nir: add divergence analysis pass.
authorDaniel Schürmann <daniel@schuermann.dev>
Mon, 20 May 2019 12:58:23 +0000 (14:58 +0200)
committerDaniel Schürmann <daniel@schuermann.dev>
Tue, 20 Aug 2019 15:40:13 +0000 (17:40 +0200)
This pass expects the shader to be in LCSSA form.
The algorithm is based on 'The Simple Divergence Analysis' from
Diogo Sampaio, Rafael De Souza, Sylvain Collange, Fernando Magno Quintão Pereira.
Divergence Analysis. ACM Transactions on Programming Languages and Systems (TOPLAS)

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
src/compiler/nir/meson.build
src/compiler/nir/nir.h
src/compiler/nir/nir_divergence_analysis.c [new file with mode: 0644]

index 541cc00df9df0245922c5007e2c2b2e7a0c19b9d..04e4f729902760368448a2141bc7a2c9aa2e54bf 100644 (file)
@@ -96,6 +96,7 @@ files_libnir = files(
   'nir_control_flow_private.h',
   'nir_deref.c',
   'nir_deref.h',
+  'nir_divergence_analysis.c',
   'nir_dominance.c',
   'nir_format_convert.h',
   'nir_from_ssa.c',
index 57326fe180e6c34b0e8f09b300e1f5dabb7a6c99..0462fd89a04f91aa6ed1316b6513ded93a95396a 100644 (file)
@@ -2506,6 +2506,13 @@ typedef enum {
    nir_lower_fp64_full_software = (1 << 11),
 } nir_lower_doubles_options;
 
+typedef enum {
+   nir_divergence_single_prim_per_subgroup = (1 << 0),
+   nir_divergence_single_patch_per_tcs_subgroup = (1 << 1),
+   nir_divergence_single_patch_per_tes_subgroup = (1 << 2),
+   nir_divergence_view_index_uniform = (1 << 3),
+} nir_divergence_options;
+
 typedef struct nir_shader_compiler_options {
    bool lower_fdiv;
    bool lower_ffma;
@@ -3924,6 +3931,7 @@ bool nir_repair_ssa(nir_shader *shader);
 
 void nir_convert_loop_to_lcssa(nir_loop *loop);
 bool nir_convert_to_lcssa(nir_shader *shader, bool skip_invariants, bool skip_bool_invariants);
+bool* nir_divergence_analysis(nir_shader *shader, nir_divergence_options options);
 
 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
  * registers.  If false, convert all values (even those not involved in a phi
diff --git a/src/compiler/nir/nir_divergence_analysis.c b/src/compiler/nir/nir_divergence_analysis.c
new file mode 100644 (file)
index 0000000..fd61c60
--- /dev/null
@@ -0,0 +1,790 @@
+/*
+ * Copyright © 2018 Valve Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+
+#include "nir.h"
+
+/* This pass computes for each ssa definition if it is uniform.
+ * That is, the variable has the same value for all invocations
+ * of the group.
+ *
+ * This divergence analysis pass expects the shader to be in LCSSA-form.
+ *
+ * This algorithm implements "The Simple Divergence Analysis" from
+ * Diogo Sampaio, Rafael De Souza, Sylvain Collange, Fernando Magno Quintão Pereira.
