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 static bool
  40 visit_cf_list(bool *divergent, struct exec_list *list,
  41               nir_divergence_options options, gl_shader_stage stage);
  42
  43 static bool
  44 visit_alu(bool *divergent, nir_alu_instr *instr)
  45 {
  46    if (divergent[instr->dest.dest.ssa.index])
  47       return false;
  48
  49    unsigned num_src = nir_op_infos[instr->op].num_inputs;
  50
  51    for (unsigned i = 0; i < num_src; i++) {
  52       if (divergent[instr->src[i].src.ssa->index]) {
  53          divergent[instr->dest.dest.ssa.index] = true;
  54          return true;
  55       }
  56    }
  57
  58    return false;
  59 }
  60
  61 static bool
  62 visit_intrinsic(bool *divergent, nir_intrinsic_instr *instr,
  63                 nir_divergence_options options, gl_shader_stage stage)
  64 {
  65    if (!nir_intrinsic_infos[instr->intrinsic].has_dest)
  66       return false;
  67
  68    if (divergent[instr->dest.ssa.index])
  69       return false;
  70
  71    bool is_divergent = false;
  72    switch (instr->intrinsic) {
  73    /* Intrinsics which are always uniform */
  74    case nir_intrinsic_shader_clock:
  75    case nir_intrinsic_ballot:
  76    case nir_intrinsic_read_invocation:
  77    case nir_intrinsic_read_first_invocation:
  78    case nir_intrinsic_vote_any:
  79    case nir_intrinsic_vote_all:
  80    case nir_intrinsic_vote_feq:
  81    case nir_intrinsic_vote_ieq:
  82    case nir_intrinsic_load_work_dim:
  83    case nir_intrinsic_load_work_group_id:
  84    case nir_intrinsic_load_num_work_groups:
  85    case nir_intrinsic_load_local_group_size:
  86    case nir_intrinsic_load_subgroup_id:
  87    case nir_intrinsic_load_num_subgroups:
  88    case nir_intrinsic_load_subgroup_size:
  89    case nir_intrinsic_load_subgroup_eq_mask:
  90    case nir_intrinsic_load_subgroup_ge_mask:
  91    case nir_intrinsic_load_subgroup_gt_mask:
  92    case nir_intrinsic_load_subgroup_le_mask:
  93    case nir_intrinsic_load_subgroup_lt_mask:
  94    case nir_intrinsic_first_invocation:
  95    case nir_intrinsic_load_base_instance:
  96    case nir_intrinsic_load_base_vertex:
  97    case nir_intrinsic_load_first_vertex:
  98    case nir_intrinsic_load_draw_id:
  99    case nir_intrinsic_load_is_indexed_draw:
 100    case nir_intrinsic_load_viewport_scale:
 101    case nir_intrinsic_load_user_clip_plane:
 102    case nir_intrinsic_load_viewport_x_scale:
 103    case nir_intrinsic_load_viewport_y_scale:
 104    case nir_intrinsic_load_viewport_z_scale:
 105    case nir_intrinsic_load_viewport_offset:
 106    case nir_intrinsic_load_viewport_z_offset:
 107    case nir_intrinsic_load_blend_const_color_a_float:
 108    case nir_intrinsic_load_blend_const_color_b_float:
 109    case nir_intrinsic_load_blend_const_color_g_float:
 110    case nir_intrinsic_load_blend_const_color_r_float:
 111    case nir_intrinsic_load_blend_const_color_rgba:
 112    case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
 113    case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
 114       is_divergent = false;
 115       break;
 116
 117    /* Intrinsics with divergence depending on shader stage and hardware */
 118    case nir_intrinsic_load_input:
 119       is_divergent = divergent[instr->src[0].ssa->index];
 120       if (stage == MESA_SHADER_FRAGMENT)
 121          is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 122       else if (stage == MESA_SHADER_TESS_EVAL)
 123          is_divergent |= !(options & nir_divergence_single_patch_per_tes_subgroup);
 124       else
 125          is_divergent = true;
 126       break;
 127    case nir_intrinsic_load_output:
 128       assert(stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_FRAGMENT);
 129       is_divergent = divergent[instr->src[0].ssa->index];
 130       if (stage == MESA_SHADER_TESS_CTRL)
 131          is_divergent |= !(options & nir_divergence_single_patch_per_tcs_subgroup);
 132       else
 133          is_divergent = true;
 134       break;
 135    case nir_intrinsic_load_layer_id:
 136    case nir_intrinsic_load_front_face:
 137       assert(stage == MESA_SHADER_FRAGMENT);
 138       is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 139       break;
 140    case nir_intrinsic_load_view_index:
 141       assert(stage != MESA_SHADER_COMPUTE && stage != MESA_SHADER_KERNEL);
 142       if (options & nir_divergence_view_index_uniform)
 143          is_divergent = false;
 144       else if (stage == MESA_SHADER_FRAGMENT)
 145          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 146       break;
 147    case nir_intrinsic_load_fs_input_interp_deltas:
 148       assert(stage == MESA_SHADER_FRAGMENT);
 149       is_divergent = divergent[instr->src[0].ssa->index];
 150       is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 151       break;
 152    case nir_intrinsic_load_primitive_id:
 153       if (stage == MESA_SHADER_FRAGMENT)
 154          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 155       else if (stage == MESA_SHADER_TESS_CTRL)
 156          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 157       else if (stage == MESA_SHADER_TESS_EVAL)
 158          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 159       else
 160          unreachable("Invalid stage for load_primitive_id");
 161       break;
 162    case nir_intrinsic_load_tess_level_inner:
 163    case nir_intrinsic_load_tess_level_outer:
 164       if (stage == MESA_SHADER_TESS_CTRL)
 165          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 166       else if (stage == MESA_SHADER_TESS_EVAL)
 167          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 168       else
 169          unreachable("Invalid stage for load_primitive_tess_level_*");
 170       break;
 171    case nir_intrinsic_load_patch_vertices_in:
 172       if (stage == MESA_SHADER_TESS_EVAL)
 173          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 174       else
 175          assert(stage == MESA_SHADER_TESS_CTRL);
 176       break;
 177
 178    /* Clustered reductions are uniform if cluster_size == subgroup_size or
 179     * the source is uniform and the operation is invariant.
 180     * Inclusive scans are uniform if
 181     * the source is uniform and the operation is invariant
 182     */
 183    case nir_intrinsic_reduce:
 184       if (nir_intrinsic_cluster_size(instr) == 0)
 185          return false;
 186       /* fallthrough */
 187    case nir_intrinsic_inclusive_scan: {
 188       nir_op op = nir_intrinsic_reduction_op(instr);
 189       is_divergent = divergent[instr->src[0].ssa->index];
 190       if (op != nir_op_umin && op != nir_op_imin && op != nir_op_fmin &&
 191           op != nir_op_umax && op != nir_op_imax && op != nir_op_fmax &&
 192           op != nir_op_iand && op != nir_op_ior)
 193          is_divergent = true;
 194       break;
 195    }
 196
 197    /* Intrinsics with divergence depending on sources */
 198    case nir_intrinsic_ballot_bitfield_extract:
 199    case nir_intrinsic_ballot_find_lsb:
 200    case nir_intrinsic_ballot_find_msb:
 201    case nir_intrinsic_ballot_bit_count_reduce:
 202    case nir_intrinsic_shuffle:
 203    case nir_intrinsic_shuffle_xor:
 204    case nir_intrinsic_shuffle_up:
 205    case nir_intrinsic_shuffle_down:
 206    case nir_intrinsic_quad_broadcast:
