[PATCH v6 0/4] MR656: vkd3d-shader/ir: Introduce a better CFG structurizer, part 1.
This first MR create an internal structured representation of the CFG and computes the domination relationship. Over a few other MRs I will implement an algorithm inspired to [this article](https://medium.com/leaningtech/solving-the-structured-control-flow-problem-o...) to compute a better structure than the one we already have, but for the moment the computed data is immediately wasted. A working prototype of the new structurizer is now available in https://gitlab.winehq.org/giomasce/vkd3d/-/commits/cfg4, though further improvement is planned on top of that too. -- v6: vkd3d-shader/ir: Dump the domination relationship. vkd3d-shader/ir: Compute the domination relationship. vkd3d-shader/ir: Dump the control flow graph in the GraphViz format. https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
From: Giovanni Mascellani <gmascellani(a)codeweavers.com> --- libs/vkd3d-shader/ir.c | 198 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 198 insertions(+) diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 886344874..18d70c78b 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3009,6 +3009,194 @@ fail: return VKD3D_ERROR_OUT_OF_MEMORY; } +struct vsir_block_list +{ + struct vsir_block **blocks; + size_t count, capacity; +}; + +static void vsir_block_list_init(struct vsir_block_list *list) +{ + memset(list, 0, sizeof(*list)); +} + +static void vsir_block_list_cleanup(struct vsir_block_list *list) +{ + vkd3d_free(list->blocks); +} + +static enum vkd3d_result vsir_block_list_add(struct vsir_block_list *list, struct vsir_block *block) +{ + size_t i; + + for (i = 0; i < list->count; ++i) + if (block == list->blocks[i]) + return VKD3D_OK; + + if (!vkd3d_array_reserve((void **)&list->blocks, &list->capacity, list->count + 1, sizeof(*list->blocks))) + { + ERR("Cannot extend block list.\n"); + return VKD3D_ERROR_OUT_OF_MEMORY; + } + + list->blocks[list->count++] = block; + + return VKD3D_OK; +} + +struct vsir_block +{ + unsigned int label; + /* `begin' points to the instruction immediately following the + * LABEL that introduces the block. `end' points to the terminator + * instruction (either BRANCH or RET). They can coincide, meaning + * that the block is empty. */ + struct vkd3d_shader_instruction *begin, *end; + struct vsir_block_list predecessors, successors; +}; + +static void vsir_block_init(struct vsir_block *block, unsigned int label) +{ + assert(label); + memset(block, 0, sizeof(*block)); + block->label = label; + vsir_block_list_init(&block->predecessors); + vsir_block_list_init(&block->successors); +} + +static void vsir_block_cleanup(struct vsir_block *block) +{ + vsir_block_list_cleanup(&block->predecessors); + vsir_block_list_cleanup(&block->successors); +} + +struct vsir_cfg +{ + struct vkd3d_shader_parser *parser; + struct vsir_block *blocks; + struct vsir_block *entry; + size_t block_count; +}; + +static void vsir_cfg_cleanup(struct vsir_cfg *cfg) +{ + size_t i; + + for (i = 0; i < cfg->block_count; ++i) + vsir_block_cleanup(&cfg->blocks[i]); + + vkd3d_free(cfg->blocks); +} + +static enum vkd3d_result vsir_cfg_add_edge(struct vsir_cfg *cfg, struct vsir_block *block, + struct vkd3d_shader_src_param *successor_param) +{ + unsigned int target = label_from_src_param(successor_param); + struct vsir_block *successor = &cfg->blocks[target - 1]; + enum vkd3d_result ret; + + assert(successor->label != 0); + + if ((ret = vsir_block_list_add(&block->successors, successor)) < 0) + return ret; + + if ((ret = vsir_block_list_add(&successor->predecessors, block)) < 0) + return ret; + + return VKD3D_OK; +} + +static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vkd3d_shader_parser *parser) +{ + struct vsir_block *current_block = NULL; + enum vkd3d_result ret; + size_t i; + + memset(cfg, 0, sizeof(*cfg)); + cfg->parser = parser; + cfg->block_count = parser->program.block_count; + + if (!