIT(branch, true, true);
2662    midgard_instruction *then_branch = mir_last_in_block(ctx->current_block);
2663    then_branch->src[0] = nir_src_index(ctx, &nif->condition);
2664    then_branch->src_types[0] = nir_type_uint32;
2665 
2666    /* Emit the two subblocks. */
2667    midgard_block *then_block = emit_cf_list(ctx, &nif->then_list);
2668    midgard_block *end_then_block = ctx->current_block;
2669 
2670    /* Emit a jump from the end of the then block to the end of the else */
2671    EMIT(branch, false, false);
2672    midgard_instruction *then_exit = mir_last_in_block(ctx->current_block);
2673 
2674    /* Emit second block, and check if it's empty */
2675 
2676    int else_idx = ctx->block_count;
2677    int count_in = ctx->instruction_count;
2678    midgard_block *else_block = emit_cf_list(ctx, &nif->else_list);
2679    midgard_block *end_else_block = ctx->current_block;
2680    int after_else_idx = ctx->block_count;
2681 
2682    /* Now that we have the subblocks emitted, fix up the branches */
2683 
2684    assert(then_block);
2685    assert(else_block);
2686 
2687    if (ctx->instruction_count == count_in) {
2688       /* The else block is empty, so don't emit an exit jump */
2689       mir_remove_instruction(then_exit);
2690       then_branch->branch.target_block = after_else_idx;
2691    } else {
2692       then_branch->branch.target_block = else_idx;
2693       then_exit->branch.target_block = after_else_idx;
2694    }
2695 
2696    /* Wire up the successors */
2697 
2698    ctx->after_block = create_empty_block(ctx);
2699 
2700    pan_block_add_successor(&before_block->base, &then_block->base);
2701    pan_block_add_successor(&before_block->base, &else_block->base);
2702 
2703    pan_block_add_successor(&end_then_block->base, &ctx->after_block->base);
2704    pan_block_add_successor(&end_else_block->base, &ctx->after_block->base);
2705 }
2706 
2707 static void
emit_loop(struct compiler_context * ctx,nir_loop * nloop)2708 emit_loop(struct compiler_context *ctx, nir_loop *nloop)
2709 {
2710    assert(!nir_loop_has_continue_construct(nloop));
2711 
2712    /* Remember where we are */
2713    midgard_block *start_block = ctx->current_block;
2714 
2715    /* Allocate a loop number, growing the current inner loop depth */
2716    int loop_idx = ++ctx->current_loop_depth;
2717 
2718    /* Get index from before the body so we can loop back later */
2719    int start_idx = ctx->block_count;
2720 
2721    /* Emit the body itself */
2722    midgard_block *loop_block = emit_cf_list(ctx, &nloop->body);
2723 
2724    /* Branch back to loop back */
2725    struct midgard_instruction br_back = v_branch(false, false);
2726    br_back.branch.target_block = start_idx;
2727    emit_mir_instruction(ctx, br_back);
2728 
2729    /* Mark down that branch in the graph. */
2730    pan_block_add_successor(&start_block->base, &loop_block->base);
2731    pan_block_add_successor(&ctx->current_block->base, &loop_block->base);
2732 
2733    /* Find the index of the block about to follow us (note: we don't add
2734     * one; blocks are 0-indexed so we get a fencepost problem) */
2735    int break_block_idx = ctx->block_count;
2736 
2737    /* Fix up the break statements we emitted to point to the right place,
2738     * now that we can allocate a block number for them */
2739    ctx->after_block = create_empty_block(ctx);
2740 
2741    mir_foreach_block_from(ctx, start_block, _block) {
2742       mir_foreach_instr_in_block(((midgard_block *)_block), ins) {
2743          if (ins->type != TAG_ALU_4)
2744             continue;
2745          if (!ins->compact_branch)
2746             continue;
2747 
2748          /* We found a branch -- check the type to see if we need to do anything
2749           */
2750          if (ins->branch.target_type != TARGET_BREAK)
2751             continue;
2752 
2753          /* It's a break! Check if it's our break */
2754          if (ins->branch.target_break != loop_idx)
2755             continue;
2756 
2757          /* Okay, cool, we're breaking out of this loop.
