xref: /third_party/mesa3d/src/freedreno/ir3/ir3_nir.c (revision bf215546)
1/*
2 * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#include "util/debug.h"
28#include "util/u_math.h"
29
30#include "ir3_compiler.h"
31#include "ir3_nir.h"
32#include "ir3_shader.h"
33
34static bool
35ir3_nir_should_vectorize_mem(unsigned align_mul, unsigned align_offset,
36                             unsigned bit_size, unsigned num_components,
37                             nir_intrinsic_instr *low,
38                             nir_intrinsic_instr *high, void *data)
39{
40   unsigned byte_size = bit_size / 8;
41
42   if (low->intrinsic != nir_intrinsic_load_ubo) {
43      return bit_size <= 32 && align_mul >= byte_size &&
44         align_offset % byte_size == 0 &&
45         num_components <= 4;
46   }
47
48   assert(bit_size >= 8);
49   if (bit_size != 32)
50      return false;
51
52   int size = num_components * byte_size;
53
54   /* Don't care about alignment past vec4. */
55   assert(util_is_power_of_two_nonzero(align_mul));
56   align_mul = MIN2(align_mul, 16);
57   align_offset &= 15;
58
59   /* Our offset alignment should aways be at least 4 bytes */
60   if (align_mul < 4)
61      return false;
62
63   unsigned worst_start_offset = 16 - align_mul + align_offset;
64   if (worst_start_offset + size > 16)
65      return false;
66
67   return true;
68}
69
70#define OPT(nir, pass, ...)                                                    \
71   ({                                                                          \
72      bool this_progress = false;                                              \
73      NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);                       \
74      this_progress;                                                           \
75   })
76
77#define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
78
79void
80ir3_optimize_loop(struct ir3_compiler *compiler, nir_shader *s)
81{
82   bool progress;
83   unsigned lower_flrp = (s->options->lower_flrp16 ? 16 : 0) |
84                         (s->options->lower_flrp32 ? 32 : 0) |
85                         (s->options->lower_flrp64 ? 64 : 0);
86
87   do {
88      progress = false;
89
90      OPT_V(s, nir_lower_vars_to_ssa);
91      progress |= OPT(s, nir_lower_alu_to_scalar, NULL, NULL);
92      progress |= OPT(s, nir_lower_phis_to_scalar, false);
93
94      progress |= OPT(s, nir_copy_prop);
95      progress |= OPT(s, nir_opt_deref);
96      progress |= OPT(s, nir_opt_dce);
97      progress |= OPT(s, nir_opt_cse);
98
99      progress |= OPT(s, nir_opt_find_array_copies);
100      progress |= OPT(s, nir_opt_copy_prop_vars);
101      progress |= OPT(s, nir_opt_dead_write_vars);
102
103      static int gcm = -1;
104      if (gcm == -1)
105         gcm = env_var_as_unsigned("GCM", 0);
106      if (gcm == 1)
107         progress |= OPT(s, nir_opt_gcm, true);
108      else if (gcm == 2)
109         progress |= OPT(s, nir_opt_gcm, false);
110      progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
111      progress |= OPT(s, nir_opt_intrinsics);
112      /* NOTE: GS lowering inserts an output var with varying slot that
113       * is larger than VARYING_SLOT_MAX (ie. GS_VERTEX_FLAGS_IR3),
114       * which triggers asserts in nir_shader_gather_info().  To work
115       * around that skip lowering phi precision for GS.
116       *
117       * Calling nir_shader_gather_info() late also seems to cause
118       * problems for tess lowering, for now since we only enable
119       * fp16/int16 for frag and compute, skip phi precision lowering
120       * for other stages.
121       */
122      if ((s->info.stage == MESA_SHADER_FRAGMENT) ||
123          (s->info.stage == MESA_SHADER_COMPUTE) ||
124          (s->info.stage == MESA_SHADER_KERNEL)) {
125         progress |= OPT(s, nir_opt_phi_precision);
126      }
127      progress |= OPT(s, nir_opt_algebraic);
128      progress |= OPT(s, nir_lower_alu);
129      progress |= OPT(s, nir_lower_pack);
130      progress |= OPT(s, nir_opt_constant_folding);
131
132      static const nir_opt_offsets_options offset_options = {
133         /* How large an offset we can encode in the instr's immediate field.
134          */
135         .uniform_max = (1 << 9) - 1,
136
137         .shared_max = (1 << 13) - 1,
138
139         .buffer_max = ~0,
140      };
141      progress |= OPT(s, nir_opt_offsets, &offset_options);
142
143      nir_load_store_vectorize_options vectorize_opts = {
144         .modes = nir_var_mem_ubo | nir_var_mem_ssbo,
145         .callback = ir3_nir_should_vectorize_mem,
146         .robust_modes = compiler->robust_buffer_access2 ? nir_var_mem_ubo | nir_var_mem_ssbo: 0,
147      };
148      progress |= OPT(s, nir_opt_load_store_vectorize, &vectorize_opts);
149
150      if (lower_flrp != 0) {
151         if (OPT(s, nir_lower_flrp, lower_flrp, false /* always_precise */)) {
152            OPT(s, nir_opt_constant_folding);
153            progress = true;
154         }
155
156         /* Nothing should rematerialize any flrps, so we only
157          * need to do this lowering once.
