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  46     * divergent because those invocations are no longer active in the loop.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