xref: /aosp_15_r20/external/mesa3d/src/panfrost/midgard/midgard_compile.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright (C) 2018-2019 Alyssa Rosenzweig <[email protected]>
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/glsl_to_nir.h"
35 #include "compiler/glsl_types.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_control_flow, 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 static nir_mem_access_size_align
mem_access_size_align_cb(nir_intrinsic_op intrin,uint8_t bytes,uint8_t bit_size,uint32_t align_mul,uint32_t align_offset,bool offset_is_const,const void * cb_data)327 mem_access_size_align_cb(nir_intrinsic_op intrin, uint8_t bytes,
328                          uint8_t bit_size, uint32_t align_mul,
329                          uint32_t align_offset, bool offset_is_const,
330                          const void *cb_data)
331 {
332    uint32_t align = nir_combined_align(align_mul, align_offset);
333    assert(util_is_power_of_two_nonzero(align));
334 
335    /* No more than 16 bytes at a time. */
336    bytes = MIN2(bytes, 16);
337 
338    /* If the number of bytes is a multiple of 4, use 32-bit loads. Else if it's
339     * a multiple of 2, use 16-bit loads. Else use 8-bit loads.
340     *
341     * But if we're only aligned to 1 byte, use 8-bit loads. If we're only
342     * aligned to 2 bytes, use 16-bit loads, unless we needed 8-bit loads due to
343     * the size.
344     */
345    if ((bytes & 1) || (align == 1))
346       bit_size = 8;
347    else if ((bytes & 2) || (align == 2))
348       bit_size = 16;
349    else if (bit_size >= 32)
350       bit_size = 32;
351 
352    unsigned num_comps = MIN2(bytes / (bit_size / 8), 4);
353 
354    /* Push constants require 32-bit loads. */
355    if (intrin == nir_intrinsic_load_push_constant) {
356       if (align_mul >= 4) {
357          /* If align_mul is bigger than 4 we can use align_offset to find
358           * the exact number of words we need to read.
359           */
360          num_comps = DIV_ROUND_UP((align_offset % 4) + bytes, 4);
361       } else {
362          /* If bytes is aligned on 32-bit, the access might still cross one
363           * word at the beginning, and one word at the end. If bytes is not
364           * aligned on 32-bit, the extra two words should cover for both the
365           * size and offset mis-alignment.
366           */
367          num_comps = (bytes / 4) + 2;
368       }
369 
370       bit_size = MIN2(bit_size, 32);
371    }
372 
373    return (nir_mem_access_size_align){
374       .num_components = num_comps,
375       .bit_size = bit_size,
376       .align = bit_size / 8,
377    };
378 }
379 
380 static uint8_t
lower_vec816_alu(const nir_instr * instr,const void * cb_data)381 lower_vec816_alu(const nir_instr *instr, const void *cb_data)
382 {
383   return 4;
384 }
385 
386 void
midgard_preprocess_nir(nir_shader * nir,unsigned gpu_id)387 midgard_preprocess_nir(nir_shader *nir, unsigned gpu_id)
388 {
389    unsigned quirks = midgard_get_quirks(gpu_id);
390 
391    /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
392     * (so we don't accidentally duplicate the epilogue since mesa/st has
393     * messed with our I/O quite a bit already).
394     */
395    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
396 
397    if (nir->info.stage == MESA_SHADER_VERTEX) {
398       NIR_PASS_V(nir, nir_lower_viewport_transform);
399       NIR_PASS_V(nir, nir_lower_point_size, 1.0, 0.0);
400    }
401 
402    NIR_PASS_V(nir, nir_lower_var_copies);
403    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
404    NIR_PASS_V(nir, nir_split_var_copies);
405    NIR_PASS_V(nir, nir_lower_var_copies);
406    NIR_PASS_V(nir, nir_lower_global_vars_to_local);
407    NIR_PASS_V(nir, nir_lower_var_copies);
408    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
409 
410    NIR_PASS_V(nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
411               glsl_type_size, 0);
412 
413    if (nir->info.stage == MESA_SHADER_VERTEX) {
414       /* nir_lower[_explicit]_io is lazy and emits mul+add chains even
415        * for offsets it could figure out are constant.  Do some
416        * constant folding before pan_nir_lower_store_component below.
417        */
418       NIR_PASS_V(nir, nir_opt_constant_folding);
419       NIR_PASS_V(nir, pan_nir_lower_store_component);
420    }
421 
422    /* Could be eventually useful for Vulkan, but we don't expect it to have
423     * the support, so limit it to compute */
424    if (gl_shader_stage_is_compute(nir->info.stage)) {
425       nir_lower_mem_access_bit_sizes_options mem_size_options = {
426          .modes = nir_var_mem_ubo | nir_var_mem_push_const | nir_var_mem_ssbo |
427                   nir_var_mem_constant | nir_var_mem_task_payload |
428                   nir_var_shader_temp | nir_var_function_temp |
429                   nir_var_mem_global | nir_var_mem_shared,
430          .callback = mem_access_size_align_cb,
431       };
432 
433       NIR_PASS_V(nir, nir_lower_mem_access_bit_sizes, &mem_size_options);
434       NIR_PASS_V(nir, nir_lower_alu_width, lower_vec816_alu, NULL);
435       NIR_PASS_V(nir, nir_lower_alu_vec8_16_srcs);
436    }
437 
438    NIR_PASS_V(nir, nir_lower_ssbo, NULL);
439    NIR_PASS_V(nir, pan_nir_lower_zs_store);
440 
441    NIR_PASS_V(nir, nir_lower_frexp);
442    NIR_PASS_V(nir, midgard_nir_lower_global_load);
443 
444    nir_lower_idiv_options idiv_options = {
445       .allow_fp16 = true,
446    };
447 
448    NIR_PASS_V(nir, nir_lower_idiv, &idiv_options);
449 
450    nir_lower_tex_options lower_tex_options = {
451       .lower_txs_lod = true,
452       .lower_txp = ~0,
453       .lower_tg4_broadcom_swizzle = true,
454       .lower_txd = true,
455       .lower_invalid_implicit_lod = true,
456    };
457 
458    NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);
459    NIR_PASS_V(nir, nir_lower_image_atomics_to_global);
460 
461    /* TEX_GRAD fails to apply sampler descriptor settings on some
462     * implementations, requiring a lowering.
463     */
464    if (quirks & MIDGARD_BROKEN_LOD)
465       NIR_PASS_V(nir, midgard_nir_lod_errata);
466 
467    /* lower MSAA image operations to 3D load before coordinate lowering */
468    NIR_PASS_V(nir, pan_nir_lower_image_ms);
469 
470    /* Midgard image ops coordinates are 16-bit instead of 32-bit */
471    NIR_PASS_V(nir, midgard_nir_lower_image_bitsize);
472 
473    if (nir->info.stage == MESA_SHADER_FRAGMENT)
474       NIR_PASS_V(nir, nir_lower_helper_writes, true);
475 
476    NIR_PASS_V(nir, pan_lower_helper_invocation);
477    NIR_PASS_V(nir, pan_lower_sample_pos);
478    NIR_PASS_V(nir, midgard_nir_lower_algebraic_early);
479    NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
480    NIR_PASS_V(nir, nir_lower_flrp, 16 | 32 | 64, false /* always_precise */);
481    NIR_PASS_V(nir, nir_lower_var_copies);
482 }
483 
484 static void
optimise_nir(nir_shader * nir,unsigned quirks,bool is_blend)485 optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
486 {
487    bool progress;
488 
489    do {
490       progress = false;
491 
492       NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
493 
494       NIR_PASS(progress, nir, nir_copy_prop);
495       NIR_PASS(progress, nir, nir_opt_remove_phis);
496       NIR_PASS(progress, nir, nir_opt_dce);
497       NIR_PASS(progress, nir, nir_opt_dead_cf);
498       NIR_PASS(progress, nir, nir_opt_cse);
499       NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
500       NIR_PASS(progress, nir, nir_opt_algebraic);
501       NIR_PASS(progress, nir, nir_opt_constant_folding);
502       NIR_PASS(progress, nir, nir_opt_undef);
503       NIR_PASS(progress, nir, nir_lower_undef_to_zero);
504 
505       NIR_PASS(progress, nir, nir_opt_loop_unroll);
506 
507       NIR_PASS(progress, nir, nir_opt_vectorize, midgard_vectorize_filter,
508                NULL);
509    } while (progress);
510 
511    NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
512 
513    /* Run after opts so it can hit more */
514    if (!is_blend)
515       NIR_PASS(progress, nir, nir_fuse_io_16);
516 
517    do {
518       progress = false;
519 
520       NIR_PASS(progress, nir, nir_opt_dce);
521       NIR_PASS(progress, nir, nir_opt_algebraic);
522       NIR_PASS(progress, nir, nir_opt_constant_folding);
523       NIR_PASS(progress, nir, nir_copy_prop);
524    } while (progress);
525 
526    NIR_PASS(progress, nir, nir_opt_algebraic_late);
527    NIR_PASS(progress, nir, nir_opt_algebraic_distribute_src_mods);
528 
529    /* We implement booleans as 32-bit 0/~0 */
530    NIR_PASS(progress, nir, nir_lower_bool_to_int32);
531 
532    /* Now that booleans are lowered, we can run out late opts */
533    NIR_PASS(progress, nir, midgard_nir_lower_algebraic_late);
534    NIR_PASS(progress, nir, midgard_nir_cancel_inot);
535    NIR_PASS_V(nir, midgard_nir_type_csel);
536 
537    /* Clean up after late opts */
538    do {
539       progress = false;
540 
541       NIR_PASS(progress, nir, nir_opt_dce);
542       NIR_PASS(progress, nir, nir_opt_constant_folding);
543       NIR_PASS(progress, nir, nir_copy_prop);
544    } while (progress);
545 
546    /* Backend scheduler is purely local, so do some global optimizations
547     * to reduce register pressure. */
548    nir_move_options move_all = nir_move_const_undef | nir_move_load_ubo |
549                                nir_move_load_input | nir_move_comparisons |
550                                nir_move_copies | nir_move_load_ssbo;
551 
552    NIR_PASS_V(nir, nir_opt_sink, move_all);
553    NIR_PASS_V(nir, nir_opt_move, move_all);
554 
555    /* Take us out of SSA */
556    NIR_PASS(progress, nir, nir_convert_from_ssa, true);
557 
558    /* We are a vector architecture; write combine where possible */
559    NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest, false);
560    NIR_PASS(progress, nir, nir_lower_vec_to_regs, NULL, NULL);
561 
562    NIR_PASS(progress, nir, nir_opt_dce);
563    NIR_PASS_V(nir, nir_trivialize_registers);
564 }
565 
566 /* Do not actually emit a load; instead, cache the constant for inlining */
567 
568 static void
emit_load_const(compiler_context * ctx,nir_load_const_instr * instr)569 emit_load_const(compiler_context *ctx, nir_load_const_instr *instr)
570 {
571    nir_def def = instr->def;
572 
573    midgard_constants *consts = rzalloc(ctx, midgard_constants);
574 
575    assert(instr->def.num_components * instr->def.bit_size <=
576           sizeof(*consts) * 8);
577 
578 #define RAW_CONST_COPY(bits)                                                   \
579    nir_const_value_to_array(consts->u##bits, instr->value,                     \
580                             instr->def.num_components, u##bits)
581 
582    switch (instr->def.bit_size) {
583    case 64:
584       RAW_CONST_COPY(64);
585       break;
586    case 32:
587       RAW_CONST_COPY(32);
588       break;
589    case 16:
590       RAW_CONST_COPY(16);
591       break;
592    case 8:
593       RAW_CONST_COPY(8);
594       break;
595    default:
596       unreachable("Invalid bit_size for load_const instruction\n");
597    }
598 
599    /* Shifted for SSA, +1 for off-by-one */
600    _mesa_hash_table_u64_insert(ctx->ssa_constants, (def.index << 1) + 1,
601                                consts);
602 }
603 
604 /* Normally constants are embedded implicitly, but for I/O and such we have to
605  * explicitly emit a move with the constant source */
606 
607 static void
emit_explicit_constant(compiler_context * ctx,unsigned node)608 emit_explicit_constant(compiler_context *ctx, unsigned node)
609 {
610    void *constant_value =
611       _mesa_hash_table_u64_search(ctx->ssa_constants, node + 1);
612 
613    if (constant_value) {
614       midgard_instruction ins =
615          v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), node);
616       attach_constants(ctx, &ins, constant_value, node + 1);
617       emit_mir_instruction(ctx, ins);
618    }
619 }
620 
621 static bool
nir_is_non_scalar_swizzle(nir_alu_src * src,unsigned nr_components)622 nir_is_non_scalar_swizzle(nir_alu_src *src, unsigned nr_components)
623 {
624    unsigned comp = src->swizzle[0];
625 
626    for (unsigned c = 1; c < nr_components; ++c) {
627       if (src->swizzle[c] != comp)
628          return true;
629    }
630 
631    return false;
632 }
633 
634 #define ALU_CASE(nir, _op)                                                     \
635    case nir_op_##nir:                                                          \
636       op = midgard_alu_op_##_op;                                               \
637       assert(src_bitsize == dst_bitsize);                                      \
638       break;
639 
640 #define ALU_CASE_RTZ(nir, _op)                                                 \
641    case nir_op_##nir:                                                          \
642       op = midgard_alu_op_##_op;                                               \
643       roundmode = MIDGARD_RTZ;                                                 \
644       break;
645 
646 #define ALU_CHECK_CMP()                                                        \
647    assert(src_bitsize == 16 || src_bitsize == 32 || src_bitsize == 64);        \
648    assert(dst_bitsize == 16 || dst_bitsize == 32);
649 
650 #define ALU_CASE_BCAST(nir, _op, count)                                        \
651    case nir_op_##nir:                                                          \
652       op = midgard_alu_op_##_op;                                               \
653       broadcast_swizzle = count;                                               \
654       ALU_CHECK_CMP();                                                         \
655       break;
656 
657 #define ALU_CASE_CMP(nir, _op)                                                 \
658    case nir_op_##nir:                                                          \
659       op = midgard_alu_op_##_op;                                               \
660       ALU_CHECK_CMP();                                                         \
661       break;
662 
663 static void
mir_copy_src(midgard_instruction * ins,nir_alu_instr * instr,unsigned i,unsigned to,bool is_int,unsigned bcast_count)664 mir_copy_src(midgard_instruction *ins, nir_alu_instr *instr, unsigned i,
665              unsigned to, bool is_int, unsigned bcast_count)
666 {
667    nir_alu_src src = instr->src[i];
668    unsigned bits = nir_src_bit_size(src.src);
669 
670    ins->src[to] = nir_src_index(NULL, &src.src);
671    ins->src_types[to] = nir_op_infos[instr->op].input_types[i] | bits;
672 
673    /* Figure out which component we should fill unused channels with. This
674     * doesn't matter too much in the non-broadcast case, but it makes
675     * should that scalar sources are packed with replicated swizzles,
676     * which works around issues seen with the combination of source
677     * expansion and destination shrinking.
