*
*/
+#include <inttypes.h>
#include "nir_search.h"
struct match_state {
+ bool inexact_match;
+ bool has_exact_alu;
unsigned variables_seen;
nir_alu_src variables[NIR_SEARCH_MAX_VARIABLES];
};
case nir_op_inot:
return src_is_bool(instr->src[0].src);
default:
- return nir_op_infos[instr->op].output_type == nir_type_bool;
+ return (nir_alu_type_get_base_type(nir_op_infos[instr->op].output_type)
+ == nir_type_bool);
}
}
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)))
+ if (nir_alu_type_get_base_type(nir_op_infos[src_alu->op].output_type) !=
+ var->type &&
+ !(nir_alu_type_get_base_type(var->type) == nir_type_bool &&
+ alu_instr_is_bool(src_alu)))
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]) {
+ switch (const_val->type) {
case nir_type_float:
for (unsigned i = 0; i < num_components; ++i) {
- if (load->value.f[new_swizzle[i]] != const_val->data.f)
+ double val;
+ switch (load->def.bit_size) {
+ case 32:
+ val = load->value.f32[new_swizzle[i]];
+ break;
+ case 64:
+ val = load->value.f64[new_swizzle[i]];
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
+
+ if (val != const_val->data.d)
return false;
}
return true;
+
case nir_type_int:
+ for (unsigned i = 0; i < num_components; ++i) {
+ int64_t val;
+ switch (load->def.bit_size) {
+ case 32:
+ val = load->value.i32[new_swizzle[i]];
+ break;
+ case 64:
+ val = load->value.i64[new_swizzle[i]];
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
+
+ if (val != const_val->data.i)
+ return false;
+ }
+ return true;
+
case nir_type_uint:
- case nir_type_bool:
+ case nir_type_bool32:
for (unsigned i = 0; i < num_components; ++i) {
- if (load->value.i[new_swizzle[i]] != const_val->data.i)
+ uint64_t val;
+ switch (load->def.bit_size) {
+ case 32:
+ val = load->value.u32[new_swizzle[i]];
+ break;
+ case 64:
+ val = load->value.u64[new_swizzle[i]];
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
+
+ if (val != const_val->data.u)
return false;
}
return true;
+
default:
unreachable("Invalid alu source type");
}
if (instr->op != expr->opcode)
return false;
+ assert(instr->dest.dest.is_ssa);
+
+ state->inexact_match = expr->inexact || state->inexact_match;
+ state->has_exact_alu = instr->exact || state->has_exact_alu;
+ if (state->inexact_match && state->has_exact_alu)
+ return false;
+
assert(!instr->dest.saturate);
assert(nir_op_infos[instr->op].num_inputs > 0);
}
}
+typedef struct bitsize_tree {
+ unsigned num_srcs;
+ struct bitsize_tree *srcs[4];
+
+ unsigned common_size;
+ bool is_src_sized[4];
+ bool is_dest_sized;
+
+ unsigned dest_size;
+ unsigned src_size[4];
+} bitsize_tree;
+
+static bitsize_tree *
+build_bitsize_tree(void *mem_ctx, struct match_state *state,
+ const nir_search_value *value)
+{
+ bitsize_tree *tree = ralloc(mem_ctx, bitsize_tree);
+
+ switch (value->type) {
+ case nir_search_value_expression: {
+ nir_search_expression *expr = nir_search_value_as_expression(value);
+ nir_op_info info = nir_op_infos[expr->opcode];
+ tree->num_srcs = info.num_inputs;
+ tree->common_size = 0;
+ for (unsigned i = 0; i < info.num_inputs; i++) {
+ tree->is_src_sized[i] = !!nir_alu_type_get_type_size(info.input_types[i]);
+ if (tree->is_src_sized[i])
+ tree->src_size[i] = nir_alu_type_get_type_size(info.input_types[i]);
+ tree->srcs[i] = build_bitsize_tree(mem_ctx, state, expr->srcs[i]);
+ }
+ tree->is_dest_sized = !!nir_alu_type_get_type_size(info.output_type);
+ if (tree->is_dest_sized)
+ tree->dest_size = nir_alu_type_get_type_size(info.output_type);
+ break;
+ }
+
+ case nir_search_value_variable: {
+ nir_search_variable *var = nir_search_value_as_variable(value);
+ tree->num_srcs = 0;
+ tree->is_dest_sized = true;
+ tree->dest_size = nir_src_bit_size(state->variables[var->variable].src);
+ break;
+ }
+
+ case nir_search_value_constant: {
+ tree->num_srcs = 0;
+ tree->is_dest_sized = false;
+ tree->common_size = 0;
+ break;
+ }
+ }
+
+ return tree;
+}
+
+static unsigned
+bitsize_tree_filter_up(bitsize_tree *tree)
+{
+ for (unsigned i = 0; i < tree->num_srcs; i++) {
+ unsigned src_size = bitsize_tree_filter_up(tree->srcs[i]);
+ if (src_size == 0)
+ continue;
+
+ if (tree->is_src_sized[i]) {
+ assert(src_size == tree->src_size[i]);
+ } else if (tree->common_size != 0) {
+ assert(src_size == tree->common_size);
+ tree->src_size[i] = src_size;
+ } else {
+ tree->common_size = src_size;
+ tree->src_size[i] = src_size;
+ }
+ }
+
+ if (tree->num_srcs && tree->common_size) {
+ if (tree->dest_size == 0)
+ tree->dest_size = tree->common_size;
+ else if (!tree->is_dest_sized)
+ assert(tree->dest_size == tree->common_size);
+
+ for (unsigned i = 0; i < tree->num_srcs; i++) {
+ if (!tree->src_size[i])
+ tree->src_size[i] = tree->common_size;
+ }
+ }
+
+ return tree->dest_size;
+}
+
+static void
