src/compiler/nir/nir_inline_functions.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
  24 #include "nir.h"
  25 #include "nir_builder.h"
  26 #include "nir_control_flow.h"
  27
  28 static bool inline_function_impl(nir_function_impl *impl, struct set *inlined);
  29
  30 static void
  31 rewrite_param_derefs(nir_instr *instr, nir_call_instr *call)
  32 {
  33    if (instr->type != nir_instr_type_intrinsic)
  34       return;
  35
  36    nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  37
  38    for (unsigned i = 0;
  39         i < nir_intrinsic_infos[intrin->intrinsic].num_variables; i++) {
  40       if (intrin->variables[i]->var->data.mode != nir_var_param)
  41          continue;
  42
  43       int param_idx = intrin->variables[i]->var->data.location;
  44
  45       nir_deref_var *call_deref;
  46       if (param_idx >= 0) {
  47          assert(param_idx < call->callee->num_params);
  48          call_deref = call->params[param_idx];
  49       } else {
  50          call_deref = call->return_deref;
  51       }
  52       assert(call_deref);
  53
  54       nir_deref_var *new_deref = nir_deref_as_var(nir_copy_deref(intrin, &call_deref->deref));
  55       nir_deref *new_tail = nir_deref_tail(&new_deref->deref);
  56       new_tail->child = intrin->variables[i]->deref.child;
  57       ralloc_steal(new_tail, new_tail->child);
  58       intrin->variables[i] = new_deref;
  59    }
  60 }
  61
  62 static void
  63 lower_param_to_local(nir_variable *param, nir_function_impl *impl, bool write)
  64 {
  65    if (param->data.mode != nir_var_param)
  66       return;
  67
  68    nir_parameter_type param_type;
  69    if (param->data.location >= 0) {
  70       assert(param->data.location < impl->num_params);
  71       param_type = impl->function->params[param->data.location].param_type;
  72    } else {
  73       /* Return variable */
  74       param_type = nir_parameter_out;
  75    }
  76
  77    if ((write && param_type == nir_parameter_in) ||
  78        (!write && param_type == nir_parameter_out)) {
  79       /* In this case, we need a shadow copy.  Turn it into a local */
  80       param->data.mode = nir_var_local;
  81       exec_list_push_tail(&impl->locals, &param->node);
  82    }
  83 }
  84
  85 static bool
  86 lower_params_to_locals_block(nir_block *block, nir_function_impl *impl)
  87 {
  88    nir_foreach_instr(instr, block) {
  89       if (instr->type != nir_instr_type_intrinsic)
  90          continue;
  91
  92       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  93
  94       switch (intrin->intrinsic) {
  95       case nir_intrinsic_store_var:
  96          lower_param_to_local(intrin->variables[0]->var, impl, true);
  97          break;
  98
  99       case nir_intrinsic_copy_var:
 100          lower_param_to_local(intrin->variables[0]->var, impl, true);
 101          lower_param_to_local(intrin->variables[1]->var, impl, false);
 102          break;
 103
 104       case nir_intrinsic_load_var:
 105          /* All other intrinsics which access variables (image_load_store)
 106           * do so in a read-only fasion.
 107           */
 108          for (unsigned i = 0;
 109               i < nir_intrinsic_infos[intrin->intrinsic].num_variables; i++) {
 110             lower_param_to_local(intrin->variables[i]->var, impl, false);
 111          }
 112          break;
 113
 114       default:
 115          continue;
 116       }
 117    }
 118
 119    return true;
 120 }
 121
 122 static bool
 123 inline_functions_block(nir_block *block, nir_builder *b,
 124                        struct set *inlined)
 125 {
 126    bool progress = false;
 127    /* This is tricky.  We're iterating over instructions in a block but, as
 128     * we go, the block and its instruction list are being split into
 129     * pieces.  However, this *should* be safe since foreach_safe always
 130     * stashes the next thing in the iteration.  That next thing will
 131     * properly get moved to the next block when it gets split, and we
 132     * continue iterating there.