678     */
679    unsigned replicate_c = 0;
680    if (bcast_count) {
681       replicate_c = bcast_count - 1;
682    } else {
683       for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c) {
684          if (nir_alu_instr_channel_used(instr, i, c))
685             replicate_c = c;
686       }
687    }
688 
689    for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c) {
690       ins->swizzle[to][c] =
691          src.swizzle[((!bcast_count || c < bcast_count) &&
692                       nir_alu_instr_channel_used(instr, i, c))
693                         ? c
694                         : replicate_c];
695    }
696 }
697 
698 static void
emit_alu(compiler_context * ctx,nir_alu_instr * instr)699 emit_alu(compiler_context *ctx, nir_alu_instr *instr)
700 {
701    unsigned nr_components = instr->def.num_components;
702    unsigned nr_inputs = nir_op_infos[instr->op].num_inputs;
703    unsigned op = 0;
704 
705    /* Number of components valid to check for the instruction (the rest
706     * will be forced to the last), or 0 to use as-is. Relevant as
707     * ball-type instructions have a channel count in NIR but are all vec4
708     * in Midgard */
709 
710    unsigned broadcast_swizzle = 0;
711 
712    /* Should we swap arguments? */
713    bool flip_src12 = false;
714 
715    ASSERTED unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
716    unsigned dst_bitsize = instr->def.bit_size;
717 
718    enum midgard_roundmode roundmode = MIDGARD_RTE;
719 
720    switch (instr->op) {
721       ALU_CASE(fadd, fadd);
722       ALU_CASE(fmul, fmul);
723       ALU_CASE(fmin, fmin);
724       ALU_CASE(fmax, fmax);
725       ALU_CASE(imin, imin);
726       ALU_CASE(imax, imax);
727       ALU_CASE(umin, umin);
728       ALU_CASE(umax, umax);
729       ALU_CASE(ffloor, ffloor);
730       ALU_CASE(fround_even, froundeven);
731       ALU_CASE(ftrunc, ftrunc);
732       ALU_CASE(fceil, fceil);
733       ALU_CASE(fdot3, fdot3);
734       ALU_CASE(fdot4, fdot4);
735       ALU_CASE(iadd, iadd);
736       ALU_CASE(isub, isub);
737       ALU_CASE(iadd_sat, iaddsat);
738       ALU_CASE(isub_sat, isubsat);
739       ALU_CASE(uadd_sat, uaddsat);
740       ALU_CASE(usub_sat, usubsat);
741       ALU_CASE(imul, imul);
742       ALU_CASE(imul_high, imul);
743       ALU_CASE(umul_high, imul);
744       ALU_CASE(uclz, iclz);
745 
746       /* Zero shoved as second-arg */
747       ALU_CASE(iabs, iabsdiff);
748 
749       ALU_CASE(uabs_isub, iabsdiff);
750       ALU_CASE(uabs_usub, uabsdiff);
751 
752       ALU_CASE(mov, imov);
753 
754       ALU_CASE_CMP(feq32, feq);
755       ALU_CASE_CMP(fneu32, fne);
756       ALU_CASE_CMP(flt32, flt);
757       ALU_CASE_CMP(ieq32, ieq);
758       ALU_CASE_CMP(ine32, ine);
759       ALU_CASE_CMP(ilt32, ilt);
760       ALU_CASE_CMP(ult32, ult);
761 
762       /* We don't have a native b2f32 instruction. Instead, like many
763        * GPUs, we exploit booleans as 0/~0 for false/true, and
764        * correspondingly AND
765        * by 1.0 to do the type conversion. For the moment, prime us
766        * to emit:
767        *
768        * iand [whatever], #0
769        *
770        * At the end of emit_alu (as MIR), we'll fix-up the constant
771        */
772 
773       ALU_CASE_CMP(b2f32, iand);
774       ALU_CASE_CMP(b2f16, iand);
775       ALU_CASE_CMP(b2i32, iand);
776       ALU_CASE_CMP(b2i16, iand);
777 
778       ALU_CASE(frcp, frcp);
779       ALU_CASE(frsq, frsqrt);
780       ALU_CASE(fsqrt, fsqrt);
781       ALU_CASE(fexp2, fexp2);
782       ALU_CASE(flog2, flog2);
783 
784       ALU_CASE_RTZ(f2i64, f2i_rte);
785       ALU_CASE_RTZ(f2u64, f2u_rte);
786       ALU_CASE_RTZ(i2f64, i2f_rte);
787       ALU_CASE_RTZ(u2f64, u2f_rte);
788 
789       ALU_CASE_RTZ(f2i32, f2i_rte);
790       ALU_CASE_RTZ(f2u32, f2u_rte);
791       ALU_CASE_RTZ(i2f32, i2f_rte);
792       ALU_CASE_RTZ(u2f32, u2f_rte);
793 
794       ALU_CASE_RTZ(f2i8, f2i_rte);
795       ALU_CASE_RTZ(f2u8, f2u_rte);
796 
797       ALU_CASE_RTZ(f2i16, f2i_rte);
798       ALU_CASE_RTZ(f2u16, f2u_rte);
799       ALU_CASE_RTZ(i2f16, i2f_rte);
800       ALU_CASE_RTZ(u2f16, u2f_rte);
801 
802       ALU_CASE(fsin_mdg, fsinpi);
803       ALU_CASE(fcos_mdg, fcospi);
804 
805       /* We'll get 0 in the second arg, so:
806        * ~a = ~(a | 0) = nor(a, 0) */
807       ALU_CASE(inot, inor);
808       ALU_CASE(iand, iand);
809       ALU_CASE(ior, ior);
810       ALU_CASE(ixor, ixor);
811       ALU_CASE(ishl, ishl);
812       ALU_CASE(ishr, iasr);
813       ALU_CASE(ushr, ilsr);
814 
815       ALU_CASE_BCAST(b32all_fequal2, fball_eq, 2);
816       ALU_CASE_BCAST(b32all_fequal3, fball_eq, 3);
817       ALU_CASE_CMP(b32all_fequal4, fball_eq);
818 
819       ALU_CASE_BCAST(b32any_fnequal2, fbany_neq, 2);
820       ALU_CASE_BCAST(b32any_fnequal3, fbany_neq, 3);
821       ALU_CASE_CMP(b32any_fnequal4, fbany_neq);
822 
823       ALU_CASE_BCAST(b32all_iequal2, iball_eq, 2);
824       ALU_CASE_BCAST(b32all_iequal3, iball_eq, 3);
825       ALU_CASE_CMP(b32all_iequal4, iball_eq);
826 
827       ALU_CASE_BCAST(b32any_inequal2, ibany_neq, 2);
828       ALU_CASE_BCAST(b32any_inequal3, ibany_neq, 3);
829       ALU_CASE_CMP(b32any_inequal4, ibany_neq);
830 
831       /* Source mods will be shoved in later */
832       ALU_CASE(fabs, fmov);
833       ALU_CASE(fneg, fmov);
834       ALU_CASE(fsat, fmov);
835       ALU_CASE(fsat_signed_mali, fmov);
836       ALU_CASE(fclamp_pos_mali, fmov);
837 
838       /* For size conversion, we use a move. Ideally though we would squash
839        * these ops together; maybe that has to happen after in NIR as part of
840        * propagation...? An earlier algebraic pass ensured we step down by
841        * only / exactly one size. If stepping down, we use a dest override to
842        * reduce the size; if stepping up, we use a larger-sized move with a
843        * half source and a sign/zero-extension modifier */
844 
845    case nir_op_i2i8:
846    case nir_op_i2i16:
847    case nir_op_i2i32:
848    case nir_op_i2i64:
849    case nir_op_u2u8:
850    case nir_op_u2u16:
851    case nir_op_u2u32:
852    case nir_op_u2u64:
853    case nir_op_f2f16:
854    case nir_op_f2f32:
855    case nir_op_f2f64: {
856       if (instr->op == nir_op_f2f16 || instr->op == nir_op_f2f32 ||
857           instr->op == nir_op_f2f64)
858          op = midgard_alu_op_fmov;
859       else
860          op = midgard_alu_op_imov;
861 
862       break;
863    }
864 
865       /* For greater-or-equal, we lower to less-or-equal and flip the
866        * arguments */
867 
868    case nir_op_fge:
869    case nir_op_fge32:
870    case nir_op_ige32:
871    case nir_op_uge32: {
872       op = instr->op == nir_op_fge     ? midgard_alu_op_fle
873            : instr->op == nir_op_fge32 ? midgard_alu_op_fle
874            : instr->op == nir_op_ige32 ? midgard_alu_op_ile
875            : instr->op == nir_op_uge32 ? midgard_alu_op_ule
876                                        : 0;
877 
878       flip_src12 = true;
879       ALU_CHECK_CMP();
880       break;
881    }
882 
883    case nir_op_b32csel:
884    case nir_op_b32fcsel_mdg: {
885       bool mixed = nir_is_non_scalar_swizzle(&instr->src[0], nr_components);
886       bool is_float = instr->op == nir_op_b32fcsel_mdg;
887       op = is_float ? (mixed ? midgard_alu_op_fcsel_v : midgard_alu_op_fcsel)
888                     : (mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel);
889 
890       int index = nir_src_index(ctx, &instr->src[0].src);
891       emit_explicit_constant(ctx, index);
892 
893       break;
894    }
895 
896    case nir_op_unpack_32_2x16:
897    case nir_op_unpack_32_4x8:
898    case nir_op_pack_32_2x16:
899    case nir_op_pack_32_4x8: {
900       op = midgard_alu_op_imov;
901       break;
902    }
903 
904    case nir_op_unpack_64_2x32:
905    case nir_op_unpack_64_4x16:
906    case nir_op_pack_64_2x32:
907    case nir_op_pack_64_4x16: {
908       op = midgard_alu_op_imov;
909       break;
910    }
911 
912    default:
913       mesa_loge("Unhandled ALU op %s\n", nir_op_infos[instr->op].name);
914       assert(0);
915       return;
916    }
917 
918    /* Promote imov to fmov if it might help inline a constant */
919    if (op == midgard_alu_op_imov && nir_src_is_const(instr->src[0].src) &&
920        nir_src_bit_size(instr->src[0].src) == 32 &&
921        nir_is_same_comp_swizzle(instr->src[0].swizzle,
922                                 nir_src_num_components(instr->src[0].src))) {
923       op = midgard_alu_op_fmov;
924    }
925 
926    /* Midgard can perform certain modifiers on output of an ALU op */
927 
928    unsigned outmod = 0;
929    bool is_int = midgard_is_integer_op(op);
930 
931    if (instr->op == nir_op_umul_high || instr->op == nir_op_imul_high) {
932       outmod = midgard_outmod_keephi;
933    } else if (midgard_is_integer_out_op(op)) {
934       outmod = midgard_outmod_keeplo;
935    } else if (instr->op == nir_op_fsat) {
936       outmod = midgard_outmod_clamp_0_1;
937    } else if (instr->op == nir_op_fsat_signed_mali) {
938       outmod = midgard_outmod_clamp_m1_1;
939    } else if (instr->op == nir_op_fclamp_pos_mali) {
940       outmod = midgard_outmod_clamp_0_inf;
941    }
942 
943    /* Fetch unit, quirks, etc information */
944    unsigned opcode_props = alu_opcode_props[op].props;
945    bool quirk_flipped_r24 = opcode_props & QUIRK_FLIPPED_R24;
946 
947    midgard_instruction ins = {
948       .type = TAG_ALU_4,
949       .dest_type = nir_op_infos[instr->op].output_type | dst_bitsize,
950       .roundmode = roundmode,
951    };
952 
953    ins.dest = nir_def_index_with_mask(&instr->def, &ins.mask);
954 
955    for (unsigned i = nr_inputs; i < ARRAY_SIZE(ins.src); ++i)
956       ins.src[i] = ~0;
957 
958    if (quirk_flipped_r24) {
959       ins.src[0] = ~0;
960       mir_copy_src(&ins, instr, 0, 1, is_int, broadcast_swizzle);
961    } else {
962       for (unsigned i = 0; i < nr_inputs; ++i) {
963          unsigned to = i;
964 
965          if (instr->op == nir_op_b32csel || instr->op == nir_op_b32fcsel_mdg) {
966             /* The condition is the first argument; move
967              * the other arguments up one to be a binary
968              * instruction for Midgard with the condition
969              * last */
970 
971             if (i == 0)
972                to = 2;
973             else if (flip_src12)
974                to = 2 - i;
975             else
976                to = i - 1;
977          } else if (flip_src12) {
978             to = 1 - to;
979          }
980 
981          mir_copy_src(&ins, instr, i, to, is_int, broadcast_swizzle);
982       }
983    }
984 
985    if (instr->op == nir_op_fneg || instr->op == nir_op_fabs) {
986       /* Lowered to move */
987       if (instr->op == nir_op_fneg)
988          ins.src_neg[1] ^= true;
989 
990       if (instr->op == nir_op_fabs)
991          ins.src_abs[1] = true;
992    }
993 
994    ins.op = op;
995    ins.outmod = outmod;
996 
997    /* Late fixup for emulated instructions */
998 
999    if (instr->op == nir_op_b2f32 || instr->op == nir_op_b2i32) {
1000       /* Presently, our second argument is an inline #0 constant.
