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  * Gets the single block that jumps back to the loop header. Already assumes
  31  * there is exactly one such block.
  32  */
  33 static nir_block*
  34 find_continue_block(nir_loop *loop)
  35 {
  36    nir_block *header_block = nir_loop_first_block(loop);
  37    nir_block *prev_block =
  38       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
  39
  40    assert(header_block->predecessors->entries == 2);
  41
  42    struct set_entry *pred_entry;
  43    set_foreach(header_block->predecessors, pred_entry) {
  44       if (pred_entry->key != prev_block)
  45          return (nir_block*)pred_entry->key;
  46    }
  47 }
  48
  49 /**
  50  * This optimization detects if statements at the tops of loops where the
  51  * condition is a phi node of two constants and moves half of the if to above
  52  * the loop and the other half of the if to the end of the loop.  A simple for
  53  * loop "for (int i = 0; i < 4; i++)", when run through the SPIR-V front-end,
  54  * ends up looking something like this:
  55  *
  56  * vec1 32 ssa_0 = load_const (0x00000000)
  57  * vec1 32 ssa_1 = load_const (0xffffffff)
  58  * loop {
  59  *    block block_1:
  60  *    vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
  61  *    vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
  62  *    if ssa_2 {
  63  *       block block_2:
  64  *       vec1 32 ssa_4 = load_const (0x00000001)
  65  *       vec1 32 ssa_5 = iadd ssa_2, ssa_4
  66  *    } else {
  67  *       block block_3:
  68  *    }
  69  *    block block_4:
  70  *    vec1 32 ssa_6 = load_const (0x00000004)
  71  *    vec1 32 ssa_7 = ilt ssa_5, ssa_6
  72  *    if ssa_7 {
  73  *       block block_5:
  74  *    } else {
  75  *       block block_6:
  76  *       break
  77  *    }
  78  *    block block_7:
  79  * }
  80  *
  81  * This turns it into something like this:
  82  *
  83  * // Stuff from block 1
  84  * // Stuff from block 3
  85  * loop {
  86  *    block block_1:
  87  *    vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
  88  *    vec1 32 ssa_6 = load_const (0x00000004)
  89  *    vec1 32 ssa_7 = ilt ssa_5, ssa_6
  90  *    if ssa_7 {
  91  *       block block_5:
  92  *    } else {
  93  *       block block_6:
  94  *       break
  95  *    }
  96  *    block block_7:
  97  *    // Stuff from block 1
  98  *    // Stuff from block 2
  99  *    vec1 32 ssa_4 = load_const (0x00000001)
 100  *    vec1 32 ssa_5 = iadd ssa_2, ssa_4
 101  * }
 102  */
 103 static bool
 104 opt_peel_loop_initial_if(nir_loop *loop)
 105 {
 106    nir_block *header_block = nir_loop_first_block(loop);
 107    nir_block *prev_block =
 108       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
 109
 110    /* It would be insane if this were not true */
 111    assert(_mesa_set_search(header_block->predecessors, prev_block));
 112
 113    /* The loop must have exactly one continue block which could be a block
 114     * ending in a continue instruction or the "natural" continue from the
 115     * last block in the loop back to the top.
 116     */
 117    if (header_block->predecessors->entries != 2)
 118       return false;
 119
 120    nir_block *continue_block = find_continue_block(loop);
 121
 122    nir_cf_node *if_node = nir_cf_node_next(&header_block->cf_node);
 123    if (!if_node || if_node->type != nir_cf_node_if)
 124       return false;
 125
 126    nir_if *nif = nir_cf_node_as_if(if_node);
 127    assert(nif->condition.is_ssa);
 128
 129    nir_ssa_def *cond = nif->condition.ssa;
 130    if (cond->parent_instr->type != nir_instr_type_phi)
 131       return false;
 132
 133    nir_phi_instr *cond_phi = nir_instr_as_phi(cond->parent_instr);
 134    if (cond->parent_instr->block != header_block)
 135       return false;
 136
 137    /* We already know we have exactly one continue */
 138    assert(exec_list_length(&cond_phi->srcs) == 2);
 139
 140    uint32_t entry_val = 0, continue_val = 0;
 141    nir_foreach_phi_src(src, cond_phi) {
 142       assert(src->src.is_ssa);
 143       nir_const_value *const_src = nir_src_as_const_value(src->src);
 144       if (!const_src)
 145          return false;
 146
 147       if (src->pred == continue_block) {
 148          continue_val = const_src->u32[0];
 149       } else {
 150          assert(src->pred == prev_block);
 151          entry_val = const_src->u32[0];
 152       }
 153    }
 154
 155    /* If they both execute or both don't execute, this is a job for
 156     * nir_dead_cf, not this pass.
