Signed-off-by: Giovanni Mascellani gmascellani@codeweavers.com --- libs/vkd3d-shader/hlsl_codegen.c | 157 +++++++++++++++++++++++++++++-- 1 file changed, 149 insertions(+), 8 deletions(-)
diff --git a/libs/vkd3d-shader/hlsl_codegen.c b/libs/vkd3d-shader/hlsl_codegen.c index b5c43754..0ecd8348 100644 --- a/libs/vkd3d-shader/hlsl_codegen.c +++ b/libs/vkd3d-shader/hlsl_codegen.c @@ -255,7 +255,18 @@ static void replace_node(struct hlsl_ir_node *old, struct hlsl_ir_node *new) * updated. When scanning through a load, it is checked if all the * registers involved in the load come from a single node. In such * case, the store can be replaced with a swizzle based on that - * node. */ + * node. + * + * All of the above works when we disregard control flow. With control + * flow it becames slightly more complicated: instead of a single map + * we keep a stack of them, pushing a new entry each time we enter an + * embedded block, and popping the entry when leaving the block. + * + * When entering a conditional block, both branches ("then" and + * "else") can inherit the variable state available just before the + * conditional block. After the conditional block, all variables that + * might have been written in either branch must be invalidated, + * because we don't know which branch has executed. */
struct copy_propagation_value { @@ -272,7 +283,9 @@ struct copy_propagation_variable
struct copy_propagation_state { - struct rb_tree variables; + struct rb_tree *variables; + unsigned int depth; + unsigned int capacity; };
static int copy_propagation_variable_compare(const void *key, const struct rb_entry *entry) @@ -293,7 +306,7 @@ static void copy_propagation_variable_destroy(struct rb_entry *entry, void *cont static struct copy_propagation_variable *copy_propagation_get_variable(struct copy_propagation_state *state, struct hlsl_ir_var *var) { - struct rb_entry *entry = rb_get(&state->variables, var); + struct rb_entry *entry = rb_get(&state->variables[state->depth], var);
if (entry) return RB_ENTRY_VALUE(entry, struct copy_propagation_variable, entry); @@ -304,7 +317,7 @@ static struct copy_propagation_variable *copy_propagation_get_variable(struct co static struct copy_propagation_variable *copy_propagation_create_variable(struct hlsl_ctx *ctx, struct copy_propagation_state *state, struct hlsl_ir_var *var) { - struct rb_entry *entry = rb_get(&state->variables, var); + struct rb_entry *entry = rb_get(&state->variables[state->depth], var); struct copy_propagation_variable *variable; int res;
@@ -323,7 +336,7 @@ static struct copy_propagation_variable *copy_propagation_create_variable(struct return NULL; }
- res = rb_put(&state->variables, var, &variable->entry); + res = rb_put(&state->variables[state->depth], var, &variable->entry); assert(!res);
return variable; @@ -354,6 +367,99 @@ static void copy_propagation_set_value(struct copy_propagation_variable *variabl } }
+static void copy_propagation_invalidate_from_block(struct hlsl_ctx *ctx, struct copy_propagation_state *state, + struct hlsl_block *block) +{ + struct hlsl_ir_node *instr; + + LIST_FOR_EACH_ENTRY(instr, &block->instrs, struct hlsl_ir_node, entry) + { + switch (instr->type) + { + case HLSL_IR_STORE: + { + struct hlsl_ir_store *store = hlsl_ir_store(instr); + struct copy_propagation_variable *variable; + struct hlsl_deref *lhs = &store->lhs; + struct hlsl_ir_var *var = lhs->var; + unsigned int offset; + + variable = copy_propagation_get_variable(state, var); + if (!variable) + continue; + + if (hlsl_offset_from_deref(lhs, &offset)) + copy_propagation_set_value(variable, offset, store->writemask, NULL); + else + copy_propagation_invalidate_whole_variable(variable); + + break; + } + + case HLSL_IR_IF: + { + struct hlsl_ir_if *iff = hlsl_ir_if(instr); + + copy_propagation_invalidate_from_block(ctx, state, &iff->then_instrs); + copy_propagation_invalidate_from_block(ctx, state, &iff->else_instrs); + + break; + } + + case HLSL_IR_LOOP: + { + struct hlsl_ir_loop *loop = hlsl_ir_loop(instr); + + copy_propagation_invalidate_from_block(ctx, state, &loop->body); + + break; + } + + default: + break; + } + } +} + +static void copy_propagation_pop(struct copy_propagation_state *state) +{ + assert(state->depth > 0); + rb_destroy(&state->variables[state->depth], copy_propagation_variable_destroy, NULL); + --state->depth; +} + +static bool copy_propagation_duplicate(struct hlsl_ctx *ctx, struct copy_propagation_state *state) +{ + struct copy_propagation_variable *var; + + if (state->depth + 1 == state->capacity) + { + unsigned int new_capacity = 2 * state->capacity; + struct rb_tree *new_vars; + + new_vars = hlsl_realloc(ctx, state->variables, sizeof(*state->variables) * new_capacity); + if (!new_vars) + return false; + state->capacity = new_capacity; + state->variables = new_vars; + } + ++state->depth; + + rb_init(&state->variables[state->depth], copy_propagation_variable_compare); + + RB_FOR_EACH_ENTRY(var, &state->variables[state->depth - 1], struct copy_propagation_variable, entry) + { + struct copy_propagation_variable *new_var = copy_propagation_create_variable(ctx, state, var->var); + + if (!new_var) + return false; + + memcpy(new_var->values, var->values, sizeof(*var->values) * var->var->data_type->reg_size); + } + + return true; +} + static struct hlsl_ir_node *copy_propagation_find_replacement(struct copy_propagation_variable *variable, unsigned int offset, unsigned int count, unsigned int *swizzle) { @@ -436,6 +542,34 @@ static void copy_propagation_record_store(struct hlsl_ctx *ctx, struct hlsl_ir_s copy_propagation_invalidate_whole_variable(variable); }
+static bool copy_propagation_transform_block(struct hlsl_ctx *ctx, struct hlsl_block *block, + struct copy_propagation_state *state); + +static bool copy_propagation_process_if(struct hlsl_ctx *ctx, struct hlsl_ir_if *iff, + struct copy_propagation_state *state) +{ + bool progress = false; + + if (!copy_propagation_duplicate(ctx, state)) + goto end; + + progress |= copy_propagation_transform_block(ctx, &iff->then_instrs, state); + + copy_propagation_pop(state); + if (!copy_propagation_duplicate(ctx, state)) + goto end; + + progress |= copy_propagation_transform_block(ctx, &iff->else_instrs, state); + + copy_propagation_pop(state); + +end: + copy_propagation_invalidate_from_block(ctx, state, &iff->then_instrs); + copy_propagation_invalidate_from_block(ctx, state, &iff->else_instrs); + + return progress; +} + static bool copy_propagation_transform_block(struct hlsl_ctx *ctx, struct hlsl_block *block, struct copy_propagation_state *state) { @@ -455,7 +589,7 @@ static bool copy_propagation_transform_block(struct hlsl_ctx *ctx, struct hlsl_b break;
case HLSL_IR_IF: - FIXME("Copy propagation doesn't support conditionals yet, leaving.\n"); + progress |= copy_propagation_process_if(ctx, hlsl_ir_if(instr), state); return progress;
case HLSL_IR_LOOP: @@ -475,11 +609,18 @@ static bool copy_propagation_execute(struct hlsl_ctx *ctx, struct hlsl_block *bl struct copy_propagation_state state; bool progress;
- rb_init(&state.variables, copy_propagation_variable_compare); + state.depth = 0; + state.capacity = 1; + state.variables = hlsl_alloc(ctx, sizeof(*state.variables) * state.capacity); + if (!state.variables) + return false; + rb_init(&state.variables[state.depth], copy_propagation_variable_compare);
progress = copy_propagation_transform_block(ctx, block, &state);
- rb_destroy(&state.variables, copy_propagation_variable_destroy, NULL); + assert(state.depth == 0); + rb_destroy(&state.variables[state.depth], copy_propagation_variable_destroy, NULL); + vkd3d_free(state.variables);
return progress; }