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    if ((options & nir_move_const_undef) && instr->type == nir_instr_type_load_const) {
  43       return true;
  44    }
  45
  46    if (instr->type == nir_instr_type_intrinsic) {
  47        nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  48       if ((options & nir_move_load_ubo) && intrin->intrinsic == nir_intrinsic_load_ubo)
  49          return true;
  50
  51       if ((options & nir_move_load_input) &&
  52           (intrin->intrinsic == nir_intrinsic_load_interpolated_input ||
  53            intrin->intrinsic == nir_intrinsic_load_input))
  54          return true;
  55    }
  56
  57    if ((options & nir_move_const_undef) && instr->type == nir_instr_type_ssa_undef) {
  58       return true;
  59    }
  60
  61    if ((options & nir_move_comparisons) && instr->type == nir_instr_type_alu &&
  62        nir_alu_instr_is_comparison(nir_instr_as_alu(instr))) {
  63       return true;
  64    }
  65
  66    return false;
  67 }
  68
  69 static nir_loop *
  70 get_innermost_loop(nir_cf_node *node)
  71 {
  72    for (; node != NULL; node = node->parent) {
  73       if (node->type == nir_cf_node_loop)
  74          return (nir_loop*)node;
  75    }
  76    return NULL;
  77 }
  78
  79 static bool
  80 loop_contains_block(nir_loop *loop, nir_block *block)
  81 {
  82    nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
  83    nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
  84
  85    return block->index > before->index && block->index < after->index;
  86 }
  87
  88 /* Given the LCA of all uses and the definition, find a block on the path
  89  * between them in the dominance tree that is outside of as many loops as
  90  * possible. If "sink_out_of_loops" is false, then we disallow sinking the
  91  * definition outside of the loop it's defined in (if any).
  92  */
  93
  94 static nir_block *
  95 adjust_block_for_loops(nir_block *use_block, nir_block *def_block,
  96                        bool sink_out_of_loops)
  97 {
  98    nir_loop *def_loop = NULL;
  99    if (!sink_out_of_loops)
 100       def_loop = get_innermost_loop(&def_block->cf_node);
 101
 102    for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
 103         cur_block = cur_block->imm_dom) {
 104       if (!sink_out_of_loops && def_loop &&
 105           !loop_contains_block(def_loop, use_block)) {
 106          use_block = cur_block;
 107          continue;
 108       }
 109
 110       nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
 111       if (next && next->type == nir_cf_node_loop) {
 112          nir_loop *following_loop = nir_cf_node_as_loop(next);
 113          if (loop_contains_block(following_loop, use_block)) {
 114              use_block = cur_block;
 115              continue;
 116          }
 117       }
 118    }
 119
 120    return use_block;
 121 }
 122
 123 /* iterate a ssa def's use's and try to find a more optimal block to
 124  * move it to, using the dominance tree.  In short, if all of the uses
 125  * are contained in a single block, the load will be moved there,
 126  * otherwise it will be move to the least common ancestor block of all
 127  * the uses
 128  */
 129 static nir_block *
 130 get_preferred_block(nir_ssa_def *def, bool sink_into_loops, bool sink_out_of_loops)
 131 {
 132    nir_block *lca = NULL;
 133
 134    nir_foreach_use(use, def) {
 135       nir_instr *instr = use->parent_instr;
 136       nir_block *use_block = instr->block;
 137
 138       /*
 139        * Kind of an ugly special-case, but phi instructions
 140        * need to appear first in the block, so by definition
 141        * we can't move an instruction into a block where it is
 142        * consumed by a phi instruction.  We could conceivably
 143        * move it into a dominator block.
 144        */
 145       if (instr->type == nir_instr_type_phi) {
 146          nir_phi_instr *phi = nir_instr_as_phi(instr);
 147          nir_block *phi_lca = NULL;
 148          nir_foreach_phi_src(src, phi) {
 149             if (&src->src == use)
 150                phi_lca = nir_dominance_lca(phi_lca, src->pred);
 151          }
 152          use_block = phi_lca;
 153       }
 154
 155       lca = nir_dominance_lca(lca, use_block);
 156    }
 157
 158    nir_foreach_if_use(use, def) {
 159       nir_block *use_block =
 160          nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
 161
 162       lca = nir_dominance_lca(lca, use_block);
 163    }
 164
 165    /* If we're moving a load_ubo or load_interpolated_input, we don't want to
 166     * sink it down into loops, which may result in accessing memory or shared
 167     * functions multiple times.  Sink it just above the start of the loop
 168     * where it's used.  For load_consts, undefs, and comparisons, we expect
 169     * the driver to be able to emit them as simple ALU ops, so sinking as far
 170     * in as we can go is probably worth it for register pressure.
 171     */
 172    if (!sink_into_loops) {
 173       lca = adjust_block_for_loops(lca, def->parent_instr->block,
 174                                    sink_out_of_loops);
 175       assert(nir_block_dominates(def->parent_instr->block, lca));
 176    } else {
 177       /* sink_into_loops = true and sink_out_of_loops = false isn't
 178        * implemented yet because it's not used.
 179        */
 180       assert(sink_out_of_loops);
 181    }
 182
 183
 184    return lca;
 185 }
 186
 187 /* insert before first non-phi instruction: */
 188 static void
 189 insert_after_phi(nir_instr *instr, nir_block *block)
 190 {
 191    nir_foreach_instr(instr2, block) {
 192       if (instr2->type == nir_instr_type_phi)
 193          continue;
 194
 195       exec_node_insert_node_before(&instr2->node,
 196                                    &instr->node);
 197
 198       return;
 199    }
 200
 201    /* if haven't inserted it, push to tail (ie. empty block or possibly
 202     * a block only containing phi's?)
 203     */
 204    exec_list_push_tail(&block->instr_list, &instr->node);
 205 }
 206
 207 bool
 208 nir_opt_sink(nir_shader *shader, nir_move_options options)
 209 {
 210    bool progress = false;
 211
 212    nir_foreach_function(function, shader) {
 213       if (!function->impl)
 214          continue;
 215
 216       nir_metadata_require(function->impl,
 217                            nir_metadata_block_index | nir_metadata_dominance);
 218
 219       nir_foreach_block_reverse(block, function->impl) {
 220          nir_foreach_instr_reverse_safe(instr, block) {
 221             if (!nir_can_move_instr(instr, options))
 222                continue;
 223
 224             nir_ssa_def *def = nir_instr_ssa_def(instr);
 225
 226             bool sink_into_loops = instr->type != nir_instr_type_intrinsic;
 227             /* Don't sink load_ubo out of loops because that can make its
 228              * resource divergent and break code like that which is generated
 229              * by nir_lower_non_uniform_access.
 230              */
 231             bool sink_out_of_loops =
 232                instr->type != nir_instr_type_intrinsic ||
 233                nir_instr_as_intrinsic(instr)->intrinsic != nir_intrinsic_load_ubo;
 234             nir_block *use_block =
 235                   get_preferred_block(def, sink_into_loops, sink_out_of_loops);
 236
 237             if (!use_block || use_block == instr->block)
 238                continue;
 239
 240             exec_node_remove(&instr->node);
 241
 242             insert_after_phi(instr, use_block);
 243
 244             instr->block = use_block;
 245
 246             progress = true;
 247          }
 248       }
 249
 250       nir_metadata_preserve(function->impl,
 251                             nir_metadata_block_index | nir_metadata_dominance);
 252    }
 253
 254    return progress;
 255 }