size_align(var->type, &var_size, &var_align);
assert(var->data.location % var_align == 0);
+ UNUSED unsigned deref_size, deref_align;
+ size_align(deref->type, &deref_size, &deref_align);
+
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_constant);
load->num_components = num_components;
nir_intrinsic_set_base(load, var->data.location);
nir_intrinsic_set_range(load, var_size);
+ nir_intrinsic_set_align(load, deref_align, 0);
load->src[0] = nir_src_for_ssa(nir_build_deref_offset(b, deref, size_align));
nir_ssa_dest_init(&load->instr, &load->dest,
num_components, bit_size, NULL);
nir_builder_instr_insert(b, &load->instr);
if (load->dest.ssa.bit_size < 8) {
- /* Booleans are special-cased to be 32-bit
- *
- * Ideally, for drivers that can handle 32-bit booleans, we wouldn't
- * emit the i2b here. However, at this point, the driver is likely to
- * still have 1-bit booleans so we need to at least convert bit sizes.
- * Unfortunately, we don't have a good way to annotate the load as
- * loading a known boolean value so the optimizer isn't going to be
- * able to get rid of the conversion. Some day, we may solve that
- * problem but not today.
- */
+ /* Booleans are special-cased to be 32-bit */
assert(glsl_type_is_boolean(deref->type));
+ assert(deref_size == num_components * 4);
load->dest.ssa.bit_size = 32;
- return nir_i2b(b, &load->dest.ssa);
+ return nir_b2b1(b, &load->dest.ssa);
} else {
+ assert(deref_size == num_components * bit_size / 8);
return &load->dest.ssa;
}
}
static void
handle_constant_store(void *mem_ctx, struct var_info *info,
nir_deref_instr *deref, nir_const_value *val,
+ unsigned writemask,
glsl_type_size_align_func size_align)
{
assert(!nir_deref_instr_has_indirect(deref));
char *dst = (char *)info->constant_data +
nir_deref_instr_get_const_offset(deref, size_align);
- switch (bit_size) {
- case 1:
- /* Booleans are special-cased to be 32-bit */
- for (unsigned i = 0; i < num_components; i++)
+ for (unsigned i = 0; i < num_components; i++) {
+ if (!(writemask & (1 << i)))
+ continue;
+
+ switch (bit_size) {
+ case 1:
+ /* Booleans are special-cased to be 32-bit */
((int32_t *)dst)[i] = -(int)val[i].b;
- break;
+ break;
- case 8:
- for (unsigned i = 0; i < num_components; i++)
+ case 8:
((uint8_t *)dst)[i] = val[i].u8;
- break;
+ break;
- case 16:
- for (unsigned i = 0; i < num_components; i++)
+ case 16:
((uint16_t *)dst)[i] = val[i].u16;
- break;
+ break;
- case 32:
- for (unsigned i = 0; i < num_components; i++)
+ case 32:
((uint32_t *)dst)[i] = val[i].u32;
- break;
+ break;
- case 64:
- for (unsigned i = 0; i < num_components; i++)
+ case 64:
((uint64_t *)dst)[i] = val[i].u64;
- break;
+ break;
- default:
- unreachable("Invalid bit size");
+ default:
+ unreachable("Invalid bit size");
+ }
}
}
/* This pass can only be run once */
assert(shader->constant_data == NULL && shader->constant_data_size == 0);
- /* The index parameter is unused for local variables so we'll use it for
- * indexing into our array of variable metadata.
- */
- unsigned num_locals = 0;
- nir_foreach_variable(var, &impl->locals)
- var->data.index = num_locals++;
+ unsigned num_locals = nir_function_impl_index_vars(impl);
- if (num_locals == 0)
+ if (num_locals == 0) {
+ nir_shader_preserve_all_metadata(shader);
return false;
+ }
struct var_info *var_infos = ralloc_array(NULL, struct var_info, num_locals);
- nir_foreach_variable(var, &impl->locals) {
- var_infos[var->data.index] = (struct var_info) {
+ nir_foreach_function_temp_variable(var, impl) {
+ var_infos[var->index] = (struct var_info) {
.var = var,
.is_constant = true,
.found_read = false,
*/
nir_foreach_block(block, impl) {
nir_foreach_instr(instr, block) {
+ if (instr->type == nir_instr_type_deref) {
+ /* If we ever see a complex use of a deref_var, we have to assume
+ * that variable is non-constant because we can't guarantee we
+ * will find all of the writers of that variable.