158          */
159         lower_flrp = 0;
160      }
161
162      progress |= OPT(s, nir_opt_dead_cf);
163      if (OPT(s, nir_opt_trivial_continues)) {
164         progress |= true;
165         /* If nir_opt_trivial_continues makes progress, then we need to clean
166          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
167          * to make progress.
168          */
169         OPT(s, nir_copy_prop);
170         OPT(s, nir_opt_dce);
171      }
172      progress |= OPT(s, nir_opt_if, nir_opt_if_optimize_phi_true_false);
173      progress |= OPT(s, nir_opt_loop_unroll);
174      progress |= OPT(s, nir_lower_64bit_phis);
175      progress |= OPT(s, nir_opt_remove_phis);
176      progress |= OPT(s, nir_opt_undef);
177   } while (progress);
178
179   OPT(s, nir_lower_var_copies);
180}
181
182static bool
183should_split_wrmask(const nir_instr *instr, const void *data)
184{
185   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
186
187   switch (intr->intrinsic) {
188   case nir_intrinsic_store_ssbo:
189   case nir_intrinsic_store_shared:
190   case nir_intrinsic_store_global:
191   case nir_intrinsic_store_scratch:
192      return true;
193   default:
194      return false;
195   }
196}
197
198static bool
199ir3_nir_lower_ssbo_size_filter(const nir_instr *instr, const void *data)
200{
201   return instr->type == nir_instr_type_intrinsic &&
202          nir_instr_as_intrinsic(instr)->intrinsic ==
203             nir_intrinsic_get_ssbo_size;
204}
205
206static nir_ssa_def *
207ir3_nir_lower_ssbo_size_instr(nir_builder *b, nir_instr *instr, void *data)
208{
209   uint8_t ssbo_size_to_bytes_shift = *(uint8_t *) data;
210   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
211   return nir_ishl(b, &intr->dest.ssa, nir_imm_int(b, ssbo_size_to_bytes_shift));
212}
213
214static bool
215ir3_nir_lower_ssbo_size(nir_shader *s, uint8_t ssbo_size_to_bytes_shift)
216{
217   return nir_shader_lower_instructions(s, ir3_nir_lower_ssbo_size_filter,
218                                        ir3_nir_lower_ssbo_size_instr,
219                                        &ssbo_size_to_bytes_shift);
220}
221
222void
223ir3_nir_lower_io_to_temporaries(nir_shader *s)
224{
225   /* Outputs consumed by the VPC, VS inputs, and FS outputs are all handled
226    * by the hardware pre-loading registers at the beginning and then reading
227    * them at the end, so we can't access them indirectly except through
228    * normal register-indirect accesses, and therefore ir3 doesn't support
229    * indirect accesses on those. Other i/o is lowered in ir3_nir_lower_tess,
230    * and indirects work just fine for those. GS outputs may be consumed by
231    * VPC, but have their own lowering in ir3_nir_lower_gs() which does
232    * something similar to nir_lower_io_to_temporaries so we shouldn't need
233    * to lower them.
234    *
235    * Note: this might be a little inefficient for VS or TES outputs which are
236    * when the next stage isn't an FS, but it probably don't make sense to
237    * depend on the next stage before variant creation.
238    *
239    * TODO: for gallium, mesa/st also does some redundant lowering, including
240    * running this pass for GS inputs/outputs which we don't want but not
241    * including TES outputs or FS inputs which we do need. We should probably
242    * stop doing that once we're sure all drivers are doing their own
243    * indirect i/o lowering.
244    */
245   bool lower_input = s->info.stage == MESA_SHADER_VERTEX ||
246                      s->info.stage == MESA_SHADER_FRAGMENT;
247   bool lower_output = s->info.stage != MESA_SHADER_TESS_CTRL &&
248                       s->info.stage != MESA_SHADER_GEOMETRY;
249   if (lower_input || lower_output) {
250      NIR_PASS_V(s, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(s),
251                 lower_output, lower_input);
252
253      /* nir_lower_io_to_temporaries() creates global variables and copy
254       * instructions which need to be cleaned up.
255       */
256      NIR_PASS_V(s, nir_split_var_copies);
257      NIR_PASS_V(s, nir_lower_var_copies);
258      NIR_PASS_V(s, nir_lower_global_vars_to_local);
259   }
260
261   /* Regardless of the above, we need to lower indirect references to
262    * compact variables such as clip/cull distances because due to how
263    * TCS<->TES IO works we cannot handle indirect accesses that "straddle"
264    * vec4 components. nir_lower_indirect_derefs has a special case for
265    * compact variables, so it will actually lower them even though we pass
266    * in 0 modes.
267    *
268    * Using temporaries would be slightly better but
269    * nir_lower_io_to_temporaries currently doesn't support TCS i/o.
270    */
271   NIR_PASS_V(s, nir_lower_indirect_derefs, 0, UINT32_MAX);
272}
273
274/**
275 * Inserts an add of 0.5 to floating point array index values in texture coordinates.
