/**
* Try to mark a portion of the given varying as used. Caller must ensure
* that the variable represents a shader input or output.
*
* 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.
*/
static bool
try_mask_partial_io(nir_shader *shader, nir_variable *var,
nir_deref_instr *deref, bool is_output_read)
{
const struct glsl_type *type = var->type;
if (nir_is_per_vertex_io(var, shader->info.stage)) {
assert(glsl_type_is_array(type));
type = glsl_get_array_element(type);
}
/* The code below only handles:
*
* - Indexing into matrices
* - Indexing into arrays of (arrays, matrices, vectors, or scalars)
*
* For now, we just give up if we see varying structs and arrays of structs
* here marking the entire variable as used.
*/
if (!(glsl_type_is_matrix(type) ||
(glsl_type_is_array(type) && !var->data.compact &&
(glsl_type_is_numeric(glsl_without_array(type)) ||
glsl_type_is_boolean(glsl_without_array(type)))))) {
/* If we don't know how to handle this case, give up and let the
* caller mark the whole variable as used.
*/
return false;
}
unsigned offset = get_io_offset(deref, false);
if (offset == -1)
return false;
unsigned num_elems;
unsigned elem_width = 1;
unsigned mat_cols = 1;
if (glsl_type_is_array(type)) {
num_elems = glsl_get_aoa_size(type);
if (glsl_type_is_matrix(glsl_without_array(type)))
mat_cols = glsl_get_matrix_columns(glsl_without_array(type));
} else {
num_elems = glsl_get_matrix_columns(type);
}
/* double element width for double types that takes two slots */
if (glsl_type_is_dual_slot(glsl_without_array(type)))
elem_width *= 2;
if (offset >= num_elems * elem_width * mat_cols) {
/* 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 set_io_mask() (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;
}
set_io_mask(shader, var, offset, elem_width, is_output_read);
return true;
}
static void
gather_intrinsic_info(nir_intrinsic_instr *instr, nir_shader *shader,
void *dead_ctx)
{
switch (instr->intrinsic) {
case nir_intrinsic_discard:
case nir_intrinsic_discard_if:
assert(shader->info.stage == MESA_SHADER_FRAGMENT);
shader->info.fs.uses_discard = true;
break;
case nir_intrinsic_interp_deref_at_centroid:
case nir_intrinsic_interp_deref_at_sample:
case nir_intrinsic_interp_deref_at_offset:
case nir_intrinsic_load_deref:
case nir_intrinsic_store_deref:{
nir_deref_instr *deref = nir_src_as_deref(instr->src[0]);
nir_variable *var = nir_deref_instr_get_variable(deref);
if (var->data.mode == nir_var_shader_in ||
var->data.mode == nir_var_shader_out) {
bool is_output_read = false;
if (var->data.mode == nir_var_shader_out &&
instr->intrinsic == nir_intrinsic_load_deref)
is_output_read = true;
if (!try_mask_partial_io(shader, var, deref, is_output_read))
mark_whole_variable(shader, var, is_output_read);
/* We need to track which input_reads bits correspond to a
* dvec3/dvec4 input attribute */
if (shader->info.stage == MESA_SHADER_VERTEX &&
var->data.mode == nir_var_shader_in &&
glsl_type_is_dual_slot(glsl_without_array(var->type))) {
for (unsigned i = 0; i < glsl_count_attribute_slots(var->type, false); i++) {
int idx = var->data.location + i;
shader->info.vs.double_inputs |= BITFIELD64_BIT(idx);
}
}
}
break;
}
case nir_intrinsic_load_draw_id:
case nir_intrinsic_load_frag_coord:
case nir_intrinsic_load_front_face:
case nir_intrinsic_load_vertex_id:
case nir_intrinsic_load_vertex_id_zero_base:
case nir_intrinsic_load_base_vertex:
case nir_intrinsic_load_first_vertex:
case nir_intrinsic_load_is_indexed_draw:
case nir_intrinsic_load_base_instance:
case nir_intrinsic_load_instance_id:
case nir_intrinsic_load_sample_id:
case nir_intrinsic_load_sample_pos:
case nir_intrinsic_load_sample_mask_in:
case nir_intrinsic_load_primitive_id:
case nir_intrinsic_load_invocation_id:
case nir_intrinsic_load_local_invocation_id:
case nir_intrinsic_load_local_invocation_index:
case nir_intrinsic_load_work_group_id:
case nir_intrinsic_load_num_work_groups:
case nir_intrinsic_load_tess_coord:
case nir_intrinsic_load_tess_level_outer:
case nir_intrinsic_load_tess_level_inner:
case nir_intrinsic_load_patch_vertices_in:
shader->info.system_values_read |=
(1ull << nir_system_value_from_intrinsic(instr->intrinsic));
break;
case nir_intrinsic_end_primitive:
case nir_intrinsic_end_primitive_with_counter:
assert(shader->info.stage == MESA_SHADER_GEOMETRY);
shader->info.gs.uses_end_primitive = 1;
/* fall through */
case nir_intrinsic_emit_vertex:
if (nir_intrinsic_stream_id(instr) > 0)
