*
* Sets the InputsRead and OutputsWritten of Mesa programs.
*
+ * Additionally, for fragment shaders, sets the InterpQualifier array, the
+ * IsCentroid and IsSample bitfields, and the UsesDFdy flag.
+ *
* Mesa programs (gl_program, not gl_shader_program) have a set of
* flags indicating which varyings are read and written. Computing
* which are actually read from some sort of backend code can be
* from the GLSL IR.
*/
-extern "C" {
-#include "main/mtypes.h"
-#include "program/hash_table.h"
-}
+#include "main/core.h" /* for struct gl_program */
#include "ir.h"
#include "ir_visitor.h"
#include "glsl_types.h"
+namespace {
+
class ir_set_program_inouts_visitor : public ir_hierarchical_visitor {
public:
- ir_set_program_inouts_visitor(struct gl_program *prog)
+ ir_set_program_inouts_visitor(struct gl_program *prog,
+ gl_shader_stage shader_stage)
{
this->prog = prog;
- this->ht = hash_table_ctor(0,
- hash_table_pointer_hash,
- hash_table_pointer_compare);
+ this->shader_stage = shader_stage;
}
~ir_set_program_inouts_visitor()
{
- hash_table_dtor(this->ht);
}
virtual ir_visitor_status visit_enter(ir_dereference_array *);
virtual ir_visitor_status visit_enter(ir_function_signature *);
+ virtual ir_visitor_status visit_enter(ir_expression *);
+ virtual ir_visitor_status visit_enter(ir_discard *);
+ virtual ir_visitor_status visit_enter(ir_texture *);
virtual ir_visitor_status visit(ir_dereference_variable *);
- virtual ir_visitor_status visit(ir_variable *);
+
+private:
+ void mark_whole_variable(ir_variable *var);
+ bool try_mark_partial_variable(ir_variable *var, ir_rvalue *index);
struct gl_program *prog;
- struct hash_table *ht;
+ gl_shader_stage shader_stage;
};
+} /* anonymous namespace */
+
+static inline bool
+is_shader_inout(ir_variable *var)
+{
+ return var->data.mode == ir_var_shader_in ||
+ var->data.mode == ir_var_shader_out ||
+ var->data.mode == ir_var_system_value;
+}
+
static void
-mark(struct gl_program *prog, ir_variable *var, int index)
+mark(struct gl_program *prog, ir_variable *var, int offset, int len,
+ gl_shader_stage stage)
{
/* As of GLSL 1.20, varyings can only be floats, floating-point
* vectors or matrices, or arrays of them. For Mesa programs using
* something doing a more clever packing would use something other
* than InputsRead/OutputsWritten.
*/
- const glsl_type *element_type;
- int element_size;
-
- if (var->type->is_array())
- element_type = var->type->fields.array;
- else
- element_type = var->type;
-
- if (element_type->is_matrix())
- element_size = element_type->matrix_columns;
- else
- element_size = 1;
-
- index *= element_size;
- for (int i = 0; i < element_size; i++) {
- if (var->mode == ir_var_in)
- prog->InputsRead |= BITFIELD64_BIT(var->location + index + i);
- else
- prog->OutputsWritten |= BITFIELD64_BIT(var->location + index + i);
+
+ for (int i = 0; i < len; i++) {
+ int idx = var->data.location + var->data.index + offset + i;
+ bool is_patch_generic = var->data.patch &&
+ idx != VARYING_SLOT_TESS_LEVEL_INNER &&
+ idx != VARYING_SLOT_TESS_LEVEL_OUTER;
+ GLbitfield64 bitfield;
+
+ if (is_patch_generic) {
+ assert(idx >= VARYING_SLOT_PATCH0 && idx < VARYING_SLOT_TESS_MAX);
+ bitfield = BITFIELD64_BIT(idx - VARYING_SLOT_PATCH0);
+ }
+ else {
+ assert(idx < VARYING_SLOT_MAX);
+ bitfield = BITFIELD64_BIT(idx);
+ }
+
+ if (var->data.mode == ir_var_shader_in) {
+ if (is_patch_generic)
+ prog->PatchInputsRead |= bitfield;
+ else
+ prog->InputsRead |= bitfield;
+
+ /* double inputs read is only for vertex inputs */
+ if (stage == MESA_SHADER_VERTEX &&
+ var->type->without_array()->is_dual_slot_double())
+ prog->DoubleInputsRead |= bitfield;
+
+ if (stage == MESA_SHADER_FRAGMENT) {
+ gl_fragment_program *fprog = (gl_fragment_program *) prog;
+ fprog->InterpQualifier[idx] =
+ (glsl_interp_qualifier) var->data.interpolation;
+ if (var->data.centroid)
+ fprog->IsCentroid |= bitfield;
+ if (var->data.sample)
+ fprog->IsSample |= bitfield;
+ }
+ } else if (var->data.mode == ir_var_system_value) {
+ prog->SystemValuesRead |= bitfield;
+ } else {
+ assert(var->data.mode == ir_var_shader_out);
+ if (is_patch_generic)
+ prog->PatchOutputsWritten |= bitfield;
+ else
+ prog->OutputsWritten |= bitfield;
+ }
+ }
+}
+
+/**
+ * Mark an entire variable as used. Caller must ensure that the variable
+ * represents a shader input or output.