+ * Divergence Analysis.  ACM Transactions on Programming Languages and Systems (TOPLAS),
+ * ACM, 2013, 35 (4), pp.13:1-13:36. <10.1145/2523815>. <hal-00909072v2>
+ */
+
+static bool
+visit_cf_list(bool *divergent, struct exec_list *list,
+              nir_divergence_options options, gl_shader_stage stage);
+
+static bool
+visit_alu(bool *divergent, nir_alu_instr *instr)
+{
+   if (divergent[instr->dest.dest.ssa.index])
+      return false;
+
+   unsigned num_src = nir_op_infos[instr->op].num_inputs;
+
+   for (unsigned i = 0; i < num_src; i++) {
+      if (divergent[instr->src[i].src.ssa->index]) {
+         divergent[instr->dest.dest.ssa.index] = true;
+         return true;
+      }
+   }
+
+   return false;
+}
+
+static bool
+visit_intrinsic(bool *divergent, nir_intrinsic_instr *instr,
+                nir_divergence_options options, gl_shader_stage stage)
+{
+   if (!nir_intrinsic_infos[instr->intrinsic].has_dest)
+      return false;
+
+   if (divergent[instr->dest.ssa.index])
+      return false;
+
+   bool is_divergent = false;
+   switch (instr->intrinsic) {
+   /* Intrinsics which are always uniform */
+   case nir_intrinsic_shader_clock:
+   case nir_intrinsic_ballot:
+   case nir_intrinsic_read_invocation:
+   case nir_intrinsic_read_first_invocation:
+   case nir_intrinsic_vote_any:
+   case nir_intrinsic_vote_all:
+   case nir_intrinsic_vote_feq:
+   case nir_intrinsic_vote_ieq:
+   case nir_intrinsic_load_work_dim:
+   case nir_intrinsic_load_work_group_id:
+   case nir_intrinsic_load_num_work_groups:
+   case nir_intrinsic_load_local_group_size:
+   case nir_intrinsic_load_subgroup_id:
+   case nir_intrinsic_load_num_subgroups:
+   case nir_intrinsic_load_subgroup_size:
+   case nir_intrinsic_load_subgroup_eq_mask:
+   case nir_intrinsic_load_subgroup_ge_mask:
+   case nir_intrinsic_load_subgroup_gt_mask:
+   case nir_intrinsic_load_subgroup_le_mask:
+   case nir_intrinsic_load_subgroup_lt_mask:
+   case nir_intrinsic_first_invocation:
+   case nir_intrinsic_load_base_instance:
+   case nir_intrinsic_load_base_vertex:
+   case nir_intrinsic_load_first_vertex:
+   case nir_intrinsic_load_draw_id:
+   case nir_intrinsic_load_is_indexed_draw:
+   case nir_intrinsic_load_viewport_scale:
+   case nir_intrinsic_load_alpha_ref_float:
+   case nir_intrinsic_load_user_clip_plane:
+   case nir_intrinsic_load_viewport_x_scale:
+   case nir_intrinsic_load_viewport_y_scale:
+   case nir_intrinsic_load_viewport_z_scale:
+   case nir_intrinsic_load_viewport_offset:
+   case nir_intrinsic_load_viewport_z_offset:
+   case nir_intrinsic_load_blend_const_color_a_float:
+   case nir_intrinsic_load_blend_const_color_b_float:
+   case nir_intrinsic_load_blend_const_color_g_float:
+   case nir_intrinsic_load_blend_const_color_r_float:
+   case nir_intrinsic_load_blend_const_color_rgba:
+   case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
+   case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
+      is_divergent = false;
+      break;
+
+   /* Intrinsics with divergence depending on shader stage and hardware */
+   case nir_intrinsic_load_input:
+      is_divergent = divergent[instr->src[0].ssa->index];
+      if (stage == MESA_SHADER_FRAGMENT)
+         is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
+      else if (stage == MESA_SHADER_TESS_EVAL)
+         is_divergent |= !(options & nir_divergence_single_patch_per_tes_subgroup);
+      else
+         is_divergent = true;
+      break;
+   case nir_intrinsic_load_output:
+      assert(stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_FRAGMENT);
+      is_divergent = divergent[instr->src[0].ssa->index];
+      if (stage == MESA_SHADER_TESS_CTRL)
+         is_divergent |= !(options & nir_divergence_single_patch_per_tcs_subgroup);
+      else
+         is_divergent = true;
+      break;
+   case nir_intrinsic_load_layer_id:
+   case nir_intrinsic_load_front_face:
+      assert(stage == MESA_SHADER_FRAGMENT);
+      is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
+      break;
+   case nir_intrinsic_load_view_index:
+      assert(stage != MESA_SHADER_COMPUTE && stage != MESA_SHADER_KERNEL);
+      if (options & nir_divergence_view_index_uniform)
+         is_divergent = false;
+      else if (stage == MESA_SHADER_FRAGMENT)
+         is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
+      break;
+   case nir_intrinsic_load_fs_input_interp_deltas:
+      assert(stage == MESA_SHADER_FRAGMENT);
+      is_divergent = divergent[instr->src[0].ssa->index];
+      is_divergent |= !(options & nir_divergence_single_prim_per_subgroup);
+      break;
+   case nir_intrinsic_load_primitive_id:
+      if (stage == MESA_SHADER_FRAGMENT)
+         is_divergent = !(options & nir_divergence_single_prim_per_subgroup);
+      else if (stage == MESA_SHADER_TESS_CTRL)
+         is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
+      else if (stage == MESA_SHADER_TESS_EVAL)
+         is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
+      else
+         unreachable("Invalid stage for load_primitive_id");
+      break;
+   case nir_intrinsic_load_tess_level_inner:
+   case nir_intrinsic_load_tess_level_outer:
+      if (stage == MESA_SHADER_TESS_CTRL)
+         is_divergent = !(options & nir_divergence_single_patch_per_tcs_subgroup);
+      else if (stage == MESA_SHADER_TESS_EVAL)
+         is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
+      else
+         unreachable("Invalid stage for load_primitive_tess_level_*");
+      break;
+   case nir_intrinsic_load_patch_vertices_in:
+      if (stage == MESA_SHADER_TESS_EVAL)
+         is_divergent = !(options & nir_divergence_single_patch_per_tes_subgroup);
+      else
+         assert(stage == MESA_SHADER_TESS_CTRL);
+      break;
+
+   /* Clustered reductions are uniform if cluster_size == subgroup_size or
+    * the source is uniform and the operation is invariant.