1001        * Switch over to an embedded 1.0 constant (that can't fit
1002        * inline, since we're 32-bit, not 16-bit like the inline
1003        * constants) */
1004 
1005       ins.has_inline_constant = false;
1006       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1007       ins.src_types[1] = nir_type_float32;
1008       ins.has_constants = true;
1009 
1010       if (instr->op == nir_op_b2f32)
1011          ins.constants.f32[0] = 1.0f;
1012       else
1013          ins.constants.i32[0] = 1;
1014 
1015       for (unsigned c = 0; c < 16; ++c)
1016          ins.swizzle[1][c] = 0;
1017    } else if (instr->op == nir_op_b2f16) {
1018       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1019       ins.src_types[1] = nir_type_float16;
1020       ins.has_constants = true;
1021       ins.constants.i16[0] = _mesa_float_to_half(1.0);
1022 
1023       for (unsigned c = 0; c < 16; ++c)
1024          ins.swizzle[1][c] = 0;
1025    } else if (nr_inputs == 1 && !quirk_flipped_r24) {
1026       /* Lots of instructions need a 0 plonked in */
1027       ins.has_inline_constant = false;
1028       ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1029       ins.src_types[1] = ins.src_types[0];
1030       ins.has_constants = true;
1031       ins.constants.u32[0] = 0;
1032 
1033       for (unsigned c = 0; c < 16; ++c)
1034          ins.swizzle[1][c] = 0;
1035    } else if (instr->op == nir_op_pack_32_2x16) {
1036       ins.dest_type = nir_type_uint16;
1037       ins.mask = mask_of(nr_components * 2);
1038       ins.is_pack = true;
1039    } else if (instr->op == nir_op_pack_32_4x8) {
1040       ins.dest_type = nir_type_uint8;
1041       ins.mask = mask_of(nr_components * 4);
1042       ins.is_pack = true;
1043    } else if (instr->op == nir_op_unpack_32_2x16) {
1044       ins.dest_type = nir_type_uint32;
1045       ins.mask = mask_of(nr_components >> 1);
1046       ins.is_pack = true;
1047    } else if (instr->op == nir_op_unpack_32_4x8) {
1048       ins.dest_type = nir_type_uint32;
1049       ins.mask = mask_of(nr_components >> 2);
1050       ins.is_pack = true;
1051    }
1052 
1053    emit_mir_instruction(ctx, ins);
1054 }
1055 
1056 #undef ALU_CASE
1057 
1058 static void
mir_set_intr_mask(nir_instr * instr,midgard_instruction * ins,bool is_read)1059 mir_set_intr_mask(nir_instr *instr, midgard_instruction *ins, bool is_read)
1060 {
1061    nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1062    unsigned nir_mask = 0;
1063    unsigned dsize = 0;
1064 
1065    if (is_read) {
1066       nir_mask = mask_of(nir_intrinsic_dest_components(intr));
1067 
1068       /* Extension is mandatory for 8/16-bit loads */
1069       dsize = intr->def.bit_size == 64 ? 64 : 32;
1070    } else {
1071       nir_mask = nir_intrinsic_write_mask(intr);
1072       dsize = OP_IS_COMMON_STORE(ins->op) ? nir_src_bit_size(intr->src[0]) : 32;
1073    }
1074 
1075    /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1076    unsigned bytemask = pan_to_bytemask(dsize, nir_mask);
1077    ins->dest_type = nir_type_uint | dsize;
1078    mir_set_bytemask(ins, bytemask);
1079 }
1080 
1081 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1082  * optimized) versions of UBO #0 */
1083 
1084 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)1085 emit_ubo_read(compiler_context *ctx, nir_instr *instr, unsigned dest,
1086               unsigned offset, nir_src *indirect_offset,
1087               unsigned indirect_shift, unsigned index, unsigned nr_comps)
1088 {
1089    midgard_instruction ins;
1090 
1091    unsigned dest_size = (instr->type == nir_instr_type_intrinsic)
1092                            ? nir_instr_as_intrinsic(instr)->def.bit_size
1093                            : 32;
1094 
1095    unsigned bitsize = dest_size * nr_comps;
1096 
1097    /* Pick the smallest intrinsic to avoid out-of-bounds reads */
1098    if (bitsize <= 8)
1099       ins = m_ld_ubo_u8(dest, 0);
1100    else if (bitsize <= 16)
1101       ins = m_ld_ubo_u16(dest, 0);
1102    else if (bitsize <= 32)
1103       ins = m_ld_ubo_32(dest, 0);
1104    else if (bitsize <= 64)
1105       ins = m_ld_ubo_64(dest, 0);
1106    else if (bitsize <= 128)
1107       ins = m_ld_ubo_128(dest, 0);
1108    else
1109       unreachable("Invalid UBO read size");
1110 
1111    ins.constants.u32[0] = offset;
1112 
1113    if (instr->type == nir_instr_type_intrinsic)
1114       mir_set_intr_mask(instr, &ins, true);
1115 
1116    if (indirect_offset) {
1117       ins.src[2] = nir_src_index(ctx, indirect_offset);
1118       ins.src_types[2] = nir_type_uint32;
1119       ins.load_store.index_shift = indirect_shift;
1120 
1121       /* X component for the whole swizzle to prevent register
1122        * pressure from ballooning from the extra components */
1123       for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[2]); ++i)
1124          ins.swizzle[2][i] = 0;
1125    } else {
1126       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1127    }
1128 
1129    if (indirect_offset && !indirect_shift)
1130       mir_set_ubo_offset(&ins, indirect_offset, offset);
1131 
1132    midgard_pack_ubo_index_imm(&ins.load_store, index);
1133 
1134    return emit_mir_instruction(ctx, ins);
1135 }
1136 
1137 /* Globals are like UBOs if you squint. And shared memory is like globals if
1138  * you squint even harder */
1139 
1140 static void
emit_global(compiler_context * ctx,nir_instr * instr,bool is_read,unsigned srcdest,nir_src * offset,unsigned seg)1141 emit_global(compiler_context *ctx, nir_instr *instr, bool is_read,
1142             unsigned srcdest, nir_src *offset, unsigned seg)
1143 {
1144    midgard_instruction ins;
1145 
1146    nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1147    if (is_read) {
1148       unsigned bitsize = intr->def.bit_size * intr->def.num_components;
1149 
1150       switch (bitsize) {
1151       case 8:
1152          ins = m_ld_u8(srcdest, 0);
1153          break;
1154       case 16:
1155          ins = m_ld_u16(srcdest, 0);
1156          break;
1157       case 32:
1158          ins = m_ld_32(srcdest, 0);
1159          break;
1160       case 64:
1161          ins = m_ld_64(srcdest, 0);
1162          break;
1163       case 128:
1164          ins = m_ld_128(srcdest, 0);
1165          break;
1166       default:
1167          unreachable("Invalid global read size");
1168       }
1169 
1170       mir_set_intr_mask(instr, &ins, is_read);
1171 
1172       /* For anything not aligned on 32bit, make sure we write full
1173        * 32 bits registers. */
1174       if (bitsize & 31) {
1175          unsigned comps_per_32b = 32 / intr->def.bit_size;
1176 
1177          for (unsigned c = 0; c < 4 * comps_per_32b; c += comps_per_32b) {
1178             if (!(ins.mask & BITFIELD_RANGE(c, comps_per_32b)))
1179                continue;
1180 
1181             unsigned base = ~0;
1182             for (unsigned i = 0; i < comps_per_32b; i++) {
1183                if (ins.mask & BITFIELD_BIT(c + i)) {
1184                   base = ins.swizzle[0][c + i];
1185                   break;
1186                }
1187             }
1188 
1189             assert(base != ~0);
1190 
1191             for (unsigned i = 0; i < comps_per_32b; i++) {
1192                if (!(ins.mask & BITFIELD_BIT(c + i))) {
1193                   ins.swizzle[0][c + i] = base + i;
1194                   ins.mask |= BITFIELD_BIT(c + i);
1195                }
1196                assert(ins.swizzle[0][c + i] == base + i);
1197             }
1198          }
1199       }
1200    } else {
1201       unsigned bitsize =
1202          nir_src_bit_size(intr->src[0]) * nir_src_num_components(intr->src[0]);
1203 
1204       if (bitsize == 8)
1205          ins = m_st_u8(srcdest, 0);
1206       else if (bitsize == 16)
1207          ins = m_st_u16(srcdest, 0);
1208       else if (bitsize <= 32)
1209          ins = m_st_32(srcdest, 0);
1210       else if (bitsize <= 64)
1211          ins = m_st_64(srcdest, 0);
1212       else if (bitsize <= 128)
1213          ins = m_st_128(srcdest, 0);
1214       else
1215          unreachable("Invalid global store size");
1216 
1217       mir_set_intr_mask(instr, &ins, is_read);
1218    }
1219 
1220    mir_set_offset(ctx, &ins, offset, seg);
1221 
1222    /* Set a valid swizzle for masked out components */
1223    assert(ins.mask);
1224    unsigned first_component = __builtin_ffs(ins.mask) - 1;
1225 
1226    for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i) {
1227       if (!(ins.mask & (1 << i)))
1228          ins.swizzle[0][i] = first_component;
1229    }
1230 
1231    emit_mir_instruction(ctx, ins);
1232 }
1233 
1234 static midgard_load_store_op
translate_atomic_op(nir_atomic_op op)1235 translate_atomic_op(nir_atomic_op op)
1236 {
1237    /* clang-format off */
1238    switch (op) {
1239    case nir_atomic_op_xchg:    return midgard_op_atomic_xchg;
1240    case nir_atomic_op_cmpxchg: return midgard_op_atomic_cmpxchg;
1241    case nir_atomic_op_iadd:    return midgard_op_atomic_add;
1242    case nir_atomic_op_iand:    return midgard_op_atomic_and;
1243    case nir_atomic_op_imax:    return midgard_op_atomic_imax;
1244    case nir_atomic_op_imin:    return midgard_op_atomic_imin;
1245    case nir_atomic_op_ior:     return midgard_op_atomic_or;
1246    case nir_atomic_op_umax:    return midgard_op_atomic_umax;
1247    case nir_atomic_op_umin:    return midgard_op_atomic_umin;
1248    case nir_atomic_op_ixor:    return midgard_op_atomic_xor;
1249    default: unreachable("Unexpected atomic");
1250    }
1251    /* clang-format on */
1252 }
1253 
1254 /* Emit an atomic to shared memory or global memory. */
1255 static void
emit_atomic(compiler_context * ctx,nir_intrinsic_instr * instr)1256 emit_atomic(compiler_context *ctx, nir_intrinsic_instr *instr)
1257 {
1258    midgard_load_store_op op =
1259       translate_atomic_op(nir_intrinsic_atomic_op(instr));
1260 
1261    nir_alu_type type =
1262       (op == midgard_op_atomic_imin || op == midgard_op_atomic_imax)
1263          ? nir_type_int
1264          : nir_type_uint;
1265 
1266    bool is_shared = (instr->intrinsic == nir_intrinsic_shared_atomic) ||
1267                     (instr->intrinsic == nir_intrinsic_shared_atomic_swap);
1268 
1269    unsigned dest = nir_def_index(&instr->def);
1270    unsigned val = nir_src_index(ctx, &instr->src[1]);
1271    unsigned bitsize = nir_src_bit_size(instr->src[1]);
1272    emit_explicit_constant(ctx, val);
1273 
1274    midgard_instruction ins = {
1275       .type = TAG_LOAD_STORE_4,
1276       .mask = 0xF,
1277       .dest = dest,
1278       .src = {~0, ~0, ~0, val},
1279       .src_types = {0, 0, 0, type | bitsize},
1280       .op = op,
1281    };
1282 
1283    nir_src *src_offset = nir_get_io_offset_src(instr);
1284 
1285    if (op == midgard_op_atomic_cmpxchg) {
1286       unsigned xchg_val = nir_src_index(ctx, &instr->src[2]);
1287       emit_explicit_constant(ctx, xchg_val);
1288 
1289       ins.src[2] = val;
1290       ins.src_types[2] = type | bitsize;
1291       ins.src[3] = xchg_val;
1292 
1293       if (is_shared) {
1294          ins.load_store.arg_reg = REGISTER_LDST_LOCAL_STORAGE_PTR;
1295          ins.load_store.arg_comp = COMPONENT_Z;
1296          ins.load_store.bitsize_toggle = true;
1297       } else {
1298          for (unsigned i = 0; i < 2; ++i)
1299             ins.swizzle[1][i] = i;
1300 
1301          ins.src[1] = nir_src_index(ctx, src_offset);
1302          ins.src_types[1] = nir_type_uint64;
1303       }
1304    } else
1305       mir_set_offset(ctx, &ins, src_offset,
1306                      is_shared ? LDST_SHARED : LDST_GLOBAL);
1307 
1308    mir_set_intr_mask(&instr->instr, &ins, true);
1309 
1310    emit_mir_instruction(ctx, ins);
1311 }
1312 
1313 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)1314 emit_varying_read(compiler_context *ctx, unsigned dest, unsigned offset,
1315                   unsigned nr_comp, unsigned component,
1316                   nir_src *indirect_offset, nir_alu_type type, bool flat)
1317 {
1318    midgard_instruction ins = m_ld_vary_32(dest, PACK_LDST_ATTRIB_OFS(offset));
1319    ins.mask = mask_of(nr_comp);
1320    ins.dest_type = type;
1321 
1322    if (type == nir_type_float16) {
1323       /* Ensure we are aligned so we can pack it later */
1324       ins.mask = mask_of(ALIGN_POT(nr_comp, 2));
1325    }
1326 
1327    for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i)
1328       ins.swizzle[0][i] = MIN2(i + component, COMPONENT_W);
1329 
1330    midgard_varying_params p = {
1331       .flat_shading = flat,
1332       .perspective_correction = 1,
1333       .interpolate_sample = true,
1334    };
1335    midgard_pack_varying_params(&ins.load_store, p);
1336 
1337    if (indirect_offset) {
1338       ins.src[2] = nir_src_index(ctx, indirect_offset);
1339       ins.src_types[2] = nir_type_uint32;
1340    } else
1341       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1342 
1343    ins.load_store.arg_reg = REGISTER_LDST_ZERO;
1344    ins.load_store.index_format = midgard_index_address_u32;
1345 
1346    /* For flat shading, for GPUs supporting auto32, we always use .u32 and
1347     * require 32-bit mode. For smooth shading, we use the appropriate
1348     * floating-point type.
1349     *
1350     * This could be optimized, but it makes it easy to check correctness.
