*
* 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).
*/
+using namespace brw;
+
namespace {
struct aeb_entry : public exec_node {
/** The instruction that generates the expression value. */
}
static bool
-is_expression(const fs_inst *const inst)
+is_expression(const fs_visitor *v, 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 BRW_OPCODE_SHR:
case BRW_OPCODE_SHL:
case BRW_OPCODE_ASR:
+ case BRW_OPCODE_CMP:
+ case BRW_OPCODE_CMPN:
case BRW_OPCODE_ADD:
case BRW_OPCODE_MUL:
+ case SHADER_OPCODE_MULH:
case BRW_OPCODE_FRC:
case BRW_OPCODE_RNDU:
case BRW_OPCODE_RNDD:
case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
case FS_OPCODE_CINTERP:
case FS_OPCODE_LINTERP:
+ case SHADER_OPCODE_FIND_LIVE_CHANNEL:
+ case SHADER_OPCODE_BROADCAST:
return true;
+ case SHADER_OPCODE_RCP:
+ case SHADER_OPCODE_RSQ:
+ case SHADER_OPCODE_SQRT:
+ case SHADER_OPCODE_EXP2:
+ case SHADER_OPCODE_LOG2:
+ case SHADER_OPCODE_POW:
+ case SHADER_OPCODE_INT_QUOTIENT:
+ case SHADER_OPCODE_INT_REMAINDER:
+ case SHADER_OPCODE_SIN:
+ case SHADER_OPCODE_COS:
+ return inst->mlen < 2;
+ case SHADER_OPCODE_LOAD_PAYLOAD:
+ return !inst->is_copy_payload(v->alloc);
default:
- return false;
+ return inst->is_send_from_grf() && !inst->has_side_effects();
+ }
+}
+
+static bool
+operands_match(const fs_inst *a, const fs_inst *b, bool *negate)
+{
+ fs_reg *xs = a->src;
+ fs_reg *ys = b->src;
+
+ 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->opcode == BRW_OPCODE_MUL && a->dst.type == BRW_REGISTER_TYPE_F) {
+ bool xs0_negate = xs[0].negate;
+ bool xs1_negate = xs[1].file == IMM ? xs[1].fixed_hw_reg.dw1.f < 0.0f
+ : xs[1].negate;
+ bool ys0_negate = ys[0].negate;
+ bool ys1_negate = ys[1].file == IMM ? ys[1].fixed_hw_reg.dw1.f < 0.0f
+ : ys[1].negate;
+ float xs1_imm = xs[1].fixed_hw_reg.dw1.f;
+ float ys1_imm = ys[1].fixed_hw_reg.dw1.f;
+
+ xs[0].negate = false;
+ xs[1].negate = false;
+ ys[0].negate = false;
+ ys[1].negate = false;
+ xs[1].fixed_hw_reg.dw1.f = fabsf(xs[1].fixed_hw_reg.dw1.f);
+ ys[1].fixed_hw_reg.dw1.f = fabsf(ys[1].fixed_hw_reg.dw1.f);
+
+ bool ret = (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
+ (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
+
+ xs[0].negate = xs0_negate;
+ xs[1].negate = xs[1].file == IMM ? false : xs1_negate;
+ ys[0].negate = ys0_negate;
+ ys[1].negate = ys[1].file == IMM ? false : ys1_negate;
+ xs[1].fixed_hw_reg.dw1.f = xs1_imm;
+ ys[1].fixed_hw_reg.dw1.f = ys1_imm;
+
+ *negate = (xs0_negate != xs1_negate) != (ys0_negate != ys1_negate);
+ return ret;
+ } else if (!a->is_commutative()) {
+ bool match = true;
+ for (int i = 0; i < a->sources; i++) {
+ if (!xs[i].equals(ys[i])) {
+ match = false;
+ break;
+ }
+ }
+ return match;
+ } else {
+ return (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
+ (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
}
}
static bool
-operands_match(fs_reg *xs, fs_reg *ys)
+instructions_match(fs_inst *a, fs_inst *b, bool *negate)
+{
+ return a->opcode == b->opcode &&
+ a->force_writemask_all == b->force_writemask_all &&
+ a->exec_size == b->exec_size &&
+ a->force_sechalf == b->force_sechalf &&
+ a->saturate == b->saturate &&
+ a->predicate == b->predicate &&
+ a->predicate_inverse == b->predicate_inverse &&
+ a->conditional_mod == b->conditional_mod &&
+ a->flag_subreg == b->flag_subreg &&
+ a->dst.type == b->dst.type &&
+ a->offset == b->offset &&
+ a->mlen == b->mlen &&
+ a->regs_written == b->regs_written &&
+ a->base_mrf == b->base_mrf &&
+ a->eot == b->eot &&
+ a->header_size == b->header_size &&
+ a->shadow_compare == b->shadow_compare &&
+ a->pi_noperspective == b->pi_noperspective &&
+ a->sources == b->sources &&
+ operands_match(a, b, negate);
+}
+
+static void
+create_copy_instr(const fs_builder &bld, fs_inst *inst, fs_reg src, bool negate)
{
- return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
+ int written = inst->regs_written;
+ int dst_width = inst->exec_size / 8;
+ fs_inst *copy;
+
+ if (written > dst_width) {
+ fs_reg *payload;
+ int sources, header_size;
+ if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