+    * Inclusive scans are uniform if
+    * the source is uniform and the operation is invariant
+    */
+   case nir_intrinsic_reduce:
+      if (nir_intrinsic_cluster_size(instr) == 0)
+         return false;
+      /* fallthrough */
+   case nir_intrinsic_inclusive_scan: {
+      nir_op op = nir_intrinsic_reduction_op(instr);
+      is_divergent = divergent[instr->src[0].ssa->index];
+      if (op != nir_op_umin && op != nir_op_imin && op != nir_op_fmin &&
+          op != nir_op_umax && op != nir_op_imax && op != nir_op_fmax &&
+          op != nir_op_iand && op != nir_op_ior)
+         is_divergent = true;
+      break;
+   }
+
+   /* Intrinsics with divergence depending on sources */
+   case nir_intrinsic_ballot_bitfield_extract:
+   case nir_intrinsic_ballot_find_lsb:
+   case nir_intrinsic_ballot_find_msb:
+   case nir_intrinsic_ballot_bit_count_reduce:
+   case nir_intrinsic_shuffle:
+   case nir_intrinsic_shuffle_xor:
+   case nir_intrinsic_shuffle_up:
+   case nir_intrinsic_shuffle_down:
+   case nir_intrinsic_quad_broadcast:
+   case nir_intrinsic_quad_swap_horizontal:
+   case nir_intrinsic_quad_swap_vertical:
+   case nir_intrinsic_quad_swap_diagonal:
+   case nir_intrinsic_load_deref:
+   case nir_intrinsic_load_ubo:
+   case nir_intrinsic_load_ssbo:
+   case nir_intrinsic_load_shared:
+   case nir_intrinsic_load_global:
+   case nir_intrinsic_load_uniform:
+   case nir_intrinsic_load_push_constant:
+   case nir_intrinsic_load_constant:
+   case nir_intrinsic_load_sample_pos_from_id:
+   case nir_intrinsic_load_kernel_input:
+   case nir_intrinsic_image_load:
+   case nir_intrinsic_image_deref_load:
+   case nir_intrinsic_bindless_image_load:
+   case nir_intrinsic_image_samples:
+   case nir_intrinsic_image_deref_samples:
+   case nir_intrinsic_bindless_image_samples:
+   case nir_intrinsic_get_buffer_size:
+   case nir_intrinsic_image_size:
+   case nir_intrinsic_image_deref_size:
+   case nir_intrinsic_bindless_image_size:
+   case nir_intrinsic_copy_deref:
+   case nir_intrinsic_deref_buffer_array_length:
+   case nir_intrinsic_vulkan_resource_index:
+   case nir_intrinsic_vulkan_resource_reindex:
+   case nir_intrinsic_load_vulkan_descriptor:
+   case nir_intrinsic_atomic_counter_read:
+   case nir_intrinsic_atomic_counter_read_deref:
+   case nir_intrinsic_quad_swizzle_amd:
+   case nir_intrinsic_masked_swizzle_amd: {
+      unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
+      for (unsigned i = 0; i < num_srcs; i++) {
+         if (divergent[instr->src[i].ssa->index]) {
+            is_divergent = true;
+            break;
+         }
+      }
+      break;
+   }
+
+   /* Intrinsics which are always divergent */
+   case nir_intrinsic_load_color0:
+   case nir_intrinsic_load_color1:
+   case nir_intrinsic_load_param:
+   case nir_intrinsic_load_sample_id:
+   case nir_intrinsic_load_sample_id_no_per_sample:
+   case nir_intrinsic_load_sample_mask_in:
+   case nir_intrinsic_load_interpolated_input:
+   case nir_intrinsic_load_barycentric_pixel:
+   case nir_intrinsic_load_barycentric_centroid:
+   case nir_intrinsic_load_barycentric_sample:
+   case nir_intrinsic_load_barycentric_at_sample:
+   case nir_intrinsic_load_barycentric_at_offset:
+   case nir_intrinsic_interp_deref_at_offset:
+   case nir_intrinsic_interp_deref_at_sample:
+   case nir_intrinsic_interp_deref_at_centroid:
+   case nir_intrinsic_load_tess_coord:
+   case nir_intrinsic_load_point_coord:
+   case nir_intrinsic_load_frag_coord:
+   case nir_intrinsic_load_sample_pos:
+   case nir_intrinsic_load_vertex_id_zero_base:
+   case nir_intrinsic_load_vertex_id:
+   case nir_intrinsic_load_per_vertex_input:
+   case nir_intrinsic_load_per_vertex_output:
+   case nir_intrinsic_load_instance_id:
+   case nir_intrinsic_load_invocation_id:
+   case nir_intrinsic_load_local_invocation_id:
+   case nir_intrinsic_load_local_invocation_index:
+   case nir_intrinsic_load_global_invocation_id:
+   case nir_intrinsic_load_global_invocation_index:
+   case nir_intrinsic_load_subgroup_invocation:
+   case nir_intrinsic_load_helper_invocation:
+   case nir_intrinsic_is_helper_invocation:
+   case nir_intrinsic_load_scratch:
+   case nir_intrinsic_deref_atomic_add:
+   case nir_intrinsic_deref_atomic_imin:
+   case nir_intrinsic_deref_atomic_umin:
+   case nir_intrinsic_deref_atomic_imax:
+   case nir_intrinsic_deref_atomic_umax:
+   case nir_intrinsic_deref_atomic_and:
+   case nir_intrinsic_deref_atomic_or:
+   case nir_intrinsic_deref_atomic_xor:
+   case nir_intrinsic_deref_atomic_exchange:
+   case nir_intrinsic_deref_atomic_comp_swap:
+   case nir_intrinsic_deref_atomic_fadd:
+   case nir_intrinsic_deref_atomic_fmin:
+   case nir_intrinsic_deref_atomic_fmax:
+   case nir_intrinsic_deref_atomic_fcomp_swap:
+   case nir_intrinsic_ssbo_atomic_add:
+   case nir_intrinsic_ssbo_atomic_imin:
+   case nir_intrinsic_ssbo_atomic_umin:
+   case nir_intrinsic_ssbo_atomic_imax:
+   case nir_intrinsic_ssbo_atomic_umax:
+   case nir_intrinsic_ssbo_atomic_and:
+   case nir_intrinsic_ssbo_atomic_or:
+   case nir_intrinsic_ssbo_atomic_xor:
+   case nir_intrinsic_ssbo_atomic_exchange:
+   case nir_intrinsic_ssbo_atomic_comp_swap:
+   case nir_intrinsic_ssbo_atomic_fadd:
+   case nir_intrinsic_ssbo_atomic_fmax:
+   case nir_intrinsic_ssbo_atomic_fmin:
+   case nir_intrinsic_ssbo_atomic_fcomp_swap:
+   case nir_intrinsic_image_deref_atomic_add:
+   case nir_intrinsic_image_deref_atomic_min:
+   case nir_intrinsic_image_deref_atomic_max:
+   case nir_intrinsic_image_deref_atomic_and:
+   case nir_intrinsic_image_deref_atomic_or:
+   case nir_intrinsic_image_deref_atomic_xor:
+   case nir_intrinsic_image_deref_atomic_exchange:
+   case nir_intrinsic_image_deref_atomic_comp_swap:
+   case nir_intrinsic_image_deref_atomic_fadd:
+   case nir_intrinsic_image_atomic_add:
+   case nir_intrinsic_image_atomic_min:
+   case nir_intrinsic_image_atomic_max:
+   case nir_intrinsic_image_atomic_and:
+   case nir_intrinsic_image_atomic_or:
+   case nir_intrinsic_image_atomic_xor:
+   case nir_intrinsic_image_atomic_exchange:
+   case nir_intrinsic_image_atomic_comp_swap:
+   case nir_intrinsic_image_atomic_fadd:
+   case nir_intrinsic_bindless_image_atomic_add:
+   case nir_intrinsic_bindless_image_atomic_min:
+   case nir_intrinsic_bindless_image_atomic_max:
+   case nir_intrinsic_bindless_image_atomic_and:
+   case nir_intrinsic_bindless_image_atomic_or:
+   case nir_intrinsic_bindless_image_atomic_xor:
+   case nir_intrinsic_bindless_image_atomic_exchange:
+   case nir_intrinsic_bindless_image_atomic_comp_swap:
+   case nir_intrinsic_bindless_image_atomic_fadd:
+   case nir_intrinsic_shared_atomic_add:
+   case nir_intrinsic_shared_atomic_imin:
+   case nir_intrinsic_shared_atomic_umin:
+   case nir_intrinsic_shared_atomic_imax:
+   case nir_intrinsic_shared_atomic_umax:
+   case nir_intrinsic_shared_atomic_and:
+   case nir_intrinsic_shared_atomic_or:
+   case nir_intrinsic_shared_atomic_xor:
+   case nir_intrinsic_shared_atomic_exchange:
+   case nir_intrinsic_shared_atomic_comp_swap:
+   case nir_intrinsic_shared_atomic_fadd:
+   case nir_intrinsic_shared_atomic_fmin:
+   case nir_intrinsic_shared_atomic_fmax:
+   case nir_intrinsic_shared_atomic_fcomp_swap:
+   case nir_intrinsic_global_atomic_add:
+   case nir_intrinsic_global_atomic_imin:
+   case nir_intrinsic_global_atomic_umin:
+   case nir_intrinsic_global_atomic_imax:
+   case nir_intrinsic_global_atomic_umax:
+   case nir_intrinsic_global_atomic_and:
+   case nir_intrinsic_global_atomic_or:
+   case nir_intrinsic_global_atomic_xor:
+   case nir_intrinsic_global_atomic_exchange:
+   case nir_intrinsic_global_atomic_comp_swap:
+   case nir_intrinsic_global_atomic_fadd:
+   case nir_intrinsic_global_atomic_fmin:
+   case nir_intrinsic_global_atomic_fmax:
+   case nir_intrinsic_global_atomic_fcomp_swap:
+   case nir_intrinsic_atomic_counter_add:
+   case nir_intrinsic_atomic_counter_min:
+   case nir_intrinsic_atomic_counter_max:
+   case nir_intrinsic_atomic_counter_and:
+   case nir_intrinsic_atomic_counter_or:
+   case nir_intrinsic_atomic_counter_xor:
+   case nir_intrinsic_atomic_counter_inc:
+   case nir_intrinsic_atomic_counter_pre_dec:
+   case nir_intrinsic_atomic_counter_post_dec:
+   case nir_intrinsic_atomic_counter_exchange:
+   case nir_intrinsic_atomic_counter_comp_swap:
+   case nir_intrinsic_atomic_counter_add_deref:
+   case nir_intrinsic_atomic_counter_min_deref:
+   case nir_intrinsic_atomic_counter_max_deref:
+   case nir_intrinsic_atomic_counter_and_deref:
+   case nir_intrinsic_atomic_counter_or_deref:
+   case nir_intrinsic_atomic_counter_xor_deref:
+   case nir_intrinsic_atomic_counter_inc_deref:
+   case nir_intrinsic_atomic_counter_pre_dec_deref:
+   case nir_intrinsic_atomic_counter_post_dec_deref:
+   case nir_intrinsic_atomic_counter_exchange_deref:
+   case nir_intrinsic_atomic_counter_comp_swap_deref:
+   case nir_intrinsic_exclusive_scan:
+   case nir_intrinsic_ballot_bit_count_exclusive:
+   case nir_intrinsic_ballot_bit_count_inclusive:
+   case nir_intrinsic_write_invocation_amd:
+   case nir_intrinsic_mbcnt_amd:
+      is_divergent = true;
+      break;
+
+   default:
+#ifdef NDEBUG
+      is_divergent = true;
+      break;
+#else
+      nir_print_instr(&instr->instr, stderr);
+      unreachable("\nNIR divergence analysis: Unhandled intrinsic.");
+#endif
+   }
+
+   divergent[instr->dest.ssa.index] = is_divergent;
+   return is_divergent;
+}
+
+static bool
+visit_tex(bool *divergent, nir_tex_instr *instr)
+{
+   if (divergent[instr->dest.ssa.index])
+      return false;
+
+   bool is_divergent = false;
+
+   for (unsigned i = 0; i < instr->num_srcs; i++) {
+      switch (instr->src[i].src_type) {
+      case nir_tex_src_sampler_deref:
+      case nir_tex_src_sampler_handle:
+      case nir_tex_src_sampler_offset:
+         is_divergent |= divergent[instr->src[i].src.ssa->index] &&
+                         instr->sampler_non_uniform;
+         break;
+      case nir_tex_src_texture_deref:
+      case nir_tex_src_texture_handle:
+      case nir_tex_src_texture_offset:
+         is_divergent |= divergent[instr->src[i].src.ssa->index] &&
+                         instr->texture_non_uniform;
+         break;
+      default:
+         is_divergent |= divergent[instr->src[i].src.ssa->index];
+         break;
+      }
+   }
+
+   divergent[instr->dest.ssa.index] = is_divergent;
+   return is_divergent;
+}
+
+static bool
+visit_phi(bool *divergent, nir_phi_instr *instr)
+{
+   /* There are 3 types of phi instructions:
+    * (1) gamma: represent the joining point of different paths
+    *     created by an “if-then-else” branch.
+    *     The resulting value is divergent if the branch condition
+    *     or any of the source values is divergent.
+    *
+    * (2) mu: which only exist at loop headers,
+    *     merge initial and loop-carried values.
+    *     The resulting value is divergent if any source value
+    *     is divergent or a divergent loop continue condition
+    *     is associated with a different ssa-def.
+    *