 157     */
 158    if ((entry_val && continue_val) || (!entry_val && !continue_val))
 159       return false;
 160
 161    struct exec_list *continue_list, *entry_list;
 162    if (continue_val) {
 163       continue_list = &nif->then_list;
 164       entry_list = &nif->else_list;
 165    } else {
 166       continue_list = &nif->else_list;
 167       entry_list = &nif->then_list;
 168    }
 169
 170    /* We want to be moving the contents of entry_list to above the loop so it
 171     * can't contain any break or continue instructions.
 172     */
 173    foreach_list_typed(nir_cf_node, cf_node, node, entry_list) {
 174       nir_foreach_block_in_cf_node(block, cf_node) {
 175          nir_instr *last_instr = nir_block_last_instr(block);
 176          if (last_instr && last_instr->type == nir_instr_type_jump)
 177             return false;
 178       }
 179    }
 180
 181    /* Before we do anything, convert the loop to LCSSA.  We're about to
 182     * replace a bunch of SSA defs with registers and this will prevent any of
 183     * it from leaking outside the loop.
 184     */
 185    nir_convert_loop_to_lcssa(loop);
 186
 187    nir_block *after_if_block =
 188       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
 189
 190    /* Get rid of phis in the header block since we will be duplicating it */
 191    nir_lower_phis_to_regs_block(header_block);
 192    /* Get rid of phis after the if since dominance will change */
 193    nir_lower_phis_to_regs_block(after_if_block);
 194
 195    /* Get rid of SSA defs in the pieces we're about to move around */
 196    nir_lower_ssa_defs_to_regs_block(header_block);
 197    nir_foreach_block_in_cf_node(block, &nif->cf_node)
 198       nir_lower_ssa_defs_to_regs_block(block);
 199
 200    nir_cf_list header, tmp;
 201    nir_cf_extract(&header, nir_before_block(header_block),
 202                            nir_after_block(header_block));
 203
 204    nir_cf_list_clone(&tmp, &header, &loop->cf_node, NULL);
 205    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
 206    nir_cf_extract(&tmp, nir_before_cf_list(entry_list),
 207                         nir_after_cf_list(entry_list));
 208    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
 209
 210    nir_cf_reinsert(&header, nir_after_block_before_jump(continue_block));
 211
 212    /* Get continue block again as the previous reinsert might have removed the block. */
 213    continue_block = find_continue_block(loop);
 214
 215    nir_cf_extract(&tmp, nir_before_cf_list(continue_list),
 216                         nir_after_cf_list(continue_list));
 217    nir_cf_reinsert(&tmp, nir_after_block_before_jump(continue_block));
 218
 219    nir_cf_node_remove(&nif->cf_node);
 220
 221    return true;
 222 }
 223
 224 static bool
 225 is_block_empty(nir_block *block)
 226 {
 227    return nir_cf_node_is_last(&block->cf_node) &&
 228           exec_list_is_empty(&block->instr_list);
 229 }
 230
 231 /**
 232  * This optimization turns:
 233  *
 234  *     if (cond) {
 235  *     } else {
 236  *         do_work();
 237  *     }
 238  *
 239  * into:
 240  *
 241  *     if (!cond) {
 242  *         do_work();
 243  *     } else {
 244  *     }
 245  */
 246 static bool
 247 opt_if_simplification(nir_builder *b, nir_if *nif)
 248 {
 249    /* Only simplify if the then block is empty and the else block is not. */
 250    if (!is_block_empty(nir_if_first_then_block(nif)) ||
 251        is_block_empty(nir_if_first_else_block(nif)))
 252       return false;
 253
 254    /* Make sure the condition is a comparison operation. */
 255    nir_instr *src_instr = nif->condition.ssa->parent_instr;
 256    if (src_instr->type != nir_instr_type_alu)
 257       return false;
 258
 259    nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
 260    if (!nir_alu_instr_is_comparison(alu_instr))
 261       return false;
 262
 263    /* Insert the inverted instruction and rewrite the condition. */
 264    b->cursor = nir_after_instr(&alu_instr->instr);
 265
 266    nir_ssa_def *new_condition =
 267       nir_inot(b, &alu_instr->dest.dest.ssa);
 268
 269    nir_if_rewrite_condition(nif, nir_src_for_ssa(new_condition));
 270
 271    /* Grab pointers to the last then/else blocks for fixing up the phis. */
 272    nir_block *then_block = nir_if_last_then_block(nif);
 273    nir_block *else_block = nir_if_last_else_block(nif);
 274
 275    /* Walk all the phis in the block immediately following the if statement and
 276     * swap the blocks.
