src/compiler/nir/nir_divergence_analysis.c

   1 /*
   2  * Copyright © 2018 Valve Corporation
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "nir.h"
  26
  27 /* This pass computes for each ssa definition if it is uniform.
  28  * That is, the variable has the same value for all invocations
  29  * of the group.
  30  *
  31  * This divergence analysis pass expects the shader to be in LCSSA-form.
  32  *
  33  * This algorithm implements "The Simple Divergence Analysis" from
  34  * Diogo Sampaio, Rafael De Souza, Sylvain Collange, Fernando Magno Quintão Pereira.
  35  * Divergence Analysis.  ACM Transactions on Programming Languages and Systems (TOPLAS),
  36  * ACM, 2013, 35 (4), pp.13:1-13:36. <10.1145/2523815>. <hal-00909072v2>
  37  */
  38
  39 struct divergence_state {
  40    const nir_divergence_options options;
  41    const gl_shader_stage stage;
  42
  43    /** current control flow state */
  44    /* True if some loop-active invocations might take a different control-flow path.
  45     * A divergent break does not cause subsequent control-flow to be considered
  46     * divergent because those invocations are no longer active in the loop.
  47     * For a divergent if, both sides are considered divergent flow because
  48     * the other side is still loop-active. */
  49    bool divergent_loop_cf;
  50    /* True if a divergent continue happened since the loop header */
  51    bool divergent_loop_continue;
  52    /* True if a divergent break happened since the loop header */
  53    bool divergent_loop_break;
  54
  55    /* True if we visit the block for the fist time */
  56    bool first_visit;
  57 };
  58
  59 static bool
  60 visit_cf_list(struct exec_list *list, struct divergence_state *state);
  61
  62 static bool
  63 visit_alu(nir_alu_instr *instr)
  64 {
  65    if (instr->dest.dest.ssa.divergent)
  66       return false;
  67
  68    unsigned num_src = nir_op_infos[instr->op].num_inputs;
  69
  70    for (unsigned i = 0; i < num_src; i++) {
  71       if (instr->src[i].src.ssa->divergent) {
  72          instr->dest.dest.ssa.divergent = true;
  73          return true;
  74       }
  75    }
  76
  77    return false;
  78 }
  79
  80 static bool
  81 visit_intrinsic(nir_intrinsic_instr *instr, struct divergence_state *state)
  82 {
  83    if (!nir_intrinsic_infos[instr->intrinsic].has_dest)
  84       return false;
  85
  86    if (instr->dest.ssa.divergent)
  87       return false;
  88
  89    nir_divergence_options options = state->options;
  90    gl_shader_stage stage = state->stage;
  91    bool is_divergent = false;
  92    switch (instr->intrinsic) {
  93    /* Intrinsics which are always uniform */
  94    case nir_intrinsic_shader_clock:
  95    case nir_intrinsic_ballot:
  96    case nir_intrinsic_read_invocation:
  97    case nir_intrinsic_read_first_invocation:
  98    case nir_intrinsic_vote_any:
  99    case nir_intrinsic_vote_all:
 100    case nir_intrinsic_vote_feq:
 101    case nir_intrinsic_vote_ieq:
 102    case nir_intrinsic_load_work_dim:
 103    case nir_intrinsic_load_work_group_id:
 104    case nir_intrinsic_load_num_work_groups:
 105    case nir_intrinsic_load_local_group_size:
 106    case nir_intrinsic_load_subgroup_id:
 107    case nir_intrinsic_load_num_subgroups:
 108    case nir_intrinsic_load_subgroup_size:
 109    case nir_intrinsic_load_subgroup_eq_mask:
 110    case nir_intrinsic_load_subgroup_ge_mask:
 111    case nir_intrinsic_load_subgroup_gt_mask:
 112    case nir_intrinsic_load_subgroup_le_mask:
 113    case nir_intrinsic_load_subgroup_lt_mask:
 114    case nir_intrinsic_first_invocation:
 115    case nir_intrinsic_load_base_instance:
 116    case nir_intrinsic_load_base_vertex:
 117    case nir_intrinsic_load_first_vertex:
 118    case nir_intrinsic_load_draw_id:
 119    case nir_intrinsic_load_is_indexed_draw:
 120    case nir_intrinsic_load_viewport_scale:
 121    case nir_intrinsic_load_alpha_ref_float:
 122    case nir_intrinsic_load_user_clip_plane:
 123    case nir_intrinsic_load_viewport_x_scale:
 124    case nir_intrinsic_load_viewport_y_scale:
 125    case nir_intrinsic_load_viewport_z_scale:
 126    case nir_intrinsic_load_viewport_offset:
 127    case nir_intrinsic_load_viewport_z_offset:
 128    case nir_intrinsic_load_blend_const_color_a_float:
 129    case nir_intrinsic_load_blend_const_color_b_float:
 130    case nir_intrinsic_load_blend_const_color_g_float:
 131    case nir_intrinsic_load_blend_const_color_r_float:
 132    case nir_intrinsic_load_blend_const_color_rgba:
 133    case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
 134    case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
 135       is_divergent = false;
 136       break;
 137
 138    /* Intrinsics with divergence depending on shader stage and hardware */
 139    case nir_intrinsic_load_input:
 140       is_divergent = instr->src[0].ssa->divergent;
 141       if (stage == MESA_SHADER_FRAGMENT)
 142          is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 143       else if (stage == MESA_SHADER_TESS_EVAL)
 144          is_divergent |= !(options & nir_divergence_single_patch_per_tes_subgroup);
 145       else
 146          is_divergent = true;
 147       break;
 148    case nir_intrinsic_load_per_vertex_input:
 149       is_divergent = instr->src[0].ssa->divergent ||
 150                      instr->src[1].ssa->divergent;
 151       if (stage == MESA_SHADER_TESS_CTRL)
 152          is_divergent |= !(options & nir_divergence_single_patch_per_tcs_subgroup);
 153       if (stage == MESA_SHADER_TESS_EVAL)
 154          is_divergent |= !(options & nir_divergence_single_patch_per_tes_subgroup);
 155       else
 156          is_divergent = true;
 157       break;
 158    case nir_intrinsic_load_input_vertex:
 159       is_divergent = instr->src[1].ssa->divergent;
 160       assert(stage == MESA_SHADER_FRAGMENT);
 161       is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 162       break;
 163    case nir_intrinsic_load_output:
 164       assert(stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_FRAGMENT);
 165       is_divergent = instr->src[0].ssa->divergent;
 166       if (stage == MESA_SHADER_TESS_CTRL)
 167          is_divergent |= !(options & nir_divergence_single_patch_per_tcs_subgroup);
 168       else
 169          is_divergent = true;
 170       break;
 171    case nir_intrinsic_load_per_vertex_output:
 172       assert(stage == MESA_SHADER_TESS_CTRL);
 173       is_divergent = instr->src[0].ssa->divergent ||
 174                      instr->src[1].ssa->divergent ||
 175                      !(options & nir_divergence_single_patch_per_tcs_subgroup);
 176       break;
 177    case nir_intrinsic_load_layer_id:
 178    case nir_intrinsic_load_front_face:
 179       assert(stage == MESA_SHADER_FRAGMENT);
 180       is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 181       break;
 182    case nir_intrinsic_load_view_index:
 183       assert(stage != MESA_SHADER_COMPUTE && stage != MESA_SHADER_KERNEL);
 184       if (options & nir_divergence_view_index_uniform)
 185          is_divergent = false;
 186       else if (stage == MESA_SHADER_FRAGMENT)
 187          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 188       break;
 189    case nir_intrinsic_load_fs_input_interp_deltas:
 190       assert(stage == MESA_SHADER_FRAGMENT);
 191       is_divergent = instr->src[0].ssa->divergent;
 192       is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 193       break;
 194    case nir_intrinsic_load_primitive_id:
 195       if (stage == MESA_SHADER_FRAGMENT)
 196          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 197       else if (stage == MESA_SHADER_TESS_CTRL)
 198          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 199       else if (stage == MESA_SHADER_TESS_EVAL)
 200          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 201       else if (stage == MESA_SHADER_GEOMETRY)
 202          is_divergent = true;
 203       else
 204          unreachable("Invalid stage for load_primitive_id");
 205       break;
 206    case nir_intrinsic_load_tess_level_inner:
 207    case nir_intrinsic_load_tess_level_outer:
 208       if (stage == MESA_SHADER_TESS_CTRL)
 209          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 210       else if (stage == MESA_SHADER_TESS_EVAL)
 211          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 212       else
 213          unreachable("Invalid stage for load_primitive_tess_level_*");
 214       break;
 215    case nir_intrinsic_load_patch_vertices_in:
 216       if (stage == MESA_SHADER_TESS_EVAL)
 217          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 218       else
 219          assert(stage == MESA_SHADER_TESS_CTRL);
 220       break;
 221
 222    /* Clustered reductions are uniform if cluster_size == subgroup_size or
 223     * the source is uniform and the operation is invariant.
