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(struct exec_list *list,
  41               nir_divergence_options options, gl_shader_stage stage);
  42
  43 static bool
  44 visit_alu(nir_alu_instr *instr)
  45 {
  46    if (instr->dest.dest.ssa.divergent)
  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 (instr->src[i].src.ssa->divergent) {
  53          instr->dest.dest.ssa.divergent = true;
  54          return true;
  55       }
  56    }
  57
  58    return false;
  59 }
  60
  61 static bool
  62 visit_intrinsic(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 (instr->dest.ssa.divergent)
  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_alpha_ref_float:
 102    case nir_intrinsic_load_user_clip_plane:
 103    case nir_intrinsic_load_viewport_x_scale:
 104    case nir_intrinsic_load_viewport_y_scale:
 105    case nir_intrinsic_load_viewport_z_scale:
 106    case nir_intrinsic_load_viewport_offset:
 107    case nir_intrinsic_load_viewport_z_offset:
 108    case nir_intrinsic_load_blend_const_color_a_float:
 109    case nir_intrinsic_load_blend_const_color_b_float:
 110    case nir_intrinsic_load_blend_const_color_g_float:
 111    case nir_intrinsic_load_blend_const_color_r_float:
 112    case nir_intrinsic_load_blend_const_color_rgba:
 113    case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
 114    case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
 115       is_divergent = false;
 116       break;
 117
 118    /* Intrinsics with divergence depending on shader stage and hardware */
 119    case nir_intrinsic_load_input:
 120       is_divergent = instr->src[0].ssa->divergent;
 121       if (stage == MESA_SHADER_FRAGMENT)
 122          is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 123       else if (stage == MESA_SHADER_TESS_EVAL)
 124          is_divergent |= !(options & nir_divergence_single_patch_per_tes_subgroup);
 125       else
 126          is_divergent = true;
 127       break;
 128    case nir_intrinsic_load_input_vertex:
 129       is_divergent = instr->src[1].ssa->divergent;
 130       assert(stage == MESA_SHADER_FRAGMENT);
 131       is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 132       break;
 133    case nir_intrinsic_load_output:
 134       assert(stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_FRAGMENT);
 135       is_divergent = instr->src[0].ssa->divergent;
 136       if (stage == MESA_SHADER_TESS_CTRL)
 137          is_divergent |= !(options & nir_divergence_single_patch_per_tcs_subgroup);
 138       else
 139          is_divergent = true;
 140       break;
 141    case nir_intrinsic_load_layer_id:
 142    case nir_intrinsic_load_front_face:
 143       assert(stage == MESA_SHADER_FRAGMENT);
 144       is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 145       break;
 146    case nir_intrinsic_load_view_index:
 147       assert(stage != MESA_SHADER_COMPUTE && stage != MESA_SHADER_KERNEL);
 148       if (options & nir_divergence_view_index_uniform)
 149          is_divergent = false;
 150       else if (stage == MESA_SHADER_FRAGMENT)
 151          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 152       break;
 153    case nir_intrinsic_load_fs_input_interp_deltas:
 154       assert(stage == MESA_SHADER_FRAGMENT);
 155       is_divergent = instr->src[0].ssa->divergent;
 156       is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
 157       break;
 158    case nir_intrinsic_load_primitive_id:
 159       if (stage == MESA_SHADER_FRAGMENT)
 160          is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
 161       else if (stage == MESA_SHADER_TESS_CTRL)
 162          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 163       else if (stage == MESA_SHADER_TESS_EVAL)
 164          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 165       else if (stage == MESA_SHADER_GEOMETRY)
 166          is_divergent = true;
 167       else
 168          unreachable("Invalid stage for load_primitive_id");
 169       break;
 170    case nir_intrinsic_load_tess_level_inner:
 171    case nir_intrinsic_load_tess_level_outer:
 172       if (stage == MESA_SHADER_TESS_CTRL)
 173          is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
 174       else if (stage == MESA_SHADER_TESS_EVAL)
 175          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 176       else
 177          unreachable("Invalid stage for load_primitive_tess_level_*");
 178       break;
 179    case nir_intrinsic_load_patch_vertices_in:
 180       if (stage == MESA_SHADER_TESS_EVAL)
 181          is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
 182       else
 183          assert(stage == MESA_SHADER_TESS_CTRL);
 184       break;
 185
 186    /* Clustered reductions are uniform if cluster_size == subgroup_size or
 187     * the source is uniform and the operation is invariant.