2758           * Rewrite from a break to a goto */
2759 
2760          ins->branch.target_type = TARGET_GOTO;
2761          ins->branch.target_block = break_block_idx;
2762 
2763          pan_block_add_successor(_block, &ctx->after_block->base);
2764       }
2765    }
2766 
2767    /* Now that we've finished emitting the loop, free up the depth again
2768     * so we play nice with recursion amid nested loops */
2769    --ctx->current_loop_depth;
2770 
2771    /* Dump loop stats */
2772    ++ctx->loop_count;
2773 }
2774 
2775 static midgard_block *
emit_cf_list(struct compiler_context * ctx,struct exec_list * list)2776 emit_cf_list(struct compiler_context *ctx, struct exec_list *list)
2777 {
2778    midgard_block *start_block = NULL;
2779 
2780    foreach_list_typed(nir_cf_node, node, node, list) {
2781       switch (node->type) {
2782       case nir_cf_node_block: {
2783          midgard_block *block = emit_block(ctx, nir_cf_node_as_block(node));
2784 
2785          if (!start_block)
2786             start_block = block;
2787 
2788          break;
2789       }
2790 
2791       case nir_cf_node_if:
2792          emit_if(ctx, nir_cf_node_as_if(node));
2793          break;
2794 
2795       case nir_cf_node_loop:
2796          emit_loop(ctx, nir_cf_node_as_loop(node));
2797          break;
2798 
2799       case nir_cf_node_function:
2800          assert(0);
2801          break;
2802       }
2803    }
2804 
2805    return start_block;
2806 }
2807 
2808 /* Due to lookahead, we need to report the first tag executed in the command
2809  * stream and in branch targets. An initial block might be empty, so iterate
2810  * until we find one that 'works' */
2811 
2812 unsigned
midgard_get_first_tag_from_block(compiler_context * ctx,unsigned block_idx)2813 midgard_get_first_tag_from_block(compiler_context *ctx, unsigned block_idx)
2814 {
2815    midgard_block *initial_block = mir_get_block(ctx, block_idx);
2816 
2817    mir_foreach_block_from(ctx, initial_block, _v) {
2818       midgard_block *v = (midgard_block *)_v;
2819       if (v->quadword_count) {
2820          midgard_bundle *initial_bundle =
2821             util_dynarray_element(&v->bundles, midgard_bundle, 0);
2822 
2823          return initial_bundle->tag;
2824       }
2825    }
2826 
2827    /* Default to a tag 1 which will break from the shader, in case we jump
2828     * to the exit block (i.e. `return` in a compute shader) */
2829 
2830    return 1;
2831 }
2832 
2833 /* For each fragment writeout instruction, generate a writeout loop to
2834  * associate with it */
2835 
2836 static void
mir_add_writeout_loops(compiler_context * ctx)2837 mir_add_writeout_loops(compiler_context *ctx)
2838 {
2839    for (unsigned rt = 0; rt < ARRAY_SIZE(ctx->writeout_branch); ++rt) {
2840       for (unsigned s = 0; s < MIDGARD_MAX_SAMPLE_ITER; ++s) {
2841          midgard_instruction *br = ctx->writeout_branch[rt][s];
2842          if (!br)
2843             continue;
2844 
2845          unsigned popped = br->branch.target_block;
2846          pan_block_add_successor(&(mir_get_block(ctx, popped - 1)->base),
2847                                  &ctx->current_block->base);
2848          br->branch.target_block = emit_fragment_epilogue(ctx, rt, s);
2849          br->branch.target_type = TARGET_GOTO;
2850 
2851          /* If we have more RTs, we'll need to restore back after our
2852           * loop terminates */
2853          midgard_instruction *next_br = NULL;
2854 
2855          if ((s + 1) < MIDGARD_MAX_SAMPLE_ITER)
2856             next_br = ctx->writeout_branch[rt][s + 1];
2857 
2858          if (!next_br && (rt + 1) < ARRAY_SIZE(ctx->writeout_branch))
2859             next_br = ctx->writeout_branch[rt + 1][0];
2860 
2861          if (next_br) {
2862             midgard_instruction uncond = v_branch(false, false);
2863             uncond.branch.target_block = popped;
2864             uncond.branch.target_type = TARGET_GOTO;
2865             emit_mir_instruction(ctx, uncond);
2866             pan_block_add_successor(&ctx->current_block->base,
2867                                     &(mir_get_block(ctx, popped)->base));
2868             schedule_barrier(ctx);
2869          } else {
2870             /* We're last, so we can terminate here */
