*
* Support for local common subexpression elimination.
*
- * See Muchnik's Advanced Compiler Design and Implementation, section
+ * See Muchnick's Advanced Compiler Design and Implementation, section
* 13.1 (p378).
*/
case BRW_OPCODE_XOR:
case BRW_OPCODE_SHR:
case BRW_OPCODE_SHL:
- case BRW_OPCODE_RSR:
- case BRW_OPCODE_RSL:
case BRW_OPCODE_ASR:
+ case BRW_OPCODE_CMP:
+ case BRW_OPCODE_CMPN:
case BRW_OPCODE_ADD:
case BRW_OPCODE_MUL:
case BRW_OPCODE_FRC:
case BRW_OPCODE_LINE:
case BRW_OPCODE_PLN:
case BRW_OPCODE_MAD:
+ case BRW_OPCODE_LRP:
case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
+ case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
+ case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
case FS_OPCODE_CINTERP:
case FS_OPCODE_LINTERP:
return true;
}
static bool
-operands_match(fs_reg *xs, fs_reg *ys)
+is_expression_commutative(enum opcode op)
{
- return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
+ 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(enum opcode op, fs_reg *xs, fs_reg *ys)
+{
+ if (!is_expression_commutative(op)) {
+ return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
+ } else {
+ return (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
+ (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
+ }
+}
+
+static bool
+instructions_match(fs_inst *a, fs_inst *b)
+{
+ return a->opcode == b->opcode &&
+ a->saturate == b->saturate &&
+ a->predicate == b->predicate &&
+ a->predicate_inverse == b->predicate_inverse &&
+ a->conditional_mod == b->conditional_mod &&
+ a->dst.type == b->dst.type &&
+ operands_match(a->opcode, a->src, b->src);
}
bool
{
bool progress = false;
- void *mem_ctx = ralloc_context(this->mem_ctx);
+ void *cse_ctx = ralloc_context(NULL);
+ int ip = block->start_ip;
for (fs_inst *inst = (fs_inst *)block->start;
inst != block->end->next;
inst = (fs_inst *) inst->next) {
/* Skip some cases. */
- if (is_expression(inst) && !inst->predicate && inst->mlen == 0 &&
- !inst->force_uncompressed && !inst->force_sechalf &&
- !inst->conditional_mod)
+ if (is_expression(inst) && !inst->is_partial_write() &&
+ (inst->dst.file != HW_REG || inst->dst.is_null()))
{
bool found = false;
entry = (aeb_entry *) entry_node;
/* Match current instruction's expression against those in AEB. */
- if (inst->opcode == entry->generator->opcode &&
- inst->saturate == entry->generator->saturate &&
- inst->dst.type == entry->generator->dst.type &&
- operands_match(entry->generator->src, inst->src)) {
-
+ if (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(mem_ctx, aeb_entry);
+ aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
entry->tmp = reg_undef;
entry->generator = inst;
aeb->push_tail(entry);
* If we don't have a temporary already, make one.
*/
bool no_existing_temp = entry->tmp.file == BAD_FILE;
- if (no_existing_temp) {
- entry->tmp = fs_reg(this, glsl_type::float_type);
- entry->tmp.type = inst->dst.type;
-
- fs_inst *copy = new(ralloc_parent(inst))
- fs_inst(BRW_OPCODE_MOV, entry->generator->dst, entry->tmp);
- entry->generator->insert_after(copy);
- entry->generator->dst = entry->tmp;
+ if (no_existing_temp && !entry->generator->dst.is_null()) {
+ int written = entry->generator->regs_written;
+
+ fs_reg orig_dst = entry->generator->dst;
+ fs_reg tmp = fs_reg(GRF, virtual_grf_alloc(written),
+ orig_dst.type);
+ entry->tmp = tmp;
+ entry->generator->dst = tmp;
+
+ for (int i = 0; i < written; i++) {
+ fs_inst *copy = MOV(orig_dst, tmp);
+ copy->force_writemask_all =
+ entry->generator->force_writemask_all;
+ entry->generator->insert_after(copy);
+
+ orig_dst.reg_offset++;
+ tmp.reg_offset++;
+ }
}
/* dest <- temp */
- assert(inst->dst.type == entry->tmp.type);
- fs_inst *copy = new(ralloc_parent(inst))
- fs_inst(BRW_OPCODE_MOV, inst->dst, entry->tmp);
- copy->force_writemask_all = inst->force_writemask_all;
- inst->replace_with(copy);
+ if (!inst->dst.is_null()) {
+ int written = inst->regs_written;
+ assert(written == entry->generator->regs_written);
+ assert(inst->dst.type == entry->tmp.type);
+ fs_reg dst = inst->dst;
+ fs_reg tmp = entry->tmp;
+ fs_inst *copy = NULL;
+ for (int i = 0; i < written; i++) {
+ copy = MOV(dst, tmp);
+ copy->force_writemask_all = inst->force_writemask_all;
+ inst->insert_before(copy);
+
+ dst.reg_offset++;
+ tmp.reg_offset++;
+ }
+ }
+
+ /* Set our iterator so that next time through the loop inst->next
+ * will get the instruction in the basic block after the one we've
+ * removed.
+ */
+ fs_inst *prev = (fs_inst *)inst->prev;
+
+ inst->remove();
/* Appending an instruction may have changed our bblock end. */
if (inst == block->end) {
- block->end = copy;
+ block->end = prev;
}
- /* Continue iteration with copy->next */
- inst = copy;
+ inst = prev;
}
}
- /* Kill all AEB entries that use the destination. */
foreach_list_safe(entry_node, aeb) {
aeb_entry *entry = (aeb_entry *)entry_node;
+ /* Kill all AEB entries that write a different value to or read from
+ * the flag register if we just wrote it.
+ */
+ if (inst->writes_flag()) {
+ if (entry->generator->reads_flag() ||
+ (entry->generator->writes_flag() &&
+ !instructions_match(inst, entry->generator))) {
+ entry->remove();
+ ralloc_free(entry);
+ continue;
+ }
+ }
+
for (int i = 0; i < 3; i++) {
+ fs_reg *src_reg = &entry->generator->src[i];
+
+ /* Kill all AEB entries that use the destination we just
+ * overwrote.
+ */
if (inst->overwrites_reg(entry->generator->src[i])) {
entry->remove();
ralloc_free(entry);
break;
}
+
+ /* Kill any AEB entries using registers that don't get reused any
+ * more -- a sure sign they'll fail operands_match().
+ */
+ if (src_reg->file == GRF && virtual_grf_end[src_reg->reg] < ip) {
+ entry->remove();
+ ralloc_free(entry);
+ break;
+ }
}
}
+
+ ip++;
}
- ralloc_free(mem_ctx);
+ ralloc_free(cse_ctx);
if (progress)
- this->live_intervals_valid = false;
+ invalidate_live_intervals();
return progress;
}
{
bool progress = false;
- cfg_t cfg(this);
+ calculate_live_intervals();
+
+ cfg_t cfg(&instructions);
for (int b = 0; b < cfg.num_blocks; b++) {
bblock_t *block = cfg.blocks[b];
projects / mesa.git / blobdiff