276 */
277static bool
278ir3_nir_lower_array_sampler_cb(struct nir_builder *b, nir_instr *instr, void *_data)
279{
280   if (instr->type != nir_instr_type_tex)
281      return false;
282
283   nir_tex_instr *tex = nir_instr_as_tex(instr);
284   if (!tex->is_array || tex->op == nir_texop_lod)
285      return false;
286
287   int coord_idx = nir_tex_instr_src_index(tex, nir_tex_src_coord);
288   if (coord_idx == -1 ||
289       nir_tex_instr_src_type(tex, coord_idx) != nir_type_float)
290      return false;
291
292   b->cursor = nir_before_instr(&tex->instr);
293
294   unsigned ncomp = tex->coord_components;
295   nir_ssa_def *src = nir_ssa_for_src(b, tex->src[coord_idx].src, ncomp);
296
297   assume(ncomp >= 1);
298   nir_ssa_def *ai = nir_channel(b, src, ncomp - 1);
299   ai = nir_fadd(b, ai, nir_imm_floatN_t(b, 0.5, src->bit_size));
300   nir_instr_rewrite_src(&tex->instr, &tex->src[coord_idx].src,
301                         nir_src_for_ssa(nir_vector_insert_imm(b, src, ai, ncomp - 1)));
302   return true;
303}
304
305static bool
306ir3_nir_lower_array_sampler(nir_shader *shader)
307{
308   return nir_shader_instructions_pass(
309      shader, ir3_nir_lower_array_sampler_cb,
310      nir_metadata_block_index | nir_metadata_dominance, NULL);
311}
312
313void
314ir3_finalize_nir(struct ir3_compiler *compiler, nir_shader *s)
315{
316   struct nir_lower_tex_options tex_options = {
317      .lower_rect = 0,
318      .lower_tg4_offsets = true,
319      .lower_invalid_implicit_lod = true,
320   };
321
322   if (compiler->gen >= 4) {
323      /* a4xx seems to have *no* sam.p */
324      tex_options.lower_txp = ~0; /* lower all txp */
325   } else {
326      /* a3xx just needs to avoid sam.p for 3d tex */
327      tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
328   }
329
330   if (ir3_shader_debug & IR3_DBG_DISASM) {
331      mesa_logi("----------------------");
332      nir_log_shaderi(s);
333      mesa_logi("----------------------");
334   }
335
336   if (s->info.stage == MESA_SHADER_GEOMETRY)
337      NIR_PASS_V(s, ir3_nir_lower_gs);
338
339   NIR_PASS_V(s, nir_lower_amul, ir3_glsl_type_size);
340
341   OPT_V(s, nir_lower_regs_to_ssa);
342   OPT_V(s, nir_lower_wrmasks, should_split_wrmask, s);
343
344   OPT_V(s, nir_lower_tex, &tex_options);
345   OPT_V(s, nir_lower_load_const_to_scalar);
346
347   if (compiler->array_index_add_half)
348      OPT_V(s, ir3_nir_lower_array_sampler);
349
350   ir3_optimize_loop(compiler, s);
351
352   /* do idiv lowering after first opt loop to get a chance to propagate
353    * constants for divide by immed power-of-two:
354    */
355   nir_lower_idiv_options idiv_options = {
356      .imprecise_32bit_lowering = true,
357      .allow_fp16 = true,
358   };
359   const bool idiv_progress = OPT(s, nir_lower_idiv, &idiv_options);
360
361   if (idiv_progress)
362      ir3_optimize_loop(compiler, s);
363
364   OPT_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
365
366   if (ir3_shader_debug & IR3_DBG_DISASM) {
367      mesa_logi("----------------------");
368      nir_log_shaderi(s);
369      mesa_logi("----------------------");
370   }
371
372   /* st_program.c's parameter list optimization requires that future nir
373    * variants don't reallocate the uniform storage, so we have to remove
374    * uniforms that occupy storage.  But we don't want to remove samplers,
375    * because they're needed for YUV variant lowering.
376    */
377   nir_foreach_uniform_variable_safe (var, s) {
378      if (var->data.mode == nir_var_uniform &&
379          (glsl_type_get_image_count(var->type) ||
380           glsl_type_get_sampler_count(var->type)))
381         continue;
382
383      exec_node_remove(&var->node);
384   }
385   nir_validate_shader(s, "after uniform var removal");
386
387   nir_sweep(s);
388}
389
390static bool
391lower_subgroup_id_filter(const nir_instr *instr, const void *unused)
392{
393   (void)unused;
394
395   if (instr->type != nir_instr_type_intrinsic)
396      return false;
397
398   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
399   return intr->intrinsic == nir_intrinsic_load_subgroup_invocation ||
400          intr->intrinsic == nir_intrinsic_load_subgroup_id ||
401          intr->intrinsic == nir_intrinsic_load_num_subgroups;
402}
403
404static nir_ssa_def *
405lower_subgroup_id(nir_builder *b, nir_instr *instr, void *unused)
406{
407   (void)unused;
408
409   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
410   if (intr->intrinsic == nir_intrinsic_load_subgroup_invocation) {
411      return nir_iand(
412         b, nir_load_local_invocation_index(b),
413         nir_isub(b, nir_load_subgroup_size(b), nir_imm_int(b, 1)));
414   } else if (intr->intrinsic == nir_intrinsic_load_subgroup_id) {
415      return nir_ishr(b, nir_load_local_invocation_index(b),
416                      nir_load_subgroup_id_shift_ir3(b));
417   } else {
418      assert(intr->intrinsic == nir_intrinsic_load_num_subgroups);
419      /* If the workgroup size is constant,
420       * nir_lower_compute_system_values() will replace local_size with a
421       * constant so this can mostly be constant folded away.
