src/compiler/nir/nir_lower_phis_to_scalar.c

   1 /*
   2  * Copyright © 2015 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 pass that lowers vector phi nodes to scalar phi nodes when
  32  * we don't think it will hurt anything.
  33  */
  34
  35 struct lower_phis_to_scalar_state {
  36    void *mem_ctx;
  37    void *dead_ctx;
  38
  39    /* Hash table marking which phi nodes are scalarizable.  The key is
  40     * pointers to phi instructions and the entry is either NULL for not
  41     * scalarizable or non-null for scalarizable.
  42     */
  43    struct hash_table *phi_table;
  44 };
  45
  46 static bool
  47 should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state);
  48
  49 static bool
  50 is_phi_src_scalarizable(nir_phi_src *src,
  51                         struct lower_phis_to_scalar_state *state)
  52 {
  53    /* Don't know what to do with non-ssa sources */
  54    if (!src->src.is_ssa)
  55       return false;
  56
  57    nir_instr *src_instr = src->src.ssa->parent_instr;
  58    switch (src_instr->type) {
  59    case nir_instr_type_alu: {
  60       nir_alu_instr *src_alu = nir_instr_as_alu(src_instr);
  61
  62       /* ALU operations with output_size == 0 should be scalarized.  We
  63        * will also see a bunch of vecN operations from scalarizing ALU
  64        * operations and, since they can easily be copy-propagated, they
  65        * are ok too.
  66        */
  67       return nir_op_infos[src_alu->op].output_size == 0 ||
  68              nir_op_is_vec(src_alu->op);
  69    }
  70
  71    case nir_instr_type_phi:
  72       /* A phi is scalarizable if we're going to lower it */
  73       return should_lower_phi(nir_instr_as_phi(src_instr), state);
  74
  75    case nir_instr_type_load_const:
  76       /* These are trivially scalarizable */
  77       return true;
  78
  79    case nir_instr_type_ssa_undef:
  80       /* The caller of this function is going to OR the results and we don't
  81        * want undefs to count so we return false.
  82        */
  83       return false;
  84
  85    case nir_instr_type_intrinsic: {
  86       nir_intrinsic_instr *src_intrin = nir_instr_as_intrinsic(src_instr);
  87
  88       switch (src_intrin->intrinsic) {
  89       case nir_intrinsic_load_deref: {
  90          nir_deref_instr *deref = nir_src_as_deref(src_intrin->src[0]);
  91          return deref->mode == nir_var_shader_in ||
  92                 deref->mode == nir_var_uniform ||
  93                 deref->mode == nir_var_mem_ubo ||
  94                 deref->mode == nir_var_mem_ssbo ||
  95                 deref->mode == nir_var_mem_global;
  96       }
  97
  98       case nir_intrinsic_interp_deref_at_centroid:
  99       case nir_intrinsic_interp_deref_at_sample:
 100       case nir_intrinsic_interp_deref_at_offset:
 101       case nir_intrinsic_interp_deref_at_vertex:
 102       case nir_intrinsic_load_uniform:
 103       case nir_intrinsic_load_ubo:
 104       case nir_intrinsic_load_ssbo:
 105       case nir_intrinsic_load_global:
 106       case nir_intrinsic_load_input:
 107          return true;
 108       default:
 109          break;
 110       }
 111    }
 112
 113    default:
 114       /* We can't scalarize this type of instruction */
 115       return false;
 116    }
 117 }
 118
 119 /**
 120  * Determines if the given phi node should be lowered.  The only phi nodes
 121  * we will scalarize at the moment are those where all of the sources are
 122  * scalarizable.
