src/compiler/nir/nir_opt_sink.c

   1 /*
   2  * Copyright © 2018 Red Hat
   3  * Copyright © 2019 Valve Corporation
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Rob Clark (robdclark@gmail.com>
  26  *    Daniel Schürmann (daniel.schuermann@campus.tu-berlin.de)
  27  *    Rhys Perry (pendingchaos02@gmail.com)
  28  *
  29  */
  30
  31 #include "nir.h"
  32
  33
  34 /*
  35  * A simple pass that moves some instructions into the least common
  36  * anscestor of consuming instructions.
  37  */
  38
  39 bool
  40 nir_can_move_instr(nir_instr *instr, nir_move_options options)
  41 {
  42    switch (instr->type) {
  43    case nir_instr_type_load_const:
  44    case nir_instr_type_ssa_undef: {
  45       return options & nir_move_const_undef;
  46    }
  47    case nir_instr_type_alu: {
  48       if (nir_op_is_vec(nir_instr_as_alu(instr)->op))
  49          return options & nir_move_copies;
  50       if (nir_alu_instr_is_comparison(nir_instr_as_alu(instr)))
  51          return options & nir_move_comparisons;
  52       return false;
  53    }
  54    case nir_instr_type_intrinsic: {
  55       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  56       switch (intrin->intrinsic) {
  57       case nir_intrinsic_load_ubo:
  58          return options & nir_move_load_ubo;
  59       case nir_intrinsic_load_input:
  60       case nir_intrinsic_load_interpolated_input:
  61       case nir_intrinsic_load_per_vertex_input:
  62          return options & nir_move_load_input;
  63       default:
  64          return false;
  65       }
  66    }
  67    default:
  68       return false;
  69    }
  70 }
  71
  72 static nir_loop *
  73 get_innermost_loop(nir_cf_node *node)
  74 {
  75    for (; node != NULL; node = node->parent) {
  76       if (node->type == nir_cf_node_loop)
  77          return (nir_loop*)node;
  78    }
  79    return NULL;
  80 }
  81
  82 static bool
  83 loop_contains_block(nir_loop *loop, nir_block *block)
  84 {
  85    nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
  86    nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
  87
  88    return block->index > before->index && block->index < after->index;
  89 }
  90
  91 /* Given the LCA of all uses and the definition, find a block on the path
  92  * between them in the dominance tree that is outside of as many loops as
  93  * possible. If "sink_out_of_loops" is false, then we disallow sinking the
  94  * definition outside of the loop it's defined in (if any).
  95  */
  96
  97 static nir_block *
  98 adjust_block_for_loops(nir_block *use_block, nir_block *def_block,
  99                        bool sink_out_of_loops)
 100 {
 101    nir_loop *def_loop = NULL;
 102    if (!sink_out_of_loops)
 103       def_loop = get_innermost_loop(&def_block->cf_node);
 104
 105    for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
 106         cur_block = cur_block->imm_dom) {
 107       if (!sink_out_of_loops && def_loop &&
 108           !loop_contains_block(def_loop, use_block)) {
 109          use_block = cur_block;
 110          continue;
 111       }
 112
 113       nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
 114       if (next && next->type == nir_cf_node_loop) {
 115          nir_loop *following_loop = nir_cf_node_as_loop(next);
 116          if (loop_contains_block(following_loop, use_block)) {
 117              use_block = cur_block;
 118              continue;
 119          }
 120       }
 121    }
 122
 123    return use_block;
 124 }
 125
 126 /* iterate a ssa def's use's and try to find a more optimal block to
 127  * move it to, using the dominance tree.  In short, if all of the uses
 128  * are contained in a single block, the load will be moved there,
 129  * otherwise it will be move to the least common ancestor block of all
 130  * the uses
 131  */
 132 static nir_block *
 133 get_preferred_block(nir_ssa_def *def, bool sink_into_loops, bool sink_out_of_loops)
 134 {
 135    nir_block *lca = NULL;
 136
 137    nir_foreach_use(use, def) {
 138       nir_instr *instr = use->parent_instr;
 139       nir_block *use_block = instr->block;
 140
 141       /*
 142        * Kind of an ugly special-case, but phi instructions
 143        * need to appear first in the block, so by definition
 144        * we can't move an instruction into a block where it is
 145        * consumed by a phi instruction.  We could conceivably
 146        * move it into a dominator block.