(cfg->blocks = vkd3d_calloc(cfg->block_count, sizeof(*cfg->blocks)))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + for (i = 0; i < parser->program.instructions.count; ++i) + { + struct vkd3d_shader_instruction *instruction = &parser->program.instructions.elements[i]; + + switch (instruction->handler_idx) + { + case VKD3DSIH_PHI: + case VKD3DSIH_SWITCH_MONOLITHIC: + vkd3d_unreachable(); + + case VKD3DSIH_LABEL: + { + unsigned int label = label_from_src_param(&instruction->src[0]); + + assert(!current_block); + assert(label > 0); + assert(label <= cfg->block_count); + current_block = &cfg->blocks[label - 1]; + vsir_block_init(current_block, label); + current_block->begin = &parser->program.instructions.elements[i + 1]; + if (!cfg->entry) + cfg->entry = current_block; + break; + } + + case VKD3DSIH_BRANCH: + case VKD3DSIH_RET: + assert(current_block); + current_block->end = instruction; + current_block = NULL; + break; + + default: + break; + } + } + + for (i = 0; i < cfg->block_count; ++i) + { + struct vsir_block *block = &cfg->blocks[i]; + + if (block->label == 0) + continue; + + switch (block->end->handler_idx) + { + case VKD3DSIH_RET: + break; + + case VKD3DSIH_BRANCH: + if (vsir_register_is_label(&block->end->src[0].reg)) + { + if ((ret = vsir_cfg_add_edge(cfg, block, &block->end->src[0])) < 0) + goto fail; + } + else + { + if ((ret = vsir_cfg_add_edge(cfg, block, &block->end->src[1])) < 0) + goto fail; + + if ((ret = vsir_cfg_add_edge(cfg, block, &block->end->src[2])) < 0) + goto fail; + } + break; + + default: + vkd3d_unreachable(); + } + } + + return VKD3D_OK; + +fail: + vsir_cfg_cleanup(cfg); + + return ret; +} + enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, const struct vkd3d_shader_compile_info *compile_info) { @@ -3022,14 +3210,24 @@ enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, if (parser->shader_desc.is_dxil) { + struct vsir_cfg cfg; + if ((result = lower_switch_to_if_ladder(&parser->program)) < 0) return result; if ((result = materialize_ssas_to_temps(parser)) < 0) return result; + if ((result = vsir_cfg_init(&cfg, parser)) < 0) + return result; + if ((result = simple_structurizer_run(parser)) < 0) + { + vsir_cfg_cleanup(&cfg); return result; + } + + vsir_cfg_cleanup(&cfg); } else { -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
From: Giovanni Mascellani <gmascellani(a)codeweavers.com> --- libs/vkd3d-shader/ir.c | 40 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 40 insertions(+) diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 18d70c78b..6fd1f53f1 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3106,6 +3106,43 @@ static enum vkd3d_result vsir_cfg_add_edge(struct vsir_cfg *cfg, struct vsir_blo return VKD3D_OK; } +static void vsir_cfg_dump_dot(struct vsir_cfg *cfg) +{ + size_t i, j; + + TRACE("digraph cfg {\n"); + + for (i = 0; i < cfg->block_count; ++i) + { + struct vsir_block *block = &cfg->blocks[i]; + const char *shape; + + if (block->label == 0) + continue; + + switch (block->end->handler_idx) + { + case VKD3DSIH_RET: + shape = "trapezium"; + break; + + case VKD3DSIH_BRANCH: + shape = vsir_register_is_label(&block->end->src[0].reg) ? "ellipse" : "box"; + break; + + default: + vkd3d_unreachable(); + } + + TRACE(" n%u [label=\"%u\", shape=\"%s\"];\n", block->label, block->label, shape); + + for (j = 0; j < block->successors.count; ++j) + TRACE(" n%u -> n%u;\n", block->label, block->successors.blocks[j]->label); + } + + TRACE("}\n"); +} + static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vkd3d_shader_parser *parser) { struct vsir_block *current_block = NULL; @@ -3189,6 +3226,9 @@ static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vkd3d_shader } } + if (TRACE_ON()) + vsir_cfg_dump_dot(cfg); + return VKD3D_OK; fail: -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
From: Giovanni Mascellani <gmascellani(a)codeweavers.com> --- libs/vkd3d-shader/ir.c | 68 ++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 66 insertions(+), 2 deletions(-) diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 6fd1f53f1..ee633165d 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3053,21 +3053,40 @@ struct vsir_block * that the block is empty. */ struct vkd3d_shader_instruction *begin, *end; struct vsir_block_list predecessors, successors; + uint32_t *dominates; }; -static void vsir_block_init(struct vsir_block *block, unsigned int label) +static enum vkd3d_result vsir_block_init(struct vsir_block *block, unsigned int label, size_t block_count) { + size_t byte_count; + + if (block_count > SIZE_MAX - (sizeof(*block->dominates) * CHAR_BIT - 1)) + return VKD3D_ERROR_OUT_OF_MEMORY; + + block_count = align(block_count, sizeof(*block->dominates) * CHAR_BIT); + byte_count = block_count / CHAR_BIT; + assert(label); memset(block, 0, sizeof(*block)); block->label = label; vsir_block_list_init(&block->predecessors); vsir_block_list_init(&block->successors); + + if (!