*/
#include "brw_fs.h"
-#include "brw_fs_cfg.h"
+#include "brw_cfg.h"
/** @file brw_fs_cse.cpp
*
* 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).
*/
};
}
+static bool
+is_copy_payload(const fs_inst *inst)
+{
+ const int reg = inst->src[0].reg;
+ if (inst->src[0].reg_offset != 0)
+ return false;
+
+ for (int i = 1; i < inst->sources; i++) {
+ if (inst->src[i].reg != reg ||
+ inst->src[i].reg_offset != i) {
+ return false;
+ }
+ }
+ return true;
+}
+
static bool
is_expression(const fs_inst *const inst)
{
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;
+ 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 == 0;
+ case SHADER_OPCODE_LOAD_PAYLOAD:
+ return !is_copy_payload(inst);
+ default:
+ 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_reg *xs, fs_reg *ys)
+operands_match(fs_inst *a, fs_inst *b)
+{
+ fs_reg *xs = a->src;
+ fs_reg *ys = b->src;
+
+ if (!is_expression_commutative(a->opcode)) {
+ 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
+instructions_match(fs_inst *a, fs_inst *b)
{
- return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
+ 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 &&
+ a->sources == b->sources &&
+ (a->is_tex() ? (a->texture_offset == b->texture_offset &&
+ a->mlen == b->mlen &&
+ a->regs_written == b->regs_written &&
+ a->base_mrf == b->base_mrf &&
+ a->eot == b->eot &&
+ a->header_present == b->header_present &&
+ a->shadow_compare == b->shadow_compare)
+ : true) &&
+ operands_match(a, b);
}
bool
-fs_visitor::opt_cse_local(fs_bblock *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);
-
- for (fs_inst *inst = block->start;
- inst != block->end->next;
- inst = (fs_inst *) inst->next) {
+ void *cse_ctx = ralloc_context(NULL);
+ int ip = block->start_ip;
+ foreach_inst_in_block(fs_inst, inst, block) {
/* Skip some cases. */
- if (is_expression(inst) && !inst->predicated && 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;
-
- 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 &&
- 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) {
- 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;
- }
-
- /* dest <- temp */
- fs_inst *copy = new(ralloc_parent(inst))
- fs_inst(BRW_OPCODE_MOV, inst->dst, entry->tmp);
- inst->replace_with(copy);
-
- /* Appending an instruction may have changed our bblock end. */
- if (inst == block->end) {
- block->end = copy;
- }
-
- /* Continue iteration with copy->next */
- inst = copy;
- }
+ bool found = false;
+
+ foreach_in_list_use_after(aeb_entry, entry, &aeb) {
+ /* Match current instruction's expression against those in AEB. */
+ 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(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;
+
+ 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;
+
+ 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] = offset(tmp, i);
+ copy = LOAD_PAYLOAD(orig_dst, sources, written);
+ } else {
+ copy = MOV(orig_dst, tmp);
+ copy->force_writemask_all =
+ entry->generator->force_writemask_all;
+ }
+ entry->generator->insert_after(block, copy);
+ }
+
+ /* dest <- temp */
+ 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;
+ if (written > 1) {
+ fs_reg *sources = ralloc_array(mem_ctx, fs_reg, written);
+ for (int i = 0; i < written; i++)
+ sources[i] = offset(tmp, i);
+ copy = LOAD_PAYLOAD(dst, sources, written);
+ } else {
+ copy = MOV(dst, tmp);
+ copy->force_writemask_all = inst->force_writemask_all;
+ }
+ inst->insert_before(block, copy);
+ }
+
+ /* 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(block);
+ inst = prev;
+ }
}
- /* Kill all AEB entries that use the destination. */
- int start_offset = inst->dst.reg_offset;
- int end_offset = start_offset + inst->regs_written();
-
- foreach_list_safe(entry_node, aeb) {
- aeb_entry *entry = (aeb_entry *)entry_node;
-
- for (int i = 0; i < 3; i++) {
- if (entry->generator->src[i].file == inst->dst.file &&
- entry->generator->src[i].reg == inst->dst.reg &&
- entry->generator->src[i].reg_offset >= start_offset &&
- entry->generator->src[i].reg_offset < end_offset) {
- entry->remove();
- ralloc_free(entry);
- break;
- }
- }
+ 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()) {
+ 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 < 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;
+ }
+
+ /* 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;
+ }
+ }
}
- }
- ralloc_free(mem_ctx);
+ ip++;
+ }
- if (progress)
- this->live_intervals_valid = false;
+ ralloc_free(cse_ctx);
return progress;
}
{
bool progress = false;
- fs_cfg cfg(this);
-
- for (int b = 0; b < cfg.num_blocks; b++) {
- fs_bblock *block = cfg.blocks[b];
- exec_list aeb;
+ calculate_live_intervals();
- 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