1351     */
1352    if (ctx->quirks & MIDGARD_NO_AUTO32) {
1353       switch (type) {
1354       case nir_type_uint32:
1355       case nir_type_bool32:
1356          ins.op = midgard_op_ld_vary_32u;
1357          break;
1358       case nir_type_int32:
1359          ins.op = midgard_op_ld_vary_32i;
1360          break;
1361       case nir_type_float32:
1362          ins.op = midgard_op_ld_vary_32;
1363          break;
1364       case nir_type_float16:
1365          ins.op = midgard_op_ld_vary_16;
1366          break;
1367       default:
1368          unreachable("Attempted to load unknown type");
1369          break;
1370       }
1371    } else if (flat) {
1372       assert(nir_alu_type_get_type_size(type) == 32);
1373       ins.op = midgard_op_ld_vary_32u;
1374    } else {
1375       assert(nir_alu_type_get_base_type(type) == nir_type_float);
1376 
1377       ins.op = (nir_alu_type_get_type_size(type) == 32) ? midgard_op_ld_vary_32
1378                                                         : midgard_op_ld_vary_16;
1379    }
1380 
1381    emit_mir_instruction(ctx, ins);
1382 }
1383 
1384 static midgard_instruction
emit_image_op(compiler_context * ctx,nir_intrinsic_instr * instr)1385 emit_image_op(compiler_context *ctx, nir_intrinsic_instr *instr)
1386 {
1387    enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
1388    unsigned nr_dim = glsl_get_sampler_dim_coordinate_components(dim);
1389    bool is_array = nir_intrinsic_image_array(instr);
1390    bool is_store = instr->intrinsic == nir_intrinsic_image_store;
1391 
1392    assert(dim != GLSL_SAMPLER_DIM_MS && "MSAA'd image not lowered");
1393 
1394    unsigned coord_reg = nir_src_index(ctx, &instr->src[1]);
1395    emit_explicit_constant(ctx, coord_reg);
1396 
1397    nir_src *index = &instr->src[0];
1398    bool is_direct = nir_src_is_const(*index);
1399 
1400    /* For image opcodes, address is used as an index into the attribute
1401     * descriptor */
1402    unsigned address = is_direct ? nir_src_as_uint(*index) : 0;
1403 
1404    midgard_instruction ins;
1405    if (is_store) { /* emit st_image_* */
1406       unsigned val = nir_src_index(ctx, &instr->src[3]);
1407       emit_explicit_constant(ctx, val);
1408 
1409       nir_alu_type type = nir_intrinsic_src_type(instr);
1410       ins = st_image(type, val, PACK_LDST_ATTRIB_OFS(address));
1411       nir_alu_type base_type = nir_alu_type_get_base_type(type);
1412       ins.src_types[0] = base_type | nir_src_bit_size(instr->src[3]);
1413    } else if (instr->intrinsic == nir_intrinsic_image_texel_address) {
1414       ins =
1415          m_lea_image(nir_def_index(&instr->def), PACK_LDST_ATTRIB_OFS(address));
1416       ins.mask = mask_of(2); /* 64-bit memory address */
1417    } else {                  /* emit ld_image_* */
1418       nir_alu_type type = nir_intrinsic_dest_type(instr);
1419       ins = ld_image(type, nir_def_index(&instr->def),
1420                      PACK_LDST_ATTRIB_OFS(address));
1421       ins.mask = mask_of(nir_intrinsic_dest_components(instr));
1422       ins.dest_type = type;
1423    }
1424 
1425    /* Coord reg */
1426    ins.src[1] = coord_reg;
1427    ins.src_types[1] = nir_type_uint16;
1428    if (nr_dim == 3 || is_array) {
1429       ins.load_store.bitsize_toggle = true;
1430    }
1431 
1432    /* Image index reg */
1433    if (!is_direct) {
1434       ins.src[2] = nir_src_index(ctx, index);
1435       ins.src_types[2] = nir_type_uint32;
1436    } else
1437       ins.load_store.index_reg = REGISTER_LDST_ZERO;
1438 
1439    emit_mir_instruction(ctx, ins);
1440 
1441    return ins;
1442 }
1443 
1444 static void
emit_attr_read(compiler_context * ctx,unsigned dest,unsigned offset,unsigned nr_comp,nir_alu_type t)1445 emit_attr_read(compiler_context *ctx, unsigned dest, unsigned offset,
1446                unsigned nr_comp, nir_alu_type t)
1447 {
1448    midgard_instruction ins = m_ld_attr_32(dest, PACK_LDST_ATTRIB_OFS(offset));
1449    ins.load_store.arg_reg = REGISTER_LDST_ZERO;
1450    ins.load_store.index_reg = REGISTER_LDST_ZERO;
1451    ins.mask = mask_of(nr_comp);
1452 
1453    /* Use the type appropriate load */
1454    switch (t) {
1455    case nir_type_uint:
1456    case nir_type_bool:
1457       ins.op = midgard_op_ld_attr_32u;
1458       break;
1459    case nir_type_int:
1460       ins.op = midgard_op_ld_attr_32i;
1461       break;
1462    case nir_type_float:
1463       ins.op = midgard_op_ld_attr_32;
1464       break;
1465    default:
1466       unreachable("Attempted to load unknown type");
1467       break;
1468    }
1469 
1470    emit_mir_instruction(ctx, ins);
1471 }
1472 
1473 static unsigned
compute_builtin_arg(nir_intrinsic_op op)1474 compute_builtin_arg(nir_intrinsic_op op)
1475 {
1476    switch (op) {
1477    case nir_intrinsic_load_workgroup_id:
1478       return REGISTER_LDST_GROUP_ID;
1479    case nir_intrinsic_load_local_invocation_id:
1480       return REGISTER_LDST_LOCAL_THREAD_ID;
1481    case nir_intrinsic_load_global_invocation_id:
1482       return REGISTER_LDST_GLOBAL_THREAD_ID;
1483    default:
1484       unreachable("Invalid compute paramater loaded");
1485    }
1486 }
1487 
1488 static void
emit_fragment_store(compiler_context * ctx,unsigned src,unsigned src_z,unsigned src_s,enum midgard_rt_id rt,unsigned sample_iter)1489 emit_fragment_store(compiler_context *ctx, unsigned src, unsigned src_z,
1490                     unsigned src_s, enum midgard_rt_id rt, unsigned sample_iter)
1491 {
1492    assert(rt < ARRAY_SIZE(ctx->writeout_branch));
1493    assert(sample_iter < ARRAY_SIZE(ctx->writeout_branch[0]));
1494 
1495    midgard_instruction *br = ctx->writeout_branch[rt][sample_iter];
1496 
1497    assert(!br);
1498 
1499    emit_explicit_constant(ctx, src);
1500 
1501    struct midgard_instruction ins = v_branch(false, false);
1502 
1503    bool depth_only = (rt == MIDGARD_ZS_RT);
1504 
1505    ins.writeout = depth_only ? 0 : PAN_WRITEOUT_C;
1506 
1507    /* Add dependencies */
1508    ins.src[0] = src;
1509    ins.src_types[0] = nir_type_uint32;
1510 
1511    if (depth_only)
1512       ins.constants.u32[0] = 0xFF;
1513    else
1514       ins.constants.u32[0] = ((rt - MIDGARD_COLOR_RT0) << 8) | sample_iter;
1515 
1516    for (int i = 0; i < 4; ++i)
1517       ins.swizzle[0][i] = i;
1518 
1519    if (~src_z) {
1520       emit_explicit_constant(ctx, src_z);
1521       ins.src[2] = src_z;
1522       ins.src_types[2] = nir_type_uint32;
1523       ins.writeout |= PAN_WRITEOUT_Z;
1524    }
1525    if (~src_s) {
1526       emit_explicit_constant(ctx, src_s);
1527       ins.src[3] = src_s;
1528       ins.src_types[3] = nir_type_uint32;
1529       ins.writeout |= PAN_WRITEOUT_S;
1530    }
1531 
1532    /* Emit the branch */
1533    br = emit_mir_instruction(ctx, ins);
1534    schedule_barrier(ctx);
1535    ctx->writeout_branch[rt][sample_iter] = br;
1536 
1537    /* Push our current location = current block count - 1 = where we'll
1538     * jump to. Maybe a bit too clever for my own good */
1539 
1540    br->branch.target_block = ctx->block_count - 1;
1541 }
1542 
1543 static void
emit_compute_builtin(compiler_context * ctx,nir_intrinsic_instr * instr)1544 emit_compute_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1545 {
1546    unsigned reg = nir_def_index(&instr->def);
1547    midgard_instruction ins = m_ldst_mov(reg, 0);
1548    ins.mask = mask_of(3);
1549    ins.swizzle[0][3] = COMPONENT_X; /* xyzx */
1550    ins.load_store.arg_reg = compute_builtin_arg(instr->intrinsic);
1551    emit_mir_instruction(ctx, ins);
1552 }
1553 
1554 static unsigned
vertex_builtin_arg(nir_intrinsic_op op)1555 vertex_builtin_arg(nir_intrinsic_op op)
1556 {
1557    switch (op) {
1558    case nir_intrinsic_load_vertex_id_zero_base:
1559       return PAN_VERTEX_ID;
1560    case nir_intrinsic_load_instance_id:
1561       return PAN_INSTANCE_ID;
1562    default:
1563       unreachable("Invalid vertex builtin");
1564    }
1565 }
1566 
1567 static void
emit_vertex_builtin(compiler_context * ctx,nir_intrinsic_instr * instr)1568 emit_vertex_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1569 {
1570    unsigned reg = nir_def_index(&instr->def);
1571    emit_attr_read(ctx, reg, vertex_builtin_arg(instr->intrinsic), 1,
1572                   nir_type_int);
1573 }
1574 
1575 static void
emit_special(compiler_context * ctx,nir_intrinsic_instr * instr,unsigned idx)1576 emit_special(compiler_context *ctx, nir_intrinsic_instr *instr, unsigned idx)
1577 {
1578    unsigned reg = nir_def_index(&instr->def);
1579 
1580    midgard_instruction ld = m_ld_tilebuffer_raw(reg, 0);
1581    ld.op = midgard_op_ld_special_32u;
1582    ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(idx);
1583    ld.load_store.index_reg = REGISTER_LDST_ZERO;
1584 
1585    for (int i = 0; i < 4; ++i)
1586       ld.swizzle[0][i] = COMPONENT_X;
1587 
1588    emit_mir_instruction(ctx, ld);
1589 }
1590 
1591 static void
emit_control_barrier(compiler_context * ctx)1592 emit_control_barrier(compiler_context *ctx)
1593 {
1594    midgard_instruction ins = {
1595       .type = TAG_TEXTURE_4,
1596       .dest = ~0,
1597       .src = {~0, ~0, ~0, ~0},
1598       .op = midgard_tex_op_barrier,
1599    };
1600 
1601    emit_mir_instruction(ctx, ins);
1602 }
1603 
1604 static uint8_t
output_load_rt_addr(compiler_context * ctx,nir_intrinsic_instr * instr)1605 output_load_rt_addr(compiler_context *ctx, nir_intrinsic_instr *instr)
1606 {
1607    unsigned loc = nir_intrinsic_io_semantics(instr).location;
1608 
1609    if (loc >= FRAG_RESULT_DATA0)
1610       return loc - FRAG_RESULT_DATA0;
1611 
1612    if (loc == FRAG_RESULT_DEPTH)
1613       return 0x1F;
1614    if (loc == FRAG_RESULT_STENCIL)
1615       return 0x1E;
1616 
1617    unreachable("Invalid RT to load from");
1618 }
1619 
1620 static void
emit_intrinsic(compiler_context * ctx,nir_intrinsic_instr * instr)1621 emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
1622 {
1623    unsigned offset = 0, reg;
1624 
1625    switch (instr->intrinsic) {
1626    case nir_intrinsic_decl_reg:
1627    case nir_intrinsic_store_reg:
1628       /* Always fully consumed */
1629       break;
1630 
1631    case nir_intrinsic_load_reg: {
1632       /* NIR guarantees that, for typical isel, this will always be fully
1633        * consumed. However, we also do our own nir_scalar chasing for
1634        * address arithmetic, bypassing the source chasing helpers. So we can end
1635        * up with unconsumed load_register instructions. Translate them here. 99%
1636        * of the time, these moves will be DCE'd away.
