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 #include "nir_vla.h"
  28
  29 void nir_inline_function_impl(struct nir_builder *b,
  30                               const nir_function_impl *impl,
  31                               nir_ssa_def **params,
  32                               struct hash_table *shader_var_remap)
  33 {
  34    nir_function_impl *copy = nir_function_impl_clone(b->shader, impl);
  35
  36    /* Insert a nop at the cursor so we can keep track of where things are as
  37     * we add/remove stuff from the CFG.
  38     */
  39    nir_intrinsic_instr *nop =
  40       nir_intrinsic_instr_create(b->shader, nir_intrinsic_nop);
  41    nir_builder_instr_insert(b, &nop->instr);
  42
  43    exec_list_append(&b->impl->locals, &copy->locals);
  44    exec_list_append(&b->impl->registers, &copy->registers);
  45
  46    nir_foreach_block(block, copy) {
  47       nir_foreach_instr_safe(instr, block) {
  48          switch (instr->type) {
  49          case nir_instr_type_deref: {
  50             nir_deref_instr *deref = nir_instr_as_deref(instr);
  51             if (deref->deref_type != nir_deref_type_var)
  52                break;
  53
  54             /* We don't need to remap function variables.  We already cloned
  55              * them as part of nir_function_impl_clone and appended them to
  56              * b->impl->locals.
  57              */
  58             if (deref->var->data.mode == nir_var_function_temp)
  59                break;
  60
  61             /* If no map is provided, we assume that there are either no
  62              * shader variables or they already live b->shader (this is the
  63              * case for function inlining within a single shader.
  64              */
  65             if (shader_var_remap == NULL)
  66                break;
  67
  68             struct hash_entry *entry =
  69                _mesa_hash_table_search(shader_var_remap, deref->var);
  70             if (entry == NULL) {
  71                nir_variable *nvar = nir_variable_clone(deref->var, b->shader);
  72                nir_shader_add_variable(b->shader, nvar);
  73                entry = _mesa_hash_table_insert(shader_var_remap,
  74                                                deref->var, nvar);
  75             }
  76             deref->var = entry->data;
  77             break;
  78          }
  79
  80          case nir_instr_type_intrinsic: {
  81             nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
  82             if (load->intrinsic != nir_intrinsic_load_param)
  83                break;
  84
  85             unsigned param_idx = nir_intrinsic_param_idx(load);
  86             assert(param_idx < impl->function->num_params);
  87             assert(load->dest.is_ssa);
  88             nir_ssa_def_rewrite_uses(&load->dest.ssa,
  89                                      nir_src_for_ssa(params[param_idx]));
  90
  91             /* Remove any left-over load_param intrinsics because they're soon
  92              * to be in another function and therefore no longer valid.
  93              */
  94             nir_instr_remove(&load->instr);
  95             break;
  96          }
  97
  98          case nir_instr_type_jump:
  99             /* Returns have to be lowered for this to work */
 100             assert(nir_instr_as_jump(instr)->type != nir_jump_return);
 101             break;
 102
 103          default:
 104             break;
 105          }
 106       }
 107    }
 108
 109    /* Pluck the body out of the function and place it here */
 110    nir_cf_list body;
 111    nir_cf_list_extract(&body, &copy->body);
 112    nir_cf_reinsert(&body, nir_before_instr(&nop->instr));
 113
 114    b->cursor = nir_instr_remove(&nop->instr);
 115 }
 116
 117 static bool inline_function_impl(nir_function_impl *impl, struct set *inlined);
 118
 119 static bool
 120 inline_functions_block(nir_block *block, nir_builder *b,
 121                        struct set *inlined)
 122 {
 123    bool progress = false;
 124    /* This is tricky.  We're iterating over instructions in a block but, as
 125     * we go, the block and its instruction list are being split into
 126     * pieces.  However, this *should* be safe since foreach_safe always
 127     * stashes the next thing in the iteration.  That next thing will
 128     * properly get moved to the next block when it gets split, and we
 129     * continue iterating there.
 130     */
 131    nir_foreach_instr_safe(instr, block) {
 132       if (instr->type != nir_instr_type_call)
 133          continue;
 134
 135       progress = true;
 136
 137       nir_call_instr *call = nir_instr_as_call(instr);
 138       assert(call->callee->impl);
 139
 140       /* Make sure that the function we're calling is already inlined */
 141       inline_function_impl(call->callee->impl, inlined);
 142
 143       b->cursor = nir_instr_remove(&call->instr);
 144
 145       /* Rewrite all of the uses of the callee's parameters to use the call
 146        * instructions sources.  In order to ensure that the "load" happens
 147        * here and not later (for register sources), we make sure to convert it
 148        * to an SSA value first.