 188     * Inclusive scans are uniform if
 189     * the source is uniform and the operation is invariant
 190     */
 191    case nir_intrinsic_reduce:
 192       if (nir_intrinsic_cluster_size(instr) == 0)
 193          return false;
 194       /* fallthrough */
 195    case nir_intrinsic_inclusive_scan: {
 196       nir_op op = nir_intrinsic_reduction_op(instr);
 197       is_divergent = instr->src[0].ssa->divergent;
 198       if (op != nir_op_umin && op != nir_op_imin && op != nir_op_fmin &&
 199           op != nir_op_umax && op != nir_op_imax && op != nir_op_fmax &&
 200           op != nir_op_iand && op != nir_op_ior)
 201          is_divergent = true;
 202       break;
 203    }
 204
 205    /* Intrinsics with divergence depending on sources */
 206    case nir_intrinsic_ballot_bitfield_extract:
 207    case nir_intrinsic_ballot_find_lsb:
 208    case nir_intrinsic_ballot_find_msb:
 209    case nir_intrinsic_ballot_bit_count_reduce:
 210    case nir_intrinsic_shuffle_xor:
 211    case nir_intrinsic_shuffle_up:
 212    case nir_intrinsic_shuffle_down:
 213    case nir_intrinsic_quad_broadcast:
 214    case nir_intrinsic_quad_swap_horizontal:
 215    case nir_intrinsic_quad_swap_vertical:
 216    case nir_intrinsic_quad_swap_diagonal:
 217    case nir_intrinsic_load_deref:
 218    case nir_intrinsic_load_ubo:
 219    case nir_intrinsic_load_ssbo:
 220    case nir_intrinsic_load_shared:
 221    case nir_intrinsic_load_global:
 222    case nir_intrinsic_load_uniform:
 223    case nir_intrinsic_load_push_constant:
 224    case nir_intrinsic_load_constant:
 225    case nir_intrinsic_load_sample_pos_from_id:
 226    case nir_intrinsic_load_kernel_input:
 227    case nir_intrinsic_image_load:
 228    case nir_intrinsic_image_deref_load:
 229    case nir_intrinsic_bindless_image_load:
 230    case nir_intrinsic_image_samples:
 231    case nir_intrinsic_image_deref_samples:
 232    case nir_intrinsic_bindless_image_samples:
 233    case nir_intrinsic_get_buffer_size:
 234    case nir_intrinsic_image_size:
 235    case nir_intrinsic_image_deref_size:
 236    case nir_intrinsic_bindless_image_size:
 237    case nir_intrinsic_copy_deref:
 238    case nir_intrinsic_deref_buffer_array_length:
 239    case nir_intrinsic_vulkan_resource_index:
 240    case nir_intrinsic_vulkan_resource_reindex:
 241    case nir_intrinsic_load_vulkan_descriptor:
 242    case nir_intrinsic_atomic_counter_read:
 243    case nir_intrinsic_atomic_counter_read_deref:
 244    case nir_intrinsic_quad_swizzle_amd:
 245    case nir_intrinsic_masked_swizzle_amd: {
 246       unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
 247       for (unsigned i = 0; i < num_srcs; i++) {
 248          if (instr->src[i].ssa->divergent) {
 249             is_divergent = true;
 250             break;
 251          }
 252       }
 253       break;
 254    }
 255
 256    case nir_intrinsic_shuffle:
 257       is_divergent = instr->src[0].ssa->divergent &&
 258                      instr->src[1].ssa->divergent;
 259       break;
 260
 261    /* Intrinsics which are always divergent */
 262    case nir_intrinsic_load_color0:
 263    case nir_intrinsic_load_color1:
 264    case nir_intrinsic_load_param:
 265    case nir_intrinsic_load_sample_id:
 266    case nir_intrinsic_load_sample_id_no_per_sample:
 267    case nir_intrinsic_load_sample_mask_in:
 268    case nir_intrinsic_load_interpolated_input:
 269    case nir_intrinsic_load_barycentric_pixel:
 270    case nir_intrinsic_load_barycentric_centroid:
 271    case nir_intrinsic_load_barycentric_sample:
 272    case nir_intrinsic_load_barycentric_model:
 273    case nir_intrinsic_load_barycentric_at_sample:
 274    case nir_intrinsic_load_barycentric_at_offset:
 275    case nir_intrinsic_interp_deref_at_offset:
 276    case nir_intrinsic_interp_deref_at_sample:
 277    case nir_intrinsic_interp_deref_at_centroid:
 278    case nir_intrinsic_interp_deref_at_vertex:
 279    case nir_intrinsic_load_tess_coord:
 280    case nir_intrinsic_load_point_coord:
 281    case nir_intrinsic_load_frag_coord:
 282    case nir_intrinsic_load_sample_pos:
 283    case nir_intrinsic_load_vertex_id_zero_base:
 284    case nir_intrinsic_load_vertex_id:
 285    case nir_intrinsic_load_per_vertex_input:
 286    case nir_intrinsic_load_per_vertex_output:
 287    case nir_intrinsic_load_instance_id:
 288    case nir_intrinsic_load_invocation_id:
 289    case nir_intrinsic_load_local_invocation_id:
 290    case nir_intrinsic_load_local_invocation_index:
 291    case nir_intrinsic_load_global_invocation_id:
 292    case nir_intrinsic_load_global_invocation_index:
 293    case nir_intrinsic_load_subgroup_invocation:
 294    case nir_intrinsic_load_helper_invocation:
 295    case nir_intrinsic_is_helper_invocation:
 296    case nir_intrinsic_load_scratch:
 297    case nir_intrinsic_deref_atomic_add:
 298    case nir_intrinsic_deref_atomic_imin:
 299    case nir_intrinsic_deref_atomic_umin:
 300    case nir_intrinsic_deref_atomic_imax:
 301    case nir_intrinsic_deref_atomic_umax:
 302    case nir_intrinsic_deref_atomic_and:
 303    case nir_intrinsic_deref_atomic_or:
 304    case nir_intrinsic_deref_atomic_xor:
 305    case nir_intrinsic_deref_atomic_exchange:
 306    case nir_intrinsic_deref_atomic_comp_swap:
 307    case nir_intrinsic_deref_atomic_fadd:
 308    case nir_intrinsic_deref_atomic_fmin:
 309    case nir_intrinsic_deref_atomic_fmax:
 310    case nir_intrinsic_deref_atomic_fcomp_swap:
 311    case nir_intrinsic_ssbo_atomic_add:
 312    case nir_intrinsic_ssbo_atomic_imin:
 313    case nir_intrinsic_ssbo_atomic_umin:
 314    case nir_intrinsic_ssbo_atomic_imax:
 315    case nir_intrinsic_ssbo_atomic_umax:
 316    case nir_intrinsic_ssbo_atomic_and:
 317    case nir_intrinsic_ssbo_atomic_or:
 318    case nir_intrinsic_ssbo_atomic_xor:
 319    case nir_intrinsic_ssbo_atomic_exchange:
 320    case nir_intrinsic_ssbo_atomic_comp_swap:
 321    case nir_intrinsic_ssbo_atomic_fadd:
 322    case nir_intrinsic_ssbo_atomic_fmax:
 323    case nir_intrinsic_ssbo_atomic_fmin:
 324    case nir_intrinsic_ssbo_atomic_fcomp_swap:
 325    case nir_intrinsic_image_deref_atomic_add:
 326    case nir_intrinsic_image_deref_atomic_imin:
 327    case nir_intrinsic_image_deref_atomic_umin:
 328    case nir_intrinsic_image_deref_atomic_imax:
 329    case nir_intrinsic_image_deref_atomic_umax:
 330    case nir_intrinsic_image_deref_atomic_and:
 331    case nir_intrinsic_image_deref_atomic_or:
 332    case nir_intrinsic_image_deref_atomic_xor:
 333    case nir_intrinsic_image_deref_atomic_exchange:
 334    case nir_intrinsic_image_deref_atomic_comp_swap:
 335    case nir_intrinsic_image_deref_atomic_fadd:
 336    case nir_intrinsic_image_atomic_add:
 337    case nir_intrinsic_image_atomic_imin:
 338    case nir_intrinsic_image_atomic_umin:
 339    case nir_intrinsic_image_atomic_imax:
 340    case nir_intrinsic_image_atomic_umax:
 341    case nir_intrinsic_image_atomic_and:
 342    case nir_intrinsic_image_atomic_or:
 343    case nir_intrinsic_image_atomic_xor:
 344    case nir_intrinsic_image_atomic_exchange:
 345    case nir_intrinsic_image_atomic_comp_swap:
 346    case nir_intrinsic_image_atomic_fadd:
 347    case nir_intrinsic_bindless_image_atomic_add:
 348    case nir_intrinsic_bindless_image_atomic_imin:
 349    case nir_intrinsic_bindless_image_atomic_umin:
 350    case nir_intrinsic_bindless_image_atomic_imax:
 351    case nir_intrinsic_bindless_image_atomic_umax:
 352    case nir_intrinsic_bindless_image_atomic_and:
 353    case nir_intrinsic_bindless_image_atomic_or:
 354    case nir_intrinsic_bindless_image_atomic_xor:
 355    case nir_intrinsic_bindless_image_atomic_exchange:
 356    case nir_intrinsic_bindless_image_atomic_comp_swap:
 357    case nir_intrinsic_bindless_image_atomic_fadd:
 358    case nir_intrinsic_shared_atomic_add:
 359    case nir_intrinsic_shared_atomic_imin:
 360    case nir_intrinsic_shared_atomic_umin:
 361    case nir_intrinsic_shared_atomic_imax:
 362    case nir_intrinsic_shared_atomic_umax:
 363    case nir_intrinsic_shared_atomic_and:
 364    case nir_intrinsic_shared_atomic_or:
 365    case nir_intrinsic_shared_atomic_xor:
 366    case nir_intrinsic_shared_atomic_exchange:
 367    case nir_intrinsic_shared_atomic_comp_swap:
 368    case nir_intrinsic_shared_atomic_fadd:
 369    case nir_intrinsic_shared_atomic_fmin:
 370    case nir_intrinsic_shared_atomic_fmax:
 371    case nir_intrinsic_shared_atomic_fcomp_swap:
 372    case nir_intrinsic_global_atomic_add:
 373    case nir_intrinsic_global_atomic_imin:
 374    case nir_intrinsic_global_atomic_umin:
 375    case nir_intrinsic_global_atomic_imax:
 376    case nir_intrinsic_global_atomic_umax:
 377    case nir_intrinsic_global_atomic_and:
 378    case nir_intrinsic_global_atomic_or:
 379    case nir_intrinsic_global_atomic_xor:
 380    case nir_intrinsic_global_atomic_exchange:
 381    case nir_intrinsic_global_atomic_comp_swap:
 382    case nir_intrinsic_global_atomic_fadd:
 383    case nir_intrinsic_global_atomic_fmin:
 384    case nir_intrinsic_global_atomic_fmax:
 385    case nir_intrinsic_global_atomic_fcomp_swap:
 386    case nir_intrinsic_atomic_counter_add:
 387    case nir_intrinsic_atomic_counter_min:
 388    case nir_intrinsic_atomic_counter_max:
 389    case nir_intrinsic_atomic_counter_and:
 390    case nir_intrinsic_atomic_counter_or:
 391    case nir_intrinsic_atomic_counter_xor:
 392    case nir_intrinsic_atomic_counter_inc:
 393    case nir_intrinsic_atomic_counter_pre_dec:
 394    case nir_intrinsic_atomic_counter_post_dec:
 395    case nir_intrinsic_atomic_counter_exchange:
 396    case nir_intrinsic_atomic_counter_comp_swap:
 397    case nir_intrinsic_atomic_counter_add_deref:
 398    case nir_intrinsic_atomic_counter_min_deref:
 399    case nir_intrinsic_atomic_counter_max_deref:
 400    case nir_intrinsic_atomic_counter_and_deref:
 401    case nir_intrinsic_atomic_counter_or_deref:
 402    case nir_intrinsic_atomic_counter_xor_deref:
 403    case nir_intrinsic_atomic_counter_inc_deref:
 404    case nir_intrinsic_atomic_counter_pre_dec_deref:
 405    case nir_intrinsic_atomic_counter_post_dec_deref:
 406    case nir_intrinsic_atomic_counter_exchange_deref:
 407    case nir_intrinsic_atomic_counter_comp_swap_deref:
 408    case nir_intrinsic_exclusive_scan:
 409    case nir_intrinsic_ballot_bit_count_exclusive:
 410    case nir_intrinsic_ballot_bit_count_inclusive:
 411    case nir_intrinsic_write_invocation_amd:
 412    case nir_intrinsic_mbcnt_amd:
 413    case nir_intrinsic_elect:
 414       is_divergent = true;
 415       break;
 416
 417    default:
 418 #ifdef NDEBUG
 419       is_divergent = true;
 420       break;
 421 #else
 422       nir_print_instr(&instr->instr, stderr);
 423       unreachable("\nNIR divergence analysis: Unhandled intrinsic.");
 424 #endif
 425    }
 426
 427    instr->dest.ssa.divergent = is_divergent;
 428    return is_divergent;
 429 }
 430
 431 static bool
 432 visit_tex(nir_tex_instr *instr)
 433 {
 434    if (instr->dest.ssa.divergent)
 435       return false;
 436
 437    bool is_divergent = false;
 438
 439    for (unsigned i = 0; i < instr->num_srcs; i++) {
 440       switch (instr->src[i].src_type) {
 441       case nir_tex_src_sampler_deref:
 442       case nir_tex_src_sampler_handle:
 443       case nir_tex_src_sampler_offset:
 444          is_divergent |= instr->src[i].src.ssa->divergent &&
 445                          instr->sampler_non_uniform;
 446          break;
 447       case nir_tex_src_texture_deref:
 448       case nir_tex_src_texture_handle:
 449       case nir_tex_src_texture_offset:
 450          is_divergent |= instr->src[i].src.ssa->divergent &&
 451                          instr->texture_non_uniform;
 452          break;
 453       default:
 454          is_divergent |= instr->src[i].src.ssa->divergent;
 455          break;
 456       }
 457    }
 458
 459    instr->dest.ssa.divergent = is_divergent;
 460    return is_divergent;
 461 }
 462
 463 static bool
 464 visit_phi(nir_phi_instr *instr)
 465 {
 466    /* There are 3 types of phi instructions:
 467     * (1) gamma: represent the joining point of different paths
 468     *     created by an “if-then-else” branch.
 469     *     The resulting value is divergent if the branch condition
 470     *     or any of the source values is divergent.
 471     *
 472     * (2) mu: which only exist at loop headers,
 473     *     merge initial and loop-carried values.
 474     *     The resulting value is divergent if any source value
 475     *     is divergent or a divergent loop continue condition
 476     *     is associated with a different ssa-def.
 477     *
 478     * (3) eta: represent values that leave a loop.
 479     *     The resulting value is divergent if the source value is divergent
 480     *     or any loop exit condition is divergent for a value which is
 481     *     not loop-invariant.
 482     *     (note: there should be no phi for loop-invariant variables.)