+bitsize_tree_filter_down(bitsize_tree *tree, unsigned size)
+{
+ if (tree->dest_size)
+ assert(tree->dest_size == size);
+ else
+ tree->dest_size = size;
+
+ if (!tree->is_dest_sized) {
+ if (tree->common_size)
+ assert(tree->common_size == size);
+ else
+ tree->common_size = size;
+ }
+
+ for (unsigned i = 0; i < tree->num_srcs; i++) {
+ if (!tree->src_size[i]) {
+ assert(tree->common_size);
+ tree->src_size[i] = tree->common_size;
+ }
+ bitsize_tree_filter_down(tree->srcs[i], tree->src_size[i]);
+ }
+}
+
static nir_alu_src
-construct_value(const nir_search_value *value, nir_alu_type type,
- unsigned num_components, struct match_state *state,
+construct_value(const nir_search_value *value,
+ unsigned num_components, bitsize_tree *bitsize,
+ struct match_state *state,
nir_instr *instr, void *mem_ctx)
{
switch (value->type) {
num_components = nir_op_infos[expr->opcode].output_size;
nir_alu_instr *alu = nir_alu_instr_create(mem_ctx, expr->opcode);
- nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components, NULL);
+ nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components,
+ bitsize->dest_size, NULL);
alu->dest.write_mask = (1 << num_components) - 1;
alu->dest.saturate = false;
+ /* We have no way of knowing what values in a given search expression
+ * map to a particular replacement value. Therefore, if the
+ * expression we are replacing has any exact values, the entire
+ * replacement should be exact.
+ */
+ alu->exact = state->has_exact_alu;
+
for (unsigned i = 0; i < nir_op_infos[expr->opcode].num_inputs; i++) {
/* If the source is an explicitly sized source, then we need to reset
* the number of components to match.
num_components = nir_op_infos[alu->op].input_sizes[i];
alu->src[i] = construct_value(expr->srcs[i],
- nir_op_infos[alu->op].input_types[i],
- num_components,
+ num_components, bitsize->srcs[i],
state, instr, mem_ctx);
}
const nir_search_constant *c = nir_search_value_as_constant(value);
nir_load_const_instr *load = nir_load_const_instr_create(mem_ctx, 1);
- switch (type) {
+ switch (c->type) {
case nir_type_float:
- load->def.name = ralloc_asprintf(mem_ctx, "%f", c->data.f);
- load->value.f[0] = c->data.f;
+ load->def.name = ralloc_asprintf(load, "%f", c->data.d);
+ switch (bitsize->dest_size) {
+ case 32:
+ load->value.f32[0] = c->data.d;
+ break;
+ case 64:
+ load->value.f64[0] = c->data.d;
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
break;
+
case nir_type_int:
- load->def.name = ralloc_asprintf(mem_ctx, "%d", c->data.i);
- load->value.i[0] = c->data.i;
+ load->def.name = ralloc_asprintf(load, "%" PRIi64, c->data.i);
+ switch (bitsize->dest_size) {
+ case 32:
+ load->value.i32[0] = c->data.i;
+ break;
+ case 64:
+ load->value.i64[0] = c->data.i;
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
break;
+
case nir_type_uint:
- case nir_type_bool:
- load->value.u[0] = c->data.u;
+ load->def.name = ralloc_asprintf(load, "%" PRIu64, c->data.u);
+ switch (bitsize->dest_size) {
+ case 32:
+ load->value.u32[0] = c->data.u;
+ break;
+ case 64:
+ load->value.u64[0] = c->data.u;
+ break;
+ default:
+ unreachable("unknown bit size");
+ }
+
+ case nir_type_bool32:
+ load->value.u32[0] = c->data.u;
break;
default:
unreachable("Invalid alu source type");
}
+ load->def.bit_size = bitsize->dest_size;
+
nir_instr_insert_before(instr, &load->instr);
nir_alu_src val;
assert(instr->dest.dest.is_ssa);
struct match_state state;
+ state.inexact_match = false;
+ state.has_exact_alu = false;
state.variables_seen = 0;
if (!match_expression(search, instr, instr->dest.dest.ssa.num_components,
swizzle, &state))
return NULL;
+ void *bitsize_ctx = ralloc_context(NULL);
+ bitsize_tree *tree = build_bitsize_tree(bitsize_ctx, &state, replace);
+ bitsize_tree_filter_up(tree);
+ bitsize_tree_filter_down(tree, instr->dest.dest.ssa.bit_size);
+
/* Inserting a mov may be unnecessary. However, it's much easier to
* simply let copy propagation clean this up than to try to go through
* and rewrite swizzles ourselves.
nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov);
mov->dest.write_mask = instr->dest.write_mask;
nir_ssa_dest_init(&mov->instr, &mov->dest.dest,
- instr->dest.dest.ssa.num_components, NULL);
+ instr->dest.dest.ssa.num_components,
+ instr->dest.dest.ssa.bit_size, NULL);
- mov->src[0] = construct_value(replace, nir_op_infos[instr->op].output_type,
- instr->dest.dest.ssa.num_components, &state,
- &instr->instr, mem_ctx);
+ mov->src[0] = construct_value(replace,
+ instr->dest.dest.ssa.num_components, tree,
+ &state, &instr->instr, mem_ctx);
nir_instr_insert_before(&instr->instr, &mov->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
*/
nir_instr_remove(&instr->instr);
+ ralloc_free(bitsize_ctx);
+
return mov;
}
projects / mesa.git / blobdiff