 224     * Inclusive scans are uniform if
 225     * the source is uniform and the operation is invariant
 226     */
 227    case nir_intrinsic_reduce:
 228       if (nir_intrinsic_cluster_size(instr) == 0)
 229          return false;
 230       /* fallthrough */
 231    case nir_intrinsic_inclusive_scan: {
 232       nir_op op = nir_intrinsic_reduction_op(instr);
 233       is_divergent = instr->src[0].ssa->divergent;
 234       if (op != nir_op_umin && op != nir_op_imin && op != nir_op_fmin &&
 235           op != nir_op_umax && op != nir_op_imax && op != nir_op_fmax &&
 236           op != nir_op_iand && op != nir_op_ior)
 237          is_divergent = true;
 238       break;
 239    }
 240
 241    /* Intrinsics with divergence depending on sources */
 242    case nir_intrinsic_ballot_bitfield_extract:
 243    case nir_intrinsic_ballot_find_lsb:
 244    case nir_intrinsic_ballot_find_msb:
 245    case nir_intrinsic_ballot_bit_count_reduce:
 246    case nir_intrinsic_shuffle_xor:
 247    case nir_intrinsic_shuffle_up:
 248    case nir_intrinsic_shuffle_down:
 249    case nir_intrinsic_quad_broadcast:
 250    case nir_intrinsic_quad_swap_horizontal:
 251    case nir_intrinsic_quad_swap_vertical:
 252    case nir_intrinsic_quad_swap_diagonal:
 253    case nir_intrinsic_load_deref:
 254    case nir_intrinsic_load_ubo:
 255    case nir_intrinsic_load_ssbo:
 256    case nir_intrinsic_load_shared:
 257    case nir_intrinsic_load_global:
 258    case nir_intrinsic_load_uniform:
 259    case nir_intrinsic_load_push_constant:
 260    case nir_intrinsic_load_constant:
 261    case nir_intrinsic_load_sample_pos_from_id:
 262    case nir_intrinsic_load_kernel_input:
 263    case nir_intrinsic_image_load:
 264    case nir_intrinsic_image_deref_load:
 265    case nir_intrinsic_bindless_image_load:
 266    case nir_intrinsic_image_samples:
 267    case nir_intrinsic_image_deref_samples:
 268    case nir_intrinsic_bindless_image_samples:
 269    case nir_intrinsic_get_buffer_size:
 270    case nir_intrinsic_image_size:
 271    case nir_intrinsic_image_deref_size:
 272    case nir_intrinsic_bindless_image_size:
 273    case nir_intrinsic_copy_deref:
 274    case nir_intrinsic_deref_buffer_array_length:
 275    case nir_intrinsic_vulkan_resource_index:
 276    case nir_intrinsic_vulkan_resource_reindex:
 277    case nir_intrinsic_load_vulkan_descriptor:
 278    case nir_intrinsic_atomic_counter_read:
 279    case nir_intrinsic_atomic_counter_read_deref:
 280    case nir_intrinsic_quad_swizzle_amd:
 281    case nir_intrinsic_masked_swizzle_amd: {
 282       unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
 283       for (unsigned i = 0; i < num_srcs; i++) {
 284          if (instr->src[i].ssa->divergent) {
 285             is_divergent = true;
 286             break;
 287          }
 288       }
 289       break;
 290    }
 291
 292    case nir_intrinsic_shuffle:
 293       is_divergent = instr->src[0].ssa->divergent &&
 294                      instr->src[1].ssa->divergent;
 295       break;
 296
 297    /* Intrinsics which are always divergent */
 298    case nir_intrinsic_load_color0:
 299    case nir_intrinsic_load_color1:
 300    case nir_intrinsic_load_param:
 301    case nir_intrinsic_load_sample_id:
 302    case nir_intrinsic_load_sample_id_no_per_sample:
 303    case nir_intrinsic_load_sample_mask_in:
 304    case nir_intrinsic_load_interpolated_input:
 305    case nir_intrinsic_load_barycentric_pixel:
 306    case nir_intrinsic_load_barycentric_centroid:
 307    case nir_intrinsic_load_barycentric_sample:
 308    case nir_intrinsic_load_barycentric_model:
 309    case nir_intrinsic_load_barycentric_at_sample:
 310    case nir_intrinsic_load_barycentric_at_offset:
 311    case nir_intrinsic_interp_deref_at_offset:
 312    case nir_intrinsic_interp_deref_at_sample:
 313    case nir_intrinsic_interp_deref_at_centroid:
 314    case nir_intrinsic_interp_deref_at_vertex:
 315    case nir_intrinsic_load_tess_coord:
 316    case nir_intrinsic_load_point_coord:
 317    case nir_intrinsic_load_line_coord:
 318    case nir_intrinsic_load_frag_coord:
 319    case nir_intrinsic_load_sample_pos:
 320    case nir_intrinsic_load_vertex_id_zero_base:
 321    case nir_intrinsic_load_vertex_id:
 322    case nir_intrinsic_load_instance_id:
 323    case nir_intrinsic_load_invocation_id:
 324    case nir_intrinsic_load_local_invocation_id:
 325    case nir_intrinsic_load_local_invocation_index:
 326    case nir_intrinsic_load_global_invocation_id:
 327    case nir_intrinsic_load_global_invocation_index:
 328    case nir_intrinsic_load_subgroup_invocation:
 329    case nir_intrinsic_load_helper_invocation:
 330    case nir_intrinsic_is_helper_invocation:
 331    case nir_intrinsic_load_scratch:
 332    case nir_intrinsic_deref_atomic_add:
 333    case nir_intrinsic_deref_atomic_imin:
 334    case nir_intrinsic_deref_atomic_umin:
 335    case nir_intrinsic_deref_atomic_imax:
 336    case nir_intrinsic_deref_atomic_umax:
 337    case nir_intrinsic_deref_atomic_and:
 338    case nir_intrinsic_deref_atomic_or:
 339    case nir_intrinsic_deref_atomic_xor:
 340    case nir_intrinsic_deref_atomic_exchange:
 341    case nir_intrinsic_deref_atomic_comp_swap:
 342    case nir_intrinsic_deref_atomic_fadd:
 343    case nir_intrinsic_deref_atomic_fmin:
 344    case nir_intrinsic_deref_atomic_fmax:
 345    case nir_intrinsic_deref_atomic_fcomp_swap:
 346    case nir_intrinsic_ssbo_atomic_add:
 347    case nir_intrinsic_ssbo_atomic_imin:
 348    case nir_intrinsic_ssbo_atomic_umin:
 349    case nir_intrinsic_ssbo_atomic_imax:
 350    case nir_intrinsic_ssbo_atomic_umax:
 351    case nir_intrinsic_ssbo_atomic_and:
 352    case nir_intrinsic_ssbo_atomic_or:
 353    case nir_intrinsic_ssbo_atomic_xor:
 354    case nir_intrinsic_ssbo_atomic_exchange:
 355    case nir_intrinsic_ssbo_atomic_comp_swap:
 356    case nir_intrinsic_ssbo_atomic_fadd:
 357    case nir_intrinsic_ssbo_atomic_fmax:
 358    case nir_intrinsic_ssbo_atomic_fmin:
 359    case nir_intrinsic_ssbo_atomic_fcomp_swap:
 360    case nir_intrinsic_image_deref_atomic_add:
 361    case nir_intrinsic_image_deref_atomic_imin:
 362    case nir_intrinsic_image_deref_atomic_umin:
 363    case nir_intrinsic_image_deref_atomic_imax:
 364    case nir_intrinsic_image_deref_atomic_umax:
 365    case nir_intrinsic_image_deref_atomic_and:
 366    case nir_intrinsic_image_deref_atomic_or:
 367    case nir_intrinsic_image_deref_atomic_xor:
 368    case nir_intrinsic_image_deref_atomic_exchange:
 369    case nir_intrinsic_image_deref_atomic_comp_swap:
 370    case nir_intrinsic_image_deref_atomic_fadd:
 371    case nir_intrinsic_image_atomic_add:
 372    case nir_intrinsic_image_atomic_imin:
 373    case nir_intrinsic_image_atomic_umin:
 374    case nir_intrinsic_image_atomic_imax:
 375    case nir_intrinsic_image_atomic_umax:
 376    case nir_intrinsic_image_atomic_and:
 377    case nir_intrinsic_image_atomic_or:
 378    case nir_intrinsic_image_atomic_xor:
 379    case nir_intrinsic_image_atomic_exchange:
 380    case nir_intrinsic_image_atomic_comp_swap:
 381    case nir_intrinsic_image_atomic_fadd:
 382    case nir_intrinsic_bindless_image_atomic_add:
 383    case nir_intrinsic_bindless_image_atomic_imin:
 384    case nir_intrinsic_bindless_image_atomic_umin:
 385    case nir_intrinsic_bindless_image_atomic_imax:
 386    case nir_intrinsic_bindless_image_atomic_umax:
 387    case nir_intrinsic_bindless_image_atomic_and:
 388    case nir_intrinsic_bindless_image_atomic_or:
 389    case nir_intrinsic_bindless_image_atomic_xor:
 390    case nir_intrinsic_bindless_image_atomic_exchange:
 391    case nir_intrinsic_bindless_image_atomic_comp_swap:
 392    case nir_intrinsic_bindless_image_atomic_fadd:
 393    case nir_intrinsic_shared_atomic_add:
 394    case nir_intrinsic_shared_atomic_imin:
 395    case nir_intrinsic_shared_atomic_umin:
 396    case nir_intrinsic_shared_atomic_imax:
 397    case nir_intrinsic_shared_atomic_umax:
 398    case nir_intrinsic_shared_atomic_and:
 399    case nir_intrinsic_shared_atomic_or:
 400    case nir_intrinsic_shared_atomic_xor:
 401    case nir_intrinsic_shared_atomic_exchange:
 402    case nir_intrinsic_shared_atomic_comp_swap:
 403    case nir_intrinsic_shared_atomic_fadd:
 404    case nir_intrinsic_shared_atomic_fmin:
 405    case nir_intrinsic_shared_atomic_fmax:
 406    case nir_intrinsic_shared_atomic_fcomp_swap:
 407    case nir_intrinsic_global_atomic_add:
 408    case nir_intrinsic_global_atomic_imin:
 409    case nir_intrinsic_global_atomic_umin:
 410    case nir_intrinsic_global_atomic_imax:
 411    case nir_intrinsic_global_atomic_umax:
 412    case nir_intrinsic_global_atomic_and:
 413    case nir_intrinsic_global_atomic_or:
 414    case nir_intrinsic_global_atomic_xor:
 415    case nir_intrinsic_global_atomic_exchange:
 416    case nir_intrinsic_global_atomic_comp_swap:
 417    case nir_intrinsic_global_atomic_fadd:
 418    case nir_intrinsic_global_atomic_fmin:
 419    case nir_intrinsic_global_atomic_fmax:
 420    case nir_intrinsic_global_atomic_fcomp_swap:
 421    case nir_intrinsic_atomic_counter_add:
 422    case nir_intrinsic_atomic_counter_min:
 423    case nir_intrinsic_atomic_counter_max:
 424    case nir_intrinsic_atomic_counter_and:
 425    case nir_intrinsic_atomic_counter_or:
 426    case nir_intrinsic_atomic_counter_xor:
 427    case nir_intrinsic_atomic_counter_inc:
 428    case nir_intrinsic_atomic_counter_pre_dec:
 429    case nir_intrinsic_atomic_counter_post_dec:
 430    case nir_intrinsic_atomic_counter_exchange:
 431    case nir_intrinsic_atomic_counter_comp_swap:
 432    case nir_intrinsic_atomic_counter_add_deref:
 433    case nir_intrinsic_atomic_counter_min_deref:
 434    case nir_intrinsic_atomic_counter_max_deref:
 435    case nir_intrinsic_atomic_counter_and_deref:
 436    case nir_intrinsic_atomic_counter_or_deref:
 437    case nir_intrinsic_atomic_counter_xor_deref:
 438    case nir_intrinsic_atomic_counter_inc_deref:
 439    case nir_intrinsic_atomic_counter_pre_dec_deref:
 440    case nir_intrinsic_atomic_counter_post_dec_deref:
 441    case nir_intrinsic_atomic_counter_exchange_deref:
 442    case nir_intrinsic_atomic_counter_comp_swap_deref:
 443    case nir_intrinsic_exclusive_scan:
 444    case nir_intrinsic_ballot_bit_count_exclusive:
 445    case nir_intrinsic_ballot_bit_count_inclusive:
 446    case nir_intrinsic_write_invocation_amd:
 447    case nir_intrinsic_mbcnt_amd:
 448    case nir_intrinsic_elect:
 449       is_divergent = true;
 450       break;
 451
 452    default:
 453 #ifdef NDEBUG
 454       is_divergent = true;
 455       break;
 456 #else
 457       nir_print_instr(&instr->instr, stderr);
 458       unreachable("\nNIR divergence analysis: Unhandled intrinsic.");
 459 #endif
 460    }
 461
 462    instr->dest.ssa.divergent = is_divergent;
 463    return is_divergent;
 464 }
 465
 466 static bool
 467 visit_tex(nir_tex_instr *instr)
 468 {
 469    if (instr->dest.ssa.divergent)
 470       return false;
 471
 472    bool is_divergent = false;
 473
 474    for (unsigned i = 0; i < instr->num_srcs; i++) {
 475       switch (instr->src[i].src_type) {
 476       case nir_tex_src_sampler_deref:
 477       case nir_tex_src_sampler_handle:
 478       case nir_tex_src_sampler_offset:
 479          is_divergent |= instr->src[i].src.ssa->divergent &&
 480                          instr->sampler_non_uniform;
 481          break;
 482       case nir_tex_src_texture_deref:
 483       case nir_tex_src_texture_handle:
 484       case nir_tex_src_texture_offset:
 485          is_divergent |= instr->src[i].src.ssa->divergent &&
 486                          instr->texture_non_uniform;
 487          break;
 488       default:
 489          is_divergent |= instr->src[i].src.ssa->divergent;
 490          break;
 491       }
 492    }
 493
 494    instr->dest.ssa.divergent = is_divergent;
 495    return is_divergent;
 496 }
 497
 498 static bool
 499 visit_load_const(nir_load_const_instr *instr)
 500 {
 501    return false;
 502 }
 503
 504 static bool
 505 visit_ssa_undef(nir_ssa_undef_instr *instr)
 506 {
 507    return false;
 508 }
 509
 510 static bool
 511 nir_variable_mode_is_uniform(nir_variable_mode mode) {
 512    switch (mode) {
 513    case nir_var_uniform:
 514    case nir_var_mem_ubo:
 515    case nir_var_mem_ssbo:
 516    case nir_var_mem_shared:
 517    case nir_var_mem_global:
 518       return true;
 519    default:
 520       return false;
 521    }
 522 }
 523
 524 static bool
 525 nir_variable_is_uniform(nir_variable *var, struct divergence_state *state)
 526 {
 527    if (nir_variable_mode_is_uniform(var->data.mode))
 528       return true;
 529
 530    if (state->stage == MESA_SHADER_FRAGMENT &&
 531        (state->options & nir_divergence_single_prim_per_subgroup) &&
 532        var->data.mode == nir_var_shader_in &&
