X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fglsl%2Flower_variable_index_to_cond_assign.cpp;h=040b0bf83576197bc65602655739672b96a2a6fa;hb=65cc68f4305a675e27feb7aae0d8a66b2710f3e4;hp=8eb1612f0a0da921fb6e3ee30b8bd56e211b4ef5;hpb=b922a0ce12916a91cfc3e56714913fcf63279ff2;p=mesa.git diff --git a/src/glsl/lower_variable_index_to_cond_assign.cpp b/src/glsl/lower_variable_index_to_cond_assign.cpp index 8eb1612f0a0..040b0bf8357 100644 --- a/src/glsl/lower_variable_index_to_cond_assign.cpp +++ b/src/glsl/lower_variable_index_to_cond_assign.cpp @@ -29,6 +29,21 @@ * * Pre-DX10 GPUs often don't have a native way to do this operation, * and this works around that. + * + * The lowering process proceeds as follows. Each non-constant index + * found in an r-value is converted to a canonical form \c array[i]. Each + * element of the array is conditionally assigned to a temporary by comparing + * \c i to a constant index. This is done by cloning the canonical form and + * replacing all occurances of \c i with a constant. Each remaining occurance + * of the canonical form in the IR is replaced with a dereference of the + * temporary variable. + * + * L-values with non-constant indices are handled similarly. In this case, + * the RHS of the assignment is assigned to a temporary. The non-constant + * index is replace with the canonical form (just like for r-values). The + * temporary is conditionally assigned to each element of the canonical form + * by comparing \c i with each index. The same clone-and-replace scheme is + * used. */ #include "ir.h" @@ -37,10 +52,140 @@ #include "glsl_types.h" #include "main/macros.h" +/** + * Generate a comparison value for a block of indices + * + * Lowering passes for non-constant indexing of arrays, matrices, or vectors + * can use this to generate blocks of index comparison values. + * + * \param instructions List where new instructions will be appended + * \param index \c ir_variable containing the desired index + * \param base Base value for this block of comparisons + * \param components Number of unique index values to compare. This must + * be on the range [1, 4]. + * \param mem_ctx ralloc memory context to be used for all allocations. + * + * \returns + * An \c ir_rvalue that \b must be cloned for each use in conditional + * assignments, etc. + */ +ir_rvalue * +compare_index_block(exec_list *instructions, ir_variable *index, + unsigned base, unsigned components, void *mem_ctx) +{ + ir_rvalue *broadcast_index = new(mem_ctx) ir_dereference_variable(index); + + assert(index->type->is_scalar()); + assert(index->type->base_type == GLSL_TYPE_INT); + assert(components >= 1 && components <= 4); + + if (components > 1) { + const ir_swizzle_mask m = { 0, 0, 0, 0, components, false }; + broadcast_index = new(mem_ctx) ir_swizzle(broadcast_index, m); + } + + /* Compare the desired index value with the next block of four indices. + */ + ir_constant_data test_indices_data; + memset(&test_indices_data, 0, sizeof(test_indices_data)); + test_indices_data.i[0] = base; + test_indices_data.i[1] = base + 1; + test_indices_data.i[2] = base + 2; + test_indices_data.i[3] = base + 3; + + ir_constant *const test_indices = + new(mem_ctx) ir_constant(broadcast_index->type, + &test_indices_data); + + ir_rvalue *const condition_val = + new(mem_ctx) ir_expression(ir_binop_equal, + &glsl_type::bool_type[components - 1], + broadcast_index, + test_indices); + + ir_variable *const condition = + new(mem_ctx) ir_variable(condition_val->type, + "dereference_condition", + ir_var_temporary); + instructions->push_tail(condition); + + ir_rvalue *const cond_deref = + new(mem_ctx) ir_dereference_variable(condition); + instructions->push_tail(new(mem_ctx) ir_assignment(cond_deref, condition_val, 0)); + + return cond_deref; +} + +static inline bool +is_array_or_matrix(const ir_rvalue *ir) +{ + return (ir->type->is_array() || ir->type->is_matrix()); +} + +/** + * Replace a dereference of a variable with a specified r-value + * + * Each time a dereference of the specified value is replaced, the r-value + * tree is cloned. + */ +class deref_replacer : public ir_rvalue_visitor { +public: + deref_replacer(const ir_variable *variable_to_replace, ir_rvalue *value) + : variable_to_replace(variable_to_replace), value(value), + progress(false) + { + assert(this->variable_to_replace != NULL); + assert(this->value != NULL); + } + + virtual void handle_rvalue(ir_rvalue **rvalue) + { + ir_dereference_variable *const dv = (*rvalue)->as_dereference_variable(); + + if ((dv != NULL) && (dv->var == this->variable_to_replace)) { + this->progress = true; + *rvalue = this->value->clone(ralloc_parent(*rvalue), NULL); + } + } + + const ir_variable *variable_to_replace; + ir_rvalue *value; + bool progress; +}; + +/** + * Find a variable index dereference of an array in an rvalue tree + */ +class find_variable_index : public ir_hierarchical_visitor { +public: + find_variable_index() + : deref(NULL) + { + /* empty */ + } + + virtual ir_visitor_status visit_enter(ir_dereference_array *ir) + { + if (is_array_or_matrix(ir->array) + && (ir->array_index->as_constant() == NULL)) { + this->deref = ir; + return visit_stop; + } + + return visit_continue; + } + + /** + * First array dereference found in the tree that has a non-constant index. + */ + ir_dereference_array *deref; +}; + struct assignment_generator { ir_instruction* base_ir; - ir_rvalue* array; + ir_dereference *rvalue; + ir_variable *old_index; bool is_write; unsigned int write_mask; ir_variable* var; @@ -55,18 +200,23 @@ struct assignment_generator * underlying variable. */ void *mem_ctx = ralloc_parent(base_ir); - ir_dereference *element = - new(mem_ctx) ir_dereference_array(this->array->clone(mem_ctx, NULL), - new(mem_ctx) ir_constant(i)); - ir_rvalue *variable = new(mem_ctx) ir_dereference_variable(this->var); - ir_assignment *assignment; - if (is_write) { - assignment = new(mem_ctx) ir_assignment(element, variable, condition, - write_mask); - } else { - assignment = new(mem_ctx) ir_assignment(variable, element, condition); - } + /* Clone the old r-value in its entirety. Then replace any occurances of + * the old variable index with the new constant index. + */ + ir_dereference *element = this->rvalue->clone(mem_ctx, NULL); + ir_constant *const index = new(mem_ctx) ir_constant(i); + deref_replacer r(this->old_index, index); + element->accept(&r); + assert(r.progress); + + /* Generate a conditional assignment to (or from) the constant indexed + * array dereference. + */ + ir_rvalue *variable = new(mem_ctx) ir_dereference_variable(this->var); + ir_assignment *const assignment = (is_write) + ? new(mem_ctx) ir_assignment(element, variable, condition, write_mask) + : new(mem_ctx) ir_assignment(variable, element, condition); list->push_tail(assignment); } @@ -118,54 +268,17 @@ struct switch_generator for (unsigned i = first; i < end; i += 4) { const unsigned comps = MIN2(condition_components, end - i); - ir_rvalue *broadcast_index = - new(this->mem_ctx) ir_dereference_variable(index); - - if (comps) { - const ir_swizzle_mask m = { 0, 0, 0, 0, comps, false }; - broadcast_index = new(this->mem_ctx) ir_swizzle(broadcast_index, m); - } - - /* Compare the desired index value with the next block of four indices. - */ - ir_constant_data test_indices_data; - memset(&test_indices_data, 0, sizeof(test_indices_data)); - test_indices_data.i[0] = i; - test_indices_data.i[1] = i + 1; - test_indices_data.i[2] = i + 2; - test_indices_data.i[3] = i + 3; - ir_constant *const test_indices = - new(this->mem_ctx) ir_constant(broadcast_index->type, - &test_indices_data); - - ir_rvalue *const condition_val = - new(this->mem_ctx) ir_expression(ir_binop_equal, - &glsl_type::bool_type[comps - 1], - broadcast_index, - test_indices); - - ir_variable *const condition = - new(this->mem_ctx) ir_variable(condition_val->type, - "dereference_array_condition", - ir_var_temporary); - list->push_tail(condition); - ir_rvalue *const cond_deref = - new(this->mem_ctx) ir_dereference_variable(condition); - list->push_tail(new(this->mem_ctx) ir_assignment(cond_deref, - condition_val, 0)); + compare_index_block(list, index, i, comps, this->mem_ctx); if (comps == 1) { - ir_rvalue *const cond_deref = - new(this->mem_ctx) ir_dereference_variable(condition); - - this->generator.generate(i, cond_deref, list); + this->generator.generate(i, cond_deref->clone(this->mem_ctx, NULL), + list); } else { for (unsigned j = 0; j < comps; j++) { - ir_rvalue *const cond_deref = - new(this->mem_ctx) ir_dereference_variable(condition); ir_rvalue *const cond_swiz = - new(this->mem_ctx) ir_swizzle(cond_deref, j, 0, 0, 0, 1); + new(this->mem_ctx) ir_swizzle(cond_deref->clone(this->mem_ctx, NULL), + j, 0, 0, 0, 1); this->generator.generate(i + j, cond_swiz, list); } @@ -233,33 +346,34 @@ public: bool lower_temps; bool lower_uniforms; - bool is_array_or_matrix(const ir_instruction *ir) const - { - return (ir->type->is_array() || ir->type->is_matrix()); - } - - bool needs_lowering(ir_dereference_array *deref) const + bool