src/compiler/nir/nir_opt_if.c

   1 /*
   2  * Copyright © 2016 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
  24 #include "nir.h"
  25 #include "nir/nir_builder.h"
  26 #include "nir_control_flow.h"
  27 #include "nir_loop_analyze.h"
  28
  29 /**
  30  * This optimization detects if statements at the tops of loops where the
  31  * condition is a phi node of two constants and moves half of the if to above
  32  * the loop and the other half of the if to the end of the loop.  A simple for
  33  * loop "for (int i = 0; i < 4; i++)", when run through the SPIR-V front-end,
  34  * ends up looking something like this:
  35  *
  36  * vec1 32 ssa_0 = load_const (0x00000000)
  37  * vec1 32 ssa_1 = load_const (0xffffffff)
  38  * loop {
  39  *    block block_1:
  40  *    vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
  41  *    vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
  42  *    if ssa_2 {
  43  *       block block_2:
  44  *       vec1 32 ssa_4 = load_const (0x00000001)
  45  *       vec1 32 ssa_5 = iadd ssa_2, ssa_4
  46  *    } else {
  47  *       block block_3:
  48  *    }
  49  *    block block_4:
  50  *    vec1 32 ssa_6 = load_const (0x00000004)
  51  *    vec1 32 ssa_7 = ilt ssa_5, ssa_6
  52  *    if ssa_7 {
  53  *       block block_5:
  54  *    } else {
  55  *       block block_6:
  56  *       break
  57  *    }
  58  *    block block_7:
  59  * }
  60  *
  61  * This turns it into something like this:
  62  *
  63  * // Stuff from block 1
  64  * // Stuff from block 3
  65  * loop {
  66  *    block block_1:
  67  *    vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
  68  *    vec1 32 ssa_6 = load_const (0x00000004)
  69  *    vec1 32 ssa_7 = ilt ssa_5, ssa_6
  70  *    if ssa_7 {
  71  *       block block_5:
  72  *    } else {
  73  *       block block_6:
  74  *       break
  75  *    }
  76  *    block block_7:
  77  *    // Stuff from block 1
  78  *    // Stuff from block 2
  79  *    vec1 32 ssa_4 = load_const (0x00000001)
  80  *    vec1 32 ssa_5 = iadd ssa_2, ssa_4
  81  * }
  82  */
  83 static bool
  84 opt_peel_loop_initial_if(nir_loop *loop)
  85 {
  86    nir_block *header_block = nir_loop_first_block(loop);
  87    nir_block *prev_block =
  88       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
  89
  90    /* It would be insane if this were not true */
  91    assert(_mesa_set_search(header_block->predecessors, prev_block));
  92
  93    /* The loop must have exactly one continue block which could be a block
  94     * ending in a continue instruction or the "natural" continue from the
  95     * last block in the loop back to the top.
  96     */
  97    if (header_block->predecessors->entries != 2)
  98       return false;
  99
 100    nir_block *continue_block = NULL;
 101    struct set_entry *pred_entry;
 102    set_foreach(header_block->predecessors, pred_entry) {
 103       if (pred_entry->key != prev_block)
 104          continue_block = (void *)pred_entry->key;
 105    }
 106
 107    nir_cf_node *if_node = nir_cf_node_next(&header_block->cf_node);
 108    if (!if_node || if_node->type != nir_cf_node_if)
 109       return false;
 110
 111    nir_if *nif = nir_cf_node_as_if(if_node);
 112    assert(nif->condition.is_ssa);
 113
 114    nir_ssa_def *cond = nif->condition.ssa;
 115    if (cond->parent_instr->type != nir_instr_type_phi)
 116       return false;
 117
 118    nir_phi_instr *cond_phi = nir_instr_as_phi(cond->parent_instr);
 119    if (cond->parent_instr->block != header_block)
 120       return false;
 121
 122    /* We already know we have exactly one continue */
 123    assert(exec_list_length(&cond_phi->srcs) == 2);
 124
 125    uint32_t entry_val = 0, continue_val = 0;
 126    nir_foreach_phi_src(src, cond_phi) {
 127       assert(src->src.is_ssa);
 128       nir_const_value *const_src = nir_src_as_const_value(src->src);
 129       if (!const_src)
 130          return false;
 131
 132       if (src->pred == continue_block) {
 133          continue_val = const_src->u32[0];
 134       } else {
 135          assert(src->pred == prev_block);
 136          entry_val = const_src->u32[0];
 137       }
 138    }
 139
 140    /* If they both execute or both don't execute, this is a job for
 141     * nir_dead_cf, not this pass.