422       */
423      nir_ssa_def *local_size = nir_load_workgroup_size(b);
424      nir_ssa_def *size =
425         nir_imul24(b, nir_channel(b, local_size, 0),
426                    nir_imul24(b, nir_channel(b, local_size, 1),
427                               nir_channel(b, local_size, 2)));
428      nir_ssa_def *one = nir_imm_int(b, 1);
429      return nir_iadd(b, one,
430                      nir_ishr(b, nir_isub(b, size, one),
431                               nir_load_subgroup_id_shift_ir3(b)));
432   }
433}
434
435static bool
436ir3_nir_lower_subgroup_id_cs(nir_shader *shader)
437{
438   return nir_shader_lower_instructions(shader, lower_subgroup_id_filter,
439                                        lower_subgroup_id, NULL);
440}
441
442/**
443 * Late passes that need to be done after pscreen->finalize_nir()
444 */
445void
446ir3_nir_post_finalize(struct ir3_shader *shader)
447{
448   struct nir_shader *s = shader->nir;
449   struct ir3_compiler *compiler = shader->compiler;
450
451   NIR_PASS_V(s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
452              ir3_glsl_type_size, (nir_lower_io_options)0);
453
454   if (s->info.stage == MESA_SHADER_FRAGMENT) {
455      /* NOTE: lower load_barycentric_at_sample first, since it
456       * produces load_barycentric_at_offset:
457       */
458      NIR_PASS_V(s, ir3_nir_lower_load_barycentric_at_sample);
459      NIR_PASS_V(s, ir3_nir_lower_load_barycentric_at_offset);
460      NIR_PASS_V(s, ir3_nir_move_varying_inputs);
461      NIR_PASS_V(s, nir_lower_fb_read);
462   }
463
464   if (compiler->gen >= 6 && s->info.stage == MESA_SHADER_FRAGMENT &&
465       !(ir3_shader_debug & IR3_DBG_NOFP16)) {
466      NIR_PASS_V(s, nir_lower_mediump_io, nir_var_shader_out, 0, false);
467   }
468
469   if ((s->info.stage == MESA_SHADER_COMPUTE) ||
470       (s->info.stage == MESA_SHADER_KERNEL) ||
471       compiler->has_getfiberid) {
472      /* If the API-facing subgroup size is forced to a particular value, lower
473       * it here. Beyond this point nir_intrinsic_load_subgroup_size will return
474       * the "real" subgroup size.
475       */
476      unsigned subgroup_size = 0, max_subgroup_size = 0;
477      switch (shader->api_wavesize) {
478      case IR3_SINGLE_ONLY:
479         subgroup_size = max_subgroup_size = compiler->threadsize_base;
480         break;
481      case IR3_DOUBLE_ONLY:
482         subgroup_size = max_subgroup_size = compiler->threadsize_base * 2;
483         break;
484      case IR3_SINGLE_OR_DOUBLE:
485         /* For vertex stages, we know the wavesize will never be doubled.
486          * Lower subgroup_size here, to avoid having to deal with it when
487          * translating from NIR. Otherwise use the "real" wavesize obtained as
488          * a driver param.
489          */
490         if (s->info.stage != MESA_SHADER_COMPUTE &&
491             s->info.stage != MESA_SHADER_FRAGMENT) {
492            subgroup_size = max_subgroup_size = compiler->threadsize_base;
493         } else {
494            subgroup_size = 0;
495            max_subgroup_size = compiler->threadsize_base * 2;
496         }
497         break;
498      }
499
500      OPT(s, nir_lower_subgroups,
501          &(nir_lower_subgroups_options){
502             .subgroup_size = subgroup_size,
503             .ballot_bit_size = 32,
504             .ballot_components = max_subgroup_size / 32,
505             .lower_to_scalar = true,
506             .lower_vote_eq = true,
507             .lower_subgroup_masks = true,
508             .lower_read_invocation_to_cond = true,
509             .lower_shuffle = true,
510             .lower_relative_shuffle = true,
511          });
512   }
513
514   if ((s->info.stage == MESA_SHADER_COMPUTE) ||
515       (s->info.stage == MESA_SHADER_KERNEL)) {
516      bool progress = false;
517      NIR_PASS(progress, s, ir3_nir_lower_subgroup_id_cs);
518
519      /* ir3_nir_lower_subgroup_id_cs creates extra compute intrinsics which
520       * we need to lower again.