 149        */
 150       const unsigned num_params = call->num_params;
 151       NIR_VLA(nir_ssa_def *, params, num_params);
 152       for (unsigned i = 0; i < num_params; i++) {
 153          params[i] = nir_ssa_for_src(b, call->params[i],
 154                                      call->callee->params[i].num_components);
 155       }
 156
 157       nir_inline_function_impl(b, call->callee->impl, params, NULL);
 158    }
 159
 160    return progress;
 161 }
 162
 163 static bool
 164 inline_function_impl(nir_function_impl *impl, struct set *inlined)
 165 {
 166    if (_mesa_set_search(inlined, impl))
 167       return false; /* Already inlined */
 168
 169    nir_builder b;
 170    nir_builder_init(&b, impl);
 171
 172    bool progress = false;
 173    nir_foreach_block_safe(block, impl) {
 174       progress |= inline_functions_block(block, &b, inlined);
 175    }
 176
 177    if (progress) {
 178       /* SSA and register indices are completely messed up now */
 179       nir_index_ssa_defs(impl);
 180       nir_index_local_regs(impl);
 181
 182       nir_metadata_preserve(impl, nir_metadata_none);
 183    } else {
 184       nir_metadata_preserve(impl, nir_metadata_all);
 185    }
 186
 187    _mesa_set_add(inlined, impl);
 188
 189    return progress;
 190 }
 191
 192 /** A pass to inline all functions in a shader into their callers
 193  *
 194  * For most use-cases, function inlining is a multi-step process.  The general
 195  * pattern employed by SPIR-V consumers and others is as follows:
 196  *
 197  *  1. nir_lower_variable_initializers(shader, nir_var_function_temp)
 198  *
 199  *     This is needed because local variables from the callee are simply added
 200  *     to the locals list for the caller and the information about where the
 201  *     constant initializer logically happens is lost.  If the callee is
 202  *     called in a loop, this can cause the variable to go from being
 203  *     initialized once per loop iteration to being initialized once at the
 204  *     top of the caller and values to persist from one invocation of the
 205  *     callee to the next.  The simple solution to this problem is to get rid
 206  *     of constant initializers before function inlining.
 207  *
 208  *  2. nir_lower_returns(shader)
 209  *
 210  *     nir_inline_functions assumes that all functions end "naturally" by
 211  *     execution reaching the end of the function without any return
 212  *     instructions causing instant jumps to the end.  Thanks to NIR being
 213  *     structured, we can't represent arbitrary jumps to various points in the
 214  *     program which is what an early return in the callee would have to turn
 215  *     into when we inline it into the caller.  Instead, we require returns to
 216  *     be lowered which lets us just copy+paste the callee directly into the
 217  *     caller.
 218  *
 219  *  3. nir_inline_functions(shader)
 220  *
 221  *     This does the actual function inlining and the resulting shader will
 222  *     contain no call instructions.
 223  *
 224  *  4. nir_opt_deref(shader)
 225  *
 226  *     Most functions contain pointer parameters where the result of a deref
 227  *     instruction is passed in as a parameter, loaded via a load_param
 228  *     intrinsic, and then turned back into a deref via a cast.  Function
 229  *     inlining will get rid of the load_param but we are still left with a
 230  *     cast.  Running nir_opt_deref gets rid of the intermediate cast and
 231  *     results in a whole deref chain again.  This is currently required by a
 232  *     number of optimizations and lowering passes at least for certain
 233  *     variable modes.
 234  *
 235  *  5. Loop over the functions and delete all but the main entrypoint.
 236  *
 237  *     In the Intel Vulkan driver this looks like this:
 238  *
 239  *        foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
 240  *           if (func != entry_point)
 241  *              exec_node_remove(&func->node);
 242  *        }
 243  *        assert(exec_list_length(&nir->functions) == 1);
 244  *
 245  *    While nir_inline_functions does get rid of all call instructions, it
 246  *    doesn't get rid of any functions because it doesn't know what the "root
 247  *    function" is.  Instead, it's up to the individual driver to know how to
 248  *    decide on a root function and delete the rest.  With SPIR-V,
 249  *    spirv_to_nir returns the root function and so we can just use == whereas
 250  *    with GL, you may have to look for a function named "main".
 251  *
 252  *  6. nir_lower_variable_initializers(shader, ~nir_var_function_temp)
 253  *
 254  *     Lowering constant initializers on inputs, outputs, global variables,
 255  *     etc. requires that we know the main entrypoint so that we know where to
 256  *     initialize them.  Otherwise, we would have to assume that anything
 257  *     could be a main entrypoint and initialize them at the start of every
 258  *     function but that would clearly be wrong if any of those functions were
 259  *     ever called within another function.  Simply requiring a single-
 260  *     entrypoint function shader is the best way to make it well-defined.
 261  */
 262 bool
 263 nir_inline_functions(nir_shader *shader)
 264 {
 265    struct set *inlined = _mesa_pointer_set_create(NULL);
 266    bool progress = false;
 267
 268    nir_foreach_function(function, shader) {
 269       if (function->impl)
 270          progress = inline_function_impl(function->impl, inlined) || progress;
 271    }
 272
 273    _mesa_set_destroy(inlined, NULL);
 274
 275    return progress;
 276 }