1637        */
1638       nir_def *handle = instr->src[0].ssa;
1639 
1640       midgard_instruction ins =
1641          v_mov(nir_reg_index(handle), nir_def_index(&instr->def));
1642 
1643       ins.dest_type = ins.src_types[1] = nir_type_uint | instr->def.bit_size;
1644 
1645       ins.mask = BITFIELD_MASK(instr->def.num_components);
1646       emit_mir_instruction(ctx, ins);
1647       break;
1648    }
1649 
1650    case nir_intrinsic_terminate_if:
1651    case nir_intrinsic_terminate: {
1652       bool conditional = instr->intrinsic == nir_intrinsic_terminate_if;
1653       struct midgard_instruction discard = v_branch(conditional, false);
1654       discard.branch.target_type = TARGET_DISCARD;
1655 
1656       if (conditional) {
1657          discard.src[0] = nir_src_index(ctx, &instr->src[0]);
1658          discard.src_types[0] = nir_type_uint32;
1659       }
1660 
1661       emit_mir_instruction(ctx, discard);
1662       schedule_barrier(ctx);
1663 
1664       break;
1665    }
1666 
1667    case nir_intrinsic_image_load:
1668    case nir_intrinsic_image_store:
1669    case nir_intrinsic_image_texel_address:
1670       emit_image_op(ctx, instr);
1671       break;
1672 
1673    case nir_intrinsic_load_ubo:
1674    case nir_intrinsic_load_global:
1675    case nir_intrinsic_load_global_constant:
1676    case nir_intrinsic_load_shared:
1677    case nir_intrinsic_load_scratch:
1678    case nir_intrinsic_load_input:
1679    case nir_intrinsic_load_interpolated_input: {
1680       bool is_ubo = instr->intrinsic == nir_intrinsic_load_ubo;
1681       bool is_global = instr->intrinsic == nir_intrinsic_load_global ||
1682                        instr->intrinsic == nir_intrinsic_load_global_constant;
1683       bool is_shared = instr->intrinsic == nir_intrinsic_load_shared;
1684       bool is_scratch = instr->intrinsic == nir_intrinsic_load_scratch;
1685       bool is_flat = instr->intrinsic == nir_intrinsic_load_input;
1686       bool is_interp =
1687          instr->intrinsic == nir_intrinsic_load_interpolated_input;
1688 
1689       /* Get the base type of the intrinsic */
1690       /* TODO: Infer type? Does it matter? */
1691       nir_alu_type t = (is_interp) ? nir_type_float
1692                        : (is_flat) ? nir_intrinsic_dest_type(instr)
1693                                    : nir_type_uint;
1694 
1695       t = nir_alu_type_get_base_type(t);
1696 
1697       if (!(is_ubo || is_global || is_scratch)) {
1698          offset = nir_intrinsic_base(instr);
1699       }
1700 
1701       unsigned nr_comp = nir_intrinsic_dest_components(instr);
1702 
1703       nir_src *src_offset = nir_get_io_offset_src(instr);
1704 
1705       bool direct = nir_src_is_const(*src_offset);
1706       nir_src *indirect_offset = direct ? NULL : src_offset;
1707 
1708       if (direct)
1709          offset += nir_src_as_uint(*src_offset);
1710 
1711       /* We may need to apply a fractional offset */
1712       int component =
1713          (is_flat || is_interp) ? nir_intrinsic_component(instr) : 0;
1714       reg = nir_def_index(&instr->def);
1715 
1716       if (is_ubo) {
1717          nir_src index = instr->src[0];
1718 
1719          /* TODO: Is indirect block number possible? */
1720          assert(nir_src_is_const(index));
1721 
1722          uint32_t uindex = nir_src_as_uint(index);
1723          emit_ubo_read(ctx, &instr->instr, reg, offset, indirect_offset, 0,
1724                        uindex, nr_comp);
1725       } else if (is_global || is_shared || is_scratch) {
1726          unsigned seg =
1727             is_global ? LDST_GLOBAL : (is_shared ? LDST_SHARED : LDST_SCRATCH);
1728          emit_global(ctx, &instr->instr, true, reg, src_offset, seg);
1729       } else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->inputs->is_blend) {
1730          emit_varying_read(ctx, reg, offset, nr_comp, component,
1731                            indirect_offset, t | instr->def.bit_size, is_flat);
1732       } else if (ctx->inputs->is_blend) {
1733          /* ctx->blend_input will be precoloured to r0/r2, where
1734           * the input is preloaded */
1735 
1736          unsigned *input = offset ? &ctx->blend_src1 : &ctx->blend_input;
1737 
1738          if (*input == ~0)
1739             *input = reg;
1740          else
1741             emit_mir_instruction(ctx, v_mov(*input, reg));
1742       } else if (ctx->stage == MESA_SHADER_VERTEX) {
1743          emit_attr_read(ctx, reg, offset, nr_comp, t);
1744       } else {
1745          unreachable("Unknown load");
1746       }
1747 
1748       break;
1749    }
1750 
1751    /* Handled together with load_interpolated_input */
1752    case nir_intrinsic_load_barycentric_pixel:
1753    case nir_intrinsic_load_barycentric_centroid:
1754    case nir_intrinsic_load_barycentric_sample:
1755       break;
1756 
1757       /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1758 
1759    case nir_intrinsic_load_raw_output_pan: {
1760       reg = nir_def_index(&instr->def);
1761 
1762       /* T720 and below use different blend opcodes with slightly
1763        * different semantics than T760 and up */
1764 
1765       midgard_instruction ld = m_ld_tilebuffer_raw(reg, 0);
1766 
1767       unsigned target = output_load_rt_addr(ctx, instr);
1768       ld.load_store.index_comp = target & 0x3;
1769       ld.load_store.index_reg = target >> 2;
1770 
1771       if (nir_src_is_const(instr->src[0])) {
1772          unsigned sample = nir_src_as_uint(instr->src[0]);
1773          ld.load_store.arg_comp = sample & 0x3;
1774          ld.load_store.arg_reg = sample >> 2;
1775       } else {
1776          /* Enable sample index via register. */
1777          ld.load_store.signed_offset |= 1;
1778          ld.src[1] = nir_src_index(ctx, &instr->src[0]);
1779          ld.src_types[1] = nir_type_int32;
1780       }
1781 
1782       if (ctx->quirks & MIDGARD_OLD_BLEND) {
1783          ld.op = midgard_op_ld_special_32u;
1784          ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(16);
1785          ld.load_store.index_reg = REGISTER_LDST_ZERO;
1786       }
1787 
1788       emit_mir_instruction(ctx, ld);
1789       break;
1790    }
1791 
1792    case nir_intrinsic_load_output: {
1793       reg = nir_def_index(&instr->def);
1794 
1795       unsigned bits = instr->def.bit_size;
1796 
1797       midgard_instruction ld;
1798       if (bits == 16)
1799          ld = m_ld_tilebuffer_16f(reg, 0);
1800       else
1801          ld = m_ld_tilebuffer_32f(reg, 0);
1802 
1803       unsigned index = output_load_rt_addr(ctx, instr);
1804       ld.load_store.index_comp = index & 0x3;
1805       ld.load_store.index_reg = index >> 2;
1806 
1807       for (unsigned c = 4; c < 16; ++c)
1808          ld.swizzle[0][c] = 0;
1809 
1810       if (ctx->quirks & MIDGARD_OLD_BLEND) {
1811          if (bits == 16)
1812             ld.op = midgard_op_ld_special_16f;
1813          else
1814             ld.op = midgard_op_ld_special_32f;
1815          ld.load_store.signed_offset = PACK_LDST_SELECTOR_OFS(1);
1816          ld.load_store.index_reg = REGISTER_LDST_ZERO;
1817       }
1818 
1819       emit_mir_instruction(ctx, ld);
1820       break;
1821    }
1822 
1823    case nir_intrinsic_store_output:
1824    case nir_intrinsic_store_combined_output_pan:
1825       assert(nir_src_is_const(instr->src[1]) && "no indirect outputs");
1826 
1827       reg = nir_src_index(ctx, &instr->src[0]);
1828 
1829       if (ctx->stage == MESA_SHADER_FRAGMENT) {
1830          bool combined =
1831             instr->intrinsic == nir_intrinsic_store_combined_output_pan;
1832 
1833          enum midgard_rt_id rt;
1834 
1835          unsigned reg_z = ~0, reg_s = ~0, reg_2 = ~0;
1836          unsigned writeout = PAN_WRITEOUT_C;
1837          if (combined) {
1838             writeout = nir_intrinsic_component(instr);
1839             if (writeout & PAN_WRITEOUT_Z)
1840                reg_z = nir_src_index(ctx, &instr->src[2]);
1841             if (writeout & PAN_WRITEOUT_S)
1842                reg_s = nir_src_index(ctx, &instr->src[3]);
1843             if (writeout & PAN_WRITEOUT_2)
1844                reg_2 = nir_src_index(ctx, &instr->src[4]);
1845          }
1846 
1847          if (writeout & PAN_WRITEOUT_C) {
1848             nir_io_semantics sem = nir_intrinsic_io_semantics(instr);
1849 
1850             rt = MIDGARD_COLOR_RT0 + (sem.location - FRAG_RESULT_DATA0);
1851          } else {
1852             rt = MIDGARD_ZS_RT;
1853          }
1854 
1855          /* Dual-source blend writeout is done by leaving the
1856           * value in r2 for the blend shader to use. */
1857          if (~reg_2) {
1858             emit_explicit_constant(ctx, reg_2);
1859 
1860             unsigned out = make_compiler_temp(ctx);
1861 
1862             midgard_instruction ins = v_mov(reg_2, out);
1863             emit_mir_instruction(ctx, ins);
1864 
1865             ctx->blend_src1 = out;
1866          }
1867 
1868          emit_fragment_store(ctx, reg, reg_z, reg_s, rt, 0);
1869       } else if (ctx->stage == MESA_SHADER_VERTEX) {
1870          assert(instr->intrinsic == nir_intrinsic_store_output);
1871 
1872          /* We should have been vectorized, though we don't
1873           * currently check that st_vary is emitted only once
1874           * per slot (this is relevant, since there's not a mask
1875           * parameter available on the store [set to 0 by the
1876           * blob]). We do respect the component by adjusting the
1877           * swizzle. If this is a constant source, we'll need to
1878           * emit that explicitly. */
1879 
1880          emit_explicit_constant(ctx, reg);
1881 
1882          offset = nir_intrinsic_base(instr) + nir_src_as_uint(instr->src[1]);
1883 
1884          unsigned dst_component = nir_intrinsic_component(instr);
1885          unsigned nr_comp = nir_src_num_components(instr->src[0]);
1886 
1887          /* ABI: Format controlled by the attribute descriptor.
1888           * This simplifies flat shading, although it prevents
1889           * certain (unimplemented) 16-bit optimizations.
1890           *
1891           * In particular, it lets the driver handle internal
1892           * TGSI shaders that set flat in the VS but smooth in
1893           * the FS. This matches our handling on Bifrost.
1894           */
1895          bool auto32 = true;
1896          assert(nir_alu_type_get_type_size(nir_intrinsic_src_type(instr)) ==
1897                 32);
1898 
1899          /* ABI: varyings in the secondary attribute table */
1900          bool secondary_table = true;
1901 
1902          midgard_instruction st =
1903             m_st_vary_32(reg, PACK_LDST_ATTRIB_OFS(offset));
1904          st.load_store.arg_reg = REGISTER_LDST_ZERO;
1905          st.load_store.index_reg = REGISTER_LDST_ZERO;
1906 
1907          /* Attribute instruction uses these 2-bits for the
1908           * a32 and table bits, pack this specially.
1909           */
1910          st.load_store.index_format =
1911             (auto32 ? (1 << 0) : 0) | (secondary_table ? (1 << 1) : 0);
1912 
1913          /* nir_intrinsic_component(store_intr) encodes the
1914           * destination component start. Source component offset
1915           * adjustment is taken care of in
1916           * install_registers_instr(), when offset_swizzle() is
1917           * called.
1918           */
1919          unsigned src_component = COMPONENT_X;
1920 
1921          assert(nr_comp > 0);
1922          for (unsigned i = 0; i < ARRAY_SIZE(st.swizzle); ++i) {
1923             st.swizzle[0][i] = src_component;
1924             if (i >= dst_component && i < dst_component + nr_comp - 1)
1925                src_component++;
1926          }
1927 
1928          emit_mir_instruction(ctx, st);
1929       } else {
1930          unreachable("Unknown store");
1931       }
1932 
1933       break;
1934 
1935    /* Special case of store_output for lowered blend shaders */
1936    case nir_intrinsic_store_raw_output_pan: {
1937       assert(ctx->stage == MESA_SHADER_FRAGMENT);
1938       reg = nir_src_index(ctx, &instr->src[0]);
1939 
1940       nir_io_semantics sem = nir_intrinsic_io_semantics(instr);
1941       assert(sem.location >= FRAG_RESULT_DATA0);
1942       unsigned rt = sem.location - FRAG_RESULT_DATA0;
1943 
1944       emit_fragment_store(ctx, reg, ~0, ~0, rt + MIDGARD_COLOR_RT0,
1945                           nir_intrinsic_base(instr));
1946       break;
1947    }
1948 
1949    case nir_intrinsic_store_global:
1950    case nir_intrinsic_store_shared:
1951    case nir_intrinsic_store_scratch:
1952       reg = nir_src_index(ctx, &instr->src[0]);
1953       emit_explicit_constant(ctx, reg);
1954 
1955       unsigned seg;
1956       if (instr->intrinsic == nir_intrinsic_store_global)
1957          seg = LDST_GLOBAL;
1958       else if (instr->intrinsic == nir_intrinsic_store_shared)
1959          seg = LDST_SHARED;
1960       else
1961          seg = LDST_SCRATCH;
1962 
1963       emit_global(ctx, &instr->instr, false, reg, &instr->src[1], seg);
1964       break;
1965 
1966    case nir_intrinsic_load_workgroup_id:
1967    case nir_intrinsic_load_local_invocation_id:
1968    case nir_intrinsic_load_global_invocation_id:
1969       emit_compute_builtin(ctx, instr);
1970       break;
1971 
1972    case nir_intrinsic_load_vertex_id_zero_base:
1973    case nir_intrinsic_load_instance_id:
1974       emit_vertex_builtin(ctx, instr);
1975       break;
1976 
1977    case nir_intrinsic_load_sample_mask_in:
1978       emit_special(ctx, instr, 96);
1979       break;
1980 
1981    case nir_intrinsic_load_sample_id:
1982       emit_special(ctx, instr, 97);
1983       break;
1984 
1985    case nir_intrinsic_barrier:
1986       if (nir_intrinsic_execution_scope(instr) != SCOPE_NONE) {
1987          schedule_barrier(ctx);
1988          emit_control_barrier(ctx);
1989          schedule_barrier(ctx);
1990       } else if (nir_intrinsic_memory_scope(instr) != SCOPE_NONE) {
1991          /* Midgard doesn't seem to want special handling, though we do need to
1992           * take care when scheduling to avoid incorrect reordering.
1993           *
1994           * Note this is an "else if" since the handling for the execution scope
1995           * case already covers the case when both scopes are present.