 277     */
 278    nir_block *after_if_block =
 279       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
 280
 281    nir_foreach_instr(instr, after_if_block) {
 282       if (instr->type != nir_instr_type_phi)
 283          continue;
 284
 285       nir_phi_instr *phi = nir_instr_as_phi(instr);
 286
 287       foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 288          if (src->pred == else_block) {
 289             src->pred = then_block;
 290          } else if (src->pred == then_block) {
 291             src->pred = else_block;
 292          }
 293       }
 294    }
 295
 296    /* Finally, move the else block to the then block. */
 297    nir_cf_list tmp;
 298    nir_cf_extract(&tmp, nir_before_cf_list(&nif->else_list),
 299                         nir_after_cf_list(&nif->else_list));
 300    nir_cf_reinsert(&tmp, nir_before_cf_list(&nif->then_list));
 301
 302    return true;
 303 }
 304
 305 /**
 306  * This optimization simplifies potential loop terminators which then allows
 307  * other passes such as opt_if_simplification() and loop unrolling to progress
 308  * further:
 309  *
 310  *     if (cond) {
 311  *        ... then block instructions ...
 312  *     } else {
 313  *         ...
 314  *        break;
 315  *     }
 316  *
 317  * into:
 318  *
 319  *     if (cond) {
 320  *     } else {
 321  *         ...
 322  *        break;
 323  *     }
 324  *     ... then block instructions ...
 325  */
 326 static bool
 327 opt_if_loop_terminator(nir_if *nif)
 328 {
 329    nir_block *break_blk = NULL;
 330    nir_block *continue_from_blk = NULL;
 331    bool continue_from_then = true;
 332
 333    nir_block *last_then = nir_if_last_then_block(nif);
 334    nir_block *last_else = nir_if_last_else_block(nif);
 335
 336    if (nir_block_ends_in_break(last_then)) {
 337       break_blk = last_then;
 338       continue_from_blk = last_else;
 339       continue_from_then = false;
 340    } else if (nir_block_ends_in_break(last_else)) {
 341       break_blk = last_else;
 342       continue_from_blk = last_then;
 343    }
 344
 345    /* Continue if the if-statement contained no jumps at all */
 346    if (!break_blk)
 347       return false;
 348
 349    /* If the continue from block is empty then return as there is nothing to
 350     * move.
 351     */
 352    nir_block *first_continue_from_blk = continue_from_then ?
 353       nir_if_first_then_block(nif) :
 354       nir_if_first_else_block(nif);
 355    if (is_block_empty(first_continue_from_blk))
 356       return false;
 357
 358    if (!nir_is_trivial_loop_if(nif, break_blk))
 359       return false;
 360
 361    /* Finally, move the continue from branch after the if-statement. */
 362    nir_cf_list tmp;
 363    nir_cf_extract(&tmp, nir_before_block(first_continue_from_blk),
 364                         nir_after_block(continue_from_blk));
 365    nir_cf_reinsert(&tmp, nir_after_cf_node(&nif->cf_node));
 366
 367    return true;
 368 }
 369
 370 static bool
 371 opt_if_cf_list(nir_builder *b, struct exec_list *cf_list)
 372 {
 373    bool progress = false;
 374    foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
 375       switch (cf_node->type) {
 376       case nir_cf_node_block:
 377          break;
 378
 379       case nir_cf_node_if: {
 380          nir_if *nif = nir_cf_node_as_if(cf_node);
 381          progress |= opt_if_cf_list(b, &nif->then_list);
 382          progress |= opt_if_cf_list(b, &nif->else_list);
 383          progress |= opt_if_loop_terminator(nif);
 384          progress |= opt_if_simplification(b, nif);
 385          break;
 386       }
 387
 388       case nir_cf_node_loop: {
 389          nir_loop *loop = nir_cf_node_as_loop(cf_node);
 390          progress |= opt_if_cf_list(b, &loop->body);
 391          progress |= opt_peel_loop_initial_if(loop);
 392          break;
 393       }
 394
 395       case nir_cf_node_function:
 396          unreachable("Invalid cf type");
 397       }
 398    }
 399
 400    return progress;
 401 }
 402
 403 bool
 404 nir_opt_if(nir_shader *shader)
 405 {
 406    bool progress = false;
 407
 408    nir_foreach_function(function, shader) {
 409       if (function->impl == NULL)
 410          continue;
 411
 412       nir_builder b;
 413       nir_builder_init(&b, function->impl);
 414
 415       if (opt_if_cf_list(&b, &function->impl->body)) {
 416          nir_metadata_preserve(function->impl, nir_metadata_none);
 417
 418          /* If that made progress, we're no longer really in SSA form.  We
 419           * need to convert registers back into SSA defs and clean up SSA defs
 420           * that don't dominate their uses.
 421           */
 422          nir_lower_regs_to_ssa_impl(function->impl);
 423
 424          /* Calling nir_convert_loop_to_lcssa() in opt_peel_loop_initial_if()
 425           * adds extra phi nodes which may not be valid if they're used for
 426           * something such as a deref.  Remove any unneeded phis.
 427           */
 428          nir_opt_remove_phis_impl(function->impl);
 429
 430          progress = true;
 431       }
 432    }
 433
 434    return progress;
 435 }