+    * (3) eta: represent values that leave a loop.
+    *     The resulting value is divergent if the source value is divergent
+    *     or any loop exit condition is divergent for a value which is
+    *     not loop-invariant.
+    *     (note: there should be no phi for loop-invariant variables.)
+    */
+
+   if (divergent[instr->dest.ssa.index])
+      return false;
+
+   nir_foreach_phi_src(src, instr) {
+      /* if any source value is divergent, the resulting value is divergent */
+      if (divergent[src->src.ssa->index]) {
+         divergent[instr->dest.ssa.index] = true;
+         return true;
+      }
+   }
+
+   nir_cf_node *prev = nir_cf_node_prev(&instr->instr.block->cf_node);
+
+   if (!prev) {
+      /* mu: if no predecessor node exists, the phi must be at a loop header */
+      nir_loop *loop = nir_cf_node_as_loop(instr->instr.block->cf_node.parent);
+      prev = nir_cf_node_prev(&loop->cf_node);
+      nir_ssa_def* same = NULL;
+      bool all_same = true;
+
+      /* first, check if all loop-carried values are from the same ssa-def */
+      nir_foreach_phi_src(src, instr) {
+         if (src->pred == nir_cf_node_as_block(prev))
+            continue;
+         if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
+            continue;
+         if (!same)
+            same = src->src.ssa;
+         else if (same != src->src.ssa)
+            all_same = false;
+      }
+
+      /* if all loop-carried values are the same, the resulting value is uniform */
+      if (all_same)
+         return false;
+
+      /* check if the loop-carried values come from different ssa-defs
+       * and the corresponding condition is divergent. */
+      nir_foreach_phi_src(src, instr) {
+         /* skip the loop preheader */
+         if (src->pred == nir_cf_node_as_block(prev))
+            continue;
+
+         /* skip the unconditional back-edge */
+         if (src->pred == nir_loop_last_block(loop))
+            continue;
+
+         /* if the value is undef, we don't need to check the condition */
+         if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef)
+            continue;
+
+         nir_cf_node *current = src->pred->cf_node.parent;
+         /* check recursively the conditions if any is divergent */
+         while (current->type != nir_cf_node_loop) {
+            assert (current->type == nir_cf_node_if);
+            nir_if *if_node = nir_cf_node_as_if(current);
+            if (divergent[if_node->condition.ssa->index]) {
+               divergent[instr->dest.ssa.index] = true;
+               return true;
+            }
+            current = current->parent;
+         }
+         assert(current == &loop->cf_node);
+      }
+
+   } else if (prev->type == nir_cf_node_if) {
+      /* if only one of the incoming values is defined, the resulting value is uniform */
+      unsigned defined_srcs = 0;
+      nir_foreach_phi_src(src, instr) {
+         if (src->src.ssa->parent_instr->type != nir_instr_type_ssa_undef)
+            defined_srcs++;
+      }
+      if (defined_srcs <= 1)
+         return false;
+
+      /* gamma: check if the condition is divergent */
+      nir_if *if_node = nir_cf_node_as_if(prev);
+      if (divergent[if_node->condition.ssa->index]) {
+         divergent[instr->dest.ssa.index] = true;
+         return true;
+      }
+
+   } else {
+      /* eta: the predecessor must be a loop */
+      assert(prev->type == nir_cf_node_loop);
+
+      /* Check if any loop exit condition is divergent:
+       * That is any break happens under divergent condition or
+       * a break is preceeded by a divergent continue
+       */
+      nir_foreach_phi_src(src, instr) {
+         nir_cf_node *current = src->pred->cf_node.parent;