 147        */
 148       if (instr->type == nir_instr_type_phi) {
 149          nir_phi_instr *phi = nir_instr_as_phi(instr);
 150          nir_block *phi_lca = NULL;
 151          nir_foreach_phi_src(src, phi) {
 152             if (&src->src == use)
 153                phi_lca = nir_dominance_lca(phi_lca, src->pred);
 154          }
 155          use_block = phi_lca;
 156       }
 157
 158       lca = nir_dominance_lca(lca, use_block);
 159    }
 160
 161    nir_foreach_if_use(use, def) {
 162       nir_block *use_block =
 163          nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
 164
 165       lca = nir_dominance_lca(lca, use_block);
 166    }
 167
 168    /* If we're moving a load_ubo or load_interpolated_input, we don't want to
 169     * sink it down into loops, which may result in accessing memory or shared
 170     * functions multiple times.  Sink it just above the start of the loop
 171     * where it's used.  For load_consts, undefs, and comparisons, we expect
 172     * the driver to be able to emit them as simple ALU ops, so sinking as far
 173     * in as we can go is probably worth it for register pressure.
 174     */
 175    if (!sink_into_loops) {
 176       lca = adjust_block_for_loops(lca, def->parent_instr->block,
 177                                    sink_out_of_loops);
 178       assert(nir_block_dominates(def->parent_instr->block, lca));
 179    } else {
 180       /* sink_into_loops = true and sink_out_of_loops = false isn't
 181        * implemented yet because it's not used.
 182        */
 183       assert(sink_out_of_loops);
 184    }
 185
 186
 187    return lca;
 188 }
 189
 190 /* insert before first non-phi instruction: */
 191 static void
 192 insert_after_phi(nir_instr *instr, nir_block *block)
 193 {
 194    nir_foreach_instr(instr2, block) {
 195       if (instr2->type == nir_instr_type_phi)
 196          continue;
 197
 198       exec_node_insert_node_before(&instr2->node,
 199                                    &instr->node);
 200
 201       return;
 202    }
 203
 204    /* if haven't inserted it, push to tail (ie. empty block or possibly
 205     * a block only containing phi's?)
 206     */
 207    exec_list_push_tail(&block->instr_list, &instr->node);
 208 }
 209
 210 bool
 211 nir_opt_sink(nir_shader *shader, nir_move_options options)
 212 {
 213    bool progress = false;
 214
 215    nir_foreach_function(function, shader) {
 216       if (!function->impl)
 217          continue;
 218
 219       nir_metadata_require(function->impl,
 220                            nir_metadata_block_index | nir_metadata_dominance);
 221
 222       nir_foreach_block_reverse(block, function->impl) {
 223          nir_foreach_instr_reverse_safe(instr, block) {
 224             if (!nir_can_move_instr(instr, options))
 225                continue;
 226
 227             nir_ssa_def *def = nir_instr_ssa_def(instr);
 228
 229             bool sink_into_loops = instr->type != nir_instr_type_intrinsic;
 230             /* Don't sink load_ubo out of loops because that can make its
 231              * resource divergent and break code like that which is generated
 232              * by nir_lower_non_uniform_access.
 233              */
 234             bool sink_out_of_loops =
 235                instr->type != nir_instr_type_intrinsic ||
 236                nir_instr_as_intrinsic(instr)->intrinsic != nir_intrinsic_load_ubo;
 237             nir_block *use_block =
 238                   get_preferred_block(def, sink_into_loops, sink_out_of_loops);
 239
 240             if (!use_block || use_block == instr->block)
 241                continue;
 242
 243             exec_node_remove(&instr->node);
 244
 245             insert_after_phi(instr, use_block);
 246
 247             instr->block = use_block;
 248
 249             progress = true;
 250          }
 251       }
 252
 253       nir_metadata_preserve(function->impl,
 254                             nir_metadata_block_index | nir_metadata_dominance);
 255    }
 256
 257    return progress;
 258 }