521       */
522      if (progress)
523         NIR_PASS_V(s, nir_lower_compute_system_values, NULL);
524   }
525
526   /* we cannot ensure that ir3_finalize_nir() is only called once, so
527    * we also need to do any run-once workarounds here:
528    */
529   OPT_V(s, ir3_nir_apply_trig_workarounds);
530
531   const nir_lower_image_options lower_image_opts = {
532      .lower_cube_size = true,
533   };
534   NIR_PASS_V(s, nir_lower_image, &lower_image_opts);
535
536   const nir_lower_idiv_options lower_idiv_options = {
537      .imprecise_32bit_lowering = true,
538      .allow_fp16 = true,
539   };
540   NIR_PASS_V(s, nir_lower_idiv, &lower_idiv_options); /* idiv generated by cube lowering */
541
542
543   /* The resinfo opcode returns the size in dwords on a4xx */
544   if (compiler->gen == 4)
545      OPT_V(s, ir3_nir_lower_ssbo_size, 2);
546
547   /* The resinfo opcode we have for getting the SSBO size on a6xx returns a
548    * byte length divided by IBO_0_FMT, while the NIR intrinsic coming in is a
549    * number of bytes. Switch things so the NIR intrinsic in our backend means
550    * dwords.
551    */
552   if (compiler->gen >= 6)
553      OPT_V(s, ir3_nir_lower_ssbo_size, compiler->storage_16bit ? 1 : 2);
554
555   ir3_optimize_loop(compiler, s);
556}
557
558static bool
559ir3_nir_lower_view_layer_id(nir_shader *nir, bool layer_zero, bool view_zero)
560{
561   unsigned layer_id_loc = ~0, view_id_loc = ~0;
562   nir_foreach_shader_in_variable (var, nir) {
563      if (var->data.location == VARYING_SLOT_LAYER)
564         layer_id_loc = var->data.driver_location;
565      if (var->data.location == VARYING_SLOT_VIEWPORT)
566         view_id_loc = var->data.driver_location;
567   }
568
569   assert(!layer_zero || layer_id_loc != ~0);
570   assert(!view_zero || view_id_loc != ~0);
571
572   bool progress = false;
573   nir_builder b;
574
575   nir_foreach_function (func, nir) {
576      nir_builder_init(&b, func->impl);
577
578      nir_foreach_block (block, func->impl) {
579         nir_foreach_instr_safe (instr, block) {
580            if (instr->type != nir_instr_type_intrinsic)
581               continue;
582
583            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
584
585            if (intrin->intrinsic != nir_intrinsic_load_input)
586               continue;
587
588            unsigned base = nir_intrinsic_base(intrin);
589            if (base != layer_id_loc && base != view_id_loc)
590               continue;
591
592            b.cursor = nir_before_instr(&intrin->instr);
593            nir_ssa_def *zero = nir_imm_int(&b, 0);
594            nir_ssa_def_rewrite_uses(&intrin->dest.ssa, zero);
595            nir_instr_remove(&intrin->instr);
596            progress = true;
597         }
598      }
599
600      if (progress) {
601         nir_metadata_preserve(
602            func->impl, nir_metadata_block_index | nir_metadata_dominance);
603      } else {
604         nir_metadata_preserve(func->impl, nir_metadata_all);
605      }
606   }
607
608   return progress;
609}
610
611static bool
612lower_ucp_vs(struct ir3_shader_variant *so)
613{
614   if (!so->key.ucp_enables)
615      return false;
616
617   gl_shader_stage last_geom_stage = MESA_SHADER_VERTEX;
618
619   if (so->key.tessellation) {
620      last_geom_stage = MESA_SHADER_TESS_EVAL;
621   } else if (so->key.has_gs) {
622      last_geom_stage = MESA_SHADER_GEOMETRY;
623   } else {
624      last_geom_stage = MESA_SHADER_VERTEX;
625   }
626
627   return so->type == last_geom_stage;
628}
629
630void
631ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s)
632{
633   if (ir3_shader_debug & IR3_DBG_DISASM) {
634      mesa_logi("----------------------");
635      nir_log_shaderi(s);
636      mesa_logi("----------------------");
637   }
638
639   bool progress = false;
640
641   if (so->key.has_gs || so->key.tessellation) {
642      switch (so->type) {
643      case MESA_SHADER_VERTEX:
644         NIR_PASS_V(s, ir3_nir_lower_to_explicit_output, so,
645                    so->key.tessellation);
646         progress = true;
647         break;
648      case MESA_SHADER_TESS_CTRL:
649         NIR_PASS_V(s, ir3_nir_lower_tess_ctrl, so, so->key.tessellation);
650         NIR_PASS_V(s, ir3_nir_lower_to_explicit_input, so);
651         progress = true;
652         break;
653      case MESA_SHADER_TESS_EVAL:
654         NIR_PASS_V(s, ir3_nir_lower_tess_eval, so, so->key.tessellation);
655         if (so->key.has_gs)
656            NIR_PASS_V(s, ir3_nir_lower_to_explicit_output, so,
657                       so->key.tessellation);
658         progress = true;
659         break;
660      case MESA_SHADER_GEOMETRY:
661         NIR_PASS_V(s, ir3_nir_lower_to_explicit_input, so);
662         progress = true;
663         break;
664      default:
665         break;
666      }
667   }
668
669   /* Note that it is intentional to use the VS lowering pass for GS, since we
670    * lower GS into something that looks more like a VS in ir3_nir_lower_gs():
671    */
672   if (lower_ucp_vs(so)) {
673      progress |= OPT(s, nir_lower_clip_vs, so->key.ucp_enables, false, true, NULL);
674   } else if (s->info.stage == MESA_SHADER_FRAGMENT) {
675      bool layer_zero =
676         so->key.layer_zero && (s->info.inputs_read & VARYING_BIT_LAYER);
677      bool view_zero =
678         so->key.view_zero && (s->info.inputs_read & VARYING_BIT_VIEWPORT);
679
680      if (so->key.ucp_enables && !so->compiler->has_clip_cull)
681         progress |= OPT(s, nir_lower_clip_fs, so->key.ucp_enables, true);
682      if (layer_zero || view_zero)
683         progress |= OPT(s, ir3_nir_lower_view_layer_id, layer_zero, view_zero);
684   }
685
686   /* Move large constant variables to the constants attached to the NIR
687    * shader, which we will upload in the immediates range.  This generates
688    * amuls, so we need to clean those up after.
