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           nir_instr_as_alu(instr)->op == nir_op_b2i32)
  50          return options & nir_move_copies;
  51       if (nir_alu_instr_is_comparison(nir_instr_as_alu(instr)))
  52          return options & nir_move_comparisons;
  53       return false;
  54    }
  55    case nir_instr_type_intrinsic: {
  56       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  57       switch (intrin->intrinsic) {
  58       case nir_intrinsic_load_ubo:
  59          return options & nir_move_load_ubo;
  60       case nir_intrinsic_load_input:
  61       case nir_intrinsic_load_interpolated_input:
  62       case nir_intrinsic_load_per_vertex_input:
  63          return options & nir_move_load_input;
  64       default:
  65          return false;
  66       }
  67    }
  68    default:
  69       return false;
  70    }
  71 }
  72
  73 static nir_loop *
  74 get_innermost_loop(nir_cf_node *node)
  75 {
  76    for (; node != NULL; node = node->parent) {
  77       if (node->type == nir_cf_node_loop)
  78          return (nir_loop*)node;
  79    }
  80    return NULL;
  81 }
  82
  83 static bool
  84 loop_contains_block(nir_loop *loop, nir_block *block)
  85 {
  86    nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
  87    nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
  88
  89    return block->index > before->index && block->index < after->index;
  90 }
  91
  92 /* Given the LCA of all uses and the definition, find a block on the path
  93  * between them in the dominance tree that is outside of as many loops as
  94  * possible. If "sink_out_of_loops" is false, then we disallow sinking the
  95  * definition outside of the loop it's defined in (if any).
  96  */
  97
  98 static nir_block *
  99 adjust_block_for_loops(nir_block *use_block, nir_block *def_block,
 100                        bool sink_out_of_loops)
 101 {
 102    nir_loop *def_loop = NULL;
 103    if (!sink_out_of_loops)
 104       def_loop = get_innermost_loop(&def_block->cf_node);
 105
 106    for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
 107         cur_block = cur_block->imm_dom) {
 108       if (!sink_out_of_loops && def_loop &&
 109           !loop_contains_block(def_loop, use_block)) {
 110          use_block = cur_block;
 111          continue;
 112       }
 113
 114       nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
 115       if (next && next->type == nir_cf_node_loop) {
 116          nir_loop *following_loop = nir_cf_node_as_loop(next);
 117          if (loop_contains_block(following_loop, use_block)) {
 118              use_block = cur_block;
 119              continue;
 120          }
 121       }
 122    }
 123
 124    return use_block;
 125 }
 126
 127 /* iterate a ssa def's use's and try to find a more optimal block to
 128  * move it to, using the dominance tree.  In short, if all of the uses
 129  * are contained in a single block, the load will be moved there,
 130  * otherwise it will be move to the least common ancestor block of all
 131  * the uses
 132  */
 133 static nir_block *
 134 get_preferred_block(nir_ssa_def *def, bool sink_into_loops, bool sink_out_of_loops)
 135 {
 136    nir_block *lca = NULL;
 137
 138    nir_foreach_use(use, def) {
 139       nir_instr *instr = use->parent_instr;
 140       nir_block *use_block = instr->block;
 141
 142       /*
 143        * Kind of an ugly special-case, but phi instructions
 144        * need to appear first in the block, so by definition
 145        * we can't move an instruction into a block where it is
 146        * consumed by a phi instruction.  We could conceivably
 147        * move it into a dominator block.
 148        */
 149       if (instr->type == nir_instr_type_phi) {
 150          nir_phi_instr *phi = nir_instr_as_phi(instr);
 151          nir_block *phi_lca = NULL;
 152          nir_foreach_phi_src(src, phi) {
 153             if (&src->src == use)
 154                phi_lca = nir_dominance_lca(phi_lca, src->pred);
 155          }
 156          use_block = phi_lca;
 157       }
 158
 159       lca = nir_dominance_lca(lca, use_block);
 160    }
 161
 162    nir_foreach_if_use(use, def) {
 163       nir_block *use_block =
 164          nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
 165
 166       lca = nir_dominance_lca(lca, use_block);
 167    }
 168
 169    /* If we're moving a load_ubo or load_interpolated_input, we don't want to
 170     * sink it down into loops, which may result in accessing memory or shared
 171     * functions multiple times.  Sink it just above the start of the loop
 172     * where it's used.  For load_consts, undefs, and comparisons, we expect
 173     * the driver to be able to emit them as simple ALU ops, so sinking as far
 174     * in as we can go is probably worth it for register pressure.
 175     */
 176    if (!sink_into_loops) {
 177       lca = adjust_block_for_loops(lca, def->parent_instr->block,
 178                                    sink_out_of_loops);
 179       assert(nir_block_dominates(def->parent_instr->block, lca));
 180    } else {
 181       /* sink_into_loops = true and sink_out_of_loops = false isn't
 182        * implemented yet because it's not used.
 183        */
 184       assert(sink_out_of_loops);
 185    }
 186
 187
 188    return lca;
 189 }
 190
 191 /* insert before first non-phi instruction: */
 192 static void
 193 insert_after_phi(nir_instr *instr, nir_block *block)
 194 {
 195    nir_foreach_instr(instr2, block) {
 196       if (instr2->type == nir_instr_type_phi)
 197          continue;
 198
 199       exec_node_insert_node_before(&instr2->node,
 200                                    &instr->node);
 201
 202       return;
 203    }
 204
 205    /* if haven't inserted it, push to tail (ie. empty block or possibly
 206     * a block only containing phi's?)
 207     */
 208    exec_list_push_tail(&block->instr_list, &instr->node);
 209 }
 210
 211 bool
 212 nir_opt_sink(nir_shader *shader, nir_move_options options)
 213 {
 214    bool progress = false;
 215
 216    nir_foreach_function(function, shader) {
 217       if (!function->impl)
 218          continue;
 219
 220       nir_metadata_require(function->impl,
 221                            nir_metadata_block_index | nir_metadata_dominance);
 222
 223       nir_foreach_block_reverse(block, function->impl) {
 224          nir_foreach_instr_reverse_safe(instr, block) {
 225             if (!nir_can_move_instr(instr, options))
 226                continue;
 227
 228             nir_ssa_def *def = nir_instr_ssa_def(instr);
 229
 230             bool sink_into_loops = instr->type != nir_instr_type_intrinsic;
 231             /* Don't sink load_ubo out of loops because that can make its
 232              * resource divergent and break code like that which is generated
 233              * by nir_lower_non_uniform_access.
 234              */
 235             bool sink_out_of_loops =
 236                instr->type != nir_instr_type_intrinsic ||
 237                nir_instr_as_intrinsic(instr)->intrinsic != nir_intrinsic_load_ubo;
 238             nir_block *use_block =
 239                   get_preferred_block(def, sink_into_loops, sink_out_of_loops);
 240
 241             if (!use_block || use_block == instr->block)
 242                continue;
 243
 244             exec_node_remove(&instr->node);
 245
 246             insert_after_phi(instr, use_block);
 247
 248             instr->block = use_block;
 249
 250             progress = true;
 251          }
 252       }
 253
 254       nir_metadata_preserve(function->impl,
 255                             nir_metadata_block_index | nir_metadata_dominance);
 256    }
 257
 258    return progress;
 259 }