 133     */
 134    nir_foreach_instr_safe(instr, block) {
 135       if (instr->type != nir_instr_type_call)
 136          continue;
 137
 138       progress = true;
 139
 140       nir_call_instr *call = nir_instr_as_call(instr);
 141       assert(call->callee->impl);
 142
 143       inline_function_impl(call->callee->impl, inlined);
 144
 145       nir_function_impl *callee_copy =
 146          nir_function_impl_clone(call->callee->impl);
 147       callee_copy->function = call->callee;
 148
 149       /* Add copies of all in parameters */
 150       assert(call->num_params == callee_copy->num_params);
 151
 152       exec_list_append(&b->impl->locals, &callee_copy->locals);
 153       exec_list_append(&b->impl->registers, &callee_copy->registers);
 154
 155       b->cursor = nir_before_instr(&call->instr);
 156
 157       /* We now need to tie the two functions together using the
 158        * parameters.  There are two ways we do this: One is to turn the
 159        * parameter into a local variable and do a shadow-copy.  The other
 160        * is to treat the parameter as a "proxy" and rewrite derefs to use
 161        * the actual variable that comes from the call instruction.  We
 162        * implement both schemes.  The first is needed in the case where we
 163        * have an in parameter that we write or similar.  The second case is
 164        * needed for handling things such as images and uniforms properly.
 165        */
 166
 167       /* Figure out when we need to lower to a shadow local */
 168       nir_foreach_block(block, callee_copy) {
 169          lower_params_to_locals_block(block, callee_copy);
 170       }
 171
 172       for (unsigned i = 0; i < callee_copy->num_params; i++) {
 173          nir_variable *param = callee_copy->params[i];
 174
 175          if (param->data.mode == nir_var_local &&
 176              call->callee->params[i].param_type != nir_parameter_out) {
 177             nir_copy_deref_var(b, nir_deref_var_create(b->shader, param),
 178                                   call->params[i]);
 179          }
 180       }
 181
 182       nir_foreach_block(block, callee_copy) {
 183          nir_foreach_instr(instr, block)
 184             rewrite_param_derefs(instr, call);
 185       }
 186
 187       /* Pluck the body out of the function and place it here */
 188       nir_cf_list body;
 189       nir_cf_list_extract(&body, &callee_copy->body);
 190       nir_cf_reinsert(&body, b->cursor);
 191
 192       b->cursor = nir_before_instr(&call->instr);
 193
 194       /* Add copies of all out parameters and the return */
 195       assert(call->num_params == callee_copy->num_params);
 196       for (unsigned i = 0; i < callee_copy->num_params; i++) {
 197          nir_variable *param = callee_copy->params[i];
 198
 199          if (param->data.mode == nir_var_local &&
 200              call->callee->params[i].param_type != nir_parameter_in) {
 201             nir_copy_deref_var(b, call->params[i],
 202                                   nir_deref_var_create(b->shader, param));
 203          }
 204       }
 205       if (!glsl_type_is_void(call->callee->return_type) &&
 206           callee_copy->return_var->data.mode == nir_var_local) {
 207          nir_copy_deref_var(b, call->return_deref,
 208                                nir_deref_var_create(b->shader,
 209                                                     callee_copy->return_var));
 210       }
 211
 212       nir_instr_remove(&call->instr);
 213    }
 214
 215    return progress;
 216 }
 217
 218 static bool
 219 inline_function_impl(nir_function_impl *impl, struct set *inlined)
 220 {
 221    if (_mesa_set_search(inlined, impl))
 222       return false; /* Already inlined */
 223
 224    nir_builder b;
 225    nir_builder_init(&b, impl);
 226
 227    bool progress = false;
 228    nir_foreach_block_safe(block, impl) {
 229       progress |= inline_functions_block(block, &b, inlined);
 230    }
 231
 232    if (progress) {
 233       /* SSA and register indices are completely messed up now */
 234       nir_index_ssa_defs(impl);
 235       nir_index_local_regs(impl);
 236
 237       nir_metadata_preserve(impl, nir_metadata_none);
 238    }
 239
 240    _mesa_set_add(inlined, impl);
 241
 242    return progress;
 243 }
 244
 245 bool
 246 nir_inline_functions(nir_shader *shader)
 247 {
 248    struct set *inlined = _mesa_set_create(NULL, _mesa_hash_pointer,
 249                                           _mesa_key_pointer_equal);
 250    bool progress = false;
 251
 252    nir_foreach_function(function, shader) {
 253       if (function->impl)
 254          progress = inline_function_impl(function->impl, inlined) || progress;
 255    }
 256
 257    _mesa_set_destroy(inlined, NULL);
 258
 259    return progress;
 260 }