1996           */
1997          schedule_barrier(ctx);
1998       }
1999       break;
2000 
2001    case nir_intrinsic_shared_atomic:
2002    case nir_intrinsic_shared_atomic_swap:
2003    case nir_intrinsic_global_atomic:
2004    case nir_intrinsic_global_atomic_swap:
2005       emit_atomic(ctx, instr);
2006       break;
2007 
2008    case nir_intrinsic_ddx:
2009    case nir_intrinsic_ddy:
2010       midgard_emit_derivatives(ctx, instr);
2011       break;
2012 
2013    default:
2014       fprintf(stderr, "Unhandled intrinsic %s\n",
2015               nir_intrinsic_infos[instr->intrinsic].name);
2016       assert(0);
2017       break;
2018    }
2019 }
2020 
2021 /* Returns dimension with 0 special casing cubemaps */
2022 static unsigned
midgard_tex_format(enum glsl_sampler_dim dim)2023 midgard_tex_format(enum glsl_sampler_dim dim)
2024 {
2025    switch (dim) {
2026    case GLSL_SAMPLER_DIM_1D:
2027    case GLSL_SAMPLER_DIM_BUF:
2028       return 1;
2029 
2030    case GLSL_SAMPLER_DIM_2D:
2031    case GLSL_SAMPLER_DIM_MS:
2032    case GLSL_SAMPLER_DIM_EXTERNAL:
2033    case GLSL_SAMPLER_DIM_RECT:
2034       return 2;
2035 
2036    case GLSL_SAMPLER_DIM_3D:
2037       return 3;
2038 
2039    case GLSL_SAMPLER_DIM_CUBE:
2040       return 0;
2041 
2042    default:
2043       unreachable("Unknown sampler dim type");
2044    }
2045 }
2046 
2047 /* Tries to attach an explicit LOD or bias as a constant. Returns whether this
2048  * was successful */
2049 
2050 static bool
pan_attach_constant_bias(compiler_context * ctx,nir_src lod,midgard_texture_word * word)2051 pan_attach_constant_bias(compiler_context *ctx, nir_src lod,
2052                          midgard_texture_word *word)
2053 {
2054    /* To attach as constant, it has to *be* constant */
2055 
2056    if (!nir_src_is_const(lod))
2057       return false;
2058 
2059    float f = nir_src_as_float(lod);
2060 
2061    /* Break into fixed-point */
2062    signed lod_int = f;
2063    float lod_frac = f - lod_int;
2064 
2065    /* Carry over negative fractions */
2066    if (lod_frac < 0.0) {
2067       lod_int--;
2068       lod_frac += 1.0;
2069    }
2070 
2071    /* Encode */
2072    word->bias = float_to_ubyte(lod_frac);
2073    word->bias_int = lod_int;
2074 
2075    return true;
2076 }
2077 
2078 static enum mali_texture_mode
mdg_texture_mode(nir_tex_instr * instr)2079 mdg_texture_mode(nir_tex_instr *instr)
2080 {
2081    if (instr->op == nir_texop_tg4 && instr->is_shadow)
2082       return TEXTURE_GATHER_SHADOW;
2083    else if (instr->op == nir_texop_tg4)
2084       return TEXTURE_GATHER_X + instr->component;
2085    else if (instr->is_shadow)
2086       return TEXTURE_SHADOW;
2087    else
2088       return TEXTURE_NORMAL;
2089 }
2090 
2091 static void
set_tex_coord(compiler_context * ctx,nir_tex_instr * instr,midgard_instruction * ins)2092 set_tex_coord(compiler_context *ctx, nir_tex_instr *instr,
2093               midgard_instruction *ins)
2094 {
2095    int coord_idx = nir_tex_instr_src_index(instr, nir_tex_src_coord);
2096 
2097    assert(coord_idx >= 0);
2098 
2099    int comparator_idx = nir_tex_instr_src_index(instr, nir_tex_src_comparator);
2100    int ms_idx = nir_tex_instr_src_index(instr, nir_tex_src_ms_index);
2101    assert(comparator_idx < 0 || ms_idx < 0);
2102    int ms_or_comparator_idx = ms_idx >= 0 ? ms_idx : comparator_idx;
2103 
2104    unsigned coords = nir_src_index(ctx, &instr->src[coord_idx].src);
2105 
2106    emit_explicit_constant(ctx, coords);
2107 
2108    ins->src_types[1] = nir_tex_instr_src_type(instr, coord_idx) |
2109                        nir_src_bit_size(instr->src[coord_idx].src);
2110 
2111    unsigned nr_comps = instr->coord_components;
2112    unsigned written_mask = 0, write_mask = 0;
2113 
2114    /* Initialize all components to coord.x which is expected to always be
2115     * present. Swizzle is updated below based on the texture dimension
2116     * and extra attributes that are packed in the coordinate argument.
2117     */
2118    for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++)
2119       ins->swizzle[1][c] = COMPONENT_X;
2120 
2121    /* Shadow ref value is part of the coordinates if there's no comparator
2122     * source, in that case it's always placed in the last component.
2123     * Midgard wants the ref value in coord.z.
2124     */
2125    if (instr->is_shadow && comparator_idx < 0) {
2126       ins->swizzle[1][COMPONENT_Z] = --nr_comps;
2127       write_mask |= 1 << COMPONENT_Z;
2128    }
2129 
2130    /* The array index is the last component if there's no shadow ref value
2131     * or second last if there's one. We already decremented the number of
2132     * components to account for the shadow ref value above.
2133     * Midgard wants the array index in coord.w.
2134     */
2135    if (instr->is_array) {
2136       ins->swizzle[1][COMPONENT_W] = --nr_comps;
2137       write_mask |= 1 << COMPONENT_W;
2138    }
2139 
2140    if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
2141       /* texelFetch is undefined on samplerCube */
2142       assert(ins->op != midgard_tex_op_fetch);
2143 
2144       ins->src[1] = make_compiler_temp_reg(ctx);
2145 
2146       /* For cubemaps, we use a special ld/st op to select the face
2147        * and copy the xy into the texture register
2148        */
2149       midgard_instruction ld = m_ld_cubemap_coords(ins->src[1], 0);
2150       ld.src[1] = coords;
2151       ld.src_types[1] = ins->src_types[1];
2152       ld.mask = 0x3; /* xy */
2153       ld.load_store.bitsize_toggle = true;
2154       ld.swizzle[1][3] = COMPONENT_X;
2155       emit_mir_instruction(ctx, ld);
2156 
2157       /* We packed cube coordiates (X,Y,Z) into (X,Y), update the
2158        * written mask accordingly and decrement the number of
2159        * components
2160        */
2161       nr_comps--;
2162       written_mask |= 3;
2163    }
2164 
2165    /* Now flag tex coord components that have not been written yet */
2166    write_mask |= mask_of(nr_comps) & ~written_mask;
2167    for (unsigned c = 0; c < nr_comps; c++)
2168       ins->swizzle[1][c] = c;
2169 
2170    /* Sample index and shadow ref are expected in coord.z */
2171    if (ms_or_comparator_idx >= 0) {
2172       assert(!((write_mask | written_mask) & (1 << COMPONENT_Z)));
2173 
2174       unsigned sample_or_ref =
2175          nir_src_index(ctx, &instr->src[ms_or_comparator_idx].src);
2176 
2177       emit_explicit_constant(ctx, sample_or_ref);
2178 
2179       if (ins->src[1] == ~0)
2180          ins->src[1] = make_compiler_temp_reg(ctx);
2181 
2182       midgard_instruction mov = v_mov(sample_or_ref, ins->src[1]);
2183 
2184       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++)
2185          mov.swizzle[1][c] = COMPONENT_X;
2186 
2187       mov.mask = 1 << COMPONENT_Z;
2188       written_mask |= 1 << COMPONENT_Z;
2189       ins->swizzle[1][COMPONENT_Z] = COMPONENT_Z;
2190       emit_mir_instruction(ctx, mov);
2191    }
2192 
2193    /* Texelfetch coordinates uses all four elements (xyz/index) regardless
2194     * of texture dimensionality, which means it's necessary to zero the
2195     * unused components to keep everything happy.
2196     */
2197    if (ins->op == midgard_tex_op_fetch && (written_mask | write_mask) != 0xF) {
2198       if (ins->src[1] == ~0)
2199          ins->src[1] = make_compiler_temp_reg(ctx);
2200 
2201       /* mov index.zw, #0, or generalized */
2202       midgard_instruction mov =
2203          v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), ins->src[1]);
2204       mov.has_constants = true;
2205       mov.mask = (written_mask | write_mask) ^ 0xF;
2206       emit_mir_instruction(ctx, mov);
2207       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++) {
2208          if (mov.mask & (1 << c))
2209             ins->swizzle[1][c] = c;
2210       }
2211    }
2212 
2213    if (ins->src[1] == ~0) {
2214       /* No temporary reg created, use the src coords directly */
2215       ins->src[1] = coords;
2216    } else if (write_mask) {
2217       /* Move the remaining coordinates to the temporary reg */
2218       midgard_instruction mov = v_mov(coords, ins->src[1]);
2219 
2220       for (unsigned c = 0; c < MIR_VEC_COMPONENTS; c++) {
2221          if ((1 << c) & write_mask) {
2222             mov.swizzle[1][c] = ins->swizzle[1][c];
2223             ins->swizzle[1][c] = c;
2224          } else {
2225             mov.swizzle[1][c] = COMPONENT_X;
2226          }
2227       }
2228 
2229       mov.mask = write_mask;
2230       emit_mir_instruction(ctx, mov);
2231    }
2232 }
2233 
2234 static void
emit_texop_native(compiler_context * ctx,nir_tex_instr * instr,unsigned midgard_texop)2235 emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
2236                   unsigned midgard_texop)
2237 {
2238    int texture_index = instr->texture_index;
2239    int sampler_index = instr->sampler_index;
2240 
2241    /* If txf is used, we assume there is a valid sampler bound at index 0. Use
2242     * it for txf operations, since there may be no other valid samplers. This is
2243     * a workaround: txf does not require a sampler in NIR (so sampler_index is
2244     * undefined) but we need one in the hardware. This is ABI with the driver.
2245     */
2246    if (!nir_tex_instr_need_sampler(instr))
2247       sampler_index = 0;
2248 
2249    midgard_instruction ins = {
2250       .type = TAG_TEXTURE_4,
2251       .mask = 0xF,
2252       .dest = nir_def_index(&instr->def),
2253       .src = {~0, ~0, ~0, ~0},
2254       .dest_type = instr->dest_type,
2255       .swizzle = SWIZZLE_IDENTITY_4,
2256       .outmod = midgard_outmod_none,
2257       .op = midgard_texop,
2258       .texture = {
2259          .format = midgard_tex_format(instr->sampler_dim),
2260          .texture_handle = texture_index,
2261          .sampler_handle = sampler_index,
2262          .mode = mdg_texture_mode(instr),
2263       }};
2264 
2265    if (instr->is_shadow && !instr->is_new_style_shadow &&
2266        instr->op != nir_texop_tg4)
2267       for (int i = 0; i < 4; ++i)
2268          ins.swizzle[0][i] = COMPONENT_X;
2269 
2270    for (unsigned i = 0; i < instr->num_srcs; ++i) {
2271       int index = nir_src_index(ctx, &instr->src[i].src);
2272       unsigned sz = nir_src_bit_size(instr->src[i].src);
2273       nir_alu_type T = nir_tex_instr_src_type(instr, i) | sz;
2274 
2275       switch (instr->src[i].src_type) {
2276       case nir_tex_src_coord:
2277          set_tex_coord(ctx, instr, &ins);
2278          break;
2279 
2280       case nir_tex_src_bias:
2281       case nir_tex_src_lod: {
2282          /* Try as a constant if we can */
2283 
2284          bool is_txf = midgard_texop == midgard_tex_op_fetch;
2285          if (!is_txf &&
2286              pan_attach_constant_bias(ctx, instr->src[i].src, &ins.texture))
2287             break;
2288 
2289          ins.texture.lod_register = true;
2290          ins.src[2] = index;
2291          ins.src_types[2] = T;
2292 
2293          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
2294             ins.swizzle[2][c] = COMPONENT_X;
2295 
2296          emit_explicit_constant(ctx, index);
2297 
2298          break;
2299       };
2300 
2301       case nir_tex_src_offset: {
2302          ins.texture.offset_register = true;
2303          ins.src[3] = index;
2304          ins.src_types[3] = T;
2305 
2306          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
2307             ins.swizzle[3][c] = (c > COMPONENT_Z) ? 0 : c;
2308 
2309          emit_explicit_constant(ctx, index);
2310          break;
2311       };
2312 
2313       case nir_tex_src_comparator:
2314       case nir_tex_src_ms_index:
2315          /* Nothing to do, handled in set_tex_coord() */
2316          break;
2317 
2318       default: {
2319          fprintf(stderr, "Unknown texture source type: %d\n",
2320                  instr->src[i].src_type);
2321          assert(0);
2322       }
2323       }
2324    }
2325 
2326    emit_mir_instruction(ctx, ins);
2327 }
2328 
2329 static void
emit_tex(compiler_context * ctx,nir_tex_instr * instr)2330 emit_tex(compiler_context *ctx, nir_tex_instr *instr)
2331 {
2332    switch (instr->op) {
2333    case nir_texop_tex:
2334    case nir_texop_txb:
2335       emit_texop_native(ctx, instr, midgard_tex_op_normal);
2336       break;
2337    case nir_texop_txl:
2338    case nir_texop_tg4:
2339       emit_texop_native(ctx, instr, midgard_tex_op_gradient);
2340       break;
2341    case nir_texop_txf:
2342    case nir_texop_txf_ms:
2343       emit_texop_native(ctx, instr, midgard_tex_op_fetch);
2344       break;
2345    default: {
2346       fprintf(stderr, "Unhandled texture op: %d\n", instr->op);
2347       assert(0);
2348    }
2349    }
2350 }
2351 
2352 static void
emit_jump(compiler_context * ctx,nir_jump_instr * instr)2353 emit_jump(compiler_context *ctx, nir_jump_instr *instr)
2354 {
2355    switch (instr->type) {
2356    case nir_jump_break: {
2357       /* Emit a branch out of the loop */
2358       struct midgard_instruction br = v_branch(false, false);
2359       br.branch.target_type = TARGET_BREAK;
2360       br.branch.target_break = ctx->current_loop_depth;
2361       emit_mir_instruction(ctx, br);
2362       break;
2363    }
2364 
2365    default:
2366       unreachable("Unhandled jump");
2367    }
2368 }
2369 
2370 static void
emit_instr(compiler_context * ctx,struct nir_instr * instr)2371 emit_instr(compiler_context *ctx, struct nir_instr *instr)
2372 {
2373    switch (instr->type) {
2374    case nir_instr_type_load_const:
2375       emit_load_const(ctx, nir_instr_as_load_const(instr));
2376       break;
2377 
2378    case nir_instr_type_intrinsic:
2379       emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
2380       break;
2381 
2382    case nir_instr_type_alu:
2383       emit_alu(ctx, nir_instr_as_alu(instr));
2384       break;
2385 
2386    case nir_instr_type_tex:
2387       emit_tex(ctx, nir_instr_as_tex(instr));
2388       break;
2389 
2390    case nir_instr_type_jump:
2391       emit_jump(ctx, nir_instr_as_jump(instr));
2392       break;
2393 
2394    case nir_instr_type_undef:
2395       /* Spurious */
2396       break;
2397 
2398    default:
2399       unreachable("Unhandled instruction type");
2400    }
2401 }
2402 
2403 /* ALU instructions can inline or embed constants, which decreases register
2404  * pressure and saves space. */
2405 
2406 #define CONDITIONAL_ATTACH(idx)                                                \
2407    {                                                                           \
2408       void *entry =                                                            \
2409          _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1);   \
2410                                                                                \
2411       if (entry) {                                                             \
2412          attach_constants(ctx, alu, entry, alu->src[idx] + 1);                 \
2413          alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);                \
2414       }                                                                        \
2415    }
2416 
2417 static void
inline_alu_constants(compiler_context * ctx,midgard_block * block)2418 inline_alu_constants(compiler_context *ctx, midgard_block *block)
2419 {
2420    mir_foreach_instr_in_block(block, alu) {
2421       /* Other instructions cannot inline constants */
2422       if (alu->type != TAG_ALU_4)
2423          continue;
2424       if (alu->compact_branch)
2425          continue;
2426 
2427       /* If there is already a constant here, we can do nothing */
2428       if (alu->has_constants)
2429          continue;
2430 
2431       CONDITIONAL_ATTACH(0);
2432 
2433       if (!alu->has_constants) {
2434          CONDITIONAL_ATTACH(1)
2435       } else if (!alu->inline_constant) {
2436          /* Corner case: _two_ vec4 constants, for instance with a
2437           * csel. For this case, we can only use a constant
2438           * register for one, we'll have to emit a move for the
2439           * other. */
2440 
2441          void *entry =
2442             _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[1] + 1);
2443          unsigned scratch = make_compiler_temp(ctx);
2444 
2445          if (entry) {
2446             midgard_instruction ins =
2447                v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), scratch);
2448             attach_constants(ctx, &ins, entry, alu->src[1] + 1);
2449 
2450             /* Set the source */
2451             alu->src[1] = scratch;
2452 
2453             /* Inject us -before- the last instruction which set r31, if
2454              * possible.