 533        var->data.interpolation == INTERP_MODE_FLAT)
 534       return true;
 535
 536    if (state->stage == MESA_SHADER_TESS_CTRL &&
 537        (state->options & nir_divergence_single_patch_per_tcs_subgroup) &&
 538        var->data.mode == nir_var_shader_out && var->data.patch)
 539       return true;
 540
 541    if (state->stage == MESA_SHADER_TESS_EVAL &&
 542        (state->options & nir_divergence_single_patch_per_tes_subgroup) &&
 543        var->data.mode == nir_var_shader_in && var->data.patch)
 544       return true;
 545
 546    return false;
 547 }
 548
 549 static bool
 550 visit_deref(nir_deref_instr *deref, struct divergence_state *state)
 551 {
 552    if (deref->dest.ssa.divergent)
 553       return false;
 554
 555    bool is_divergent = false;
 556    switch (deref->deref_type) {
 557    case nir_deref_type_var:
 558       is_divergent = !nir_variable_is_uniform(deref->var, state);
 559       break;
 560    case nir_deref_type_array:
 561    case nir_deref_type_ptr_as_array:
 562       is_divergent = deref->arr.index.ssa->divergent;
 563       /* fallthrough */
 564    case nir_deref_type_struct:
 565    case nir_deref_type_array_wildcard:
 566       is_divergent |= deref->parent.ssa->divergent;
 567       break;
 568    case nir_deref_type_cast:
 569       is_divergent = !nir_variable_mode_is_uniform(deref->var->data.mode) ||
 570                      deref->parent.ssa->divergent;
 571       break;
 572    }
 573
 574    deref->dest.ssa.divergent = is_divergent;
 575    return is_divergent;
 576 }
 577
 578 static bool
 579 visit_jump(nir_jump_instr *jump, struct divergence_state *state)
 580 {
 581    switch (jump->type) {
 582    case nir_jump_continue:
 583       if (state->divergent_loop_continue)
 584          return false;
 585       if (state->divergent_loop_cf)
 586          state->divergent_loop_continue = true;
 587       return state->divergent_loop_continue;
 588    case nir_jump_break:
 589       if (state->divergent_loop_break)
 590          return false;
 591       if (state->divergent_loop_cf)
 592          state->divergent_loop_break = true;
 593       return state->divergent_loop_break;
 594    case nir_jump_return:
 595       unreachable("NIR divergence analysis: Unsupported return instruction.");
 596    }
 597    return false;
 598 }
 599
 600 static bool
 601 set_ssa_def_not_divergent(nir_ssa_def *def, UNUSED void *_state)
 602 {
 603    def->divergent = false;
 604    return true;
 605 }
 606
 607 static bool
 608 visit_block(nir_block *block, struct divergence_state *state)
 609 {
 610    bool has_changed = false;
 611
 612    nir_foreach_instr(instr, block) {
 613       /* phis are handled when processing the branches */
 614       if (instr->type == nir_instr_type_phi)
 615          continue;
 616
 617       if (state->first_visit)
 618          nir_foreach_ssa_def(instr, set_ssa_def_not_divergent, NULL);
 619
 620       switch (instr->type) {
 621       case nir_instr_type_alu:
 622          has_changed |= visit_alu(nir_instr_as_alu(instr));
 623          break;
 624       case nir_instr_type_intrinsic:
 625          has_changed |= visit_intrinsic(nir_instr_as_intrinsic(instr), state);
 626          break;
 627       case nir_instr_type_tex:
 628          has_changed |= visit_tex(nir_instr_as_tex(instr));
 629          break;
 630       case nir_instr_type_load_const:
 631          has_changed |= visit_load_const(nir_instr_as_load_const(instr));
 632          break;
 633       case nir_instr_type_ssa_undef:
 634          has_changed |= visit_ssa_undef(nir_instr_as_ssa_undef(instr));
 635          break;
 636       case nir_instr_type_deref:
 637          has_changed |= visit_deref(nir_instr_as_deref(instr), state);
 638          break;
 639       case nir_instr_type_jump:
 640          has_changed |= visit_jump(nir_instr_as_jump(instr), state);
 641          break;
 642       case nir_instr_type_phi:
 643       case nir_instr_type_call:
 644       case nir_instr_type_parallel_copy:
 645          unreachable("NIR divergence analysis: Unsupported instruction type.");
 646       }
 647    }
 648
 649    return has_changed;
 650 }
 651
 652 /* There are 3 types of phi instructions:
 653  * (1) gamma: represent the joining point of different paths
 654  *     created by an “if-then-else” branch.
