compiler/nir/nir_split_var_copies.c

/*
 * Copyright © 2014 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 * Authors:
 *    Jason Ekstrand (jason@jlekstrand.net)
 *
 */

#include "nir.h"

/*
 * Implements "copy splitting" which is similar to structure splitting only
 * it works on copy operations rather than the datatypes themselves.  The
 * GLSL language allows you to copy one variable to another an entire
 * structure (which may contain arrays or other structures) at a time.
 * Normally, in a language such as C this would be handled by a "structure
 * splitting" pass that breaks up the structures.  Unfortunately for us,
 * structures used in inputs or outputs can't be split.  Therefore,
 * regardlesss of what we do, we have to be able to copy to/from
 * structures.
 *
 * The primary purpose of structure splitting is to allow you to better
 * optimize variable access and lower things to registers where you can.
 * The primary issue here is that, if you lower the copy to a bunch of
 * loads and stores, you loose a lot of information about the copy
 * operation that you would like to keep around.  To solve this problem, we
 * have a "copy splitting" pass that, instead of splitting the structures
 * or lowering the copy into loads and storres, splits the copy operation
 * into a bunch of copy operations one for each leaf of the structure tree.
 * If an intermediate array is encountered, it is referenced with a
 * wildcard reference to indicate that the entire array is to be copied.
 *
 * As things become direct, array copies may be able to be losslessly
 * lowered to having fewer and fewer wildcards.  However, until that
 * happens we want to keep the information about the arrays intact.
 *
 * Prior to the copy splitting pass, there are no wildcard references but
 * there may be incomplete references where the tail of the deref chain is
 * an array or a structure and not a specific element.  After the copy
 * splitting pass has completed, every variable deref will be a full-length
 * dereference pointing to a single leaf in the structure type tree with
 * possibly a few wildcard array dereferences.
 */

struct split_var_copies_state {
   nir_shader *shader;
   void *dead_ctx;
   bool progress;
};

/* Recursively constructs deref chains to split a copy instruction into
 * multiple (if needed) copy instructions with full-length deref chains.
 * External callers of this function should pass the tail and head of the
 * deref chains found as the source and destination of the copy instruction
 * into this function.
 *
 * \param  old_copy  The copy instruction we are splitting
 * \param  dest_head The head of the destination deref chain we are building
 * \param  src_head  The head of the source deref chain we are building
 * \param  dest_tail The tail of the destination deref chain we are building
 * \param  src_tail  The tail of the source deref chain we are building
 * \param  state     The current split_var_copies_state object
 */
static void
split_var_copy_instr(nir_intrinsic_instr *old_copy,
                     nir_deref_var *dest_head, nir_deref_var *src_head,
                     nir_deref *dest_tail, nir_deref *src_tail,
                     struct split_var_copies_state *state)
{
   assert(src_tail->type == dest_tail->type);

   /* Make sure these really are the tails of the deref chains */
   assert(dest_tail->child == NULL);
   assert(src_tail->child == NULL);

   switch (glsl_get_base_type(src_tail->type)) {
   case GLSL_TYPE_ARRAY: {
      /* Make a wildcard dereference */
      nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
      deref->deref.type = glsl_get_array_element(src_tail->type);
      deref->deref_array_type = nir_deref_array_type_wildcard;

      /* Set the tail of both as the newly created wildcard deref.  It is
       * safe to use the same wildcard in both places because a) we will be
       * copying it before we put it in an actual instruction and b)
       * everything that will potentially add another link in the deref
       * chain will also add the same thing to both chains.
       */
      src_tail->child = &deref->deref;
      dest_tail->child = &deref->deref;

      split_var_copy_instr(old_copy, dest_head, src_head,
                           dest_tail->child, src_tail->child, state);

      /* Set it back to the way we found it */
      src_tail->child = NULL;
      dest_tail->child = NULL;
      break;
   }

   case GLSL_TYPE_STRUCT:
      /* This is the only part that actually does any interesting
       * splitting.  For array types, we just use wildcards and resolve
       * them later.  For structure types, we need to emit one copy
       * instruction for every structure element.  Because we may have
       * structs inside structs, we just recurse and let the next level
       * take care of any additional structures.
       */
      for (unsigned i = 0; i < glsl_get_length(src_tail->type); i++) {
         nir_deref_struct *deref = nir_deref_struct_create(state->dead_ctx, i);
         deref->deref.type = glsl_get_struct_field(src_tail->type, i);