+ */
+void
+ir_set_program_inouts_visitor::mark_whole_variable(ir_variable *var)
+{
+ const glsl_type *type = var->type;
+ bool vertex_input = false;
+ if (this->shader_stage == MESA_SHADER_GEOMETRY &&
+ var->data.mode == ir_var_shader_in && type->is_array()) {
+ type = type->fields.array;
+ }
+
+ if (this->shader_stage == MESA_SHADER_TESS_CTRL &&
+ var->data.mode == ir_var_shader_in) {
+ assert(type->is_array());
+ type = type->fields.array;
+ }
+
+ if (this->shader_stage == MESA_SHADER_TESS_CTRL &&
+ var->data.mode == ir_var_shader_out && !var->data.patch) {
+ assert(type->is_array());
+ type = type->fields.array;
+ }
+
+ if (this->shader_stage == MESA_SHADER_TESS_EVAL &&
+ var->data.mode == ir_var_shader_in && !var->data.patch) {
+ assert(type->is_array());
+ type = type->fields.array;
}
+
+ if (this->shader_stage == MESA_SHADER_VERTEX &&
+ var->data.mode == ir_var_shader_in)
+ vertex_input = true;
+
+ mark(this->prog, var, 0, type->count_attribute_slots(vertex_input),
+ this->shader_stage);
}
/* Default handler: Mark all the locations in the variable as used. */
ir_visitor_status
ir_set_program_inouts_visitor::visit(ir_dereference_variable *ir)
{
- if (hash_table_find(this->ht, ir->var) == NULL)
+ if (!is_shader_inout(ir->var))
return visit_continue;
- if (ir->type->is_array()) {
- for (unsigned int i = 0; i < ir->type->length; i++) {
- mark(this->prog, ir->var, i);
- }
- } else {
- mark(this->prog, ir->var, 0);
- }
+ mark_whole_variable(ir->var);
return visit_continue;
}
-ir_visitor_status
-ir_set_program_inouts_visitor::visit_enter(ir_dereference_array *ir)
+/**
+ * Try to mark a portion of the given variable as used. Caller must ensure
+ * that the variable represents a shader input or output which can be indexed
+ * into in array fashion (an array or matrix). For the purpose of geometry
+ * shader inputs (which are always arrays*), this means that the array element
+ * must be something that can be indexed into in array fashion.
+ *
+ * *Except gl_PrimitiveIDIn, as noted below.
+ *
+ * For tessellation control shaders all inputs and non-patch outputs are
+ * arrays. For tessellation evaluation shaders non-patch inputs are arrays.
+ *
+ * If the index can't be interpreted as a constant, or some other problem
+ * occurs, then nothing will be marked and false will be returned.
+ */
+bool
+ir_set_program_inouts_visitor::try_mark_partial_variable(ir_variable *var,
+ ir_rvalue *index)
{
- ir_dereference_variable *deref_var;
- ir_constant *index = ir->array_index->as_constant();
- deref_var = ir->array->as_dereference_variable();
- ir_variable *var = NULL;
+ const glsl_type *type = var->type;
- /* Check that we're dereferencing a shader in or out */
- if (deref_var)
- var = (ir_variable *)hash_table_find(this->ht, deref_var->var);
+ if (this->shader_stage == MESA_SHADER_GEOMETRY &&
+ var->data.mode == ir_var_shader_in) {
+ /* The only geometry shader input that is not an array is
+ * gl_PrimitiveIDIn, and in that case, this code will never be reached,
+ * because gl_PrimitiveIDIn can't be indexed into in array fashion.