2455              */
2456             midgard_instruction *first = list_first_entry(
2457                &block->base.instructions, midgard_instruction, link);
2458 
2459             if (alu == first) {
2460                mir_insert_instruction_before(ctx, alu, ins);
2461             } else {
2462                mir_insert_instruction_before(ctx, mir_prev_op(alu), ins);
2463             }
2464          }
2465       }
2466    }
2467 }
2468 
2469 unsigned
max_bitsize_for_alu(midgard_instruction * ins)2470 max_bitsize_for_alu(midgard_instruction *ins)
2471 {
2472    unsigned max_bitsize = 0;
2473    for (int i = 0; i < MIR_SRC_COUNT; i++) {
2474       if (ins->src[i] == ~0)
2475          continue;
2476       unsigned src_bitsize = nir_alu_type_get_type_size(ins->src_types[i]);
2477       max_bitsize = MAX2(src_bitsize, max_bitsize);
2478    }
2479    unsigned dst_bitsize = nir_alu_type_get_type_size(ins->dest_type);
2480    max_bitsize = MAX2(dst_bitsize, max_bitsize);
2481 
2482    /* We emulate 8-bit as 16-bit for simplicity of packing */
2483    max_bitsize = MAX2(max_bitsize, 16);
2484 
2485    /* We don't have fp16 LUTs, so we'll want to emit code like:
2486     *
2487     *      vlut.fsinr hr0, hr0
2488     *
2489     * where both input and output are 16-bit but the operation is carried
2490     * out in 32-bit
2491     */
2492 
2493    switch (ins->op) {
2494    case midgard_alu_op_fsqrt:
2495    case midgard_alu_op_frcp:
2496    case midgard_alu_op_frsqrt:
2497    case midgard_alu_op_fsinpi:
2498    case midgard_alu_op_fcospi:
2499    case midgard_alu_op_fexp2:
2500    case midgard_alu_op_flog2:
2501       max_bitsize = MAX2(max_bitsize, 32);
2502       break;
2503 
2504    default:
2505       break;
2506    }
2507 
2508    /* High implies computing at a higher bitsize, e.g umul_high of 32-bit
2509     * requires computing at 64-bit */
2510    if (midgard_is_integer_out_op(ins->op) &&
2511        ins->outmod == midgard_outmod_keephi) {
2512       max_bitsize *= 2;
2513       assert(max_bitsize <= 64);
2514    }
2515 
2516    return max_bitsize;
2517 }
2518 
2519 midgard_reg_mode
reg_mode_for_bitsize(unsigned bitsize)2520 reg_mode_for_bitsize(unsigned bitsize)
2521 {
2522    switch (bitsize) {
2523       /* use 16 pipe for 8 since we don't support vec16 yet */
2524    case 8:
2525    case 16:
2526       return midgard_reg_mode_16;
2527    case 32:
2528       return midgard_reg_mode_32;
2529    case 64:
2530       return midgard_reg_mode_64;
2531    default:
2532       unreachable("invalid bit size");
2533    }
2534 }
2535 
2536 /* Midgard supports two types of constants, embedded constants (128-bit) and
2537  * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2538  * constants can be demoted to inline constants, for space savings and
2539  * sometimes a performance boost */
2540 
2541 static void
embedded_to_inline_constant(compiler_context * ctx,midgard_block * block)2542 embedded_to_inline_constant(compiler_context *ctx, midgard_block *block)
2543 {
2544    mir_foreach_instr_in_block(block, ins) {
2545       if (!ins->has_constants)
2546          continue;
2547       if (ins->has_inline_constant)
2548          continue;
2549 
2550       unsigned max_bitsize = max_bitsize_for_alu(ins);
2551 
2552       /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2553       bool is_16 = max_bitsize == 16;
2554       bool is_32 = max_bitsize == 32;
2555 
2556       if (!(is_16 || is_32))
2557          continue;
2558 
2559       /* src1 cannot be an inline constant due to encoding
2560        * restrictions. So, if possible we try to flip the arguments
2561        * in that case */
2562 
2563       int op = ins->op;
2564 
2565       if (ins->src[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT) &&
2566           alu_opcode_props[op].props & OP_COMMUTES) {
2567          mir_flip(ins);
2568       }
2569 
2570       if (ins->src[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
2571          /* Component is from the swizzle. Take a nonzero component */
2572          assert(ins->mask);
2573          unsigned first_comp = ffs(ins->mask) - 1;
2574          unsigned component = ins->swizzle[1][first_comp];
2575 
2576          /* Scale constant appropriately, if we can legally */
2577          int16_t scaled_constant = 0;
2578 
2579          if (is_16) {
2580             scaled_constant = ins->constants.u16[component];
2581          } else if (midgard_is_integer_op(op)) {
2582             scaled_constant = ins->constants.u32[component];
2583 
2584             /* Constant overflow after resize */
2585             if (scaled_constant != ins->constants.u32[component])
2586                continue;
2587          } else {
2588             float original = ins->constants.f32[component];
2589             scaled_constant = _mesa_float_to_half(original);
2590 
2591             /* Check for loss of precision. If this is
2592              * mediump, we don't care, but for a highp
2593              * shader, we need to pay attention. NIR
2594              * doesn't yet tell us which mode we're in!
2595              * Practically this prevents most constants
2596              * from being inlined, sadly. */
2597 
2598             float fp32 = _mesa_half_to_float(scaled_constant);
2599 
2600             if (fp32 != original)
2601                continue;
2602          }
2603 
2604          /* Should've been const folded */
2605          if (ins->src_abs[1] || ins->src_neg[1])
2606             continue;
2607 
2608          /* Make sure that the constant is not itself a vector
2609           * by checking if all accessed values are the same. */
2610 
2611          const midgard_constants *cons = &ins->constants;
2612          uint32_t value = is_16 ? cons->u16[component] : cons->u32[component];
2613 
2614          bool is_vector = false;
2615          unsigned mask = effective_writemask(ins->op, ins->mask);
2616 
2617          for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c) {
2618             /* We only care if this component is actually used */
2619             if (!(mask & (1 << c)))
2620                continue;
2621 
2622             uint32_t test = is_16 ? cons->u16[ins->swizzle[1][c]]
2623                                   : cons->u32[ins->swizzle[1][c]];
2624 
2625             if (test != value) {
2626                is_vector = true;
2627                break;
2628             }
2629          }
2630 
2631          if (is_vector)
2632             continue;
2633 
2634          /* Get rid of the embedded constant */
2635          ins->has_constants = false;
2636          ins->src[1] = ~0;
2637          ins->has_inline_constant = true;
2638          ins->inline_constant = scaled_constant;
2639       }
2640    }
2641 }
2642 
2643 /* Dead code elimination for branches at the end of a block - only one branch
2644  * per block is legal semantically */
2645 
2646 static void
midgard_cull_dead_branch(compiler_context * ctx,midgard_block * block)2647 midgard_cull_dead_branch(compiler_context *ctx, midgard_block *block)
2648 {
2649    bool branched = false;
2650 
2651    mir_foreach_instr_in_block_safe(block, ins) {
2652       if (!midgard_is_branch_unit(ins->unit))
2653          continue;
2654 
2655       if (branched)
2656          mir_remove_instruction(ins);
2657 
2658       branched = true;
2659    }
2660 }
2661 
2662 /* We want to force the invert on AND/OR to the second slot to legalize into
2663  * iandnot/iornot. The relevant patterns are for AND (and OR respectively)
2664  *
2665  *   ~a & #b = ~a & ~(#~b)
2666  *   ~a & b = b & ~a
2667  */
2668 
2669 static void
midgard_legalize_invert(compiler_context * ctx,midgard_block * block)2670 midgard_legalize_invert(compiler_context *ctx, midgard_block *block)
2671 {
2672    mir_foreach_instr_in_block(block, ins) {
2673       if (ins->type != TAG_ALU_4)
2674          continue;
2675 
2676       if (ins->op != midgard_alu_op_iand && ins->op != midgard_alu_op_ior)
2677          continue;
2678 
2679       if (ins->src_invert[1] || !ins->src_invert[0])
2680          continue;
2681 
2682       if (ins->has_inline_constant) {
2683          /* ~(#~a) = ~(~#a) = a, so valid, and forces both
2684           * inverts on */
2685          ins->inline_constant = ~ins->inline_constant;
2686          ins->src_invert[1] = true;
2687       } else {
2688          /* Flip to the right invert order. Note
2689           * has_inline_constant false by assumption on the
2690           * branch, so flipping makes sense. */
2691          mir_flip(ins);
2692       }
2693    }
2694 }
2695 
2696 static unsigned
emit_fragment_epilogue(compiler_context * ctx,unsigned rt,unsigned sample_iter)2697 emit_fragment_epilogue(compiler_context *ctx, unsigned rt, unsigned sample_iter)
2698 {
2699    /* Loop to ourselves */
2700    midgard_instruction *br = ctx->writeout_branch[rt][sample_iter];
2701    struct midgard_instruction ins = v_branch(false, false);
2702    ins.writeout = br->writeout;
2703    ins.branch.target_block = ctx->block_count - 1;
2704    ins.constants.u32[0] = br->constants.u32[0];
2705    memcpy(&ins.src_types, &br->src_types, sizeof(ins.src_types));
2706    emit_mir_instruction(ctx, ins);
2707 
2708    ctx->current_block->epilogue = true;
2709    schedule_barrier(ctx);
2710    return ins.branch.target_block;
2711 }
2712 
2713 static midgard_block *
emit_block_init(compiler_context * ctx)2714 emit_block_init(compiler_context *ctx)
2715 {
2716    midgard_block *this_block = ctx->after_block;
2717    ctx->after_block = NULL;
2718 
2719    if (!this_block)
2720       this_block = create_empty_block(ctx);
2721 
2722    list_addtail(&this_block->base.link, &ctx->blocks);
2723 
2724    this_block->scheduled = false;
2725    ++ctx->block_count;
2726 
2727    /* Set up current block */
2728    list_inithead(&this_block->base.instructions);
2729    ctx->current_block = this_block;
2730 
2731    return this_block;
2732 }
2733 
2734 static midgard_block *
emit_block(compiler_context * ctx,nir_block * block)2735 emit_block(compiler_context *ctx, nir_block *block)
2736 {
2737    midgard_block *this_block = emit_block_init(ctx);
2738 
2739    nir_foreach_instr(instr, block) {
2740       emit_instr(ctx, instr);
2741       ++ctx->instruction_count;
2742    }
2743 
2744    return this_block;
2745 }
2746 
2747 static midgard_block *emit_cf_list(struct compiler_context *ctx,
2748                                    struct exec_list *list);
2749 
2750 static void
emit_if(struct compiler_context * ctx,nir_if * nif)2751 emit_if(struct compiler_context *ctx, nir_if *nif)
2752 {
2753    midgard_block *before_block = ctx->current_block;
2754 
2755    /* Speculatively emit the branch, but we can't fill it in until later */
2756    EMIT(branch, true, true);
2757    midgard_instruction *then_branch = mir_last_in_block(ctx->current_block);
2758    then_branch->src[0] = nir_src_index(ctx, &nif->condition);
2759    then_branch->src_types[0] = nir_type_uint32;
2760 
2761    /* Emit the two subblocks. */
2762    midgard_block *then_block = emit_cf_list(ctx, &nif->then_list);
2763    midgard_block *end_then_block = ctx->current_block;
2764 
2765    /* Emit a jump from the end of the then block to the end of the else */
2766    EMIT(branch, false, false);
2767    midgard_instruction *then_exit = mir_last_in_block(ctx->current_block);
2768 
2769    /* Emit second block, and check if it's empty */
2770 
2771    int else_idx = ctx->block_count;
2772    int count_in = ctx->instruction_count;
2773    midgard_block *else_block = emit_cf_list(ctx, &nif->else_list);
2774    midgard_block *end_else_block = ctx->current_block;
2775    int after_else_idx = ctx->block_count;
2776 
2777    /* Now that we have the subblocks emitted, fix up the branches */
2778 
2779    assert(then_block);
2780    assert(else_block);
2781 
2782    if (ctx->instruction_count == count_in) {
2783       /* The else block is empty, so don't emit an exit jump */
2784       mir_remove_instruction(then_exit);
2785       then_branch->branch.target_block = after_else_idx;
2786    } else {
2787       then_branch->branch.target_block = else_idx;
2788       then_exit->branch.target_block = after_else_idx;
2789    }
2790 
2791    /* Wire up the successors */
2792 
2793    ctx->after_block = create_empty_block(ctx);
2794 
2795    pan_block_add_successor(&before_block->base, &then_block->base);
2796    pan_block_add_successor(&before_block->base, &else_block->base);
2797 
2798    pan_block_add_successor(&end_then_block->base, &ctx->after_block->base);
2799    pan_block_add_successor(&end_else_block->base, &ctx->after_block->base);
2800 }
2801 
2802 static void
emit_loop(struct compiler_context * ctx,nir_loop * nloop)2803 emit_loop(struct compiler_context *ctx, nir_loop *nloop)
2804 {
2805    assert(!nir_loop_has_continue_construct(nloop));
2806 
2807    /* Remember where we are */
2808    midgard_block *start_block = ctx->current_block;
2809 
2810    /* Allocate a loop number, growing the current inner loop depth */
2811    int loop_idx = ++ctx->current_loop_depth;
2812 
2813    /* Get index from before the body so we can loop back later */
2814    int start_idx = ctx->block_count;
2815 
2816    /* Emit the body itself */
2817    midgard_block *loop_block = emit_cf_list(ctx, &nloop->body);
2818 
2819    /* Branch back to loop back */
2820    struct midgard_instruction br_back = v_branch(false, false);
2821    br_back.branch.target_block = start_idx;
2822    emit_mir_instruction(ctx, br_back);
2823 
2824    /* Mark down that branch in the graph. */
2825    pan_block_add_successor(&start_block->base, &loop_block->base);
2826    pan_block_add_successor(&ctx->current_block->base, &loop_block->base);
2827 
2828    /* Find the index of the block about to follow us (note: we don't add
2829     * one; blocks are 0-indexed so we get a fencepost problem) */
2830    int break_block_idx = ctx->block_count;
2831 
2832    /* Fix up the break statements we emitted to point to the right place,
2833     * now that we can allocate a block number for them */
2834    ctx->after_block = create_empty_block(ctx);
2835 
2836    mir_foreach_block_from(ctx, start_block, _block) {
2837       mir_foreach_instr_in_block(((midgard_block *)_block), ins) {
2838          if (ins->type != TAG_ALU_4)
2839             continue;
2840          if (!ins->compact_branch)
2841             continue;
2842 
2843          /* We found a branch -- check the type to see if we need to do anything
2844           */
2845          if (ins->branch.target_type != TARGET_BREAK)
2846             continue;
2847 
2848          /* It's a break! Check if it's our break */
2849          if (ins->branch.target_break != loop_idx)
2850             continue;
2851 
2852          /* Okay, cool, we're breaking out of this loop.
