101e04c3fSmrg/*
201e04c3fSmrg * Copyright © 2015 Intel Corporation
301e04c3fSmrg *
401e04c3fSmrg * Permission is hereby granted, free of charge, to any person obtaining a
501e04c3fSmrg * copy of this software and associated documentation files (the "Software"),
601e04c3fSmrg * to deal in the Software without restriction, including without limitation
701e04c3fSmrg * the rights to use, copy, modify, merge, publish, distribute, sublicense,
801e04c3fSmrg * and/or sell copies of the Software, and to permit persons to whom the
901e04c3fSmrg * Software is furnished to do so, subject to the following conditions:
1001e04c3fSmrg *
1101e04c3fSmrg * The above copyright notice and this permission notice (including the next
1201e04c3fSmrg * paragraph) shall be included in all copies or substantial portions of the
1301e04c3fSmrg * Software.
1401e04c3fSmrg *
1501e04c3fSmrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
1601e04c3fSmrg * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
1701e04c3fSmrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
1801e04c3fSmrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
1901e04c3fSmrg * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
2001e04c3fSmrg * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
2101e04c3fSmrg * IN THE SOFTWARE.
2201e04c3fSmrg */
2301e04c3fSmrg
2401e04c3fSmrg#include "nir.h"
2501e04c3fSmrg
2601e04c3fSmrg/**
2701e04c3fSmrg * \file nir_sweep.c
2801e04c3fSmrg *
2901e04c3fSmrg * The nir_sweep() pass performs a mark and sweep pass over a nir_shader's associated
3001e04c3fSmrg * memory - anything still connected to the program will be kept, and any dead memory
3101e04c3fSmrg * we dropped on the floor will be freed.
3201e04c3fSmrg *
3301e04c3fSmrg * The expectation is that drivers should call this when finished compiling the shader
3401e04c3fSmrg * (after any optimization, lowering, and so on).  However, it's also fine to call it
3501e04c3fSmrg * earlier, and even many times, trading CPU cycles for memory savings.
3601e04c3fSmrg */
3701e04c3fSmrg
3801e04c3fSmrg#define steal_list(mem_ctx, type, list) \
3901e04c3fSmrg   foreach_list_typed(type, obj, node, list) { ralloc_steal(mem_ctx, obj); }
4001e04c3fSmrg
4101e04c3fSmrgstatic void sweep_cf_node(nir_shader *nir, nir_cf_node *cf_node);
4201e04c3fSmrg
4301e04c3fSmrgstatic void
4401e04c3fSmrgsweep_block(nir_shader *nir, nir_block *block)
4501e04c3fSmrg{
4601e04c3fSmrg   ralloc_steal(nir, block);
4701e04c3fSmrg
487e102996Smaya   /* sweep_impl will mark all metadata invalid.  We can safely release all of
497e102996Smaya    * this here.
507e102996Smaya    */
517e102996Smaya   ralloc_free(block->live_in);
527e102996Smaya   block->live_in = NULL;
537e102996Smaya
547e102996Smaya   ralloc_free(block->live_out);
557e102996Smaya   block->live_out = NULL;
567e102996Smaya
5701e04c3fSmrg   nir_foreach_instr(instr, block) {
587ec681f3Smrg      list_del(&instr->gc_node);
597ec681f3Smrg      list_add(&instr->gc_node, &nir->gc_list);
6001e04c3fSmrg   }
6101e04c3fSmrg}
6201e04c3fSmrg
6301e04c3fSmrgstatic void
6401e04c3fSmrgsweep_if(nir_shader *nir, nir_if *iff)
6501e04c3fSmrg{
6601e04c3fSmrg   ralloc_steal(nir, iff);
6701e04c3fSmrg
6801e04c3fSmrg   foreach_list_typed(nir_cf_node, cf_node, node, &iff->then_list) {
6901e04c3fSmrg      sweep_cf_node(nir, cf_node);
7001e04c3fSmrg   }
7101e04c3fSmrg
7201e04c3fSmrg   foreach_list_typed(nir_cf_node, cf_node, node, &iff->else_list) {
7301e04c3fSmrg      sweep_cf_node(nir, cf_node);
7401e04c3fSmrg   }
7501e04c3fSmrg}
7601e04c3fSmrg
7701e04c3fSmrgstatic void
7801e04c3fSmrgsweep_loop(nir_shader *nir, nir_loop *loop)