 483     */
 484
 485    if (instr->dest.ssa.divergent)
 486       return false;
 487
 488    nir_foreach_phi_src(src, instr) {
 489       /* if any source value is divergent, the resulting value is divergent */
 490       if (src->src.ssa->divergent) {
 491          instr->dest.ssa.divergent = true;
 492          return true;
 493       }
 494    }
 495
 496    nir_cf_node *prev = nir_cf_node_prev(&instr->instr.block->cf_node);
 497
 498    if (!prev) {
 499       /* mu: if no predecessor node exists, the phi must be at a loop header */
 500       nir_loop *loop = nir_cf_node_as_loop(instr->instr.block->cf_node.parent);
 501       prev = nir_cf_node_prev(&loop->cf_node);
 502       nir_ssa_def* same = NULL;
 503       bool all_same = true;
 504
 505       /* first, check if all loop-carried values are from the same ssa-def */
 506       nir_foreach_phi_src(src, instr) {
 507          if (src->pred == nir_cf_node_as_block(prev))
 508             continue;
 509          if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
 510             continue;
 511          if (!same)
 512             same = src->src.ssa;
 513          else if (same != src->src.ssa)
 514             all_same = false;
 515       }
 516
 517       /* if all loop-carried values are the same, the resulting value is uniform */
 518       if (all_same)
 519          return false;
 520
 521       /* check if the loop-carried values come from different ssa-defs
 522        * and the corresponding condition is divergent. */
 523       nir_foreach_phi_src(src, instr) {
 524          /* skip the loop preheader */
 525          if (src->pred == nir_cf_node_as_block(prev))
 526             continue;
 527
 528          /* skip the unconditional back-edge */
 529          if (src->pred == nir_loop_last_block(loop))
 530             continue;
 531
 532          /* if the value is undef, we don't need to check the condition */
 533          if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
 534             continue;
 535
 536          nir_cf_node *current = src->pred->cf_node.parent;
 537          /* check recursively the conditions if any is divergent */
 538          while (current->type != nir_cf_node_loop) {
 539             assert (current->type == nir_cf_node_if);
 540             nir_if *if_node = nir_cf_node_as_if(current);
 541             if (if_node->condition.ssa->divergent) {
 542                instr->dest.ssa.divergent = true;
 543                return true;
 544             }
 545             current = current->parent;
 546          }
 547          assert(current == &loop->cf_node);
 548       }
 549
 550    } else if (prev->type == nir_cf_node_if) {
 551       /* if only one of the incoming values is defined, the resulting value is uniform */
 552       unsigned defined_srcs = 0;
 553       nir_foreach_phi_src(src, instr) {
 554          if (src->src.ssa->parent_instr->type != nir_instr_type_ssa_undef)
 555             defined_srcs++;
 556       }
 557       if (defined_srcs <= 1)
 558          return false;
 559
 560       /* gamma: check if the condition is divergent */
 561       nir_if *if_node = nir_cf_node_as_if(prev);
 562       if (if_node->condition.ssa->divergent) {
 563          instr->dest.ssa.divergent = true;
 564          return true;
 565       }
 566
 567    } else {
 568       /* eta: the predecessor must be a loop */
 569       assert(prev->type == nir_cf_node_loop);
 570
 571       /* Check if any loop exit condition is divergent:
 572        * That is any break happens under divergent condition or
 573        * a break is preceeded by a divergent continue
 574        */
 575       nir_foreach_phi_src(src, instr) {
 576          nir_cf_node *current = src->pred->cf_node.parent;
 577
 578          /* check recursively the conditions if any is divergent */
 579          while (current->type != nir_cf_node_loop) {
 580             assert(current->type == nir_cf_node_if);
 581             nir_if *if_node = nir_cf_node_as_if(current);