 123  *
 124  * The reason for this comes down to coalescing.  Since phi sources can't
 125  * swizzle, swizzles on phis have to be resolved by inserting a mov right
 126  * before the phi.  The choice then becomes between movs to pick off
 127  * components for a scalar phi or potentially movs to recombine components
 128  * for a vector phi.  The problem is that the movs generated to pick off
 129  * the components are almost uncoalescable.  We can't coalesce them in NIR
 130  * because we need them to pick off components and we can't coalesce them
 131  * in the backend because the source register is a vector and the
 132  * destination is a scalar that may be used at other places in the program.
 133  * On the other hand, if we have a bunch of scalars going into a vector
 134  * phi, the situation is much better.  In this case, if the SSA def is
 135  * generated in the predecessor block to the corresponding phi source, the
 136  * backend code will be an ALU op into a temporary and then a mov into the
 137  * given vector component;  this move can almost certainly be coalesced
 138  * away.
 139  */
 140 static bool
 141 should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state)
 142 {
 143    /* Already scalar */
 144    if (phi->dest.ssa.num_components == 1)
 145       return false;
 146
 147    struct hash_entry *entry = _mesa_hash_table_search(state->phi_table, phi);
 148    if (entry)
 149       return entry->data != NULL;
 150
 151    /* Insert an entry and mark it as scalarizable for now. That way
 152     * we don't recurse forever and a cycle in the dependence graph
 153     * won't automatically make us fail to scalarize.
 154     */
 155    entry = _mesa_hash_table_insert(state->phi_table, phi, (void *)(intptr_t)1);
 156
 157    bool scalarizable = false;
 158
 159    nir_foreach_phi_src(src, phi) {
 160       /* This loop ignores srcs that are not scalarizable because its likely
 161        * still worth copying to temps if another phi source is scalarizable.
 162        * This reduces register spilling by a huge amount in the i965 driver for
 163        * Deus Ex: MD.
 164        */
 165       scalarizable = is_phi_src_scalarizable(src, state);
 166       if (scalarizable)
 167          break;
 168    }
 169
 170    /* The hash table entry for 'phi' may have changed while recursing the
 171     * dependence graph, so we need to reset it */
 172    entry = _mesa_hash_table_search(state->phi_table, phi);
 173    assert(entry);
 174
 175    entry->data = (void *)(intptr_t)scalarizable;
 176
 177    return scalarizable;
 178 }
 179
 180 static bool
 181 lower_phis_to_scalar_block(nir_block *block,
 182                            struct lower_phis_to_scalar_state *state)
 183 {
 184    bool progress = false;
 185
 186    /* Find the last phi node in the block */
 187    nir_phi_instr *last_phi = NULL;
 188    nir_foreach_instr(instr, block) {
 189       if (instr->type != nir_instr_type_phi)
 190          break;
 191
 192       last_phi = nir_instr_as_phi(instr);
 193    }
 194
 195    /* We have to handle the phi nodes in their own pass due to the way
 196     * we're modifying the linked list of instructions.
 197     */
 198    nir_foreach_instr_safe(instr, block) {
 199       if (instr->type != nir_instr_type_phi)
 200          break;
 201
 202       nir_phi_instr *phi = nir_instr_as_phi(instr);
 203
 204       if (!should_lower_phi(phi, state))
 205          continue;
 206
 207       unsigned bit_size = phi->dest.ssa.bit_size;
 208
 209       /* Create a vecN operation to combine the results.  Most of these
 210        * will be redundant, but copy propagation should clean them up for
 211        * us.  No need to add the complexity here.