 142     */
 143    if ((entry_val && continue_val) || (!entry_val && !continue_val))
 144       return false;
 145
 146    struct exec_list *continue_list, *entry_list;
 147    if (continue_val) {
 148       continue_list = &nif->then_list;
 149       entry_list = &nif->else_list;
 150    } else {
 151       continue_list = &nif->else_list;
 152       entry_list = &nif->then_list;
 153    }
 154
 155    /* We want to be moving the contents of entry_list to above the loop so it
 156     * can't contain any break or continue instructions.
 157     */
 158    foreach_list_typed(nir_cf_node, cf_node, node, entry_list) {
 159       nir_foreach_block_in_cf_node(block, cf_node) {
 160          nir_instr *last_instr = nir_block_last_instr(block);
 161          if (last_instr && last_instr->type == nir_instr_type_jump)
 162             return false;
 163       }
 164    }
 165
 166    /* Before we do anything, convert the loop to LCSSA.  We're about to
 167     * replace a bunch of SSA defs with registers and this will prevent any of
 168     * it from leaking outside the loop.
 169     */
 170    nir_convert_loop_to_lcssa(loop);
 171
 172    nir_block *after_if_block =
 173       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
 174
 175    /* Get rid of phis in the header block since we will be duplicating it */
 176    nir_lower_phis_to_regs_block(header_block);
 177    /* Get rid of phis after the if since dominance will change */
 178    nir_lower_phis_to_regs_block(after_if_block);
 179
 180    /* Get rid of SSA defs in the pieces we're about to move around */
 181    nir_lower_ssa_defs_to_regs_block(header_block);
 182    nir_foreach_block_in_cf_node(block, &nif->cf_node)
 183       nir_lower_ssa_defs_to_regs_block(block);
 184
 185    nir_cf_list header, tmp;
 186    nir_cf_extract(&header, nir_before_block(header_block),
 187                            nir_after_block(header_block));
 188
 189    nir_cf_list_clone(&tmp, &header, &loop->cf_node, NULL);
 190    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
 191    nir_cf_extract(&tmp, nir_before_cf_list(entry_list),
 192                         nir_after_cf_list(entry_list));
 193    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
 194
 195    nir_cf_reinsert(&header, nir_after_block_before_jump(continue_block));
 196    nir_cf_extract(&tmp, nir_before_cf_list(continue_list),
 197                         nir_after_cf_list(continue_list));
 198    nir_cf_reinsert(&tmp, nir_after_block_before_jump(continue_block));
 199
 200    nir_cf_node_remove(&nif->cf_node);
 201
 202    return true;
 203 }
 204
 205 static bool
 206 is_block_empty(nir_block *block)
 207 {
 208    return nir_cf_node_is_last(&block->cf_node) &&
 209           exec_list_is_empty(&block->instr_list);
 210 }
 211
 212 /**
 213  * This optimization turns:
 214  *
 215  *     if (cond) {
 216  *     } else {
 217  *         do_work();
 218  *     }
 219  *
 220  * into:
 221  *
 222  *     if (!cond) {
 223  *         do_work();
 224  *     } else {
 225  *     }
 226  */
 227 static bool
 228 opt_if_simplification(nir_builder *b, nir_if *nif)
 229 {
 230    /* Only simplify if the then block is empty and the else block is not. */
 231    if (!is_block_empty(nir_if_first_then_block(nif)) ||
 232        is_block_empty(nir_if_first_else_block(nif)))
 233       return false;
 234
 235    /* Make sure the condition is a comparison operation. */
 236    nir_instr *src_instr = nif->condition.ssa->parent_instr;
 237    if (src_instr->type != nir_instr_type_alu)
 238       return false;
 239
 240    nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
 241    if (!nir_alu_instr_is_comparison(alu_instr))
 242       return false;
 243
 244    /* Insert the inverted instruction and rewrite the condition. */
 245    b->cursor = nir_after_instr(&alu_instr->instr);
 246
 247    nir_ssa_def *new_condition =
 248       nir_inot(b, &alu_instr->dest.dest.ssa);
 249
 250    nir_if_rewrite_condition(nif, nir_src_for_ssa(new_condition));
 251
 252    /* Grab pointers to the last then/else blocks for fixing up the phis. */
 253    nir_block *then_block = nir_if_last_then_block(nif);
 254    nir_block *else_block = nir_if_last_else_block(nif);
 255
 256    /* Walk all the phis in the block immediately following the if statement and
 257     * swap the blocks.