 582             if (if_node->condition.ssa->divergent) {
 583                instr->dest.ssa.divergent = true;
 584                return true;
 585             }
 586             current = current->parent;
 587          }
 588          assert(current == prev);
 589
 590          /* check if any divergent continue happened before the break */
 591          nir_foreach_block_in_cf_node(block, prev) {
 592             if (block == src->pred)
 593                break;
 594             if (!nir_block_ends_in_jump(block))
 595                continue;
 596
 597             nir_jump_instr *jump = nir_instr_as_jump(nir_block_last_instr(block));
 598             if (jump->type != nir_jump_continue)
 599                continue;
 600
 601             current = block->cf_node.parent;
 602             bool is_divergent = false;
 603             while (current != prev) {
 604                /* the continue belongs to an inner loop */
 605                if (current->type == nir_cf_node_loop) {
 606                   is_divergent = false;
 607                   break;
 608                }
 609                assert(current->type == nir_cf_node_if);
 610                nir_if *if_node = nir_cf_node_as_if(current);
 611                is_divergent |= if_node->condition.ssa->divergent;
 612                current = current->parent;
 613             }
 614
 615             if (is_divergent) {
 616                instr->dest.ssa.divergent = true;
 617                return true;
 618             }
 619          }
 620       }
 621    }
 622
 623    return false;
 624 }
 625
 626 static bool
 627 visit_load_const(nir_load_const_instr *instr)
 628 {
 629    return false;
 630 }
 631
 632 static bool
 633 visit_ssa_undef(nir_ssa_undef_instr *instr)
 634 {
 635    return false;
 636 }
 637
 638 static bool
 639 nir_variable_mode_is_uniform(nir_variable_mode mode) {
 640    switch (mode) {
 641    case nir_var_uniform:
 642    case nir_var_mem_ubo:
 643    case nir_var_mem_ssbo:
 644    case nir_var_mem_shared:
 645    case nir_var_mem_global:
 646       return true;
 647    default:
 648       return false;
 649    }
 650 }
 651
 652 static bool
 653 nir_variable_is_uniform(nir_variable *var, nir_divergence_options options,
 654                         gl_shader_stage stage)
 655 {
 656    if (nir_variable_mode_is_uniform(var->data.mode))
 657       return true;
 658
 659    if (stage == MESA_SHADER_FRAGMENT &&
 660        (options & nir_divergence_single_prim_per_subgroup) &&
 661        var->data.mode == nir_var_shader_in &&
 662        var->data.interpolation == INTERP_MODE_FLAT)
 663       return true;
 664
 665    if (stage == MESA_SHADER_TESS_CTRL &&
 666        (options & nir_divergence_single_patch_per_tcs_subgroup) &&
 667        var->data.mode == nir_var_shader_out && var->data.patch)
 668       return true;
 669
 670    if (stage == MESA_SHADER_TESS_EVAL &&
 671        (options & nir_divergence_single_patch_per_tes_subgroup) &&
 672        var->data.mode == nir_var_shader_in && var->data.patch)
 673       return true;
 674
 675    return false;
 676 }
 677
 678 static bool
 679 visit_deref(nir_deref_instr *deref,
 680             nir_divergence_options options, gl_shader_stage stage)
 681 {
 682    if (deref->dest.ssa.divergent)
 683       return false;
 684
 685    bool is_divergent = false;
 686    switch (deref->deref_type) {
 687    case nir_deref_type_var:
 688       is_divergent = !nir_variable_is_uniform(deref->var, options, stage);
 689       break;
 690    case nir_deref_type_array:
 691    case nir_deref_type_ptr_as_array:
 692       is_divergent = deref->arr.index.ssa->divergent;
 693       /* fallthrough */
 694    case nir_deref_type_struct:
 695    case nir_deref_type_array_wildcard:
 696       is_divergent |= deref->parent.ssa->divergent;
 697       break;
 698    case nir_deref_type_cast:
 699       is_divergent = !nir_variable_mode_is_uniform(deref->var->data.mode) ||
 700                      deref->parent.ssa->divergent;