689    *
690    * Passing no size_align, we would get packed values, which if we end up
691    * having to load with LDC would result in extra reads to unpack from
692    * straddling loads.  Align everything to vec4 to avoid that, though we
693    * could theoretically do better.
694    */
695   OPT_V(s, nir_opt_large_constants, glsl_get_vec4_size_align_bytes,
696         32 /* bytes */);
697   OPT_V(s, ir3_nir_lower_load_constant, so);
698
699   /* Lower large temporaries to scratch, which in Qualcomm terms is private
700    * memory, to avoid excess register pressure. This should happen after
701    * nir_opt_large_constants, because loading from a UBO is much, much less
702    * expensive.
703    */
704   if (so->compiler->has_pvtmem) {
705      progress |= OPT(s, nir_lower_vars_to_scratch, nir_var_function_temp,
706                      16 * 16 /* bytes */, glsl_get_natural_size_align_bytes);
707   }
708
709   /* Lower scratch writemasks */
710   progress |= OPT(s, nir_lower_wrmasks, should_split_wrmask, s);
711
712   progress |= OPT(s, ir3_nir_lower_wide_load_store);
713   progress |= OPT(s, ir3_nir_lower_64b_global);
714   progress |= OPT(s, ir3_nir_lower_64b_intrinsics);
715   progress |= OPT(s, ir3_nir_lower_64b_undef);
716   progress |= OPT(s, nir_lower_int64);
717
718   /* Cleanup code leftover from lowering passes before opt_preamble */
719   if (progress) {
720      progress |= OPT(s, nir_opt_constant_folding);
721   }
722
723   /* Do the preamble before analysing UBO ranges, because it's usually
724    * higher-value and because it can result in eliminating some indirect UBO
725    * accesses where otherwise we'd have to push the whole range. However we
726    * have to lower the preamble after UBO lowering so that UBO lowering can
727    * insert instructions in the preamble to push UBOs.
728    */
729   if (so->compiler->has_preamble &&
730       !(ir3_shader_debug & IR3_DBG_NOPREAMBLE))
731      progress |= OPT(s, ir3_nir_opt_preamble, so);
732
733   if (!so->binning_pass)
734      OPT_V(s, ir3_nir_analyze_ubo_ranges, so);
735
736   progress |= OPT(s, ir3_nir_lower_ubo_loads, so);
737
738   progress |= OPT(s, ir3_nir_lower_preamble, so);
739
740   OPT_V(s, nir_lower_amul, ir3_glsl_type_size);
741
742   /* UBO offset lowering has to come after we've decided what will
743    * be left as load_ubo
744    */
745   if (so->compiler->gen >= 6)
746      progress |= OPT(s, nir_lower_ubo_vec4);
747
748   OPT_V(s, ir3_nir_lower_io_offsets);
749
750   if (progress)
751      ir3_optimize_loop(so->compiler, s);
752
753   /* Fixup indirect load_uniform's which end up with a const base offset
754    * which is too large to encode.  Do this late(ish) so we actually
755    * can differentiate indirect vs non-indirect.
756    */
757   if (OPT(s, ir3_nir_fixup_load_uniform))
758      ir3_optimize_loop(so->compiler, s);
759
760   /* Do late algebraic optimization to turn add(a, neg(b)) back into
761    * subs, then the mandatory cleanup after algebraic.  Note that it may
762    * produce fnegs, and if so then we need to keep running to squash
763    * fneg(fneg(a)).
764    */
765   bool more_late_algebraic = true;
766   while (more_late_algebraic) {
767      more_late_algebraic = OPT(s, nir_opt_algebraic_late);
768      if (!more_late_algebraic && so->compiler->gen >= 5) {
769         /* Lowers texture operations that have only f2f16 or u2u16 called on
770          * them to have a 16-bit destination.  Also, lower 16-bit texture
771          * coordinates that had been upconverted to 32-bits just for the
772          * sampler to just be 16-bit texture sources.
