static bool
constant_fold_deref(nir_instr *instr, nir_deref_var *deref)
{
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;
progress = true;
}
}
return progress;
}
static bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
nir_const_value src[4];
if (!instr->dest.dest.is_ssa)
return false;
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
if (!instr->src[i].src.is_ssa)
return false;
nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
if (src_instr->type != nir_instr_type_load_const)
return false;
nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
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]];
}
/* We shouldn't have any source modifiers in the optimization loop. */
assert(!instr->src[i].abs && !instr->src[i].negate);
}
/* 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_load_const_instr *new_instr =
nir_load_const_instr_create(mem_ctx,
instr->dest.dest.ssa.num_components);
new_instr->value = dest;
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(&new_instr->def),
mem_ctx);
nir_instr_remove(&instr->instr);
ralloc_free(instr);
return true;
}
static void
validate_instr(nir_instr *instr, validate_state *state)
{
assert(instr->block == state->block);
state->instr = instr;
switch (instr->type) {
case nir_instr_type_alu:
validate_alu_instr(nir_instr_as_alu(instr), state);
break;
case nir_instr_type_call:
validate_call_instr(nir_instr_as_call(instr), state);
break;
case nir_instr_type_intrinsic:
validate_intrinsic_instr(nir_instr_as_intrinsic(instr), state);
break;
case nir_instr_type_tex:
validate_tex_instr(nir_instr_as_tex(instr), state);
break;
case nir_instr_type_load_const:
validate_load_const_instr(nir_instr_as_load_const(instr), state);
break;
case nir_instr_type_phi:
validate_phi_instr(nir_instr_as_phi(instr), state);
break;
case nir_instr_type_ssa_undef:
validate_ssa_undef_instr(nir_instr_as_ssa_undef(instr), state);
break;
case nir_instr_type_jump:
break;
default:
assert(!"Invalid ALU instruction type");
break;
}
state->instr = NULL;
}
nir_foreach_instr(block, instr) {
switch (instr->type) {
case nir_instr_type_alu: {
/* For ALU instructions, the resolve status is handled in a
* three-step process.
*
* 1) Look at the instruction type and sources and determine if it
* can be left unresolved.
*
* 2) Look at the destination and see if we have to resolve
* anyway. (This is the case if this instruction is not the
* only instruction writing to a given register.)
*
* 3) If the instruction has a resolve status other than
* BOOL_UNRESOLVED or BOOL_NEEDS_RESOLVE then we walk through
* the sources and ensure that they are also resolved. This
* ensures that we don't end up with any stray unresolved
* booleans going into ADDs or something like that.
*/
uint8_t resolve_status;
nir_alu_instr *alu = nir_instr_as_alu(instr);
switch (alu->op) {
case nir_op_bany2:
case nir_op_bany3:
case nir_op_bany4:
case nir_op_ball_fequal2:
case nir_op_ball_iequal2:
case nir_op_ball_fequal3:
case nir_op_ball_iequal3:
case nir_op_ball_fequal4:
case nir_op_ball_iequal4:
case nir_op_bany_fnequal2:
case nir_op_bany_inequal2:
case nir_op_bany_fnequal3:
case nir_op_bany_inequal3:
case nir_op_bany_fnequal4:
case nir_op_bany_inequal4:
/* These are only implemented by the vec4 backend and its
* implementation emits resolved booleans. At some point in the
* future, this may change and we'll have to remove some of the
* above cases.
*/
resolve_status = BRW_NIR_BOOLEAN_NO_RESOLVE;
break;
case nir_op_imov:
case nir_op_inot:
/* This is a single-source instruction. Just copy the resolve
* status from the source.
*/
resolve_status = get_resolve_status_for_src(&alu->src[0].src);
break;
case nir_op_iand:
case nir_op_ior:
case nir_op_ixor: {
uint8_t src0_status = get_resolve_status_for_src(&alu->src[0].src);
uint8_t src1_status = get_resolve_status_for_src(&alu->src[1].src);
if (src0_status == src1_status) {
resolve_status = src0_status;
} else if (src0_status == BRW_NIR_NON_BOOLEAN ||
src1_status == BRW_NIR_NON_BOOLEAN) {
/* If one of the sources is a non-boolean then the whole
* thing is a non-boolean.
*/
resolve_status = BRW_NIR_NON_BOOLEAN;
} else {
/* At this point one of them is a true boolean and one is a
* boolean that needs a resolve. We could either resolve the
* unresolved source or we could resolve here. If we resolve
* the unresolved source then we get two resolves for the price
* of one. Just set this one to BOOLEAN_NO_RESOLVE and we'll
* let the code below force a resolve on the unresolved source.
*/
resolve_status = BRW_NIR_BOOLEAN_NO_RESOLVE;
}
break;
}
default:
if (nir_op_infos[alu->op].output_type == nir_type_bool) {
/* This instructions will turn into a CMP when we actually emit
* them so the result will have to be resolved before it can be
* used.
*/
resolve_status = BRW_NIR_BOOLEAN_UNRESOLVED;
/* Even though the destination is allowed to be left
* unresolved, the sources are treated as regular integers or
* floats so they need to be resolved.
*/
nir_foreach_src(instr, src_mark_needs_resolve, NULL);
} else {
resolve_status = BRW_NIR_NON_BOOLEAN;
}
}
/* If the destination is SSA, go ahead allow unresolved booleans.
* If the destination register doesn't have a well-defined parent_instr
* we need to resolve immediately.