 655  *     The resulting value is divergent if the branch condition
 656  *     or any of the source values is divergent. */
 657 static bool
 658 visit_if_merge_phi(nir_phi_instr *phi, bool if_cond_divergent)
 659 {
 660    if (phi->dest.ssa.divergent)
 661       return false;
 662
 663    unsigned defined_srcs = 0;
 664    nir_foreach_phi_src(src, phi) {
 665       /* if any source value is divergent, the resulting value is divergent */
 666       if (src->src.ssa->divergent) {
 667          phi->dest.ssa.divergent = true;
 668          return true;
 669       }
 670       if (src->src.ssa->parent_instr->type != nir_instr_type_ssa_undef) {
 671          defined_srcs++;
 672       }
 673    }
 674
 675    /* if the condition is divergent and two sources defined, the definition is divergent */
 676    if (defined_srcs > 1 && if_cond_divergent) {
 677       phi->dest.ssa.divergent = true;
 678       return true;
 679    }
 680
 681    return false;
 682 }
 683
 684 /* There are 3 types of phi instructions:
 685  * (2) mu: which only exist at loop headers,
 686  *     merge initial and loop-carried values.
 687  *     The resulting value is divergent if any source value
 688  *     is divergent or a divergent loop continue condition
 689  *     is associated with a different ssa-def. */
 690 static bool
 691 visit_loop_header_phi(nir_phi_instr *phi, nir_block *preheader, bool divergent_continue)
 692 {
 693    if (phi->dest.ssa.divergent)
 694       return false;
 695
 696    nir_ssa_def* same = NULL;
 697    nir_foreach_phi_src(src, phi) {
 698       /* if any source value is divergent, the resulting value is divergent */
 699       if (src->src.ssa->divergent) {
 700          phi->dest.ssa.divergent = true;
 701          return true;
 702       }
 703       /* if this loop is uniform, we're done here */
 704       if (!divergent_continue)
 705          continue;
 706       /* skip the loop preheader */
 707       if (src->pred == preheader)
 708          continue;
 709       /* skip undef values */
 710       if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
 711          continue;
 712
 713       /* check if all loop-carried values are from the same ssa-def */
 714       if (!same)
 715          same = src->src.ssa;
 716       else if (same != src->src.ssa) {
 717          phi->dest.ssa.divergent = true;
 718          return true;
 719       }
 720    }
 721
 722    return false;
 723 }
 724
 725 /* There are 3 types of phi instructions:
 726  * (3) eta: represent values that leave a loop.
 727  *     The resulting value is divergent if the source value is divergent
 728  *     or any loop exit condition is divergent for a value which is
 729  *     not loop-invariant.
 730  *     (note: there should be no phi for loop-invariant variables.) */
 731 static bool
 732 visit_loop_exit_phi(nir_phi_instr *phi, bool divergent_break)
 733 {
 734    if (phi->dest.ssa.divergent)
 735       return false;
 736
 737    if (divergent_break) {
 738       phi->dest.ssa.divergent = true;
 739       return true;
 740    }
 741
 742    /* if any source value is divergent, the resulting value is divergent */
 743    nir_foreach_phi_src(src, phi) {
 744       if (src->src.ssa->divergent) {
 745          phi->dest.ssa.divergent = true;
 746          return true;
 747       }
 748    }
 749
 750    return false;
 751 }
 752
 753 static bool
 754 visit_if(nir_if *if_stmt, struct divergence_state *state)
 755 {
 756    bool progress = false;
 757
 758    struct divergence_state then_state = *state;
 759    then_state.divergent_loop_cf |= if_stmt->condition.ssa->divergent;
 760    progress |= visit_cf_list(&if_stmt->then_list, &then_state);
 761
 762    struct divergence_state else_state = *state;
 763    else_state.divergent_loop_cf |= if_stmt->condition.ssa->divergent;
 764    progress |= visit_cf_list(&if_stmt->else_list, &else_state);
 765
 766    /* handle phis after the IF */
 767    nir_foreach_instr(instr, nir_cf_node_cf_tree_next(&if_stmt->cf_node)) {
 768       if (instr->type != nir_instr_type_phi)
 769          break;
 770
 771       if (state->first_visit)
 772          nir_instr_as_phi(instr)->dest.ssa.divergent = false;
 773       progress |= visit_if_merge_phi(nir_instr_as_phi(instr),
 774                                      if_stmt->condition.ssa->divergent);
 775    }
 776
 777    /* join loop divergence information from both branch legs */
