*/
struct constant_fold_state {
- void *mem_ctx;
- nir_function_impl *impl;
- bool progress;
+ nir_shader *shader;
+ unsigned execution_mode;
+ bool has_load_constant;
+ bool has_indirect_load_const;
};
static bool
-constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
+constant_fold_alu_instr(struct constant_fold_state *state, nir_alu_instr *instr)
{
- nir_const_value src[4];
+ nir_const_value src[NIR_MAX_VEC_COMPONENTS][NIR_MAX_VEC_COMPONENTS];
if (!instr->dest.dest.is_ssa)
return false;
+ /* In the case that any outputs/inputs have unsized types, then we need to
+ * guess the bit-size. In this case, the validator ensures that all
+ * bit-sizes match so we can just take the bit-size from first
+ * output/input with an unsized type. If all the outputs/inputs are sized
+ * then we don't need to guess the bit-size at all because the code we
+ * generate for constant opcodes in this case already knows the sizes of
+ * the types involved and does not need the provided bit-size for anything
+ * (although it still requires to receive a valid bit-size).
+ */
+ unsigned bit_size = 0;
+ if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
+ bit_size = instr->dest.dest.ssa.bit_size;
+
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
if (!instr->src[i].src.is_ssa)
return false;
+ if (bit_size == 0 &&
+ !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_types[i]))
+ bit_size = instr->src[i].src.ssa->bit_size;
+
nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
if (src_instr->type != nir_instr_type_load_const)
for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
j++) {
- src[i].u[j] = load_const->value.u[instr->src[i].swizzle[j]];
+ src[i][j] = load_const->value[instr->src[i].swizzle[j]];
}
/* We shouldn't have any source modifiers in the optimization loop. */
assert(!instr->src[i].abs && !instr->src[i].negate);
}
+ if (bit_size == 0)
+ bit_size = 32;
+
/* We shouldn't have any saturate modifiers in the optimization loop. */
assert(!instr->dest.saturate);
- nir_const_value dest =
- nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
- src);
+ nir_const_value dest[NIR_MAX_VEC_COMPONENTS];
+ nir_const_value *srcs[NIR_MAX_VEC_COMPONENTS];
+ memset(dest, 0, sizeof(dest));
+ for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; ++i)
+ srcs[i] = src[i];
+ nir_eval_const_opcode(instr->op, dest, instr->dest.dest.ssa.num_components,
+ bit_size, srcs, state->execution_mode);
nir_load_const_instr *new_instr =
- nir_load_const_instr_create(mem_ctx,
- instr->dest.dest.ssa.num_components);
+ nir_load_const_instr_create(state->shader,
+ instr->dest.dest.ssa.num_components,
+ instr->dest.dest.ssa.bit_size);
- new_instr->value = dest;
+ memcpy(new_instr->value, dest, sizeof(*new_instr->value) * new_instr->def.num_components);
nir_instr_insert_before(&instr->instr, &new_instr->instr);
}
static bool
-constant_fold_deref(nir_instr *instr, nir_deref_var *deref)
+constant_fold_intrinsic_instr(struct constant_fold_state *state, nir_intrinsic_instr *instr)
{
bool progress = false;
- for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
- if (tail->deref_type != nir_deref_type_array)
- continue;
-
- nir_deref_array *arr = nir_deref_as_array(tail);
-
- if (arr->deref_array_type == nir_deref_array_type_indirect &&
- arr->indirect.is_ssa &&
- arr->indirect.ssa->parent_instr->type == nir_instr_type_load_const) {
- nir_load_const_instr *indirect =
- nir_instr_as_load_const(arr->indirect.ssa->parent_instr);
-
- arr->base_offset += indirect->value.u[0];
-
- /* Clear out the source */
- nir_instr_rewrite_src(instr, &arr->indirect, nir_src_for_ssa(NULL));
-
- arr->deref_array_type = nir_deref_array_type_direct;
-
+ if ((instr->intrinsic == nir_intrinsic_demote_if ||
+ instr->intrinsic == nir_intrinsic_discard_if) &&
+ nir_src_is_const(instr->src[0])) {
+ if (nir_src_as_bool(instr->src[0])) {
+ nir_intrinsic_op op = instr->intrinsic == nir_intrinsic_discard_if ?