+
+         /* check recursively the conditions if any is divergent */
+         while (current->type != nir_cf_node_loop) {
+            assert(current->type == nir_cf_node_if);
+            nir_if *if_node = nir_cf_node_as_if(current);
+            if (divergent[if_node->condition.ssa->index]) {
+               divergent[instr->dest.ssa.index] = true;
+               return true;
+            }
+            current = current->parent;
+         }
+         assert(current == prev);
+
+         /* check if any divergent continue happened before the break */
+         nir_foreach_block_in_cf_node(block, prev) {
+            if (block == src->pred)
+               break;
+            if (!nir_block_ends_in_jump(block))
+               continue;
+
+            nir_jump_instr *jump = nir_instr_as_jump(nir_block_last_instr(block));
+            if (jump->type != nir_jump_continue)
+               continue;
+
+            current = block->cf_node.parent;
+            bool is_divergent = false;
+            while (current != prev) {
+               /* the continue belongs to an inner loop */
+               if (current->type == nir_cf_node_loop) {
+                  is_divergent = false;
+                  break;
+               }
+               assert(current->type == nir_cf_node_if);
+               nir_if *if_node = nir_cf_node_as_if(current);
+               is_divergent |= divergent[if_node->condition.ssa->index];
+               current = current->parent;
+            }
+
+            if (is_divergent) {
+               divergent[instr->dest.ssa.index] = true;
+               return true;
+            }
+         }
+      }
+   }
+
+   return false;
+}
+
+static bool
+visit_load_const(bool *divergent, nir_load_const_instr *instr)
+{
+   return false;
+}
+
+static bool
+visit_ssa_undef(bool *divergent, nir_ssa_undef_instr *instr)
+{
+   return false;
+}
+
+static bool
+nir_variable_mode_is_uniform(nir_variable_mode mode) {
+   switch (mode) {
+   case nir_var_uniform:
+   case nir_var_mem_ubo:
+   case nir_var_mem_ssbo:
+   case nir_var_mem_shared:
+   case nir_var_mem_global:
+      return true;
+   default:
+      return false;
+   }
+}
+
+static bool
+nir_variable_is_uniform(nir_variable *var, nir_divergence_options options,
+                        gl_shader_stage stage)
+{
+   if (nir_variable_mode_is_uniform(var->data.mode))
+      return true;
+
+   if (stage == MESA_SHADER_FRAGMENT &&
+       (options & nir_divergence_single_prim_per_subgroup) &&
+       var->data.mode == nir_var_shader_in &&
+       var->data.interpolation == INTERP_MODE_FLAT)
+      return true;
+
+   if (stage == MESA_SHADER_TESS_CTRL &&
+       (options & nir_divergence_single_patch_per_tcs_subgroup) &&
+       var->data.mode == nir_var_shader_out && var->data.patch)
+      return true;
+
+   if (stage == MESA_SHADER_TESS_EVAL &&
+       (options & nir_divergence_single_patch_per_tes_subgroup) &&
+       var->data.mode == nir_var_shader_in && var->data.patch)
+      return true;
+
+   return false;
+}
+
+static bool
+visit_deref(bool *divergent, nir_deref_instr *deref,
+            nir_divergence_options options, gl_shader_stage stage)
+{
+   if (divergent[deref->dest.ssa.index])
+      return false;
+
+   bool is_divergent = false;
+   switch (deref->deref_type) {
+   case nir_deref_type_var:
+      is_divergent = !nir_variable_is_uniform(deref->var, options, stage);
+      break;
+   case nir_deref_type_array:
+   case nir_deref_type_ptr_as_array:
+      is_divergent = divergent[deref->arr.index.ssa->index];
+      /* fallthrough */
+   case nir_deref_type_struct:
+   case nir_deref_type_array_wildcard:
+      is_divergent |= divergent[deref->parent.ssa->index];