+ */
+ assert(type->is_array());
+ type = type->fields.array;
+ }
- if (index && var) {
- mark(this->prog, var, index->value.i[0]);
- return visit_continue_with_parent;
+ if (this->shader_stage == MESA_SHADER_TESS_CTRL &&
+ var->data.mode == ir_var_shader_in) {
+ assert(type->is_array());
+ type = type->fields.array;
}
- return visit_continue;
+ if (this->shader_stage == MESA_SHADER_TESS_CTRL &&
+ var->data.mode == ir_var_shader_out && !var->data.patch) {
+ assert(type->is_array());
+ type = type->fields.array;
+ }
+
+ if (this->shader_stage == MESA_SHADER_TESS_EVAL &&
+ var->data.mode == ir_var_shader_in && !var->data.patch) {
+ assert(type->is_array());
+ type = type->fields.array;
+ }
+
+ /* TODO: implement proper arrays of arrays support
+ * for now let the caller mark whole variable as used.
+ */
+ if (type->is_array() && type->fields.array->is_array())
+ return false;
+
+ /* The code below only handles:
+ *
+ * - Indexing into matrices
+ * - Indexing into arrays of (matrices, vectors, or scalars)
+ *
+ * All other possibilities are either prohibited by GLSL (vertex inputs and
+ * fragment outputs can't be structs) or should have been eliminated by
+ * lowering passes (do_vec_index_to_swizzle() gets rid of indexing into
+ * vectors, and lower_packed_varyings() gets rid of structs that occur in
+ * varyings).
+ */
+ if (!(type->is_matrix() ||
+ (type->is_array() &&
+ (type->fields.array->is_numeric() ||
+ type->fields.array->is_boolean())))) {
+ assert(!"Unexpected indexing in ir_set_program_inouts");
+
+ /* For safety in release builds, in case we ever encounter unexpected
+ * indexing, give up and let the caller mark the whole variable as used.
+ */
+ return false;
+ }
+
+ ir_constant *index_as_constant = index->as_constant();
+ if (!index_as_constant)
+ return false;
+
+ unsigned elem_width;
+ unsigned num_elems;
+ if (type->is_array()) {
+ num_elems = type->length;
+ if (type->fields.array->is_matrix())
+ elem_width = type->fields.array->matrix_columns;
+ else
+ elem_width = 1;
+ } else {
+ num_elems = type->matrix_columns;
+ elem_width = 1;
+ }
+
+ if (index_as_constant->value.u[0] >= num_elems) {
+ /* Constant index outside the bounds of the matrix/array. This could
+ * arise as a result of constant folding of a legal GLSL program.
+ *
+ * Even though the spec says that indexing outside the bounds of a
+ * matrix/array results in undefined behaviour, we don't want to pass
+ * out-of-range values to mark() (since this could result in slots that
+ * don't exist being marked as used), so just let the caller mark the
+ * whole variable as used.
+ */
+ return false;
+ }
+
+ /* double element width for double types that takes two slots */
+ if (this->shader_stage != MESA_SHADER_VERTEX ||
+ var->data.mode != ir_var_shader_in) {
+ if (type->without_array()->is_dual_slot_double())
+ elem_width *= 2;
+ }
+
+ mark(this->prog, var, index_as_constant->value.u[0] * elem_width,
+ elem_width, this->shader_stage);
+ return true;
+}
+
+static bool
+is_multiple_vertices(gl_shader_stage stage, ir_variable *var)
+{
+ if (var->data.patch)
+ return false;
+
+ if (var->data.mode == ir_var_shader_in)
+ return stage == MESA_SHADER_GEOMETRY ||
+ stage == MESA_SHADER_TESS_CTRL ||
+ stage == MESA_SHADER_TESS_EVAL;
+ if (var->data.mode == ir_var_shader_out)
+ return stage == MESA_SHADER_TESS_CTRL;
+
+ return false;
}
ir_visitor_status
-ir_set_program_inouts_visitor::visit(ir_variable *ir)
+ir_set_program_inouts_visitor::visit_enter(ir_dereference_array *ir)
{
- if (ir->mode == ir_var_in ||
- ir->mode == ir_var_out) {
- hash_table_insert(this->ht, ir, ir);
+ /* Note: for geometry shader inputs, lower_named_interface_blocks may
+ * create 2D arrays, so we need to be able to handle those. 2D arrays
+ * shouldn't be able to crop up for any other reason.
