src/glsl/nir/nir_opt_peephole_select.c

   1 /*
   2  * Copyright © 2014 Intel 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  * Authors:
  24  *    Jason Ekstrand (jason@jlekstrand.net)
  25  *
  26  */
  27
  28 #include "nir.h"
  29
  30 /*
  31  * Implements a small peephole optimization that looks for
  32  *
  33  * if (cond) {
  34  *    <empty>
  35  * } else {
  36  *    <empty>
  37  * }
  38  * phi
  39  * ...
  40  * phi
  41  *
  42  * and replaces it with a series of selects.  It can also handle the case
  43  * where, instead of being empty, the if may contain some move operations
  44  * whose only use is one of the following phi nodes.  This happens all the
  45  * time when the SSA form comes from a conditional assignment with a
  46  * swizzle.
  47  */
  48
  49 struct peephole_select_state {
  50    void *mem_ctx;
  51    nir_function_impl *impl;
  52    bool progress;
  53 };
  54
  55 static bool
  56 are_all_move_to_phi(nir_block *block)
  57 {
  58    nir_foreach_instr(block, instr) {
  59       if (instr->type != nir_instr_type_alu)
  60          return false;
  61
  62       /* It must be a move operation */
  63       nir_alu_instr *mov = nir_instr_as_alu(instr);
  64       if (mov->op != nir_op_fmov && mov->op != nir_op_imov)
  65          return false;
  66
  67       /* Can't handle saturate */
  68       if (mov->dest.saturate)
  69          return false;
  70
  71       /* It must be SSA */
  72       if (!mov->dest.dest.is_ssa)
  73          return false;
  74
  75       /* It cannot have any if-uses */
  76       if (mov->dest.dest.ssa.if_uses->entries != 0)
  77          return false;
  78
  79       /* The only uses of this definition must be phi's in the successor */
  80       struct set_entry *entry;
  81       set_foreach(mov->dest.dest.ssa.uses, entry) {
  82          const nir_instr *dest_instr = entry->key;
  83          if (dest_instr->type != nir_instr_type_phi ||
  84              dest_instr->block != block->successors[0])
  85             return false;
  86       }
  87    }
  88
  89    return true;
  90 }
  91
  92 static bool
  93 nir_opt_peephole_select_block(nir_block *block, void *void_state)
  94 {
  95    struct peephole_select_state *state = void_state;
  96
  97    /* If the block is empty, then it certainly doesn't have any phi nodes,
  98     * so we can skip it.  This also ensures that we do an early skip on the
  99     * end block of the function which isn't actually attached to the CFG.
 100     */
 101    if (exec_list_is_empty(&block->instr_list))
 102       return true;
 103
 104    if (nir_cf_node_is_first(&block->cf_node))
 105       return true;
 106
 107    nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node);
 108    if (prev_node->type != nir_cf_node_if)
 109       return true;
 110
 111    nir_if *if_stmt = nir_cf_node_as_if(prev_node);
 112    nir_cf_node *then_node = nir_if_first_then_node(if_stmt);
 113    nir_cf_node *else_node = nir_if_first_else_node(if_stmt);
 114
 115    /* We can only have one block in each side ... */
 116    if (nir_if_last_then_node(if_stmt) != then_node ||
 117        nir_if_last_else_node(if_stmt) != else_node)
 118       return true;
 119
 120    nir_block *then_block = nir_cf_node_as_block(then_node);
 121    nir_block *else_block = nir_cf_node_as_block(else_node);
 122
 123    /* ... and those blocks must only contain move-to-phi. */
 124    if (!are_all_move_to_phi(then_block) || !are_all_move_to_phi(else_block))
 125       return true;
 126
 127    /* At this point, we know that the previous CFG node is an if-then
 128     * statement containing only moves to phi nodes in this block.  We can
 129     * just remove that entire CF node and replace all of the phi nodes with
 130     * selects.
 131     */
 132
 133    nir_foreach_instr_safe(block, instr) {
 134       if (instr->type != nir_instr_type_phi)
 135          break;
 136
 137       nir_phi_instr *phi = nir_instr_as_phi(instr);
 138       nir_alu_instr *sel = nir_alu_instr_create(state->mem_ctx, nir_op_bcsel);
 139       sel->src[0].src = nir_src_copy(if_stmt->condition, state->mem_ctx);
 140
 141       assert(exec_list_length(&phi->srcs) == 2);
 142       foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 143          assert(src->pred == then_block || src->pred == else_block);
 144          assert(src->src.is_ssa);
 145
 146          unsigned idx = src->pred == then_block ? 1 : 2;
 147
 148          if (src->src.ssa->parent_instr->block == src->pred) {
 149             /* We already know that this instruction must be a move with
 150              * this phi's in this block as its only users.
 151              */
 152             nir_alu_instr *mov = nir_instr_as_alu(src->src.ssa->parent_instr);
 153             assert(mov->instr.type == nir_instr_type_alu);
 154             assert(mov->op == nir_op_fmov || mov->op == nir_op_imov);
 155
 156             sel->src[idx].src = nir_src_copy(mov->src[0].src, state->mem_ctx);
 157             sel->src[idx].negate = mov->src[0].negate;
 158             sel->src[idx].abs = mov->src[0].abs;
 159             memcpy(sel->src[idx].swizzle, mov->src[0].swizzle, 4);
 160          } else {
 161             sel->src[idx].src = nir_src_copy(src->src, state->mem_ctx);
 162          }
 163       }
 164
 165       sel->dest.dest.is_ssa = true;
 166       nir_ssa_def_init(state->impl, &sel->instr, &sel->dest.dest.ssa,
 167                        phi->dest.ssa.num_components, phi->dest.ssa.name);
 168       sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
 169
 170       nir_src sel_dest_src = {
 171          .is_ssa = true,
 172          .ssa = &sel->dest.dest.ssa,
 173       };
 174       nir_ssa_def_rewrite_uses(&phi->dest.ssa, sel_dest_src, state->mem_ctx);
 175
 176       nir_instr_insert_before(&phi->instr, &sel->instr);
 177       nir_instr_remove(&phi->instr);
 178    }
 179
 180    nir_cf_node_remove(&if_stmt->cf_node);
 181    state->progress = true;
 182
 183    return true;
 184 }
 185
 186 static bool
 187 nir_opt_peephole_select_impl(nir_function_impl *impl)
 188 {
 189    struct peephole_select_state state;
 190
 191    state.mem_ctx = ralloc_parent(impl);
 192    state.progress = false;
 193    state.impl = impl;
 194
 195    nir_foreach_block(impl, nir_opt_peephole_select_block, &state);
 196
 197    if (state.progress)
 198       nir_metadata_dirty(impl, nir_metadata_none);
 199
 200    return state.progress;
 201 }
 202
 203 bool
 204 nir_opt_peephole_select(nir_shader *shader)
 205 {
 206    bool progress = false;
 207
 208    nir_foreach_overload(shader, overload) {
 209       if (overload->impl)
 210          progress |= nir_opt_peephole_select_impl(overload->impl);
 211    }
 212
 213    return progress;
 214 }