2871             br->last_writeout = true;
2872          }
2873       }
2874    }
2875 }
2876 
2877 void
midgard_compile_shader_nir(nir_shader * nir,const struct panfrost_compile_inputs * inputs,struct util_dynarray * binary,struct pan_shader_info * info)2878 midgard_compile_shader_nir(nir_shader *nir,
2879                            const struct panfrost_compile_inputs *inputs,
2880                            struct util_dynarray *binary,
2881                            struct pan_shader_info *info)
2882 {
2883    midgard_debug = debug_get_option_midgard_debug();
2884 
2885    /* TODO: Bound against what? */
2886    compiler_context *ctx = rzalloc(NULL, compiler_context);
2887 
2888    ctx->inputs = inputs;
2889    ctx->nir = nir;
2890    ctx->info = info;
2891    ctx->stage = nir->info.stage;
2892    ctx->blend_input = ~0;
2893    ctx->blend_src1 = ~0;
2894    ctx->quirks = midgard_get_quirks(inputs->gpu_id);
2895 
2896    /* Initialize at a global (not block) level hash tables */
2897 
2898    ctx->ssa_constants = _mesa_hash_table_u64_create(ctx);
2899 
2900    /* Collect varyings after lowering I/O */
2901    pan_nir_collect_varyings(nir, info);
2902 
2903    /* Optimisation passes */
2904    optimise_nir(nir, ctx->quirks, inputs->is_blend);
2905 
2906    bool skip_internal = nir->info.internal;
2907    skip_internal &= !(midgard_debug & MIDGARD_DBG_INTERNAL);
2908 
2909    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
2910       nir_print_shader(nir, stdout);
2911 
2912    info->tls_size = nir->scratch_size;
2913 
2914    nir_foreach_function_with_impl(func, impl, nir) {
2915       list_inithead(&ctx->blocks);
2916       ctx->block_count = 0;
2917       ctx->func = func;
2918 
2919       if (nir->info.outputs_read && !inputs->is_blend) {
2920          emit_block_init(ctx);
2921 
2922          struct midgard_instruction wait = v_branch(false, false);
2923          wait.branch.target_type = TARGET_TILEBUF_WAIT;
2924 
2925          emit_mir_instruction(ctx, wait);
2926 
2927          ++ctx->instruction_count;
2928       }
2929 
2930       emit_cf_list(ctx, &impl->body);
2931       break; /* TODO: Multi-function shaders */
2932    }
2933 
2934    /* Per-block lowering before opts */
2935 
2936    mir_foreach_block(ctx, _block) {
2937       midgard_block *block = (midgard_block *)_block;
2938       inline_alu_constants(ctx, block);
2939       embedded_to_inline_constant(ctx, block);
2940    }
2941    /* MIR-level optimizations */
2942 
2943    bool progress = false;
2944 
2945    do {
2946       progress = false;
2947       progress |= midgard_opt_dead_code_eliminate(ctx);
2948       progress |= midgard_opt_prop(ctx);
2949 
2950       mir_foreach_block(ctx, _block) {
2951          midgard_block *block = (midgard_block *)_block;
2952          progress |= midgard_opt_copy_prop(ctx, block);
2953          progress |= midgard_opt_combine_projection(ctx, block);
2954          progress |= midgard_opt_varying_projection(ctx, block);
2955       }
2956    } while (progress);
2957 
2958    mir_foreach_block(ctx, _block) {
2959       midgard_block *block = (midgard_block *)_block;
2960       midgard_lower_derivatives(ctx, block);
2961       midgard_legalize_invert(ctx, block);
2962       midgard_cull_dead_branch(ctx, block);
2963    }
2964 
2965    if (ctx->stage == MESA_SHADER_FRAGMENT)
2966       mir_add_writeout_loops(ctx);
2967 
2968    /* Analyze now that the code is known but before scheduling creates
2969     * pipeline registers which are harder to track */
2970    mir_analyze_helper_requirements(ctx);
2971 
2972    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
2973       mir_print_shader(ctx);
2974 
2975    /* Schedule! */
2976    midgard_schedule_program(ctx);
2977    mir_ra(ctx);
2978 
2979    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
2980       mir_print_shader(ctx);
2981 
2982    /* Analyze after scheduling since this is order-dependent */
2983    mir_analyze_helper_terminate(ctx);
2984 
2985    /* Emit flat binary from the instruction arrays. Iterate each block in
2986     * sequence. Save instruction boundaries such that lookahead tags can
2987     * be assigned easily */
2988 