(block->dominates = vkd3d_malloc(byte_count))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + memset(block->dominates, 0xff, byte_count); + + return VKD3D_OK; } static void vsir_block_cleanup(struct vsir_block *block) { + if (block->label == 0) + return; vsir_block_list_cleanup(&block->predecessors); vsir_block_list_cleanup(&block->successors); + vkd3d_free(block->dominates); } struct vsir_cfg @@ -3174,7 +3193,9 @@ static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vkd3d_shader assert(label > 0); assert(label <= cfg->block_count); current_block = &cfg->blocks[label - 1]; - vsir_block_init(current_block, label); + assert(current_block->label == 0); + if ((ret = vsir_block_init(current_block, label, parser->program.block_count)) < 0) + goto fail; current_block->begin = &parser->program.instructions.elements[i + 1]; if (!cfg->entry) cfg->entry = current_block; @@ -3237,6 +3258,47 @@ fail: return ret; } +/* Block A dominates block B if every path from the entry point to B + * must pass through A. Naively computer the set of blocks that are + * dominated by `reference' by running a graph visit starting from the + * entry point (which must be the initial value of `current') and + * avoiding `reference'. Running this for all the blocks takes + * quadratic time: if in the future something better is sought after, + * the standard tool seems to be the Lengauer-Tarjan algorithm. */ +static void vsir_cfg_compute_dominators_recurse(uint32_t *dominates, struct vsir_block *current, + struct vsir_block *reference) +{ + size_t i; + + assert(current->label != 0); + + if (current == reference) + return; + + if (!bitmap_is_set(dominates, current->label - 1)) + return; + + bitmap_clear(dominates, current->label - 1); + + for (i = 0; i < current->successors.count; ++i) + vsir_cfg_compute_dominators_recurse(dominates, current->successors.blocks[i], reference); +} + +static void vsir_cfg_compute_dominators(struct vsir_cfg *cfg) +{ + size_t i; + + for (i = 0; i < cfg->block_count; ++i) + { + struct vsir_block *block = &cfg->blocks[i]; + + if (block->label == 0) + continue; + + vsir_cfg_compute_dominators_recurse(block->dominates, cfg->entry, block); + } +} + enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, const struct vkd3d_shader_compile_info *compile_info) { @@ -3261,6 +3323,8 @@ enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, if ((result = vsir_cfg_init(&cfg, parser)) < 0) return result; + vsir_cfg_compute_dominators(&cfg); + if ((result = simple_structurizer_run(parser)) < 0) { vsir_cfg_cleanup(&cfg); -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
From: Giovanni Mascellani <gmascellani(a)codeweavers.com> --- libs/vkd3d-shader/ir.c | 26 +++++++++++++++++++++++- libs/vkd3d-shader/vkd3d_shader_main.c | 2 +- libs/vkd3d-shader/vkd3d_shader_private.h | 1 + 3 files changed, 27 insertions(+), 2 deletions(-) diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index ee633165d..837418c4f 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3286,7 +3286,11 @@ static void vsir_cfg_compute_dominators_recurse(uint32_t *dominates, struct vsir static void vsir_cfg_compute_dominators(struct vsir_cfg *cfg) { - size_t i; + struct vkd3d_string_buffer buf; + size_t i, j; + + if (TRACE_ON()) + vkd3d_string_buffer_init(&buf); for (i = 0; i < cfg->block_count; ++i) { @@ -3296,7 +3300,27 @@ static void vsir_cfg_compute_dominators(struct vsir_cfg *cfg) continue; vsir_cfg_compute_dominators_recurse(block->dominates, cfg->entry, block); + + if (TRACE_ON()) + { + vkd3d_string_buffer_printf(&buf, "Block %u dominates:", block->label); + for (j = 0; j < cfg->block_count; j++) + { + struct vsir_block *block2 = &cfg->blocks[j]; + + if (block2->label == 0) + continue; + + if (bitmap_is_set(block->dominates, j)) + vkd3d_string_buffer_printf(&buf, " %u", block2->label); + } + TRACE("%s\n", buf.buffer); + vkd3d_string_buffer_clear(&buf); + } } + + if (TRACE_ON()) + vkd3d_string_buffer_cleanup(&buf); } enum vkd3d_result vkd3d_shader_normalise(struct vkd3d_shader_parser *parser, diff --git a/libs/vkd3d-shader/vkd3d_shader_main.c b/libs/vkd3d-shader/vkd3d_shader_main.c index 20c4c2e2d..462a5c25e 100644 --- a/libs/vkd3d-shader/vkd3d_shader_main.c +++ b/libs/vkd3d-shader/vkd3d_shader_main.c @@ -71,7 +71,7 @@ void vkd3d_string_buffer_cleanup(struct vkd3d_string_buffer *buffer) vkd3d_free(buffer->buffer); } -static void vkd3d_string_buffer_clear(struct vkd3d_string_buffer *buffer) +void vkd3d_string_buffer_clear(struct vkd3d_string_buffer *buffer) { buffer->buffer[0] = '\0'; buffer->content_size = 0; diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h index 7239beaf7..4b322b95b 100644 --- a/libs/vkd3d-shader/vkd3d_shader_private.h +++ b/libs/vkd3d-shader/vkd3d_shader_private.h @@ -1399,6 +1399,7 @@ struct vkd3d_string_buffer *vkd3d_string_buffer_get(struct vkd3d_string_buffer_c void vkd3d_string_buffer_init(struct vkd3d_string_buffer *buffer); void vkd3d_string_buffer_cache_cleanup(struct vkd3d_string_buffer_cache *list); void vkd3d_string_buffer_cache_init(struct vkd3d_string_buffer_cache *list); +void vkd3d_string_buffer_clear(struct vkd3d_string_buffer *buffer); int vkd3d_string_buffer_print_f32(struct vkd3d_string_buffer *buffer, float f); int vkd3d_string_buffer_print_f64(struct vkd3d_string_buffer *buffer, double d); int vkd3d_string_buffer_printf(struct vkd3d_string_buffer *buffer, const char *format, ...) VKD3D_PRINTF_FUNC(2, 3); -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
On Wed Feb 21 06:12:57 2024 +0000, Conor McCarthy wrote:
Is `block` ever `null` here? Looks like it shouldn't be. Yeah, I thought I had fixed all of them but this one slipped through. The correct test is to check whether the block as label 0 (meaning that it doesn't exist).
-- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656#note_62166
On Wed Feb 21 08:28:01 2024 +0000, Giovanni Mascellani wrote:
changed this line in [version 6 of the diff](/wine/vkd3d/-/merge_requests/656/diffs?diff_id=100878&start_sha=a17ec61426e50a37f0ebac67d07160dcc7309987#4fae87cf1e8d06a5c7871f244e3dd75ef2dac299_3063_3063) I ended up with a solution which I think makes the code more readable. It relies on `size_t` being no smaller than `unsigned int`, but I think we already assume this more or less everywhere.
-- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656#note_62167
This merge request was approved by Conor McCarthy. -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656
```diff +static enum vkd3d_result vsir_cfg_init(struct vsir_cfg *cfg, struct vkd3d_shader_parser *parser) +{ + struct vsir_block *current_block = NULL; + enum vkd3d_result ret; + size_t i; + + memset(cfg, 0, sizeof(*cfg)); + cfg->parser = parser; + cfg->block_count = parser->program.block_count; ```
Do we need "parser"? It looks like we only use "parser->program".
```diff + * must pass through A. Naively computer the set of blocks that are ```
"compute" -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656#note_62198
Conor McCarthy (@cmccarthy) commented about libs/vkd3d-shader/ir.c:
return ret; }
+/* Block A dominates block B if every path from the entry point to B + * must pass through A. Naively computer the set of blocks that are + * dominated by `reference' by running a graph visit starting from the + * entry point (which must be the initial value of `current') and + * avoiding `reference'. Running this for all the blocks takes + * quadratic time: if in the future something better is sought after, + * the standard tool seems to be the Lengauer-Tarjan algorithm. */ +static void vsir_cfg_compute_dominators_recurse(uint32_t *dominates, struct vsir_block *current, + struct vsir_block *reference)
Any particular reason for passing `dominates` to this function? Looks like it can be accessed from `reference` on each recursion. -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/656#note_62202
participants (4)
-
Conor McCarthy (＠cmccarthy) -
Giovanni Mascellani
-
Giovanni Mascellani (＠giomasce)
-
Henri Verbeet (＠hverbeet)