2853           * Rewrite from a break to a goto */
2854 
2855          ins->branch.target_type = TARGET_GOTO;
2856          ins->branch.target_block = break_block_idx;
2857 
2858          pan_block_add_successor(_block, &ctx->after_block->base);
2859       }
2860    }
2861 
2862    /* Now that we've finished emitting the loop, free up the depth again
2863     * so we play nice with recursion amid nested loops */
2864    --ctx->current_loop_depth;
2865 
2866    /* Dump loop stats */
2867    ++ctx->loop_count;
2868 }
2869 
2870 static midgard_block *
emit_cf_list(struct compiler_context * ctx,struct exec_list * list)2871 emit_cf_list(struct compiler_context *ctx, struct exec_list *list)
2872 {
2873    midgard_block *start_block = NULL;
2874 
2875    foreach_list_typed(nir_cf_node, node, node, list) {
2876       switch (node->type) {
2877       case nir_cf_node_block: {
2878          midgard_block *block = emit_block(ctx, nir_cf_node_as_block(node));
2879 
2880          if (!start_block)
2881             start_block = block;
2882 
2883          break;
2884       }
2885 
2886       case nir_cf_node_if:
2887          emit_if(ctx, nir_cf_node_as_if(node));
2888          break;
2889 
2890       case nir_cf_node_loop:
2891          emit_loop(ctx, nir_cf_node_as_loop(node));
2892          break;
2893 
2894       case nir_cf_node_function:
2895          assert(0);
2896          break;
2897       }
2898    }
2899 
2900    return start_block;
2901 }
2902 
2903 /* Due to lookahead, we need to report the first tag executed in the command
2904  * stream and in branch targets. An initial block might be empty, so iterate
2905  * until we find one that 'works' */
2906 
2907 unsigned
midgard_get_first_tag_from_block(compiler_context * ctx,unsigned block_idx)2908 midgard_get_first_tag_from_block(compiler_context *ctx, unsigned block_idx)
2909 {
2910    midgard_block *initial_block = mir_get_block(ctx, block_idx);
2911 
2912    mir_foreach_block_from(ctx, initial_block, _v) {
2913       midgard_block *v = (midgard_block *)_v;
2914       if (v->quadword_count) {
2915          midgard_bundle *initial_bundle =
2916             util_dynarray_element(&v->bundles, midgard_bundle, 0);
2917 
2918          return initial_bundle->tag;
2919       }
2920    }
2921 
2922    /* Default to a tag 1 which will break from the shader, in case we jump
2923     * to the exit block (i.e. `return` in a compute shader) */
2924 
2925    return 1;
2926 }
2927 
2928 /* For each fragment writeout instruction, generate a writeout loop to
2929  * associate with it */
2930 
2931 static void
mir_add_writeout_loops(compiler_context * ctx)2932 mir_add_writeout_loops(compiler_context *ctx)
2933 {
2934    for (unsigned rt = 0; rt < ARRAY_SIZE(ctx->writeout_branch); ++rt) {
2935       for (unsigned s = 0; s < MIDGARD_MAX_SAMPLE_ITER; ++s) {
2936          midgard_instruction *br = ctx->writeout_branch[rt][s];
2937          if (!br)
2938             continue;
2939 
2940          unsigned popped = br->branch.target_block;
2941          pan_block_add_successor(&(mir_get_block(ctx, popped - 1)->base),
2942                                  &ctx->current_block->base);
2943          br->branch.target_block = emit_fragment_epilogue(ctx, rt, s);
2944          br->branch.target_type = TARGET_GOTO;
2945 
2946          /* If we have more RTs, we'll need to restore back after our
2947           * loop terminates */
2948          midgard_instruction *next_br = NULL;
2949 
2950          if ((s + 1) < MIDGARD_MAX_SAMPLE_ITER)
2951             next_br = ctx->writeout_branch[rt][s + 1];
2952 
2953          if (!next_br && (rt + 1) < ARRAY_SIZE(ctx->writeout_branch))
2954             next_br = ctx->writeout_branch[rt + 1][0];
2955 
2956          if (next_br) {
2957             midgard_instruction uncond = v_branch(false, false);
2958             uncond.branch.target_block = popped;
2959             uncond.branch.target_type = TARGET_GOTO;
2960             emit_mir_instruction(ctx, uncond);
2961             pan_block_add_successor(&ctx->current_block->base,
2962                                     &(mir_get_block(ctx, popped)->base));
2963             schedule_barrier(ctx);
2964          } else {
2965             /* We're last, so we can terminate here */
2966             br->last_writeout = true;
2967          }
2968       }
2969    }
2970 }
2971 
2972 void
midgard_compile_shader_nir(nir_shader * nir,const struct panfrost_compile_inputs * inputs,struct util_dynarray * binary,struct pan_shader_info * info)2973 midgard_compile_shader_nir(nir_shader *nir,
2974                            const struct panfrost_compile_inputs *inputs,
2975                            struct util_dynarray *binary,
2976                            struct pan_shader_info *info)
2977 {
2978    midgard_debug = debug_get_option_midgard_debug();
2979 
2980    /* TODO: Bound against what? */
2981    compiler_context *ctx = rzalloc(NULL, compiler_context);
2982 
2983    ctx->inputs = inputs;
2984    ctx->nir = nir;
2985    ctx->info = info;
2986    ctx->stage = nir->info.stage;
2987    ctx->blend_input = ~0;
2988    ctx->blend_src1 = ~0;
2989    ctx->quirks = midgard_get_quirks(inputs->gpu_id);
2990 
2991    /* Initialize at a global (not block) level hash tables */
2992 
2993    ctx->ssa_constants = _mesa_hash_table_u64_create(ctx);
2994 
2995    /* Collect varyings after lowering I/O */
2996    info->quirk_no_auto32 = (ctx->quirks & MIDGARD_NO_AUTO32);
2997    pan_nir_collect_varyings(nir, info);
2998 
2999    /* Optimisation passes */
3000    optimise_nir(nir, ctx->quirks, inputs->is_blend);
3001 
3002    bool skip_internal = nir->info.internal;
3003    skip_internal &= !(midgard_debug & MIDGARD_DBG_INTERNAL);
3004 
3005    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
3006       nir_print_shader(nir, stdout);
3007 
3008    info->tls_size = nir->scratch_size;
3009 
3010    nir_foreach_function_with_impl(func, impl, nir) {
3011       list_inithead(&ctx->blocks);
3012       ctx->block_count = 0;
3013       ctx->func = func;
3014 
3015       if (nir->info.outputs_read && !inputs->is_blend) {
3016          emit_block_init(ctx);
3017 
3018          struct midgard_instruction wait = v_branch(false, false);
3019          wait.branch.target_type = TARGET_TILEBUF_WAIT;
3020 
3021          emit_mir_instruction(ctx, wait);
3022 
3023          ++ctx->instruction_count;
3024       }
3025 
3026       emit_cf_list(ctx, &impl->body);
3027       break; /* TODO: Multi-function shaders */
3028    }
3029 
3030    /* Per-block lowering before opts */
3031 
3032    mir_foreach_block(ctx, _block) {
3033       midgard_block *block = (midgard_block *)_block;
3034       inline_alu_constants(ctx, block);
3035       embedded_to_inline_constant(ctx, block);
3036    }
3037    /* MIR-level optimizations */
3038 
3039    bool progress = false;
3040 
3041    do {
3042       progress = false;
3043       progress |= midgard_opt_dead_code_eliminate(ctx);
3044       progress |= midgard_opt_prop(ctx);
3045 
3046       mir_foreach_block(ctx, _block) {
3047          midgard_block *block = (midgard_block *)_block;
3048          progress |= midgard_opt_copy_prop(ctx, block);
3049          progress |= midgard_opt_combine_projection(ctx, block);
3050          progress |= midgard_opt_varying_projection(ctx, block);
3051       }
3052    } while (progress);
3053 
3054    mir_foreach_block(ctx, _block) {
3055       midgard_block *block = (midgard_block *)_block;
3056       midgard_lower_derivatives(ctx, block);
3057       midgard_legalize_invert(ctx, block);
3058       midgard_cull_dead_branch(ctx, block);
3059    }
3060 
3061    if (ctx->stage == MESA_SHADER_FRAGMENT)
3062       mir_add_writeout_loops(ctx);
3063 
3064    /* Analyze now that the code is known but before scheduling creates
3065     * pipeline registers which are harder to track */
3066    mir_analyze_helper_requirements(ctx);
3067 
3068    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
3069       mir_print_shader(ctx);
3070 
3071    /* Schedule! */
3072    midgard_schedule_program(ctx);
3073    mir_ra(ctx);
3074 
3075    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal)
3076       mir_print_shader(ctx);
3077 
3078    /* Analyze after scheduling since this is order-dependent */
3079    mir_analyze_helper_terminate(ctx);
3080 
3081    /* Emit flat binary from the instruction arrays. Iterate each block in
3082     * sequence. Save instruction boundaries such that lookahead tags can
3083     * be assigned easily */
3084 
3085    /* Cache _all_ bundles in source order for lookahead across failed branches */
3086 
3087    int bundle_count = 0;
3088    mir_foreach_block(ctx, _block) {
3089       midgard_block *block = (midgard_block *)_block;
3090       bundle_count += block->bundles.size / sizeof(midgard_bundle);
3091    }
3092    midgard_bundle **source_order_bundles =
3093       malloc(sizeof(midgard_bundle *) * bundle_count);
3094    int bundle_idx = 0;
3095    mir_foreach_block(ctx, _block) {
3096       midgard_block *block = (midgard_block *)_block;
3097       util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
3098          source_order_bundles[bundle_idx++] = bundle;
3099       }
3100    }
3101 
3102    int current_bundle = 0;
3103 
3104    /* Midgard prefetches instruction types, so during emission we
3105     * need to lookahead. Unless this is the last instruction, in
3106     * which we return 1. */
3107 
3108    mir_foreach_block(ctx, _block) {
3109       midgard_block *block = (midgard_block *)_block;
3110       mir_foreach_bundle_in_block(block, bundle) {
3111          int lookahead = 1;
3112 
3113          if (!bundle->last_writeout && (current_bundle + 1 < bundle_count))
3114             lookahead = source_order_bundles[current_bundle + 1]->tag;
3115 
3116          emit_binary_bundle(ctx, block, bundle, binary, lookahead);
3117          ++current_bundle;
3118       }
3119 
3120       /* TODO: Free deeper */
3121       // util_dynarray_fini(&block->instructions);
3122    }
3123 
3124    free(source_order_bundles);
3125 
3126    /* Report the very first tag executed */
3127    info->midgard.first_tag = midgard_get_first_tag_from_block(ctx, 0);
3128 
3129    info->ubo_mask = ctx->ubo_mask & ((1 << ctx->nir->info.num_ubos) - 1);
3130 
3131    if (midgard_debug & MIDGARD_DBG_SHADERS && !skip_internal) {
3132       disassemble_midgard(stdout, binary->data, binary->size, inputs->gpu_id,
3133                           midgard_debug & MIDGARD_DBG_VERBOSE);
3134       fflush(stdout);
3135    }
3136 
3137    /* A shader ending on a 16MB boundary causes INSTR_INVALID_PC faults,
3138     * workaround by adding some padding to the end of the shader. (The
3139     * kernel makes sure shader BOs can't cross 16MB boundaries.) */
3140    if (binary->size)
3141       memset(util_dynarray_grow(binary, uint8_t, 16), 0, 16);
3142 
3143    if ((midgard_debug & MIDGARD_DBG_SHADERDB || inputs->debug) &&
3144        !nir->info.internal) {
3145       unsigned nr_bundles = 0, nr_ins = 0;
3146 
3147       /* Count instructions and bundles */
3148 
3149       mir_foreach_block(ctx, _block) {
3150          midgard_block *block = (midgard_block *)_block;
3151          nr_bundles +=
3152             util_dynarray_num_elements(&block->bundles, midgard_bundle);
3153 
3154          mir_foreach_bundle_in_block(block, bun)
3155             nr_ins += bun->instruction_count;
3156       }
3157 
3158       /* Calculate thread count. There are certain cutoffs by
3159        * register count for thread count */
3160 
3161       unsigned nr_registers = info->work_reg_count;
3162 
3163       unsigned nr_threads = (nr_registers <= 4)   ? 4
3164                             : (nr_registers <= 8) ? 2
3165                                                   : 1;
3166 
3167       char *shaderdb = NULL;
3168 
3169       /* Dump stats */
3170 
3171       asprintf(&shaderdb,
3172                "%s shader: "
3173                "%u inst, %u bundles, %u quadwords, "
3174                "%u registers, %u threads, %u loops, "
3175                "%u:%u spills:fills",
3176                ctx->inputs->is_blend ? "PAN_SHADER_BLEND"
3177                                      : gl_shader_stage_name(ctx->stage),
3178                nr_ins, nr_bundles, ctx->quadword_count, nr_registers,
3179                nr_threads, ctx->loop_count, ctx->spills, ctx->fills);
3180 
3181       if (midgard_debug & MIDGARD_DBG_SHADERDB)
3182          fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
3183 
3184       if (inputs->debug)
3185          util_debug_message(inputs->debug, SHADER_INFO, "%s", shaderdb);
3186 
3187       free(shaderdb);
3188    }
3189 
3190    _mesa_hash_table_u64_destroy(ctx->ssa_constants);
3191    ralloc_free(ctx);
3192 }
3193