From: Giovanni Mascellani gmascellani@codeweavers.com
--- include/vkd3d_types.h | 2 + libs/vkd3d-shader/ir.c | 143 ++++++++++++++++++++++++++++++++++++++++- 2 files changed, 143 insertions(+), 2 deletions(-)
diff --git a/include/vkd3d_types.h b/include/vkd3d_types.h index 4a7aca236..9060fa087 100644 --- a/include/vkd3d_types.h +++ b/include/vkd3d_types.h @@ -41,6 +41,8 @@ enum vkd3d_result { /** Success. */ VKD3D_OK = 0, + /** Success as a result of there being nothing to do. */ + VKD3D_FALSE = 1, /** An unspecified failure occurred. */ VKD3D_ERROR = -1, /** There are not enough resources available to complete the operation. */ diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 2e05d19ca..691632798 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3038,7 +3038,7 @@ static enum vkd3d_result vsir_block_list_add(struct vsir_block_list *list, struc
for (i = 0; i < list->count; ++i) if (block == list->blocks[i]) - return VKD3D_OK; + return VKD3D_FALSE;
if (!vkd3d_array_reserve((void **)&list->blocks, &list->capacity, list->count + 1, sizeof(*list->blocks))) { @@ -3051,6 +3051,13 @@ static enum vkd3d_result vsir_block_list_add(struct vsir_block_list *list, struc return VKD3D_OK; }
+static int block_compare(const void *ptr1, const void *ptr2); + +static void vsir_block_list_sort(struct vsir_block_list *list) +{ + qsort(list->blocks, list->count, sizeof(*list->blocks), block_compare); +} + struct vsir_block { unsigned int label; @@ -3063,6 +3070,14 @@ struct vsir_block uint32_t *dominates; };
+static int block_compare(const void *ptr1, const void *ptr2) +{ + const struct vsir_block *block1 = *(const struct vsir_block **)ptr1; + const struct vsir_block *block2 = *(const struct vsir_block **)ptr2; + + return vkd3d_u32_compare(block1->label, block2->label); +} + static enum vkd3d_result vsir_block_init(struct vsir_block *block, unsigned int label, size_t block_count) { size_t byte_count; @@ -3103,6 +3118,10 @@ struct vsir_cfg struct vsir_block *entry; size_t block_count; struct vkd3d_string_buffer debug_buffer; + + struct vsir_block_list *loops; + size_t loops_count, loops_capacity; + struct vsir_block_list **loops_by_header; };
static void vsir_cfg_cleanup(struct vsir_cfg *cfg) @@ -3112,12 +3131,22 @@ static void vsir_cfg_cleanup(struct vsir_cfg *cfg) for (i = 0; i < cfg->block_count; ++i) vsir_block_cleanup(&cfg->blocks[i]);
+ for (i = 0; i < cfg->loops_count; ++i) + vsir_block_list_cleanup(&cfg->loops[i]); + vkd3d_free(cfg->blocks); + vkd3d_free(cfg->loops); + vkd3d_free(cfg->loops_by_header);
if (TRACE_ON()) vkd3d_string_buffer_cleanup(&cfg->debug_buffer); }
+static bool vsir_block_dominates(struct vsir_block *b1, struct vsir_block *b2) +{ + return bitmap_is_set(b1->dominates, b2->label - 1); +} + static enum vkd3d_result vsir_cfg_add_edge(struct vsir_cfg *cfg, struct vsir_block *block, struct vkd3d_shader_src_param *successor_param) { @@ -3320,7 +3349,7 @@ static void vsir_cfg_compute_dominators(struct vsir_cfg *cfg) if (block2->label == 0) continue;
- if (bitmap_is_set(block->dominates, j)) + if (vsir_block_dominates(block, block2)) vkd3d_string_buffer_printf(&cfg->debug_buffer, " %u", block2->label); } TRACE("%s\n", cfg->debug_buffer.buffer); @@ -3329,6 +3358,110 @@ static void vsir_cfg_compute_dominators(struct vsir_cfg *cfg) } }