storage_type_needs_lowering(ir_dereference_array *deref) const { - if (deref == NULL || deref->array_index->as_constant() - || !is_array_or_matrix(deref->array)) - return false; - - if (deref->array->ir_type == ir_type_constant) + /* If a variable isn't eventually the target of this dereference, then + * it must be a constant or some sort of anonymous temporary storage. + * + * FINISHME: Is this correct? Most drivers treat arrays of constants as + * FINISHME: uniforms. It seems like this should do the same. + */ + const ir_variable *const var = deref->array->variable_referenced(); + if (var == NULL) return this->lower_temps; - const ir_variable *const var = deref->array->variable_referenced(); switch (var->mode) { case ir_var_auto: case ir_var_temporary: return this->lower_temps; case ir_var_uniform: return this->lower_uniforms; - case ir_var_in: + case ir_var_function_in: case ir_var_const_in: - return (var->location == -1) ? this->lower_temps : this->lower_inputs; - case ir_var_out: - return (var->location == -1) ? this->lower_temps : this->lower_outputs; - case ir_var_inout: + return this->lower_temps; + case ir_var_shader_in: + return this->lower_inputs; + case ir_var_function_out: + return this->lower_temps; + case ir_var_shader_out: + return this->lower_outputs; + case ir_var_function_inout: return this->lower_temps; } @@ -267,8 +381,18 @@ public: return false; } + bool needs_lowering(ir_dereference_array *deref) const + { + if (deref == NULL || deref->array_index->as_constant() + || !is_array_or_matrix(deref->array)) + return false; + + return this->storage_type_needs_lowering(deref); + } + ir_variable *convert_dereference_array(ir_dereference_array *orig_deref, - ir_assignment* orig_assign) + ir_assignment* orig_assign, + ir_dereference *orig_base) { assert(is_array_or_matrix(orig_deref->array)); @@ -314,9 +438,12 @@ public: new(mem_ctx) ir_assignment(lhs, orig_deref->array_index, NULL); base_ir->insert_before(assign); + orig_deref->array_index = lhs->clone(mem_ctx, NULL); + assignment_generator ag; - ag.array = orig_deref->array; + ag.rvalue = orig_base; ag.base_ir = base_ir; + ag.old_index = index; ag.var = var; if (orig_assign) { ag.is_write = true; @@ -327,21 +454,40 @@ public: switch_generator sg(ag, index, 4, 4); - exec_list list; - sg.generate(0, length, &list); - base_ir->insert_before(&list); + /* If the original assignment has a condition, respect that original + * condition! This is acomplished by wrapping the new conditional + * assignments in an if-statement that uses the original condition. + */ + if ((orig_assign != NULL) && (orig_assign->condition != NULL)) { + /* No need to clone the condition because the IR that it hangs on is + * going to be removed from the instruction sequence. + */ + ir_if *if_stmt = new(mem_ctx) ir_if(orig_assign->condition); + + sg.generate(0, length, &if_stmt->then_instructions); + base_ir->insert_before(if_stmt); + } else { + exec_list list; + + sg.generate(0, length, &list); + base_ir->insert_before(&list); + } return var; } virtual void handle_rvalue(ir_rvalue **pir) { + if (this->in_assignee) + return; + if (!*pir) return; ir_dereference_array* orig_deref = (*pir)->as_dereference_array(); if (needs_lowering(orig_deref)) { - ir_variable* var = convert_dereference_array(orig_deref, 0); + ir_variable *var = + convert_dereference_array(orig_deref, NULL, orig_deref); assert(var); *pir = new(ralloc_parent(base_ir)) ir_dereference_variable(var); this->progress = true; @@ -353,10 +499,11 @@ public: { ir_rvalue_visitor::visit_leave(ir); - ir_dereference_array *orig_deref = ir->lhs->as_dereference_array(); + find_variable_index f; + ir->lhs->accept(&f); - if (needs_lowering(orig_deref)) { - convert_dereference_array(orig_deref, ir); + if ((f.deref != NULL) && storage_type_needs_lowering(f.deref)) { + convert_dereference_array(f.deref, ir, ir->lhs); ir->remove(); this->progress = true; } @@ -377,7 +524,17 @@ lower_variable_index_to_cond_assign(exec_list *instructions, lower_temp, lower_uniform); - visit_list_elements(&v, instructions); - - return v.progress; + /* Continue lowering until no progress is made. If there are multiple + * levels of indirection (e.g., non-constant indexing of array elements and + * matrix columns of an array of matrix), each pass will only lower one + * level of indirection. + */ + bool progress_ever = false; + do { + v.progress = false; + visit_list_elements(&v, instructions); + progress_ever = v.progress || progress_ever; + } while (v.progress); + + return progress_ever; }