+ sources = inst->sources;
+ header_size = inst->header_size;
+ } else {
+ assert(written % dst_width == 0);
+ sources = written / dst_width;
+ header_size = 0;
+ }
+
+ assert(src.file == GRF);
+ payload = ralloc_array(bld.shader->mem_ctx, fs_reg, sources);
+ for (int i = 0; i < header_size; i++) {
+ payload[i] = src;
+ src.reg_offset++;
+ }
+ for (int i = header_size; i < sources; i++) {
+ payload[i] = src;
+ src = offset(src, bld, 1);
+ }
+ copy = bld.LOAD_PAYLOAD(inst->dst, payload, sources, header_size);
+ } else {
+ copy = bld.MOV(inst->dst, src);
+ copy->src[0].negate = negate;
+ }
+ assert(copy->regs_written == written);
}
bool
-fs_visitor::opt_cse_local(bblock_t *block, exec_list *aeb)
+fs_visitor::opt_cse_local(bblock_t *block)
{
bool progress = false;
+ exec_list aeb;
- 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) {
-
+ foreach_inst_in_block(fs_inst, inst, block) {
/* Skip some cases. */
- if (is_expression(inst) &&
- !inst->predicate &&
- !inst->is_partial_write() &&
- !inst->conditional_mod &&
- inst->dst.file != HW_REG)
+ if (is_expression(this, inst) && !inst->is_partial_write() &&
+ (inst->dst.file != HW_REG || inst->dst.is_null()))
{
- bool found = false;
-
- aeb_entry *entry;
- foreach_list(entry_node, aeb) {
- 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)) {
-
- found = true;
- progress = true;
- break;
- }
- }
-
- if (!found) {
- /* Our first sighting of this expression. Create an entry. */
- aeb_entry *entry = ralloc(mem_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) {
+ bool found = false;
+ bool negate = false;
+
+ foreach_in_list_use_after(aeb_entry, entry, &aeb) {
+ /* Match current instruction's expression against those in AEB. */
+ if (!(entry->generator->dst.is_null() && !inst->dst.is_null()) &&
+ instructions_match(inst, entry->generator, &negate)) {
+ found = true;
+ progress = true;
+ break;
+ }
+ }
+
+ if (!found) {
+ 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()) {
+ const fs_builder ibld = fs_builder(this, block, entry->generator)
+ .at(block, entry->generator->next);
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 */
- 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++;
+ entry->tmp = fs_reg(GRF, alloc.allocate(written),
+ entry->generator->dst.type);
+
+ create_copy_instr(ibld, entry->generator, entry->tmp, false);
+
+ entry->generator->dst = entry->tmp;
}
- inst->remove();
- /* Appending an instruction may have changed our bblock end. */
- if (inst == block->end) {
- block->end = copy;
- }
+ /* dest <- temp */
+ if (!inst->dst.is_null()) {
+ assert(inst->regs_written == entry->generator->regs_written);
+ assert(inst->dst.type == entry->tmp.type);
+ const fs_builder ibld(this, block, inst);
+
+ create_copy_instr(ibld, inst, entry->tmp, negate);
+ }
+
+ /* 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;
- /* Continue iteration with copy->next */
- inst = copy;
- }
+ inst->remove(block);
+ inst = prev;
+ }
}
- foreach_list_safe(entry_node, aeb) {
- aeb_entry *entry = (aeb_entry *)entry_node;
+ foreach_in_list_safe(aeb_entry, entry, &aeb) {
+ /* 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()) {
+ bool negate; /* dummy */
+ if (entry->generator->reads_flag() ||
+ (entry->generator->writes_flag() &&
+ !instructions_match(inst, entry->generator, &negate))) {
+ entry->remove();
+ ralloc_free(entry);
+ continue;
+ }
+ }
- for (int i = 0; i < 3; i++) {
+ for (int i = 0; i < entry->generator->sources; 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;
- }
+ 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;
+ break;
}
- }
+ }
}
ip++;
}
- ralloc_free(mem_ctx);
-
- if (progress)
- invalidate_live_intervals();
+ ralloc_free(cse_ctx);
return progress;
}
calculate_live_intervals();
- cfg_t cfg(this);
-
- for (int b = 0; b < cfg.num_blocks; b++) {
- bblock_t *block = cfg.blocks[b];
- exec_list aeb;
-
- progress = opt_cse_local(block, &aeb) || progress;
+ foreach_block (block, cfg) {
+ progress = opt_cse_local(block) || progress;
}
+ if (progress)
+ invalidate_live_intervals();
+
return progress;
}
projects / mesa.git / blobdiff