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    bool progress;
  52 };
  53
  54 static bool
  55 block_check_for_allowed_instrs(nir_block *block)
  56 {
  57    nir_foreach_instr(block, instr) {
  58       switch (instr->type) {
  59       case nir_instr_type_alu: {
  60          /* It must be a move operation */
  61          nir_alu_instr *mov = nir_instr_as_alu(instr);
  62          if (mov->op != nir_op_fmov && mov->op != nir_op_imov)
  63             return false;
  64
  65          /* Can't handle saturate */
  66          if (mov->dest.saturate)
  67             return false;
  68
  69          /* It must be SSA */
  70          if (!mov->dest.dest.is_ssa)
  71             return false;
  72
  73          /* It cannot have any if-uses */
  74          if (mov->dest.dest.ssa.if_uses->entries != 0)
  75             return false;
  76
  77          /* The only uses of this definition must be phi's in the successor */
  78          struct set_entry *entry;
  79          set_foreach(mov->dest.dest.ssa.uses, entry) {
  80             const nir_instr *dest_instr = entry->key;
  81             if (dest_instr->type != nir_instr_type_phi ||
  82                 dest_instr->block != block->successors[0])
  83                return false;
  84          }
  85          break;
  86       }
  87
  88       default:
  89          return false;
  90       }
  91    }
  92
  93    return true;
  94 }
  95
  96 static bool
  97 nir_opt_peephole_select_block(nir_block *block, void *void_state)
  98 {
  99    struct peephole_select_state *state = void_state;
 100
 101    /* If the block is empty, then it certainly doesn't have any phi nodes,
 102     * so we can skip it.  This also ensures that we do an early skip on the
 103     * end block of the function which isn't actually attached to the CFG.
 104     */
 105    if (exec_list_is_empty(&block->instr_list))
 106       return true;
 107
 108    if (nir_cf_node_is_first(&block->cf_node))
 109       return true;
 110
 111    nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node);
 112    if (prev_node->type != nir_cf_node_if)
 113       return true;
 114
 115    nir_if *if_stmt = nir_cf_node_as_if(prev_node);
 116    nir_cf_node *then_node = nir_if_first_then_node(if_stmt);
 117    nir_cf_node *else_node = nir_if_first_else_node(if_stmt);
 118
 119    /* We can only have one block in each side ... */
 120    if (nir_if_last_then_node(if_stmt) != then_node ||
 121        nir_if_last_else_node(if_stmt) != else_node)
 122       return true;
 123
 124    nir_block *then_block = nir_cf_node_as_block(then_node);
 125    nir_block *else_block = nir_cf_node_as_block(else_node);
 126
 127    /* ... and those blocks must only contain "allowed" instructions. */
 128    if (!block_check_for_allowed_instrs(then_block) ||
 129        !block_check_for_allowed_instrs(else_block))
 130       return true;
 131
 132    /* At this point, we know that the previous CFG node is an if-then
 133     * statement containing only moves to phi nodes in this block.  We can
 134     * just remove that entire CF node and replace all of the phi nodes with
 135     * selects.
 136     */
 137
 138    nir_foreach_instr_safe(block, instr) {
 139       if (instr->type != nir_instr_type_phi)
 140          break;
 141
 142       nir_phi_instr *phi = nir_instr_as_phi(instr);
 143       nir_alu_instr *sel = nir_alu_instr_create(state->mem_ctx, nir_op_bcsel);
 144       nir_src_copy(&sel->src[0].src, &if_stmt->condition, state->mem_ctx);
 145       /* Splat the condition to all channels */
 146       memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle);
 147
 148       assert(exec_list_length(&phi->srcs) == 2);
 149       nir_foreach_phi_src(phi, src) {
 150          assert(src->pred == then_block || src->pred == else_block);
 151          assert(src->src.is_ssa);
 152
 153          unsigned idx = src->pred == then_block ? 1 : 2;
 154
 155          if (src->src.ssa->parent_instr->block == src->pred) {
 156             /* We already know that this instruction must be a move with
 157              * this phi's in this block as its only users.
 158              */
 159             nir_alu_instr *mov = nir_instr_as_alu(src->src.ssa->parent_instr);
 160             assert(mov->instr.type == nir_instr_type_alu);
 161             assert(mov->op == nir_op_fmov || mov->op == nir_op_imov);
 162
 163             nir_alu_src_copy(&sel->src[idx], &mov->src[0], state->mem_ctx);
 164          } else {
 165             nir_src_copy(&sel->src[idx].src, &src->src, state->mem_ctx);
 166          }
 167       }
 168
 169       nir_ssa_dest_init(&sel->instr, &sel->dest.dest,
 170                         phi->dest.ssa.num_components, phi->dest.ssa.name);
 171       sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
 172
 173       nir_ssa_def_rewrite_uses(&phi->dest.ssa,
 174                                nir_src_for_ssa(&sel->dest.dest.ssa),
 175                                state->mem_ctx);
 176
 177       nir_instr_insert_before(&phi->instr, &sel->instr);
 178       nir_instr_remove(&phi->instr);
 179    }
 180
 181    nir_cf_node_remove(&if_stmt->cf_node);
 182    state->progress = true;
 183
 184    return true;
 185 }
 186
 187 static bool
 188 nir_opt_peephole_select_impl(nir_function_impl *impl)
 189 {
 190    struct peephole_select_state state;
 191
 192    state.mem_ctx = ralloc_parent(impl);
 193    state.progress = false;
 194
 195    nir_foreach_block(impl, nir_opt_peephole_select_block, &state);
 196
 197    if (state.progress)
 198       nir_metadata_preserve(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 }