+      break;
+   case nir_deref_type_cast:
+      is_divergent = !nir_variable_mode_is_uniform(deref->var->data.mode) ||
+                     divergent[deref->parent.ssa->index];
+      break;
+   }
+
+   divergent[deref->dest.ssa.index] = is_divergent;
+   return is_divergent;
+}
+
+static bool
+visit_block(bool *divergent, nir_block *block, nir_divergence_options options,
+            gl_shader_stage stage)
+{
+   bool has_changed = false;
+
+   nir_foreach_instr(instr, block) {
+      switch (instr->type) {
+      case nir_instr_type_alu:
+         has_changed |= visit_alu(divergent, nir_instr_as_alu(instr));
+         break;
+      case nir_instr_type_intrinsic:
+         has_changed |= visit_intrinsic(divergent, nir_instr_as_intrinsic(instr),
+                                        options, stage);
+         break;
+      case nir_instr_type_tex:
+         has_changed |= visit_tex(divergent, nir_instr_as_tex(instr));
+         break;
+      case nir_instr_type_phi:
+         has_changed |= visit_phi(divergent, nir_instr_as_phi(instr));
+         break;
+      case nir_instr_type_load_const:
+         has_changed |= visit_load_const(divergent, nir_instr_as_load_const(instr));
+         break;
+      case nir_instr_type_ssa_undef:
+         has_changed |= visit_ssa_undef(divergent, nir_instr_as_ssa_undef(instr));
+         break;
+      case nir_instr_type_deref:
+         has_changed |= visit_deref(divergent, nir_instr_as_deref(instr),
+                                    options, stage);
+         break;
+      case nir_instr_type_jump:
+         break;
+      case nir_instr_type_call:
+      case nir_instr_type_parallel_copy:
+         unreachable("NIR divergence analysis: Unsupported instruction type.");
+      }
+   }
+
+   return has_changed;
+}
+
+static bool
+visit_if(bool *divergent, nir_if *if_stmt, nir_divergence_options options, gl_shader_stage stage)
+{
+   return visit_cf_list(divergent, &if_stmt->then_list, options, stage) |
+          visit_cf_list(divergent, &if_stmt->else_list, options, stage);
+}
+
+static bool
+visit_loop(bool *divergent, nir_loop *loop, nir_divergence_options options, gl_shader_stage stage)
+{
+   bool has_changed = false;
+   bool repeat = true;
+
+   /* TODO: restructure this and the phi handling more efficiently */
+   while (repeat) {
+      repeat = visit_cf_list(divergent, &loop->body, options, stage);
+      has_changed |= repeat;
+   }
+
+   return has_changed;
+}
+
+static bool
+visit_cf_list(bool *divergent, struct exec_list *list,
+              nir_divergence_options options, gl_shader_stage stage)
+{
+   bool has_changed = false;
+
+   foreach_list_typed(nir_cf_node, node, node, list) {
+      switch (node->type) {
+      case nir_cf_node_block:
+         has_changed |= visit_block(divergent, nir_cf_node_as_block(node),
+                                    options, stage);
+         break;
+      case nir_cf_node_if:
+         has_changed |= visit_if(divergent, nir_cf_node_as_if(node),
+                                 options, stage);
+         break;
+      case nir_cf_node_loop:
+         has_changed |= visit_loop(divergent, nir_cf_node_as_loop(node),
+                                   options, stage);
+         break;
+      case nir_cf_node_function:
+         unreachable("NIR divergence analysis: Unsupported cf_node type.");
+      }
+   }
+
+   return has_changed;
+}
+
+
+bool*
+nir_divergence_analysis(nir_shader *shader, nir_divergence_options options)
+{
+   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+   bool *t = rzalloc_array(shader, bool, impl->ssa_alloc);
+
+   visit_cf_list(t, &impl->body, options, shader->info.stage);
+
+   return t;
+}