773          */
774         struct nir_fold_tex_srcs_options fold_srcs_options = {
775            .sampler_dims = ~0,
776            .src_types = (1 << nir_tex_src_coord) |
777                         (1 << nir_tex_src_lod) |
778                         (1 << nir_tex_src_bias) |
779                         (1 << nir_tex_src_offset) |
780                         (1 << nir_tex_src_comparator) |
781                         (1 << nir_tex_src_min_lod) |
782                         (1 << nir_tex_src_ms_index) |
783                         (1 << nir_tex_src_ddx) |
784                         (1 << nir_tex_src_ddy),
785         };
786         struct nir_fold_16bit_tex_image_options fold_16bit_options = {
787            .rounding_mode = nir_rounding_mode_rtz,
788            .fold_tex_dest = true,
789            /* blob dumps have no half regs on pixel 2's ldib or stib, so only enable for a6xx+. */
790            .fold_image_load_store_data = so->compiler->gen >= 6,
791            .fold_srcs_options_count = 1,
792            .fold_srcs_options = &fold_srcs_options,
793         };
794         OPT(s, nir_fold_16bit_tex_image, &fold_16bit_options);
795      }
796      OPT_V(s, nir_opt_constant_folding);
797      OPT_V(s, nir_copy_prop);
798      OPT_V(s, nir_opt_dce);
799      OPT_V(s, nir_opt_cse);
800   }
801
802   OPT_V(s, nir_opt_sink, nir_move_const_undef);
803
804   if (ir3_shader_debug & IR3_DBG_DISASM) {
805      mesa_logi("----------------------");
806      nir_log_shaderi(s);
807      mesa_logi("----------------------");
808   }
809
810   nir_sweep(s);
811
812   /* Binning pass variants re-use  the const_state of the corresponding
813    * draw pass shader, so that same const emit can be re-used for both
814    * passes:
815    */
816   if (!so->binning_pass)
817      ir3_setup_const_state(s, so, ir3_const_state(so));
818}
819
820static void
821ir3_nir_scan_driver_consts(struct ir3_compiler *compiler, nir_shader *shader, struct ir3_const_state *layout)
822{
823   nir_foreach_function (function, shader) {
824      if (!function->impl)
825         continue;
826
827      nir_foreach_block (block, function->impl) {
828         nir_foreach_instr (instr, block) {
829            if (instr->type != nir_instr_type_intrinsic)
830               continue;
831
832            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
833            unsigned idx;
834
835            switch (intr->intrinsic) {
836            case nir_intrinsic_image_atomic_add:
837            case nir_intrinsic_image_atomic_imin:
838            case nir_intrinsic_image_atomic_umin:
839            case nir_intrinsic_image_atomic_imax:
840            case nir_intrinsic_image_atomic_umax:
841            case nir_intrinsic_image_atomic_and:
842            case nir_intrinsic_image_atomic_or:
843            case nir_intrinsic_image_atomic_xor:
844            case nir_intrinsic_image_atomic_exchange:
845            case nir_intrinsic_image_atomic_comp_swap:
846            case nir_intrinsic_image_load:
847            case nir_intrinsic_image_store:
848            case nir_intrinsic_image_size:
849               /* a4xx gets these supplied by the hw directly (maybe CP?) */
850               if (compiler->gen == 5 &&
851                   !(intr->intrinsic == nir_intrinsic_image_load &&
852                     !(nir_intrinsic_access(intr) & ACCESS_COHERENT))) {
853                  idx = nir_src_as_uint(intr->src[0]);
854                  if (layout->image_dims.mask & (1 << idx))
855                     break;
856                  layout->image_dims.mask |= (1 << idx);
857                  layout->image_dims.off[idx] = layout->image_dims.count;
858                  layout->image_dims.count += 3; /* three const per */
859               }
860               break;
861            case nir_intrinsic_load_base_vertex:
862            case nir_intrinsic_load_first_vertex:
863               layout->num_driver_params =
864                  MAX2(layout->num_driver_params, IR3_DP_VTXID_BASE + 1);
865               break;
866            case nir_intrinsic_load_base_instance:
867               layout->num_driver_params =
868                  MAX2(layout->num_driver_params, IR3_DP_INSTID_BASE + 1);
869               break;
870            case nir_intrinsic_load_user_clip_plane:
871               idx = nir_intrinsic_ucp_id(intr);
872               layout->num_driver_params = MAX2(layout->num_driver_params,
873                                                IR3_DP_UCP0_X + (idx + 1) * 4);
874               break;
875            case nir_intrinsic_load_num_workgroups:
876               layout->num_driver_params =
877                  MAX2(layout->num_driver_params, IR3_DP_NUM_WORK_GROUPS_Z + 1);
878               break;
879            case nir_intrinsic_load_workgroup_id:
880               if (!compiler->has_shared_regfile) {
881                  layout->num_driver_params =
882                     MAX2(layout->num_driver_params, IR3_DP_WORKGROUP_ID_Z + 1);
883               }
884               break;
885            case nir_intrinsic_load_workgroup_size:
886               layout->num_driver_params = MAX2(layout->num_driver_params,
887                                                IR3_DP_LOCAL_GROUP_SIZE_Z + 1);
888               break;
889            case nir_intrinsic_load_base_workgroup_id:
890               layout->num_driver_params =
891                  MAX2(layout->num_driver_params, IR3_DP_BASE_GROUP_Z + 1);
892               break;
893            case nir_intrinsic_load_subgroup_size: {
894               assert(shader->info.stage == MESA_SHADER_COMPUTE ||
895                      shader->info.stage == MESA_SHADER_FRAGMENT);
896               enum ir3_driver_param size = shader->info.stage == MESA_SHADER_COMPUTE ?