+ */
+ if (ir_dereference_array * const inner_array =
+ ir->array->as_dereference_array()) {
+ /* ir => foo[i][j]
+ * inner_array => foo[i]
+ */
+ if (ir_dereference_variable * const deref_var =
+ inner_array->array->as_dereference_variable()) {
+ if (is_multiple_vertices(this->shader_stage, deref_var->var)) {
+ /* foo is a geometry or tessellation shader input, so i is
+ * the vertex, and j the part of the input we're accessing.
+ */
+ if (try_mark_partial_variable(deref_var->var, ir->array_index))
+ {
+ /* We've now taken care of foo and j, but i might contain a
+ * subexpression that accesses shader inputs. So manually
+ * visit i and then continue with the parent.
+ */
+ inner_array->array_index->accept(this);
+ return visit_continue_with_parent;
+ }
+ }
+ }
+ } else if (ir_dereference_variable * const deref_var =
+ ir->array->as_dereference_variable()) {
+ /* ir => foo[i], where foo is a variable. */
+ if (is_multiple_vertices(this->shader_stage, deref_var->var)) {
+ /* foo is a geometry or tessellation shader input, so i is
+ * the vertex, and we're accessing the entire input.
+ */
+ mark_whole_variable(deref_var->var);
+ /* We've now taken care of foo, but i might contain a subexpression
+ * that accesses shader inputs. So manually visit i and then
+ * continue with the parent.
+ */
+ ir->array_index->accept(this);
+ return visit_continue_with_parent;
+ } else if (is_shader_inout(deref_var->var)) {
+ /* foo is a shader input/output, but not a geometry shader input,
+ * so i is the part of the input we're accessing.
+ */
+ if (try_mark_partial_variable(deref_var->var, ir->array_index))
+ return visit_continue_with_parent;
+ }
}
+ /* The expression is something we don't recognize. Just visit its
+ * subexpressions.
+ */
return visit_continue;
}
return visit_continue_with_parent;
}
+ir_visitor_status
+ir_set_program_inouts_visitor::visit_enter(ir_expression *ir)
+{
+ if (this->shader_stage == MESA_SHADER_FRAGMENT &&
+ (ir->operation == ir_unop_dFdy ||
+ ir->operation == ir_unop_dFdy_coarse ||
+ ir->operation == ir_unop_dFdy_fine)) {
+ gl_fragment_program *fprog = (gl_fragment_program *) prog;
+ fprog->UsesDFdy = true;
+ }
+ return visit_continue;
+}
+
+ir_visitor_status
+ir_set_program_inouts_visitor::visit_enter(ir_discard *)
+{
+ /* discards are only allowed in fragment shaders. */
+ assert(this->shader_stage == MESA_SHADER_FRAGMENT);
+
+ gl_fragment_program *fprog = (gl_fragment_program *) prog;
+ fprog->UsesKill = true;
+
+ return visit_continue;
+}
+
+ir_visitor_status
+ir_set_program_inouts_visitor::visit_enter(ir_texture *ir)
+{
+ if (ir->op == ir_tg4)
+ prog->UsesGather = true;
+ return visit_continue;
+}
+
void
-do_set_program_inouts(exec_list *instructions, struct gl_program *prog)
+do_set_program_inouts(exec_list *instructions, struct gl_program *prog,
+ gl_shader_stage shader_stage)
{
- ir_set_program_inouts_visitor v(prog);
+ ir_set_program_inouts_visitor v(prog, shader_stage);
prog->InputsRead = 0;
prog->OutputsWritten = 0;
+ prog->PatchInputsRead = 0;
+ prog->PatchOutputsWritten = 0;
+ prog->SystemValuesRead = 0;
+ if (shader_stage == MESA_SHADER_FRAGMENT) {
+ gl_fragment_program *fprog = (gl_fragment_program *) prog;
+ memset(fprog->InterpQualifier, 0, sizeof(fprog->InterpQualifier));
+ fprog->IsCentroid = 0;
+ fprog->IsSample = 0;
+ fprog->UsesDFdy = false;
+ fprog->UsesKill = false;
+ }
visit_list_elements(&v, instructions);
}