 207    case nir_intrinsic_quad_swap_horizontal:
 208    case nir_intrinsic_quad_swap_vertical:
 209    case nir_intrinsic_quad_swap_diagonal:
 210    case nir_intrinsic_load_deref:
 211    case nir_intrinsic_load_ubo:
 212    case nir_intrinsic_load_ssbo:
 213    case nir_intrinsic_load_shared:
 214    case nir_intrinsic_load_global:
 215    case nir_intrinsic_load_uniform:
 216    case nir_intrinsic_load_push_constant:
 217    case nir_intrinsic_load_constant:
 218    case nir_intrinsic_load_sample_pos_from_id:
 219    case nir_intrinsic_load_kernel_input:
 220    case nir_intrinsic_image_load:
 221    case nir_intrinsic_image_deref_load:
 222    case nir_intrinsic_bindless_image_load:
 223    case nir_intrinsic_image_samples:
 224    case nir_intrinsic_image_deref_samples:
 225    case nir_intrinsic_bindless_image_samples:
 226    case nir_intrinsic_get_buffer_size:
 227    case nir_intrinsic_image_size:
 228    case nir_intrinsic_image_deref_size:
 229    case nir_intrinsic_bindless_image_size:
 230    case nir_intrinsic_copy_deref:
 231    case nir_intrinsic_deref_buffer_array_length:
 232    case nir_intrinsic_vulkan_resource_index:
 233    case nir_intrinsic_vulkan_resource_reindex:
 234    case nir_intrinsic_load_vulkan_descriptor:
 235    case nir_intrinsic_atomic_counter_read:
 236    case nir_intrinsic_atomic_counter_read_deref:
 237    case nir_intrinsic_quad_swizzle_amd:
 238    case nir_intrinsic_masked_swizzle_amd: {
 239       unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
 240       for (unsigned i = 0; i < num_srcs; i++) {
 241          if (divergent[instr->src[i].ssa->index]) {
 242             is_divergent = true;
 243             break;
 244          }
 245       }
 246       break;
 247    }
 248
 249    /* Intrinsics which are always divergent */
 250    case nir_intrinsic_load_color0:
 251    case nir_intrinsic_load_color1:
 252    case nir_intrinsic_load_param:
 253    case nir_intrinsic_load_sample_id:
 254    case nir_intrinsic_load_sample_id_no_per_sample:
 255    case nir_intrinsic_load_sample_mask_in:
 256    case nir_intrinsic_load_interpolated_input:
 257    case nir_intrinsic_load_barycentric_pixel:
 258    case nir_intrinsic_load_barycentric_centroid:
 259    case nir_intrinsic_load_barycentric_sample:
 260    case nir_intrinsic_load_barycentric_at_sample:
 261    case nir_intrinsic_load_barycentric_at_offset:
 262    case nir_intrinsic_interp_deref_at_offset:
 263    case nir_intrinsic_interp_deref_at_sample:
 264    case nir_intrinsic_interp_deref_at_centroid:
 265    case nir_intrinsic_load_tess_coord:
 266    case nir_intrinsic_load_point_coord:
 267    case nir_intrinsic_load_frag_coord:
 268    case nir_intrinsic_load_sample_pos:
 269    case nir_intrinsic_load_vertex_id_zero_base:
 270    case nir_intrinsic_load_vertex_id:
 271    case nir_intrinsic_load_per_vertex_input:
 272    case nir_intrinsic_load_per_vertex_output:
 273    case nir_intrinsic_load_instance_id:
 274    case nir_intrinsic_load_invocation_id:
 275    case nir_intrinsic_load_local_invocation_id:
 276    case nir_intrinsic_load_local_invocation_index:
 277    case nir_intrinsic_load_global_invocation_id:
 278    case nir_intrinsic_load_global_invocation_index:
 279    case nir_intrinsic_load_subgroup_invocation:
 280    case nir_intrinsic_load_helper_invocation:
 281    case nir_intrinsic_is_helper_invocation:
 282    case nir_intrinsic_load_scratch:
 283    case nir_intrinsic_deref_atomic_add:
 284    case nir_intrinsic_deref_atomic_imin:
 285    case nir_intrinsic_deref_atomic_umin:
 286    case nir_intrinsic_deref_atomic_imax:
 287    case nir_intrinsic_deref_atomic_umax:
