is_expression(const fs_inst *const inst)
{
switch (inst->opcode) {
+ case BRW_OPCODE_MOV:
case BRW_OPCODE_SEL:
case BRW_OPCODE_NOT:
case BRW_OPCODE_AND:
case SHADER_OPCODE_INT_REMAINDER:
case SHADER_OPCODE_SIN:
case SHADER_OPCODE_COS:
- return inst->mlen == 0;
+ return inst->mlen < 2;
case SHADER_OPCODE_LOAD_PAYLOAD:
return !is_copy_payload(inst);
default:
- return inst->is_tex();
+ return inst->is_send_from_grf() && !inst->has_side_effects();
}
}
static bool
-is_expression_commutative(enum opcode op)
-{
- switch (op) {
- case BRW_OPCODE_AND:
- case BRW_OPCODE_OR:
- case BRW_OPCODE_XOR:
- case BRW_OPCODE_ADD:
- case BRW_OPCODE_MUL:
- return true;
- default:
- return false;
- }
-}
-
-static bool
-operands_match(fs_inst *a, fs_inst *b)
+operands_match(const fs_inst *a, const fs_inst *b)
{
fs_reg *xs = a->src;
fs_reg *ys = b->src;
- if (!is_expression_commutative(a->opcode)) {
+ if (a->opcode == BRW_OPCODE_MAD) {
+ return xs[0].equals(ys[0]) &&
+ ((xs[1].equals(ys[1]) && xs[2].equals(ys[2])) ||
+ (xs[2].equals(ys[1]) && xs[1].equals(ys[2])));
+ } else if (!a->is_commutative()) {
bool match = true;
for (int i = 0; i < a->sources; i++) {
if (!xs[i].equals(ys[i])) {
a->conditional_mod == b->conditional_mod &&
a->dst.type == b->dst.type &&
a->sources == b->sources &&
- (a->is_tex() ? (a->texture_offset == b->texture_offset &&
+ (a->is_tex() ? (a->offset == b->offset &&
a->mlen == b->mlen &&
a->regs_written == b->regs_written &&
a->base_mrf == b->base_mrf &&
- a->sampler == b->sampler &&
a->eot == b->eot &&
a->header_present == b->header_present &&
a->shadow_compare == b->shadow_compare)
foreach_in_list_use_after(aeb_entry, entry, &aeb) {
/* Match current instruction's expression against those in AEB. */
- if (instructions_match(inst, entry->generator)) {
+ if (!(entry->generator->dst.is_null() && !inst->dst.is_null()) &&
+ instructions_match(inst, entry->generator)) {
found = true;
progress = true;
break;
}
if (!found) {
- /* Our first sighting of this expression. Create an entry. */
- aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
- entry->tmp = reg_undef;
- entry->generator = inst;
- aeb.push_tail(entry);
+ if (inst->opcode != BRW_OPCODE_MOV ||
+ (inst->opcode == BRW_OPCODE_MOV &&
+ inst->src[0].file == IMM &&
+ inst->src[0].type == BRW_REGISTER_TYPE_VF)) {
+ /* Our first sighting of this expression. Create an entry. */
+ aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
+ entry->tmp = reg_undef;
+ entry->generator = inst;
+ aeb.push_tail(entry);
+ }
} else {
/* This is at least our second sighting of this expression.
* If we don't have a temporary already, make one.
bool no_existing_temp = entry->tmp.file == BAD_FILE;
if (no_existing_temp && !entry->generator->dst.is_null()) {
int written = entry->generator->regs_written;
+ int dst_width = entry->generator->dst.width / 8;
+ assert(written % dst_width == 0);
fs_reg orig_dst = entry->generator->dst;
- fs_reg tmp = fs_reg(GRF, virtual_grf_alloc(written),
- orig_dst.type);
+ fs_reg tmp = fs_reg(GRF, alloc.allocate(written),
+ orig_dst.type, orig_dst.width);
entry->tmp = tmp;
entry->generator->dst = tmp;
fs_inst *copy;
- if (written > 1) {
- fs_reg *sources = ralloc_array(mem_ctx, fs_reg, written);
- for (int i = 0; i < written; i++) {
- sources[i] = tmp;
- sources[i].reg_offset = i;
- }
- copy = LOAD_PAYLOAD(orig_dst, sources, written);
+ if (written > dst_width) {
+ fs_reg *sources = ralloc_array(mem_ctx, fs_reg, written / dst_width);
+ for (int i = 0; i < written / dst_width; i++)
+ sources[i] = offset(tmp, i);
+ copy = LOAD_PAYLOAD(orig_dst, sources, written / dst_width);
} else {
copy = MOV(orig_dst, tmp);
copy->force_writemask_all =
entry->generator->force_writemask_all;
}
- entry->generator->insert_after(copy);
+ entry->generator->insert_after(block, copy);
}
/* dest <- temp */
if (!inst->dst.is_null()) {
int written = inst->regs_written;
+ int dst_width = inst->dst.width / 8;
assert(written == entry->generator->regs_written);
+ assert(dst_width == entry->generator->dst.width / 8);
assert(inst->dst.type == entry->tmp.type);
fs_reg dst = inst->dst;
fs_reg tmp = entry->tmp;
fs_inst *copy;
- if (written > 1) {
- fs_reg *sources = ralloc_array(mem_ctx, fs_reg, written);
- for (int i = 0; i < written; i++) {
- sources[i] = tmp;
- sources[i].reg_offset = i;
- }
- copy = LOAD_PAYLOAD(dst, sources, written);
+ if (written > dst_width) {
+ fs_reg *sources = ralloc_array(mem_ctx, fs_reg, written / dst_width);
+ for (int i = 0; i < written / dst_width; i++)
+ sources[i] = offset(tmp, i);
+ copy = LOAD_PAYLOAD(dst, sources, written / dst_width);
} else {
copy = MOV(dst, tmp);
copy->force_writemask_all = inst->force_writemask_all;
}
- inst->insert_before(copy);
+ inst->insert_before(block, copy);
}
/* Set our iterator so that next time through the loop inst->next
*/
fs_inst *prev = (fs_inst *)inst->prev;
- inst->remove();
-
- /* Appending an instruction may have changed our bblock end. */
- if (inst == block->end) {
- block->end = prev;
- }
-
+ inst->remove(block);
inst = prev;
}
}
ralloc_free(cse_ctx);
- if (progress)
- invalidate_live_intervals();
-
return progress;
}
{
bool progress = false;
- cfg_t cfg(&instructions);
- calculate_live_intervals(&cfg);
-
- for (int b = 0; b < cfg.num_blocks; b++) {
- bblock_t *block = cfg.blocks[b];
+ calculate_live_intervals();
+ foreach_block (block, cfg) {
progress = opt_cse_local(block) || progress;
}
+ if (progress)
+ invalidate_live_intervals();
+
return progress;
}