 778    state->divergent_loop_continue |= then_state.divergent_loop_continue ||
 779                                      else_state.divergent_loop_continue;
 780    state->divergent_loop_break |= then_state.divergent_loop_break ||
 781                                   else_state.divergent_loop_break;
 782
 783    /* A divergent continue makes succeeding loop CF divergent:
 784     * not all loop-active invocations participate in the remaining loop-body
 785     * which means that a following break might be taken by some invocations, only */
 786    state->divergent_loop_cf |= state->divergent_loop_continue;
 787
 788    return progress;
 789 }
 790
 791 static bool
 792 visit_loop(nir_loop *loop, struct divergence_state *state)
 793 {
 794    bool progress = false;
 795    nir_block *loop_header = nir_loop_first_block(loop);
 796    nir_block *loop_preheader = nir_block_cf_tree_prev(loop_header);
 797
 798    /* handle loop header phis first: we have no knowledge yet about
 799     * the loop's control flow or any loop-carried sources. */
 800    nir_foreach_instr(instr, loop_header) {
 801       if (instr->type != nir_instr_type_phi)
 802          break;
 803
 804       nir_phi_instr *phi = nir_instr_as_phi(instr);
 805       if (!state->first_visit && phi->dest.ssa.divergent)
 806          continue;
 807
 808       nir_foreach_phi_src(src, phi) {
 809          if (src->pred == loop_preheader) {
 810             phi->dest.ssa.divergent = src->src.ssa->divergent;
 811             break;
 812          }
 813       }
 814       progress |= phi->dest.ssa.divergent;
 815    }
 816
 817    /* setup loop state */
 818    struct divergence_state loop_state = *state;
 819    loop_state.divergent_loop_cf = false;
 820    loop_state.divergent_loop_continue = false;
 821    loop_state.divergent_loop_break = false;
 822
 823    /* process loop body until no further changes are made */
 824    bool repeat;
 825    do {
 826       progress |= visit_cf_list(&loop->body, &loop_state);
 827       repeat = false;
 828
 829       /* revisit loop header phis to see if something has changed */
 830       nir_foreach_instr(instr, loop_header) {
 831          if (instr->type != nir_instr_type_phi)
 832             break;
 833
 834          repeat |= visit_loop_header_phi(nir_instr_as_phi(instr),
 835                                          loop_preheader,
 836                                          loop_state.divergent_loop_continue);
 837       }
 838
 839       loop_state.divergent_loop_cf = false;
 840       loop_state.first_visit = false;
 841    } while (repeat);
 842
 843    /* handle phis after the loop */
 844    nir_foreach_instr(instr, nir_cf_node_cf_tree_next(&loop->cf_node)) {
 845       if (instr->type != nir_instr_type_phi)
 846          break;
 847
 848       if (state->first_visit)
 849          nir_instr_as_phi(instr)->dest.ssa.divergent = false;
 850       progress |= visit_loop_exit_phi(nir_instr_as_phi(instr),
 851                                       loop_state.divergent_loop_break);
 852    }
 853
 854    return progress;
 855 }
 856
 857 static bool
 858 visit_cf_list(struct exec_list *list, struct divergence_state *state)
 859 {
 860    bool has_changed = false;
 861
 862    foreach_list_typed(nir_cf_node, node, node, list) {
 863       switch (node->type) {
 864       case nir_cf_node_block:
 865          has_changed |= visit_block(nir_cf_node_as_block(node), state);
 866          break;
 867       case nir_cf_node_if:
 868          has_changed |= visit_if(nir_cf_node_as_if(node), state);
 869          break;
 870       case nir_cf_node_loop:
 871          has_changed |= visit_loop(nir_cf_node_as_loop(node), state);
 872          break;
 873       case nir_cf_node_function:
 874          unreachable("NIR divergence analysis: Unsupported cf_node type.");
 875       }
 876    }
 877
 878    return has_changed;
 879 }
 880
 881 void
 882 nir_divergence_analysis(nir_shader *shader, nir_divergence_options options)
 883 {
 884    struct divergence_state state = {
 885       .options = options,
 886       .stage = shader->info.stage,
 887       .divergent_loop_cf = false,
 888       .divergent_loop_continue = false,
 889       .divergent_loop_break = false,
 890       .first_visit = true,
 891    };
 892
 893    visit_cf_list(&nir_shader_get_entrypoint(shader)->body, &state);
 894 }
 895