 288    case nir_intrinsic_deref_atomic_and:
 289    case nir_intrinsic_deref_atomic_or:
 290    case nir_intrinsic_deref_atomic_xor:
 291    case nir_intrinsic_deref_atomic_exchange:
 292    case nir_intrinsic_deref_atomic_comp_swap:
 293    case nir_intrinsic_deref_atomic_fadd:
 294    case nir_intrinsic_deref_atomic_fmin:
 295    case nir_intrinsic_deref_atomic_fmax:
 296    case nir_intrinsic_deref_atomic_fcomp_swap:
 297    case nir_intrinsic_ssbo_atomic_add:
 298    case nir_intrinsic_ssbo_atomic_imin:
 299    case nir_intrinsic_ssbo_atomic_umin:
 300    case nir_intrinsic_ssbo_atomic_imax:
 301    case nir_intrinsic_ssbo_atomic_umax:
 302    case nir_intrinsic_ssbo_atomic_and:
 303    case nir_intrinsic_ssbo_atomic_or:
 304    case nir_intrinsic_ssbo_atomic_xor:
 305    case nir_intrinsic_ssbo_atomic_exchange:
 306    case nir_intrinsic_ssbo_atomic_comp_swap:
 307    case nir_intrinsic_ssbo_atomic_fadd:
 308    case nir_intrinsic_ssbo_atomic_fmax:
 309    case nir_intrinsic_ssbo_atomic_fmin:
 310    case nir_intrinsic_ssbo_atomic_fcomp_swap:
 311    case nir_intrinsic_image_deref_atomic_add:
 312    case nir_intrinsic_image_deref_atomic_imin:
 313    case nir_intrinsic_image_deref_atomic_umin:
 314    case nir_intrinsic_image_deref_atomic_imax:
 315    case nir_intrinsic_image_deref_atomic_umax:
 316    case nir_intrinsic_image_deref_atomic_and:
 317    case nir_intrinsic_image_deref_atomic_or:
 318    case nir_intrinsic_image_deref_atomic_xor:
 319    case nir_intrinsic_image_deref_atomic_exchange:
 320    case nir_intrinsic_image_deref_atomic_comp_swap:
 321    case nir_intrinsic_image_deref_atomic_fadd:
 322    case nir_intrinsic_image_atomic_add:
 323    case nir_intrinsic_image_atomic_imin:
 324    case nir_intrinsic_image_atomic_umin:
 325    case nir_intrinsic_image_atomic_imax:
 326    case nir_intrinsic_image_atomic_umax:
 327    case nir_intrinsic_image_atomic_and:
 328    case nir_intrinsic_image_atomic_or:
 329    case nir_intrinsic_image_atomic_xor:
 330    case nir_intrinsic_image_atomic_exchange:
 331    case nir_intrinsic_image_atomic_comp_swap:
 332    case nir_intrinsic_image_atomic_fadd:
 333    case nir_intrinsic_bindless_image_atomic_add:
 334    case nir_intrinsic_bindless_image_atomic_imin:
 335    case nir_intrinsic_bindless_image_atomic_umin:
 336    case nir_intrinsic_bindless_image_atomic_imax:
 337    case nir_intrinsic_bindless_image_atomic_umax:
 338    case nir_intrinsic_bindless_image_atomic_and:
 339    case nir_intrinsic_bindless_image_atomic_or:
 340    case nir_intrinsic_bindless_image_atomic_xor:
 341    case nir_intrinsic_bindless_image_atomic_exchange:
 342    case nir_intrinsic_bindless_image_atomic_comp_swap:
 343    case nir_intrinsic_bindless_image_atomic_fadd:
 344    case nir_intrinsic_shared_atomic_add:
 345    case nir_intrinsic_shared_atomic_imin:
 346    case nir_intrinsic_shared_atomic_umin:
 347    case nir_intrinsic_shared_atomic_imax:
 348    case nir_intrinsic_shared_atomic_umax:
 349    case nir_intrinsic_shared_atomic_and:
 350    case nir_intrinsic_shared_atomic_or:
 351    case nir_intrinsic_shared_atomic_xor:
 352    case nir_intrinsic_shared_atomic_exchange:
 353    case nir_intrinsic_shared_atomic_comp_swap:
 354    case nir_intrinsic_shared_atomic_fadd:
 355    case nir_intrinsic_shared_atomic_fmin:
 356    case nir_intrinsic_shared_atomic_fmax:
 357    case nir_intrinsic_shared_atomic_fcomp_swap:
 358    case nir_intrinsic_global_atomic_add:
 359    case nir_intrinsic_global_atomic_imin:
 360    case nir_intrinsic_global_atomic_umin:
 361    case nir_intrinsic_global_atomic_imax:
 362    case nir_intrinsic_global_atomic_umax:
 363    case nir_intrinsic_global_atomic_and:
 364    case nir_intrinsic_global_atomic_or:
 365    case nir_intrinsic_global_atomic_xor:
 366    case nir_intrinsic_global_atomic_exchange:
 367    case nir_intrinsic_global_atomic_comp_swap:
 368    case nir_intrinsic_global_atomic_fadd:
 369    case nir_intrinsic_global_atomic_fmin:
 370    case nir_intrinsic_global_atomic_fmax:
 371    case nir_intrinsic_global_atomic_fcomp_swap:
 372    case nir_intrinsic_atomic_counter_add:
 373    case nir_intrinsic_atomic_counter_min:
 374    case nir_intrinsic_atomic_counter_max:
 375    case nir_intrinsic_atomic_counter_and:
 376    case nir_intrinsic_atomic_counter_or:
 377    case nir_intrinsic_atomic_counter_xor:
 378    case nir_intrinsic_atomic_counter_inc:
 379    case nir_intrinsic_atomic_counter_pre_dec:
 380    case nir_intrinsic_atomic_counter_post_dec:
 381    case nir_intrinsic_atomic_counter_exchange:
 382    case nir_intrinsic_atomic_counter_comp_swap:
 383    case nir_intrinsic_atomic_counter_add_deref:
 384    case nir_intrinsic_atomic_counter_min_deref:
 385    case nir_intrinsic_atomic_counter_max_deref:
 386    case nir_intrinsic_atomic_counter_and_deref:
 387    case nir_intrinsic_atomic_counter_or_deref:
 388    case nir_intrinsic_atomic_counter_xor_deref:
 389    case nir_intrinsic_atomic_counter_inc_deref:
 390    case nir_intrinsic_atomic_counter_pre_dec_deref:
 391    case nir_intrinsic_atomic_counter_post_dec_deref:
 392    case nir_intrinsic_atomic_counter_exchange_deref:
 393    case nir_intrinsic_atomic_counter_comp_swap_deref:
 394    case nir_intrinsic_exclusive_scan:
 395    case nir_intrinsic_ballot_bit_count_exclusive:
 396    case nir_intrinsic_ballot_bit_count_inclusive:
 397    case nir_intrinsic_write_invocation_amd:
 398    case nir_intrinsic_mbcnt_amd:
 399       is_divergent = true;
 400       break;
 401
 402    default:
 403 #ifdef NDEBUG
 404       is_divergent = true;
 405       break;
 406 #else
 407       nir_print_instr(&instr->instr, stderr);
 408       unreachable("\nNIR divergence analysis: Unhandled intrinsic.");
 409 #endif
 410    }
 411
 412    divergent[instr->dest.ssa.index] = is_divergent;
 413    return is_divergent;
 414 }
 415
 416 static bool
 417 visit_tex(bool *divergent, nir_tex_instr *instr)
 418 {
 419    if (divergent[instr->dest.ssa.index])
 420       return false;
 421
 422    bool is_divergent = false;
 423
 424    for (unsigned i = 0; i < instr->num_srcs; i++) {
 425       switch (instr->src[i].src_type) {
 426       case nir_tex_src_sampler_deref:
 427       case nir_tex_src_sampler_handle:
 428       case nir_tex_src_sampler_offset:
 429          is_divergent |= divergent[instr->src[i].src.ssa->index] &&
 430                          instr->sampler_non_uniform;
 431          break;
 432       case nir_tex_src_texture_deref:
 433       case nir_tex_src_texture_handle:
 434       case nir_tex_src_texture_offset:
 435          is_divergent |= divergent[instr->src[i].src.ssa->index] &&
 436                          instr->texture_non_uniform;
 437          break;
 438       default:
 439          is_divergent |= divergent[instr->src[i].src.ssa->index];
 440          break;
 441       }
 442    }
 443
 444    divergent[instr->dest.ssa.index] = is_divergent;
 445    return is_divergent;
 446 }
 447
 448 static bool
 449 visit_phi(bool *divergent, nir_phi_instr *instr)
 450 {
 451    /* There are 3 types of phi instructions:
 452     * (1) gamma: represent the joining point of different paths
 453     *     created by an “if-then-else” branch.
 454     *     The resulting value is divergent if the branch condition
 455     *     or any of the source values is divergent.