+/* A back edge is an edge X -> Y for which block Y dominates block + * X. All the other edges are forward edges, and it is required that + * the input CFG is reducible, i.e., it is acyclic once you strip away + * the back edges. + * + * Each back edge X -> Y defines a loop: block X is the header block, + * block Y is the back edge block, and the loop consists of all the + * blocks which are dominated by the header block and have a path to + * the back edge block that doesn't pass through the header block + * (including the header block itself). It can be proved that all the + * blocks in such a path (connecting a loop block to the back edge + * block without passing through the header block) belong to the same + * loop. + * + * If the input CFG is reducible, each two loops are either disjoint + * or one is a strict subset of the other, provided that each block + * has at most one incoming back edge. If this condition does not + * hold, a synthetic block can be introduced as the only back edge + * block for the given header block, with all the previous back edge + * now being forward edges to the synthetic block. This is not + * currently implemented (but it is rarely found in practice + * anyway). */ +static enum vkd3d_result vsir_cfg_scan_loop(struct vsir_block_list *loop, struct vsir_block *block, + struct vsir_block *header) +{ + enum vkd3d_result ret; + size_t i; + + if (!vsir_block_dominates(header, block)) + return VKD3D_OK; + + if ((ret = vsir_block_list_add(loop, block)) < 0) + return ret; + + if (ret == VKD3D_FALSE || block == header) + return VKD3D_OK; + + for (i = 0; i < block->predecessors.count; ++i) + { + if ((ret = vsir_cfg_scan_loop(loop, block->predecessors.blocks[i], header)) < 0) + return ret; + } + + return VKD3D_OK; +} + +static enum vkd3d_result vsir_cfg_compute_loops(struct vsir_cfg *cfg) +{ + size_t i, j, k; + + if (!(cfg->loops_by_header = vkd3d_calloc(cfg->block_count, sizeof(*cfg->loops_by_header)))) + 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) + continue; + + for (j = 0; j < block->successors.count; ++j) + { + struct vsir_block *header = block->successors.blocks[j]; + struct vsir_block_list *loop; + enum vkd3d_result ret; + + if (!vsir_block_dominates(header, block)) + continue; + + if (!vkd3d_array_reserve((void **)&cfg->loops, &cfg->loops_capacity, cfg->loops_count + 1, sizeof(*cfg->loops))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + loop = &cfg->loops[cfg->loops_count++]; + vsir_block_list_init(loop); + + if ((ret = vsir_cfg_scan_loop(loop, block, header)) < 0) + return ret; + + vsir_block_list_sort(loop); + + if (TRACE_ON()) + { + vkd3d_string_buffer_printf(&cfg->debug_buffer, "Back edge %u -> %u with loop:", block->label, header->label); + + for (k = 0; k < loop->count; ++k) + vkd3d_string_buffer_printf(&cfg->debug_buffer, " %u", loop->blocks[k]->label); + + TRACE("%s\n", cfg->debug_buffer.buffer); + vkd3d_string_buffer_clear(&cfg->debug_buffer); + } + + if (cfg->loops_by_header[header->label - 1]) + { + FIXME("Block %u is header to more than one loop, this is not implemented.", header->label); + return VKD3D_ERROR_NOT_IMPLEMENTED; + } + + cfg->loops_by_header[header->label - 1] = loop; + } + } + + return VKD3D_OK; +} + enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, const struct vkd3d_shader_compile_info *compile_info) { @@ -3355,6 +3488,12 @@ enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser,
vsir_cfg_compute_dominators(&cfg);
+ if ((result = vsir_cfg_compute_loops(&cfg)) < 0) + { + vsir_cfg_cleanup(&cfg); + return result; + } + if ((result = simple_structurizer_run(parser)) < 0) { vsir_cfg_cleanup(&cfg);