From: Giovanni Mascellani gmascellani@codeweavers.com
--- libs/vkd3d-shader/ir.c | 135 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 135 insertions(+)
diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 06388c3ee..63242539e 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3056,6 +3056,13 @@ static enum vkd3d_result vsir_block_list_add(struct vsir_block_list *list, struc return vsir_block_list_add_unchecked(list, block); }
+/* It is guaranteed that the relative order is kept. */ +static void vsir_block_list_remove_index(struct vsir_block_list *list, size_t idx) +{ + memmove(&list->blocks[idx], &list->blocks[idx + 1], (list->count - idx - 1) * sizeof(*list->blocks)); + --list->count; +} + static int block_compare(const void *ptr1, const void *ptr2);
static void vsir_block_list_sort(struct vsir_block_list *list) @@ -3127,6 +3134,8 @@ struct vsir_cfg struct vsir_block_list *loops; size_t loops_count, loops_capacity; struct vsir_block_list **loops_by_header; + + struct vsir_block_list order; };
static void vsir_cfg_cleanup(struct vsir_cfg *cfg) @@ -3139,6 +3148,8 @@ static void vsir_cfg_cleanup(struct vsir_cfg *cfg) for (i = 0; i < cfg->loops_count; ++i) vsir_block_list_cleanup(&cfg->loops[i]);
+ vsir_block_list_cleanup(&cfg->order); + vkd3d_free(cfg->blocks); vkd3d_free(cfg->loops); vkd3d_free(cfg->loops_by_header); @@ -3225,6 +3236,8 @@ static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vsir_program cfg->program = program; cfg->block_count = program->block_count;
+ vsir_block_list_init(&cfg->order); + if (!(cfg->blocks = vkd3d_calloc(cfg->block_count, sizeof(*cfg->blocks)))) return VKD3D_ERROR_OUT_OF_MEMORY;
@@ -3475,6 +3488,122 @@ static enum vkd3d_result vsir_cfg_compute_loops(struct vsir_cfg *cfg) return VKD3D_OK; }
+struct vsir_cfg_node_sorter +{ + struct vsir_cfg *cfg; + struct vsir_block_list available_blocks; +}; + +static enum vkd3d_result vsir_cfg_make_node_available(struct vsir_cfg_node_sorter *sorter, struct vsir_block *block) +{ + enum vkd3d_result ret; + + if ((ret = vsir_block_list_add_unchecked(&sorter->available_blocks, block)) < 0) + return ret; + + return VKD3D_OK; +} + +/* Topologically sort the blocks according to the forward edges. By + * definition if the input CFG is reducible then its forward edges + * form a DAG, so a topological sorting exists. In order to compute it + * we keep an array with the incoming degree for each block and an + * available list of all the blocks whose incoming degree has reached + * zero. At each step we pick a block from the available list and + * strip it away from the graph, updating the incoming degrees and + * available list. We prefer picking the most recently added block + * (i.e., from the end of the available list) because it will keep + * together related blocks in the order, which allow us to generate + * fewer control flow primitives. */ +static enum vkd3d_result vsir_cfg_sort_nodes(struct vsir_cfg *cfg) +{ + enum vkd3d_result ret; + struct vsir_cfg_node_sorter sorter = { .cfg = cfg }; + unsigned int *in_degrees = NULL; + size_t i; + + if (!(in_degrees = vkd3d_calloc(cfg->block_count, sizeof(*in_degrees)))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + for (i = 0; i < cfg->block_count; ++i) + { + struct vsir_block *block = &cfg->blocks[i]; + + if (block->label == 0) + { + in_degrees[i] = UINT_MAX; + continue; + } + + in_degrees[i] = block->predecessors.count; + + /* Do not count back edges. */ + if (cfg->loops_by_header[i]) + { + assert(in_degrees[i] > 0); + in_degrees[i] -= 1; + } + + if (in_degrees[i] == 0 && block != cfg->entry) + { + ERR("Unexpected entry point %u.\n", block->label); + ret = VKD3D_ERROR_INVALID_SHADER; + goto fail; + } + } + + if (in_degrees[cfg->entry->label - 1] != 0) + { + ERR("Entry point has %u incoming forward edges.\n", in_degrees[cfg->entry->label - 1]); + ret = VKD3D_ERROR_INVALID_SHADER; + goto fail; + } + + vsir_block_list_init(&sorter.available_blocks); + + if ((ret = vsir_cfg_make_node_available(&sorter, cfg->entry)) < 0) + goto fail; + + while (sorter.available_blocks.count != 0) + { + struct vsir_block *block = sorter.available_blocks.blocks[sorter.available_blocks.count - 1]; + + vsir_block_list_remove_index(&sorter.available_blocks, sorter.available_blocks.count - 1); + if ((ret = vsir_block_list_add_unchecked(&cfg->order, block)) < 0) + goto fail; + + /* Remove (forward) edges and make new nodes available. */ + for (i = 0; i < block->successors.count; ++i) + { + struct vsir_block *successor = block->successors.blocks[i]; + + if (vsir_block_is_back_edge(block, successor)) + continue; + + assert(in_degrees[successor->label - 1] > 0); + --in_degrees[successor->label - 1]; + + if (in_degrees[successor->label - 1] == 0) + { + if ((ret = vsir_cfg_make_node_available(&sorter, successor)) < 0) + goto fail; + } + } + } + + vkd3d_free(in_degrees); + vsir_block_list_cleanup(&sorter.available_blocks); + + return VKD3D_OK; + +fail: + vkd3d_free(in_degrees); + vsir_block_list_cleanup(&sorter.available_blocks); + vsir_block_list_cleanup(&cfg->order); + + return ret; +} + enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, const struct vkd3d_shader_compile_info *compile_info) { @@ -3507,6 +3636,12 @@ enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, return result; }
+ if ((result = vsir_cfg_sort_nodes(&cfg)) < 0) + { + vsir_cfg_cleanup(&cfg); + return result; + } + if ((result = simple_structurizer_run(parser)) < 0) { vsir_cfg_cleanup(&cfg);