shader->info.gs.uses_streams = true;
break;
default:
break;
}
}
static void
var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
const struct vtn_decoration *dec, void *void_var)
{
struct vtn_variable *vtn_var = void_var;
/* Handle decorations that apply to a vtn_variable as a whole */
switch (dec->decoration) {
case SpvDecorationNonWritable:
/* Do nothing with this for now */
return;
case SpvDecorationBinding:
vtn_var->binding = dec->literals[0];
return;
case SpvDecorationDescriptorSet:
vtn_var->descriptor_set = dec->literals[0];
return;
case SpvDecorationLocation: {
unsigned location = dec->literals[0];
bool is_vertex_input;
if (b->shader->stage == MESA_SHADER_FRAGMENT &&
vtn_var->mode == vtn_variable_mode_output) {
is_vertex_input = false;
location += FRAG_RESULT_DATA0;
} else if (b->shader->stage == MESA_SHADER_VERTEX &&
vtn_var->mode == vtn_variable_mode_input) {
is_vertex_input = true;
location += VERT_ATTRIB_GENERIC0;
} else if (vtn_var->mode == vtn_variable_mode_input ||
vtn_var->mode == vtn_variable_mode_output) {
is_vertex_input = false;
location += VARYING_SLOT_VAR0;
} else {
assert(!"Location must be on input or output variable");
}
if (vtn_var->var) {
vtn_var->var->data.location = location;
vtn_var->var->data.explicit_location = true;
} else {
assert(vtn_var->members);
unsigned length =
glsl_get_length(glsl_without_array(vtn_var->type->type));
for (unsigned i = 0; i < length; i++) {
vtn_var->members[i]->data.location = location;
vtn_var->members[i]->data.explicit_location = true;
location +=
glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
is_vertex_input);
}
}
return;
}
default:
break;
}
/* Now we handle decorations that apply to a particular nir_variable */
nir_variable *nir_var = vtn_var->var;
if (val->value_type == vtn_value_type_access_chain) {
assert(val->access_chain->length == 0);
assert(val->access_chain->var == void_var);
assert(member == -1);
} else {
assert(val->value_type == vtn_value_type_type);
if (member != -1)
nir_var = vtn_var->members[member];
}
if (nir_var == NULL)
return;
switch (dec->decoration) {
case SpvDecorationRelaxedPrecision:
break; /* FIXME: Do nothing with this for now. */
case SpvDecorationNoPerspective:
nir_var->data.interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
break;
case SpvDecorationFlat:
nir_var->data.interpolation = INTERP_QUALIFIER_FLAT;
break;
case SpvDecorationCentroid:
nir_var->data.centroid = true;
break;
case SpvDecorationSample:
nir_var->data.sample = true;
break;
case SpvDecorationInvariant:
nir_var->data.invariant = true;
break;
case SpvDecorationConstant:
assert(nir_var->constant_initializer != NULL);
nir_var->data.read_only = true;
break;
case SpvDecorationNonWritable:
nir_var->data.read_only = true;
break;
case SpvDecorationComponent:
nir_var->data.location_frac = dec->literals[0];
break;
case SpvDecorationIndex:
nir_var->data.explicit_index = true;
nir_var->data.index = dec->literals[0];
break;
case SpvDecorationBuiltIn: {
SpvBuiltIn builtin = dec->literals[0];
if (builtin == SpvBuiltInWorkgroupSize) {
/* This shouldn't be a builtin. It's actually a constant. */
nir_var->data.mode = nir_var_global;
nir_var->data.read_only = true;
nir_constant *c = rzalloc(nir_var, nir_constant);
c->value.u[0] = b->shader->info.cs.local_size[0];
c->value.u[1] = b->shader->info.cs.local_size[1];
c->value.u[2] = b->shader->info.cs.local_size[2];
nir_var->constant_initializer = c;
break;
}
nir_variable_mode mode = nir_var->data.mode;
vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
nir_var->data.explicit_location = true;
nir_var->data.mode = mode;
if (builtin == SpvBuiltInFragCoord || builtin == SpvBuiltInSamplePosition)
nir_var->data.origin_upper_left = b->origin_upper_left;
break;
}
case SpvDecorationRowMajor:
case SpvDecorationColMajor:
case SpvDecorationGLSLShared:
case SpvDecorationPatch:
case SpvDecorationRestrict:
case SpvDecorationAliased:
case SpvDecorationVolatile:
case SpvDecorationCoherent:
case SpvDecorationNonReadable:
case SpvDecorationUniform:
/* This is really nice but we have no use for it right now. */
case SpvDecorationCPacked:
case SpvDecorationSaturatedConversion:
case SpvDecorationStream:
case SpvDecorationOffset:
case SpvDecorationXfbBuffer:
case SpvDecorationFuncParamAttr:
case SpvDecorationFPRoundingMode:
case SpvDecorationFPFastMathMode:
case SpvDecorationLinkageAttributes:
case SpvDecorationSpecId:
break;
default:
unreachable("Unhandled variable decoration");
}
}