+ */
+ nir_deref_instr *deref = nir_instr_as_deref(instr);
+ if (deref->deref_type == nir_deref_type_var &&
+ deref->mode == nir_var_function_temp &&
+ nir_deref_instr_has_complex_use(deref))
+ var_infos[deref->var->index].is_constant = false;
+ continue;
+ }
+
if (instr->type != nir_instr_type_intrinsic)
continue;
bool src_is_const = false;
nir_deref_instr *src_deref = NULL, *dst_deref = NULL;
+ unsigned writemask = 0;
switch (intrin->intrinsic) {
case nir_intrinsic_store_deref:
dst_deref = nir_src_as_deref(intrin->src[0]);
src_is_const = nir_src_is_const(intrin->src[1]);
+ writemask = nir_intrinsic_write_mask(intrin);
break;
case nir_intrinsic_load_deref:
break;
case nir_intrinsic_copy_deref:
- /* We always assume the src and therefore the dst are not
- * constants here. Copy and constant propagation passes should
- * have taken care of this in most cases anyway.
- */
- dst_deref = nir_src_as_deref(intrin->src[0]);
- src_deref = nir_src_as_deref(intrin->src[1]);
- src_is_const = false;
+ assert(!"Lowering of copy_deref with large constants is prohibited");
break;
default:
if (dst_deref && dst_deref->mode == nir_var_function_temp) {
nir_variable *var = nir_deref_instr_get_variable(dst_deref);
+ if (var == NULL)
+ continue;
+
assert(var->data.mode == nir_var_function_temp);
- struct var_info *info = &var_infos[var->data.index];
+ struct var_info *info = &var_infos[var->index];
if (!info->is_constant)
continue;
info->is_constant = false;
} else {
nir_const_value *val = nir_src_as_const_value(intrin->src[1]);
- handle_constant_store(var_infos, info, dst_deref, val, size_align);
+ handle_constant_store(var_infos, info, dst_deref, val, writemask,
+ size_align);
}
}
if (src_deref && src_deref->mode == nir_var_function_temp) {
nir_variable *var = nir_deref_instr_get_variable(src_deref);
+ if (var == NULL)
+ continue;
+
assert(var->data.mode == nir_var_function_temp);
/* We only consider variables constant if all the reads are
* dominated by the block that writes to it.
*/
- struct var_info *info = &var_infos[var->data.index];
+ struct var_info *info = &var_infos[var->index];
if (!info->is_constant)
continue;
struct var_info *info = &var_infos[i];
/* Fix up indices after we sorted. */
- info->var->data.index = i;
+ info->var->index = i;
if (!info->is_constant)
continue;
}
if (shader->constant_data_size == 0) {
+ nir_shader_preserve_all_metadata(shader);
ralloc_free(var_infos);
return false;
}
shader->constant_data = rzalloc_size(shader, shader->constant_data_size);
for (int i = 0; i < num_locals; i++) {
struct var_info *info = &var_infos[i];
- if (!info->duplicate) {
+ if (!info->duplicate && info->is_constant) {
memcpy((char *)shader->constant_data + info->var->data.location,
info->constant_data, info->constant_data_size);
}
continue;
nir_variable *var = nir_deref_instr_get_variable(deref);
- struct var_info *info = &var_infos[var->data.index];
+ if (var == NULL)
+ continue;
+
+ struct var_info *info = &var_infos[var->index];
if (info->is_constant) {
b.cursor = nir_after_instr(&intrin->instr);
nir_ssa_def *val = build_constant_load(&b, deref, size_align);
continue;
nir_variable *var = nir_deref_instr_get_variable(deref);
- struct var_info *info = &var_infos[var->data.index];
- if (info->is_constant) {
- nir_instr_remove(&intrin->instr);
- nir_deref_instr_remove_if_unused(deref);
- }
- break;
- }
-
- case nir_intrinsic_copy_deref: {
- nir_deref_instr *deref = nir_src_as_deref(intrin->src[1]);
- if (deref->mode != nir_var_function_temp)
+ if (var == NULL)
continue;
- nir_variable *var = nir_deref_instr_get_variable(deref);
- struct var_info *info = &var_infos[var->data.index];
+ struct var_info *info = &var_infos[var->index];
if (info->is_constant) {
- b.cursor = nir_after_instr(&intrin->instr);
- nir_ssa_def *val = build_constant_load(&b, deref, size_align);
- nir_store_deref(&b, nir_src_as_deref(intrin->src[0]), val, ~0);
nir_instr_remove(&intrin->instr);
nir_deref_instr_remove_if_unused(deref);
}
break;
}
-
+ case nir_intrinsic_copy_deref:
default:
continue;
}