+ nir_intrinsic_discard :
+ nir_intrinsic_demote;
+ nir_intrinsic_instr *new_instr = nir_intrinsic_instr_create(state->shader, op);
+ nir_instr_insert_before(&instr->instr, &new_instr->instr);
+ nir_instr_remove(&instr->instr);
+ progress = true;
+ } else {
+ /* We're not discarding, just delete the instruction */
+ nir_instr_remove(&instr->instr);
progress = true;
}
- }
+ } else if (instr->intrinsic == nir_intrinsic_load_constant) {
+ state->has_load_constant = true;
- return progress;
-}
+ if (!nir_src_is_const(instr->src[0])) {
+ state->has_indirect_load_const = true;
+ return progress;
+ }
-static bool
-constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
-{
- bool progress = false;
+ unsigned offset = nir_src_as_uint(instr->src[0]);
+ unsigned base = nir_intrinsic_base(instr);
+ unsigned range = nir_intrinsic_range(instr);
+ assert(base + range <= state->shader->constant_data_size);
+
+ nir_instr *new_instr = NULL;
+ if (offset >= range) {
+ nir_ssa_undef_instr *undef =
+ nir_ssa_undef_instr_create(state->shader,
+ instr->num_components,
+ instr->dest.ssa.bit_size);
+
+ nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&undef->def));
+ new_instr = &undef->instr;
+ } else {
+ nir_load_const_instr *load_const =
+ nir_load_const_instr_create(state->shader,
+ instr->num_components,
+ instr->dest.ssa.bit_size);
+
+ uint8_t *data = (uint8_t*)state->shader->constant_data + base;
+ for (unsigned i = 0; i < instr->num_components; i++) {
+ unsigned bytes = instr->dest.ssa.bit_size / 8;
+ bytes = MIN2(bytes, range - offset);
+
+ memcpy(&load_const->value[i].u64, data + offset, bytes);
+ offset += bytes;
+ }
+
+ nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&load_const->def));
+ new_instr = &load_const->instr;
+ }
- unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables;
- for (unsigned i = 0; i < num_vars; i++) {
- progress |= constant_fold_deref(&instr->instr, instr->variables[i]);
+ nir_instr_insert_before(&instr->instr, new_instr);
+ nir_instr_remove(&instr->instr);
+ progress = true;
}
return progress;
}
static bool
-constant_fold_tex_instr(nir_tex_instr *instr)
+constant_fold_block(struct constant_fold_state *state, nir_block *block)
{
- if (instr->sampler)
- return constant_fold_deref(&instr->instr, instr->sampler);
- else
- return false;
-}
-
-static bool
-constant_fold_block(nir_block *block, void *void_state)
-{
- struct constant_fold_state *state = void_state;
+ bool progress = false;
- nir_foreach_instr_safe(block, instr) {
+ nir_foreach_instr_safe(instr, block) {
switch (instr->type) {
case nir_instr_type_alu:
- state->progress |= constant_fold_alu_instr(nir_instr_as_alu(instr),
- state->mem_ctx);
+ progress |= constant_fold_alu_instr(state, nir_instr_as_alu(instr));
break;
case nir_instr_type_intrinsic:
- state->progress |=
- constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr));
- break;
- case nir_instr_type_tex:
- state->progress |= constant_fold_tex_instr(nir_instr_as_tex(instr));
+ progress |=
+ constant_fold_intrinsic_instr(state, nir_instr_as_intrinsic(instr));
break;
default:
/* Don't know how to constant fold */
}
}
- return true;
+ return progress;
}
static bool
-nir_opt_constant_folding_impl(nir_function_impl *impl)
+nir_opt_constant_folding_impl(struct constant_fold_state *state, nir_function_impl *impl)
{
- struct constant_fold_state state;
-
- state.mem_ctx = ralloc_parent(impl);
- state.impl = impl;
- state.progress = false;
+ bool progress = false;
- nir_foreach_block(impl, constant_fold_block, &state);
+ nir_foreach_block(block, impl) {
+ progress |= constant_fold_block(state, block);
+ }
- if (state.progress)
+ if (progress) {
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
+ } else {
+ nir_metadata_preserve(impl, nir_metadata_all);
+ }
- return state.progress;
+ return progress;
}
bool
nir_opt_constant_folding(nir_shader *shader)
{
bool progress = false;
+ struct constant_fold_state state;
+ state.shader = shader;
+ state.execution_mode = shader->info.float_controls_execution_mode;
+ state.has_load_constant = false;
+ state.has_indirect_load_const = false;
- nir_foreach_function(shader, function) {
+ nir_foreach_function(function, shader) {
if (function->impl)
- progress |= nir_opt_constant_folding_impl(function->impl);
+ progress |= nir_opt_constant_folding_impl(&state, function->impl);
+ }
+
+ /* This doesn't free the constant data if there are no constant loads because
+ * the data might still be used but the loads have been lowered to load_ubo
+ */
+ if (state.has_load_constant && !state.has_indirect_load_const &&
+ shader->constant_data_size) {
+ ralloc_free(shader->constant_data);
+ shader->constant_data = NULL;
+ shader->constant_data_size = 0;
}
return progress;