 701       break;
 702    }
 703
 704    deref->dest.ssa.divergent = is_divergent;
 705    return is_divergent;
 706 }
 707
 708 static bool
 709 visit_block(nir_block *block, nir_divergence_options options,
 710             gl_shader_stage stage)
 711 {
 712    bool has_changed = false;
 713
 714    nir_foreach_instr(instr, block) {
 715       switch (instr->type) {
 716       case nir_instr_type_alu:
 717          has_changed |= visit_alu(nir_instr_as_alu(instr));
 718          break;
 719       case nir_instr_type_intrinsic:
 720          has_changed |= visit_intrinsic(nir_instr_as_intrinsic(instr),
 721                                         options, stage);
 722          break;
 723       case nir_instr_type_tex:
 724          has_changed |= visit_tex(nir_instr_as_tex(instr));
 725          break;
 726       case nir_instr_type_phi:
 727          has_changed |= visit_phi(nir_instr_as_phi(instr));
 728          break;
 729       case nir_instr_type_load_const:
 730          has_changed |= visit_load_const(nir_instr_as_load_const(instr));
 731          break;
 732       case nir_instr_type_ssa_undef:
 733          has_changed |= visit_ssa_undef(nir_instr_as_ssa_undef(instr));
 734          break;
 735       case nir_instr_type_deref:
 736          has_changed |= visit_deref(nir_instr_as_deref(instr),
 737                                     options, stage);
 738          break;
 739       case nir_instr_type_jump:
 740          break;
 741       case nir_instr_type_call:
 742       case nir_instr_type_parallel_copy:
 743          unreachable("NIR divergence analysis: Unsupported instruction type.");
 744       }
 745    }
 746
 747    return has_changed;
 748 }
 749
 750 static bool
 751 visit_if(nir_if *if_stmt, nir_divergence_options options, gl_shader_stage stage)
 752 {
 753    return visit_cf_list(&if_stmt->then_list, options, stage) |
 754           visit_cf_list(&if_stmt->else_list, options, stage);
 755 }
 756
 757 static bool
 758 visit_loop(nir_loop *loop, nir_divergence_options options, gl_shader_stage stage)
 759 {
 760    bool has_changed = false;
 761    bool repeat = true;
 762
 763    /* TODO: restructure this and the phi handling more efficiently */
 764    while (repeat) {
 765       repeat = visit_cf_list(&loop->body, options, stage);
 766       has_changed |= repeat;
 767    }
 768
 769    return has_changed;
 770 }
 771
 772 static bool
 773 visit_cf_list(struct exec_list *list,
 774               nir_divergence_options options, gl_shader_stage stage)
 775 {
 776    bool has_changed = false;
 777
 778    foreach_list_typed(nir_cf_node, node, node, list) {
 779       switch (node->type) {
 780       case nir_cf_node_block:
 781          has_changed |= visit_block(nir_cf_node_as_block(node),
 782                                     options, stage);
 783          break;
 784       case nir_cf_node_if:
 785          has_changed |= visit_if(nir_cf_node_as_if(node),
 786                                  options, stage);
 787          break;
 788       case nir_cf_node_loop:
 789          has_changed |= visit_loop(nir_cf_node_as_loop(node),
 790                                    options, stage);
 791          break;
 792       case nir_cf_node_function:
 793          unreachable("NIR divergence analysis: Unsupported cf_node type.");
 794       }
 795    }
 796
 797    return has_changed;
 798 }
 799
 800 static bool
 801 set_ssa_def_not_divergent(nir_ssa_def *def, UNUSED void *_state)
 802 {
 803    def->divergent = false;
 804    return true;
 805 }
 806
 807 void
 808 nir_divergence_analysis(nir_shader *shader, nir_divergence_options options)
 809 {
 810    nir_function_impl *impl = nir_shader_get_entrypoint(shader);
 811
 812    /* Set all SSA defs to non-divergent to start off */
 813    nir_foreach_block(block, impl) {
 814       nir_foreach_instr(instr, block)
 815          nir_foreach_ssa_def(instr, set_ssa_def_not_divergent, NULL);
 816    }
 817
 818    visit_cf_list(&impl->body, options, shader->info.stage);
 819 }