 258     */
 259    nir_block *after_if_block =
 260       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
 261
 262    nir_foreach_instr(instr, after_if_block) {
 263       if (instr->type != nir_instr_type_phi)
 264          continue;
 265
 266       nir_phi_instr *phi = nir_instr_as_phi(instr);
 267
 268       foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 269          if (src->pred == else_block) {
 270             src->pred = then_block;
 271          } else if (src->pred == then_block) {
 272             src->pred = else_block;
 273          }
 274       }
 275    }
 276
 277    /* Finally, move the else block to the then block. */
 278    nir_cf_list tmp;
 279    nir_cf_extract(&tmp, nir_before_cf_list(&nif->else_list),
 280                         nir_after_cf_list(&nif->else_list));
 281    nir_cf_reinsert(&tmp, nir_before_cf_list(&nif->then_list));
 282    nir_cf_delete(&tmp);
 283
 284    return true;
 285 }
 286
 287 /**
 288  * This optimization simplifies potential loop terminators which then allows
 289  * other passes such as opt_if_simplification() and loop unrolling to progress
 290  * further:
 291  *
 292  *     if (cond) {
 293  *        ... then block instructions ...
 294  *     } else {
 295  *         ...
 296  *        break;
 297  *     }
 298  *
 299  * into:
 300  *
 301  *     if (cond) {
 302  *     } else {
 303  *         ...
 304  *        break;
 305  *     }
 306  *     ... then block instructions ...
 307  */
 308 static bool
 309 opt_if_loop_terminator(nir_if *nif)
 310 {
 311    nir_block *break_blk = NULL;
 312    nir_block *continue_from_blk = NULL;
 313    bool continue_from_then = true;
 314
 315    nir_block *last_then = nir_if_last_then_block(nif);
 316    nir_block *last_else = nir_if_last_else_block(nif);
 317
 318    if (nir_block_ends_in_break(last_then)) {
 319       break_blk = last_then;
 320       continue_from_blk = last_else;
 321       continue_from_then = false;
 322    } else if (nir_block_ends_in_break(last_else)) {
 323       break_blk = last_else;
 324       continue_from_blk = last_then;
 325    }
 326
 327    /* Continue if the if-statement contained no jumps at all */
 328    if (!break_blk)
 329       return false;
 330
 331    /* If the continue from block is empty then return as there is nothing to
 332     * move.
 333     */
 334    nir_block *first_continue_from_blk = continue_from_then ?
 335       nir_if_first_then_block(nif) :
 336       nir_if_first_else_block(nif);
 337    if (is_block_empty(first_continue_from_blk))
 338       return false;
 339
 340    if (!nir_is_trivial_loop_if(nif, break_blk))
 341       return false;
 342
 343    /* Finally, move the continue from branch after the if-statement. */
 344    nir_cf_list tmp;
 345    nir_cf_extract(&tmp, nir_before_block(first_continue_from_blk),
 346                         nir_after_block(continue_from_blk));
 347    nir_cf_reinsert(&tmp, nir_after_cf_node(&nif->cf_node));
 348    nir_cf_delete(&tmp);
 349
 350    return true;
 351 }
 352
 353 static bool
 354 opt_if_cf_list(nir_builder *b, struct exec_list *cf_list)
 355 {
 356    bool progress = false;
 357    foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
 358       switch (cf_node->type) {
 359       case nir_cf_node_block:
 360          break;
 361
 362       case nir_cf_node_if: {
 363          nir_if *nif = nir_cf_node_as_if(cf_node);
 364          progress |= opt_if_cf_list(b, &nif->then_list);
 365          progress |= opt_if_cf_list(b, &nif->else_list);
 366          progress |= opt_if_loop_terminator(nif);
 367          progress |= opt_if_simplification(b, nif);
 368          break;
 369       }
 370
 371       case nir_cf_node_loop: {
 372          nir_loop *loop = nir_cf_node_as_loop(cf_node);
 373          progress |= opt_if_cf_list(b, &loop->body);
 374          progress |= opt_peel_loop_initial_if(loop);
 375          break;
 376       }
 377
 378       case nir_cf_node_function:
 379          unreachable("Invalid cf type");
 380       }
 381    }
 382
 383    return progress;
 384 }
 385
 386 bool
 387 nir_opt_if(nir_shader *shader)
 388 {
 389    bool progress = false;
 390
 391    nir_foreach_function(function, shader) {
 392       if (function->impl == NULL)
 393          continue;
 394
 395       nir_builder b;
 396       nir_builder_init(&b, function->impl);
 397
 398       if (opt_if_cf_list(&b, &function->impl->body)) {
 399          nir_metadata_preserve(function->impl, nir_metadata_none);
 400
 401          /* If that made progress, we're no longer really in SSA form.  We
 402           * need to convert registers back into SSA defs and clean up SSA defs
 403           * that don't dominate their uses.
 404           */
 405          nir_lower_regs_to_ssa_impl(function->impl);
 406
 407          /* Calling nir_convert_loop_to_lcssa() in opt_peel_loop_initial_if()
 408           * adds extra phi nodes which may not be valid if they're used for
 409           * something such as a deref.  Remove any unneeded phis.
 410           */
 411          nir_opt_remove_phis_impl(function->impl);
 412
 413          progress = true;
 414       }
 415    }
 416
 417    return progress;
 418 }