897                  IR3_DP_CS_SUBGROUP_SIZE : IR3_DP_FS_SUBGROUP_SIZE;
898               layout->num_driver_params =
899                  MAX2(layout->num_driver_params, size + 1);
900               break;
901            }
902            case nir_intrinsic_load_subgroup_id_shift_ir3:
903               layout->num_driver_params =
904                  MAX2(layout->num_driver_params, IR3_DP_SUBGROUP_ID_SHIFT + 1);
905               break;
906            case nir_intrinsic_load_draw_id:
907               layout->num_driver_params =
908                  MAX2(layout->num_driver_params, IR3_DP_DRAWID + 1);
909               break;
910            default:
911               break;
912            }
913         }
914      }
915   }
916
917   /* TODO: Provide a spot somewhere to safely upload unwanted values, and a way
918    * to determine if they're wanted or not. For now we always make the whole
919    * driver param range available, since the driver will always instruct the
920    * hardware to upload these.
921    */
922   if (!compiler->has_shared_regfile &&
923         shader->info.stage == MESA_SHADER_COMPUTE) {
924      layout->num_driver_params =
925         MAX2(layout->num_driver_params, IR3_DP_WORKGROUP_ID_Z + 1);
926   }
927}
928
929/* Sets up the variant-dependent constant state for the ir3_shader.  Note
930 * that it is also used from ir3_nir_analyze_ubo_ranges() to figure out the
931 * maximum number of driver params that would eventually be used, to leave
932 * space for this function to allocate the driver params.
933 */
934void
935ir3_setup_const_state(nir_shader *nir, struct ir3_shader_variant *v,
936                      struct ir3_const_state *const_state)
937{
938   struct ir3_compiler *compiler = v->compiler;
939
940   memset(&const_state->offsets, ~0, sizeof(const_state->offsets));
941
942   ir3_nir_scan_driver_consts(compiler, nir, const_state);
943
944   if ((compiler->gen < 5) && (v->stream_output.num_outputs > 0)) {
945      const_state->num_driver_params =
946         MAX2(const_state->num_driver_params, IR3_DP_VTXCNT_MAX + 1);
947   }
948
949   const_state->num_ubos = nir->info.num_ubos;
950
951   assert((const_state->ubo_state.size % 16) == 0);
952   unsigned constoff = v->num_reserved_user_consts +
953      const_state->ubo_state.size / 16 +
954      const_state->preamble_size;
955   unsigned ptrsz = ir3_pointer_size(compiler);
956
957   if (const_state->num_ubos > 0) {
958      const_state->offsets.ubo = constoff;
959      constoff += align(const_state->num_ubos * ptrsz, 4) / 4;
960   }
961
962   if (const_state->image_dims.count > 0) {
963      unsigned cnt = const_state->image_dims.count;
964      const_state->offsets.image_dims = constoff;
965      constoff += align(cnt, 4) / 4;
966   }
967
968   if (v->type == MESA_SHADER_KERNEL) {
969      const_state->offsets.kernel_params = constoff;
970      constoff += align(v->cs.req_input_mem, 4) / 4;
971   }
972
973   if (const_state->num_driver_params > 0) {
974      /* num_driver_params in dwords.  we only need to align to vec4s for the
975       * common case of immediate constant uploads, but for indirect dispatch
976       * the constants may also be indirect and so we have to align the area in
977       * const space to that requirement.
978       */
979      const_state->num_driver_params = align(const_state->num_driver_params, 4);
980      unsigned upload_unit = 1;
981      if (v->type == MESA_SHADER_COMPUTE ||
982          (const_state->num_driver_params >= IR3_DP_VTXID_BASE)) {
983         upload_unit = compiler->const_upload_unit;
984      }
985
986      /* offset cannot be 0 for vs params loaded by CP_DRAW_INDIRECT_MULTI */
987      if (v->type == MESA_SHADER_VERTEX && compiler->gen >= 6)
988         constoff = MAX2(constoff, 1);
989      constoff = align(constoff, upload_unit);
990      const_state->offsets.driver_param = constoff;
991
992      constoff += align(const_state->num_driver_params / 4, upload_unit);
993   }
994
995   if ((v->type == MESA_SHADER_VERTEX) && (compiler->gen < 5) &&
996       v->stream_output.num_outputs > 0) {
997      const_state->offsets.tfbo = constoff;
998      constoff += align(IR3_MAX_SO_BUFFERS * ptrsz, 4) / 4;
999   }
1000
1001   switch (v->type) {
1002   case MESA_SHADER_VERTEX:
1003      const_state->offsets.primitive_param = constoff;
1004      constoff += 1;
1005      break;
1006   case MESA_SHADER_TESS_CTRL:
1007   case MESA_SHADER_TESS_EVAL:
1008      const_state->offsets.primitive_param = constoff;
1009      constoff += 2;
1010
1011      const_state->offsets.primitive_map = constoff;
1012      constoff += DIV_ROUND_UP(v->input_size, 4);
1013      break;
1014   case MESA_SHADER_GEOMETRY:
1015      const_state->offsets.primitive_param = constoff;
1016      constoff += 1;
1017
1018      const_state->offsets.primitive_map = constoff;
1019      constoff += DIV_ROUND_UP(v->input_size, 4);
1020      break;
1021   default:
1022      break;
1023   }
1024
1025   const_state->offsets.immediate = constoff;
1026
1027   assert(constoff <= ir3_max_const(v));
1028}
1029