 212        */
 213       nir_op vec_op = nir_op_vec(phi->dest.ssa.num_components);
 214
 215       nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, vec_op);
 216       nir_ssa_dest_init(&vec->instr, &vec->dest.dest,
 217                         phi->dest.ssa.num_components,
 218                         bit_size, NULL);
 219       vec->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
 220
 221       for (unsigned i = 0; i < phi->dest.ssa.num_components; i++) {
 222          nir_phi_instr *new_phi = nir_phi_instr_create(state->mem_ctx);
 223          nir_ssa_dest_init(&new_phi->instr, &new_phi->dest, 1,
 224                            phi->dest.ssa.bit_size, NULL);
 225
 226          vec->src[i].src = nir_src_for_ssa(&new_phi->dest.ssa);
 227
 228          nir_foreach_phi_src(src, phi) {
 229             /* We need to insert a mov to grab the i'th component of src */
 230             nir_alu_instr *mov = nir_alu_instr_create(state->mem_ctx,
 231                                                       nir_op_mov);
 232             nir_ssa_dest_init(&mov->instr, &mov->dest.dest, 1, bit_size, NULL);
 233             mov->dest.write_mask = 1;
 234             nir_src_copy(&mov->src[0].src, &src->src, state->mem_ctx);
 235             mov->src[0].swizzle[0] = i;
 236
 237             /* Insert at the end of the predecessor but before the jump */
 238             nir_instr *pred_last_instr = nir_block_last_instr(src->pred);
 239             if (pred_last_instr && pred_last_instr->type == nir_instr_type_jump)
 240                nir_instr_insert_before(pred_last_instr, &mov->instr);
 241             else
 242                nir_instr_insert_after_block(src->pred, &mov->instr);
 243
 244             nir_phi_src *new_src = ralloc(new_phi, nir_phi_src);
 245             new_src->pred = src->pred;
 246             new_src->src = nir_src_for_ssa(&mov->dest.dest.ssa);
 247
 248             exec_list_push_tail(&new_phi->srcs, &new_src->node);
 249          }
 250
 251          nir_instr_insert_before(&phi->instr, &new_phi->instr);
 252       }
 253
 254       nir_instr_insert_after(&last_phi->instr, &vec->instr);
 255
 256       nir_ssa_def_rewrite_uses(&phi->dest.ssa,
 257                                nir_src_for_ssa(&vec->dest.dest.ssa));
 258
 259       ralloc_steal(state->dead_ctx, phi);
 260       nir_instr_remove(&phi->instr);
 261
 262       progress = true;
 263
 264       /* We're using the safe iterator and inserting all the newly
 265        * scalarized phi nodes before their non-scalarized version so that's
 266        * ok.  However, we are also inserting vec operations after all of
 267        * the last phi node so once we get here, we can't trust even the
 268        * safe iterator to stop properly.  We have to break manually.
 269        */
 270       if (instr == &last_phi->instr)
 271          break;
 272    }
 273
 274    return progress;
 275 }
 276
 277 static bool
 278 lower_phis_to_scalar_impl(nir_function_impl *impl)
 279 {
 280    struct lower_phis_to_scalar_state state;
 281    bool progress = false;
 282
 283    state.mem_ctx = ralloc_parent(impl);
 284    state.dead_ctx = ralloc_context(NULL);
 285    state.phi_table = _mesa_pointer_hash_table_create(state.dead_ctx);
 286
 287    nir_foreach_block(block, impl) {
 288       progress = lower_phis_to_scalar_block(block, &state) || progress;
 289    }
 290
 291    nir_metadata_preserve(impl, nir_metadata_block_index |
 292                                nir_metadata_dominance);
 293
 294    ralloc_free(state.dead_ctx);
 295    return progress;
 296 }
 297
 298 /** A pass that lowers vector phi nodes to scalar
 299  *
 300  * This pass loops through the blocks and lowers looks for vector phi nodes
 301  * it can lower to scalar phi nodes.  Not all phi nodes are lowered.  For
 302  * instance, if one of the sources is a non-scalarizable vector, then we
 303  * don't bother lowering because that would generate hard-to-coalesce movs.
 304  */
 305 bool
 306 nir_lower_phis_to_scalar(nir_shader *shader)
 307 {
 308    bool progress = false;
 309
 310    nir_foreach_function(function, shader) {
 311       if (function->impl)
 312          progress = lower_phis_to_scalar_impl(function->impl) || progress;
 313    }
 314
 315    return progress;
 316 }