2989    /* Cache _all_ bundles in source order for lookahead across failed branches */
2990 
2991    int bundle_count = 0;
2992    mir_foreach_block(ctx, _block) {
2993       midgard_block *block = (midgard_block *)_block;
2994       bundle_count += block->bundles.size / sizeof(midgard_bundle);
2995    }
2996    midgard_bundle **source_order_bundles =
2997       malloc(sizeof(midgard_bundle *) * bundle_count);
2998    int bundle_idx = 0;
2999    mir_foreach_block(ctx, _block) {
3000       midgard_block *block = (midgard_block *)_block;
3001       util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
3002          source_order_bundles[bundle_idx++] = bundle;
3003       }
3004    }
3005 
3006    int current_bundle = 0;
3007 
3008    /* Midgard prefetches instruction types, so during emission we
3009     * need to lookahead. Unless this is the last instruction, in
3010     * which we return 1. */
3011 
3012    mir_foreach_block(ctx, _block) {
3013       midgard_block *block = (midgard_block *)_block;
3014       mir_foreach_bundle_in_block(block, bundle) {
3015          int lookahead = 1;
3016 
3017          if (!bundle->last_writeout && (current_bundle + 1 < bundle_count))
3018             lookahead = source_order_bundles[current_bundle + 1]->tag;
3019 
3020          emit_binary_bundle(ctx, block, bundle, binary, lookahead);
3021          ++current_bundle;
3022       }
3023 
3024       /* TODO: Free deeper */
3025       // util_dynarray_fini(&block->instructions);
3026    }
3027 
3028    free(source_order_bundles);
3029 
3030    /* Report the very first tag executed */
3031    info->midgard.first_tag = midgard_get_first_tag_from_block(ctx, 0);
3032 
3033    info->ubo_mask = ctx->ubo_mask & ((1 << ctx->nir->info.num_ubos) - 1);
3034 
3035    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal) {
3036       disassemble_midgard(stdout, binary->data, binary->size, inputs->gpu_id,
3037                           midgard_debug & MIDGARD_DBG_VERBOSE);
3038       fflush(stdout);
3039    }
3040 
3041    /* A shader ending on a 16MB boundary causes INSTR_INVALID_PC faults,
3042     * workaround by adding some padding to the end of the shader. (The
3043     * kernel makes sure shader BOs can't cross 16MB boundaries.) */
3044    if (binary->size)
3045       memset(util_dynarray_grow(binary, uint8_t, 16), 0, 16);
3046 
3047    if ((midgard_debug & MIDGARD_DBG_SHADERDB || inputs->debug) &&
3048        !nir->info.internal) {
3049       unsigned nr_bundles = 0, nr_ins = 0;
3050 
3051       /* Count instructions and bundles */
3052 
3053       mir_foreach_block(ctx, _block) {
3054          midgard_block *block = (midgard_block *)_block;
3055          nr_bundles +=
3056             util_dynarray_num_elements(&block->bundles, midgard_bundle);
3057 
3058          mir_foreach_bundle_in_block(block, bun)
3059             nr_ins += bun->instruction_count;
3060       }
3061 
3062       /* Calculate thread count. There are certain cutoffs by
3063        * register count for thread count */
3064 
3065       unsigned nr_registers = info->work_reg_count;
3066 
3067       unsigned nr_threads = (nr_registers <= 4)   ? 4
3068                             : (nr_registers <= 8) ? 2
3069                                                   : 1;
3070 
3071       char *shaderdb = NULL;
3072 
3073       /* Dump stats */
3074 
3075       asprintf(&shaderdb,
3076                "%s shader: "
3077                "%u inst, %u bundles, %u quadwords, "
3078                "%u registers, %u threads, %u loops, "
3079                "%u:%u spills:fills",
3080                ctx->inputs->is_blend ? "PAN_SHADER_BLEND"
3081                                      : gl_shader_stage_name(ctx->stage),
3082                nr_ins, nr_bundles, ctx->quadword_count, nr_registers,
3083                nr_threads, ctx->loop_count, ctx->spills, ctx->fills);
3084 
3085       if (midgard_debug & MIDGARD_DBG_SHADERDB)
3086          fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
3087 
3088       if (inputs->debug)
3089          util_debug_message(inputs->debug, SHADER_INFO, "%s", shaderdb);
3090 
3091       free(shaderdb);
3092    }
3093 
3094    _mesa_hash_table_u64_destroy(ctx->ssa_constants);
3095    ralloc_free(ctx);
3096 }
3097