         /* Set the tail of both as the newly created structure deref.  It
          * is safe to use the same wildcard in both places because a) we
          * will be copying it before we put it in an actual instruction
          * and b) everything that will potentially add another link in the
          * deref chain will also add the same thing to both chains.
          */
         src_tail->child = &deref->deref;
         dest_tail->child = &deref->deref;

         split_var_copy_instr(old_copy, dest_head, src_head,
                              dest_tail->child, src_tail->child, state);
      }
      /* Set it back to the way we found it */
      src_tail->child = NULL;
      dest_tail->child = NULL;
      break;

   case GLSL_TYPE_UINT:
   case GLSL_TYPE_UINT16:
   case GLSL_TYPE_UINT64:
   case GLSL_TYPE_INT:
   case GLSL_TYPE_INT16:
   case GLSL_TYPE_INT64:
   case GLSL_TYPE_FLOAT:
   case GLSL_TYPE_FLOAT16:
   case GLSL_TYPE_DOUBLE:
   case GLSL_TYPE_BOOL:
      if (glsl_type_is_matrix(src_tail->type)) {
         nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
         deref->deref.type = glsl_get_column_type(src_tail->type);
         deref->deref_array_type = nir_deref_array_type_wildcard;

         /* Set the tail of both as the newly created wildcard deref.  It
          * is safe to use the same wildcard in both places because a) we
          * will be copying it before we put it in an actual instruction
          * and b) everything that will potentially add another link in the
          * deref chain will also add the same thing to both chains.
          */
         src_tail->child = &deref->deref;
         dest_tail->child = &deref->deref;

         split_var_copy_instr(old_copy, dest_head, src_head,
                              dest_tail->child, src_tail->child, state);

         /* Set it back to the way we found it */
         src_tail->child = NULL;
         dest_tail->child = NULL;
      } else {
         /* At this point, we have fully built our deref chains and can
          * actually add the new copy instruction.
          */
         nir_intrinsic_instr *new_copy =
            nir_intrinsic_instr_create(state->shader, nir_intrinsic_copy_var);

         /* We need to make copies because a) this deref chain actually
          * belongs to the copy instruction and b) the deref chains may
          * have some of the same links due to the way we constructed them
          */
         new_copy->variables[0] = nir_deref_var_clone(dest_head, new_copy);
         new_copy->variables[1] = nir_deref_var_clone(src_head, new_copy);

         /* Emit the copy instruction after the old instruction.  We'll
          * remove the old one later.
          */
         nir_instr_insert_after(&old_copy->instr, &new_copy->instr);
         state->progress = true;
      }
      break;

   case GLSL_TYPE_SAMPLER:
   case GLSL_TYPE_IMAGE:
   case GLSL_TYPE_ATOMIC_UINT:
   case GLSL_TYPE_INTERFACE:
   default:
      unreachable("Cannot copy these types");
   }
}

static bool
split_var_copies_block(nir_block *block, struct split_var_copies_state *state)
{
   nir_foreach_instr_safe(instr, block) {
      if (instr->type != nir_instr_type_intrinsic)
         continue;

      nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
      if (intrinsic->intrinsic != nir_intrinsic_copy_var)
         continue;

      nir_deref_var *dest_head = intrinsic->variables[0];
      nir_deref_var *src_head = intrinsic->variables[1];
      nir_deref *dest_tail = nir_deref_tail(&dest_head->deref);
      nir_deref *src_tail = nir_deref_tail(&src_head->deref);

      switch (glsl_get_base_type(src_tail->type)) {
      case GLSL_TYPE_ARRAY:
      case GLSL_TYPE_STRUCT:
         split_var_copy_instr(intrinsic, dest_head, src_head,
                              dest_tail, src_tail, state);
         nir_instr_remove(&intrinsic->instr);
         ralloc_steal(state->dead_ctx, instr);
         break;
      case GLSL_TYPE_FLOAT:
      case GLSL_TYPE_FLOAT16:
      case GLSL_TYPE_DOUBLE:
         if (glsl_type_is_matrix(src_tail->type)) {
            split_var_copy_instr(intrinsic, dest_head, src_head,
                                 dest_tail, src_tail, state);
            nir_instr_remove(&intrinsic->instr);
            ralloc_steal(state->dead_ctx, instr);
         }
         break;
      case GLSL_TYPE_INT:
      case GLSL_TYPE_UINT:
      case GLSL_TYPE_INT16:
      case GLSL_TYPE_UINT16:
      case GLSL_TYPE_INT64:
      case GLSL_TYPE_UINT64:
      case GLSL_TYPE_BOOL:
         assert(!glsl_type_is_matrix(src_tail->type));
         break;
      default:
         unreachable("Invalid type");
         break;
      }
   }

   return true;
}

static bool
split_var_copies_impl(nir_function_impl *impl)
{
   struct split_var_copies_state state;

   state.shader = impl->function->shader;
   state.dead_ctx = ralloc_context(NULL);
   state.progress = false;

   nir_foreach_block(block, impl) {
      split_var_copies_block(block, &state);
   }

   ralloc_free(state.dead_ctx);

   if (state.progress) {
      nir_metadata_preserve(impl, nir_metadata_block_index |
                                  nir_metadata_dominance);
   }

   return state.progress;
}

bool
nir_split_var_copies(nir_shader *shader)
{
   bool progress = false;

   nir_foreach_function(function, shader) {
      if (function->impl)
         progress = split_var_copies_impl(function->impl) || progress;
   }

   return progress;
}