7901e04c3fSmrg{
8001e04c3fSmrg   ralloc_steal(nir, loop);
8101e04c3fSmrg
8201e04c3fSmrg   foreach_list_typed(nir_cf_node, cf_node, node, &loop->body) {
8301e04c3fSmrg      sweep_cf_node(nir, cf_node);
8401e04c3fSmrg   }
8501e04c3fSmrg}
8601e04c3fSmrg
8701e04c3fSmrgstatic void
8801e04c3fSmrgsweep_cf_node(nir_shader *nir, nir_cf_node *cf_node)
8901e04c3fSmrg{
9001e04c3fSmrg   switch (cf_node->type) {
9101e04c3fSmrg   case nir_cf_node_block:
9201e04c3fSmrg      sweep_block(nir, nir_cf_node_as_block(cf_node));
9301e04c3fSmrg      break;
9401e04c3fSmrg   case nir_cf_node_if:
9501e04c3fSmrg      sweep_if(nir, nir_cf_node_as_if(cf_node));
9601e04c3fSmrg      break;
9701e04c3fSmrg   case nir_cf_node_loop:
9801e04c3fSmrg      sweep_loop(nir, nir_cf_node_as_loop(cf_node));
9901e04c3fSmrg      break;
10001e04c3fSmrg   default:
10101e04c3fSmrg      unreachable("Invalid CF node type");
10201e04c3fSmrg   }
10301e04c3fSmrg}
10401e04c3fSmrg
10501e04c3fSmrgstatic void
10601e04c3fSmrgsweep_impl(nir_shader *nir, nir_function_impl *impl)
10701e04c3fSmrg{
10801e04c3fSmrg   ralloc_steal(nir, impl);
10901e04c3fSmrg
11001e04c3fSmrg   steal_list(nir, nir_variable, &impl->locals);
11101e04c3fSmrg   steal_list(nir, nir_register, &impl->registers);
11201e04c3fSmrg
11301e04c3fSmrg   foreach_list_typed(nir_cf_node, cf_node, node, &impl->body) {
11401e04c3fSmrg      sweep_cf_node(nir, cf_node);
11501e04c3fSmrg   }
11601e04c3fSmrg
11701e04c3fSmrg   sweep_block(nir, impl->end_block);
11801e04c3fSmrg
11901e04c3fSmrg   /* Wipe out all the metadata, if any. */
12001e04c3fSmrg   nir_metadata_preserve(impl, nir_metadata_none);
12101e04c3fSmrg}
12201e04c3fSmrg
12301e04c3fSmrgstatic void
12401e04c3fSmrgsweep_function(nir_shader *nir, nir_function *f)
12501e04c3fSmrg{
12601e04c3fSmrg   ralloc_steal(nir, f);
12701e04c3fSmrg   ralloc_steal(nir, f->params);
12801e04c3fSmrg
12901e04c3fSmrg   if (f->impl)
13001e04c3fSmrg      sweep_impl(nir, f->impl);
13101e04c3fSmrg}
13201e04c3fSmrg
13301e04c3fSmrgvoid
13401e04c3fSmrgnir_sweep(nir_shader *nir)
13501e04c3fSmrg{
13601e04c3fSmrg   void *rubbish = ralloc_context(NULL);
13701e04c3fSmrg
1387ec681f3Smrg   struct list_head instr_gc_list;
1397ec681f3Smrg   list_inithead(&instr_gc_list);
1407ec681f3Smrg
1417ec681f3Smrg   list_replace(&nir->gc_list, &instr_gc_list);
1427ec681f3Smrg   list_inithead(&nir->gc_list);
1437ec681f3Smrg
14401e04c3fSmrg   /* First, move ownership of all the memory to a temporary context; assume dead. */
14501e04c3fSmrg   ralloc_adopt(rubbish, nir);
14601e04c3fSmrg
14701e04c3fSmrg   ralloc_steal(nir, (char *)nir->info.name);
14801e04c3fSmrg   if (nir->info.label)
14901e04c3fSmrg      ralloc_steal(nir, (char *)nir->info.label);
15001e04c3fSmrg
15101e04c3fSmrg   /* Variables and registers are not dead.  Steal them back. */
1527ec681f3Smrg   steal_list(nir, nir_variable, &nir->variables);
15301e04c3fSmrg
15401e04c3fSmrg   /* Recurse into functions, stealing their contents back. */
15501e04c3fSmrg   foreach_list_typed(nir_function, func, node, &nir->functions) {
15601e04c3fSmrg      sweep_function(nir, func);
15701e04c3fSmrg   }
15801e04c3fSmrg
1597ec681f3Smrg   /* Sweep instrs not found while walking the shader. */
1607ec681f3Smrg   list_for_each_entry_safe(nir_instr, instr, &instr_gc_list, gc_node) {
1617ec681f3Smrg      nir_instr_free(instr);
1627ec681f3Smrg   }
1637ec681f3Smrg   assert(list_is_empty(&instr_gc_list));
1647ec681f3Smrg
16501e04c3fSmrg   ralloc_steal(nir, nir->constant_data);
16601e04c3fSmrg
16701e04c3fSmrg   /* Free everything we didn't steal back. */
16801e04c3fSmrg   ralloc_free(rubbish);
16901e04c3fSmrg}
170