 456     *
 457     * (2) mu: which only exist at loop headers,
 458     *     merge initial and loop-carried values.
 459     *     The resulting value is divergent if any source value
 460     *     is divergent or a divergent loop continue condition
 461     *     is associated with a different ssa-def.
 462     *
 463     * (3) eta: represent values that leave a loop.
 464     *     The resulting value is divergent if the source value is divergent
 465     *     or any loop exit condition is divergent for a value which is
 466     *     not loop-invariant.
 467     *     (note: there should be no phi for loop-invariant variables.)
 468     */
 469
 470    if (divergent[instr->dest.ssa.index])
 471       return false;
 472
 473    nir_foreach_phi_src(src, instr) {
 474       /* if any source value is divergent, the resulting value is divergent */
 475       if (divergent[src->src.ssa->index]) {
 476          divergent[instr->dest.ssa.index] = true;
 477          return true;
 478       }
 479    }
 480
 481    nir_cf_node *prev = nir_cf_node_prev(&instr->instr.block->cf_node);
 482
 483    if (!prev) {
 484       /* mu: if no predecessor node exists, the phi must be at a loop header */
 485       nir_loop *loop = nir_cf_node_as_loop(instr->instr.block->cf_node.parent);
 486       prev = nir_cf_node_prev(&loop->cf_node);
 487       nir_ssa_def* same = NULL;
 488       bool all_same = true;
 489
 490       /* first, check if all loop-carried values are from the same ssa-def */
 491       nir_foreach_phi_src(src, instr) {
 492          if (src->pred == nir_cf_node_as_block(prev))
 493             continue;
 494          if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
 495             continue;
 496          if (!same)
 497             same = src->src.ssa;
 498          else if (same != src->src.ssa)
 499             all_same = false;
 500       }
 501
 502       /* if all loop-carried values are the same, the resulting value is uniform */
 503       if (all_same)
 504          return false;
 505
 506       /* check if the loop-carried values come from different ssa-defs
 507        * and the corresponding condition is divergent. */
 508       nir_foreach_phi_src(src, instr) {
 509          /* skip the loop preheader */
 510          if (src->pred == nir_cf_node_as_block(prev))
 511             continue;
 512
 513          /* skip the unconditional back-edge */
 514          if (src->pred == nir_loop_last_block(loop))
 515             continue;
 516
 517          /* if the value is undef, we don't need to check the condition */
 518          if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
 519             continue;
 520
 521          nir_cf_node *current = src->pred->cf_node.parent;
 522          /* check recursively the conditions if any is divergent */
 523          while (current->type != nir_cf_node_loop) {
 524             assert (current->type == nir_cf_node_if);
 525             nir_if *if_node = nir_cf_node_as_if(current);
 526             if (divergent[if_node->condition.ssa->index]) {
 527                divergent[instr->dest.ssa.index] = true;
 528                return true;
 529             }
 530             current = current->parent;
 531          }
 532          assert(current == &loop->cf_node);
 533       }
 534
 535    } else if (prev->type == nir_cf_node_if) {
 536       /* if only one of the incoming values is defined, the resulting value is uniform */
 537       unsigned defined_srcs = 0;
 538       nir_foreach_phi_src(src, instr) {
 539          if (src->src.ssa->parent_instr->type != nir_instr_type_ssa_undef)
 540             defined_srcs++;
 541       }
 542       if (defined_srcs <= 1)
 543          return false;
 544
 545       /* gamma: check if the condition is divergent */
 546       nir_if *if_node = nir_cf_node_as_if(prev);
 547       if (divergent[if_node->condition.ssa->index]) {
 548          divergent[instr->dest.ssa.index] = true;
 549          return true;
 550       }
 551
 552    } else {
 553       /* eta: the predecessor must be a loop */
 554       assert(prev->type == nir_cf_node_loop);
 555
 556       /* Check if any loop exit condition is divergent:
 557        * That is any break happens under divergent condition or