*/
if (!alu->dest.dest.is_ssa &&
resolve_status == BRW_NIR_BOOLEAN_UNRESOLVED) {
resolve_status = BRW_NIR_BOOLEAN_NEEDS_RESOLVE;
}
instr->pass_flags = (instr->pass_flags & ~BRW_NIR_BOOLEAN_MASK) |
resolve_status;
/* Finally, resolve sources if it's needed */
switch (resolve_status) {
case BRW_NIR_BOOLEAN_NEEDS_RESOLVE:
case BRW_NIR_BOOLEAN_UNRESOLVED:
/* This instruction is either unresolved or we're doing the
* resolve here; leave the sources alone.
*/
break;
case BRW_NIR_BOOLEAN_NO_RESOLVE:
case BRW_NIR_NON_BOOLEAN:
nir_foreach_src(instr, src_mark_needs_resolve, NULL);
break;
default:
unreachable("Invalid boolean flag");
}
break;
}
case nir_instr_type_load_const: {
nir_load_const_instr *load = nir_instr_as_load_const(instr);
/* For load_const instructions, it's a boolean exactly when it holds
* one of the values NIR_TRUE or NIR_FALSE.
*
* Since load_const instructions don't have any sources, we don't
* have to worry about resolving them.
*/
instr->pass_flags &= ~BRW_NIR_BOOLEAN_MASK;
if (load->value.u[0] == NIR_TRUE || load->value.u[0] == NIR_FALSE) {
instr->pass_flags |= BRW_NIR_BOOLEAN_NO_RESOLVE;
} else {
instr->pass_flags |= BRW_NIR_NON_BOOLEAN;
}
continue;
}
default:
/* Everything else is an unknown non-boolean value and needs to
* have all sources resolved.
*/
instr->pass_flags = (instr->pass_flags & ~BRW_NIR_BOOLEAN_MASK) |
BRW_NIR_NON_BOOLEAN;
nir_foreach_src(instr, src_mark_needs_resolve, NULL);
continue;
}
}
static bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
nir_const_value src[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_sizes[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)
return false;
nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
j++) {
switch(load_const->def.bit_size) {
case 64:
src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
break;
case 32:
src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
break;
case 16:
src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
break;
case 8:
src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
break;
default:
unreachable("Invalid bit size");
}
}
/* 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,
bit_size, src);
nir_load_const_instr *new_instr =
nir_load_const_instr_create(mem_ctx,
instr->dest.dest.ssa.num_components,
instr->dest.dest.ssa.bit_size);
new_instr->value = dest;
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
nir_src_for_ssa(&new_instr->def));
nir_instr_remove(&instr->instr);
ralloc_free(instr);
return true;
}
nir_foreach_instr_safe(block, instr) {
if (instr->type == nir_instr_type_load_const)
lower_load_const_instr_scalar(nir_instr_as_load_const(instr));
}
static bool
match_value(const nir_search_value *value, nir_alu_instr *instr, unsigned src,
unsigned num_components, const uint8_t *swizzle,
struct match_state *state)
{
uint8_t new_swizzle[4];
/* If the source is an explicitly sized source, then we need to reset
* both the number of components and the swizzle.
*/
if (nir_op_infos[instr->op].input_sizes[src] != 0) {
num_components = nir_op_infos[instr->op].input_sizes[src];
swizzle = identity_swizzle;
}
for (int i = 0; i < num_components; ++i)
new_swizzle[i] = instr->src[src].swizzle[swizzle[i]];
switch (value->type) {
case nir_search_value_expression:
if (!instr->src[src].src.is_ssa)
return false;
if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu)
return false;
return match_expression(nir_search_value_as_expression(value),
nir_instr_as_alu(instr->src[src].src.ssa->parent_instr),
num_components, new_swizzle, state);
case nir_search_value_variable: {
nir_search_variable *var = nir_search_value_as_variable(value);
assert(var->variable < NIR_SEARCH_MAX_VARIABLES);
if (state->variables_seen & (1 << var->variable)) {
if (!nir_srcs_equal(state->variables[var->variable].src,
instr->src[src].src))
return false;
assert(!instr->src[src].abs && !instr->src[src].negate);
for (int i = 0; i < num_components; ++i) {
if (state->variables[var->variable].swizzle[i] != new_swizzle[i])
return false;
}
return true;
} else {
if (var->is_constant &&
instr->src[src].src.ssa->parent_instr->type != nir_instr_type_load_const)
return false;
if (var->type != nir_type_invalid) {
if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu)
return false;
nir_alu_instr *src_alu =
nir_instr_as_alu(instr->src[src].src.ssa->parent_instr);
if (nir_op_infos[src_alu->op].output_type != var->type &&
!(var->type == nir_type_bool && alu_instr_is_bool(src_alu)))
return false;
}
state->variables_seen |= (1 << var->variable);
state->variables[var->variable].src = instr->src[src].src;
state->variables[var->variable].abs = false;
state->variables[var->variable].negate = false;
for (int i = 0; i < 4; ++i) {
if (i < num_components)
state->variables[var->variable].swizzle[i] = new_swizzle[i];
else
state->variables[var->variable].swizzle[i] = 0;
}
return true;
}
}
case nir_search_value_constant: {
nir_search_constant *const_val = nir_search_value_as_constant(value);
if (!instr->src[src].src.is_ssa)
return false;
if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_load_const)
return false;
nir_load_const_instr *load =
nir_instr_as_load_const(instr->src[src].src.ssa->parent_instr);
switch (nir_op_infos[instr->op].input_types[src]) {
case nir_type_float:
for (unsigned i = 0; i < num_components; ++i) {
if (load->value.f[new_swizzle[i]] != const_val->data.f)
return false;
}
return true;
case nir_type_int:
case nir_type_unsigned:
case nir_type_bool:
for (unsigned i = 0; i < num_components; ++i) {
if (load->value.i[new_swizzle[i]] != const_val->data.i)
return false;
}
return true;
default:
unreachable("Invalid alu source type");
}
}
default:
unreachable("Invalid search value type");
}
}