 558        * a break is preceeded by a divergent continue
 559        */
 560       nir_foreach_phi_src(src, instr) {
 561          nir_cf_node *current = src->pred->cf_node.parent;
 562
 563          /* check recursively the conditions if any is divergent */
 564          while (current->type != nir_cf_node_loop) {
 565             assert(current->type == nir_cf_node_if);
 566             nir_if *if_node = nir_cf_node_as_if(current);
 567             if (divergent[if_node->condition.ssa->index]) {
 568                divergent[instr->dest.ssa.index] = true;
 569                return true;
 570             }
 571             current = current->parent;
 572          }
 573          assert(current == prev);
 574
 575          /* check if any divergent continue happened before the break */
 576          nir_foreach_block_in_cf_node(block, prev) {
 577             if (block == src->pred)
 578                break;
 579             if (!nir_block_ends_in_jump(block))
 580                continue;
 581
 582             nir_jump_instr *jump = nir_instr_as_jump(nir_block_last_instr(block));
 583             if (jump->type != nir_jump_continue)
 584                continue;
 585
 586             current = block->cf_node.parent;
 587             bool is_divergent = false;
 588             while (current != prev) {
 589                /* the continue belongs to an inner loop */
 590                if (current->type == nir_cf_node_loop) {
 591                   is_divergent = false;
 592                   break;
 593                }
 594                assert(current->type == nir_cf_node_if);
 595                nir_if *if_node = nir_cf_node_as_if(current);
 596                is_divergent |= divergent[if_node->condition.ssa->index];
 597                current = current->parent;
 598             }
 599
 600             if (is_divergent) {
 601                divergent[instr->dest.ssa.index] = true;
 602                return true;
 603             }
 604          }
 605       }
 606    }
 607
 608    return false;
 609 }
 610
 611 static bool
 612 visit_load_const(bool *divergent, nir_load_const_instr *instr)
 613 {
 614    return false;
 615 }
 616
 617 static bool
 618 visit_ssa_undef(bool *divergent, nir_ssa_undef_instr *instr)
 619 {
 620    return false;
 621 }
 622
 623 static bool
 624 nir_variable_mode_is_uniform(nir_variable_mode mode) {
 625    switch (mode) {
 626    case nir_var_uniform:
 627    case nir_var_mem_ubo:
 628    case nir_var_mem_ssbo:
 629    case nir_var_mem_shared:
 630    case nir_var_mem_global:
 631       return true;
 632    default:
 633       return false;
 634    }
 635 }
 636
 637 static bool
 638 nir_variable_is_uniform(nir_variable *var, nir_divergence_options options,
 639                         gl_shader_stage stage)
 640 {
 641    if (nir_variable_mode_is_uniform(var->data.mode))
 642       return true;
 643
 644    if (stage == MESA_SHADER_FRAGMENT &&
 645        (options & nir_divergence_single_prim_per_subgroup) &&
 646        var->data.mode == nir_var_shader_in &&
 647        var->data.interpolation == INTERP_MODE_FLAT)
 648       return true;
 649
 650    if (stage == MESA_SHADER_TESS_CTRL &&
 651        (options & nir_divergence_single_patch_per_tcs_subgroup) &&
 652        var->data.mode == nir_var_shader_out && var->data.patch)
 653       return true;
 654
 655    if (stage == MESA_SHADER_TESS_EVAL &&
 656        (options & nir_divergence_single_patch_per_tes_subgroup) &&
 657        var->data.mode == nir_var_shader_in && var->data.patch)
 658       return true;
 659
 660    return false;
 661 }
 662
 663 static bool
 664 visit_deref(bool *divergent, nir_deref_instr *deref,
 665             nir_divergence_options options, gl_shader_stage stage)
 666 {
 667    if (divergent[deref->dest.ssa.index])
 668       return false;
 669
 670    bool is_divergent = false;
 671    switch (deref->deref_type) {
 672    case nir_deref_type_var:
 673       is_divergent = !nir_variable_is_uniform(deref->var, options, stage);
 674       break;
 675    case nir_deref_type_array:
 676    case nir_deref_type_ptr_as_array:
 677       is_divergent = divergent[deref->arr.index.ssa->index];
 678       /* fallthrough */
 679    case nir_deref_type_struct:
 680    case nir_deref_type_array_wildcard:
 681       is_divergent |= divergent[deref->parent.ssa->index];
 682       break;
 683    case nir_deref_type_cast:
 684       is_divergent = !nir_variable_mode_is_uniform(deref->var->data.mode) ||
 685                      divergent[deref->parent.ssa->index];
 686       break;
 687    }
 688
 689    divergent[deref->dest.ssa.index] = is_divergent;
 690    return is_divergent;
 691 }
 692
 693 static bool
 694 visit_block(bool *divergent, nir_block *block, nir_divergence_options options,
 695             gl_shader_stage stage)
 696 {
 697    bool has_changed = false;
 698
 699    nir_foreach_instr(instr, block) {
 700       switch (instr->type) {
 701       case nir_instr_type_alu:
 702          has_changed |= visit_alu(divergent, nir_instr_as_alu(instr));
 703          break;
 704       case nir_instr_type_intrinsic:
 705          has_changed |= visit_intrinsic(divergent, nir_instr_as_intrinsic(instr),
 706                                         options, stage);
 707          break;
 708       case nir_instr_type_tex:
 709          has_changed |= visit_tex(divergent, nir_instr_as_tex(instr));
 710          break;
 711       case nir_instr_type_phi:
 712          has_changed |= visit_phi(divergent, nir_instr_as_phi(instr));
 713          break;
 714       case nir_instr_type_load_const:
 715          has_changed |= visit_load_const(divergent, nir_instr_as_load_const(instr));
 716          break;
 717       case nir_instr_type_ssa_undef:
 718          has_changed |= visit_ssa_undef(divergent, nir_instr_as_ssa_undef(instr));
 719          break;
 720       case nir_instr_type_deref:
 721          has_changed |= visit_deref(divergent, nir_instr_as_deref(instr),
 722                                     options, stage);
 723          break;
 724       case nir_instr_type_jump:
 725          break;
 726       case nir_instr_type_call:
 727       case nir_instr_type_parallel_copy:
 728          unreachable("NIR divergence analysis: Unsupported instruction type.");
 729       }
 730    }
 731
 732    return has_changed;
 733 }
 734
 735 static bool
 736 visit_if(bool *divergent, nir_if *if_stmt, nir_divergence_options options, gl_shader_stage stage)
 737 {
 738    return visit_cf_list(divergent, &if_stmt->then_list, options, stage) |
 739           visit_cf_list(divergent, &if_stmt->else_list, options, stage);
 740 }
 741
 742 static bool
 743 visit_loop(bool *divergent, nir_loop *loop, nir_divergence_options options, gl_shader_stage stage)
 744 {
 745    bool has_changed = false;
 746    bool repeat = true;
 747
 748    /* TODO: restructure this and the phi handling more efficiently */
 749    while (repeat) {
 750       repeat = visit_cf_list(divergent, &loop->body, options, stage);
 751       has_changed |= repeat;
 752    }
 753
 754    return has_changed;
 755 }
 756
 757 static bool
 758 visit_cf_list(bool *divergent, struct exec_list *list,
 759               nir_divergence_options options, gl_shader_stage stage)
 760 {
 761    bool has_changed = false;
 762
 763    foreach_list_typed(nir_cf_node, node, node, list) {
 764       switch (node->type) {
 765       case nir_cf_node_block:
 766          has_changed |= visit_block(divergent, nir_cf_node_as_block(node),
 767                                     options, stage);
 768          break;
 769       case nir_cf_node_if:
 770          has_changed |= visit_if(divergent, nir_cf_node_as_if(node),
 771                                  options, stage);
 772          break;
 773       case nir_cf_node_loop:
 774          has_changed |= visit_loop(divergent, nir_cf_node_as_loop(node),
 775                                    options, stage);
 776          break;
 777       case nir_cf_node_function:
 778          unreachable("NIR divergence analysis: Unsupported cf_node type.");
 779       }
 780    }
 781
 782    return has_changed;
 783 }
 784
 785
 786 bool*
 787 nir_divergence_analysis(nir_shader *shader, nir_divergence_options options)
 788 {
 789    nir_function_impl *impl = nir_shader_get_entrypoint(shader);
 790    bool *t = rzalloc_array(shader, bool, impl->ssa_alloc);
 791
 792    visit_cf_list(t, &impl->body, options, shader->info.stage);
 793
 794    return t;
 795 }