[Mesa-dev] [PATCH 5/5] i965/fs: Add basic-block-level dead code elimination.
Eric Anholt
eric at anholt.net
Wed Apr 10 11:54:42 PDT 2013
This is a poor substitute for proper global dead code elimination that
could replace both our current paths, but it was very easy to write. It
particularly helps with Valve's shaders that are translated out of DX
assembly, which has been register allocated and thus have a bunch of
unrelated uses of the same variable (some of which get copy-propagated
from and then left for dead).
shader-db results:
total instructions in shared programs: 1735753 -> 1731698 (-0.23%)
instructions in affected programs: 492620 -> 488565 (-0.82%)
---
src/mesa/drivers/dri/i965/brw_fs.cpp | 160 ++++++++++++++++++++++++++++++++++
src/mesa/drivers/dri/i965/brw_fs.h | 1 +
2 files changed, 161 insertions(+)
diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index 3917dba..96e8a99 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -32,6 +32,7 @@ extern "C" {
#include <sys/types.h>
+#include "main/hash_table.h"
#include "main/macros.h"
#include "main/shaderobj.h"
#include "main/uniforms.h"
@@ -1823,6 +1824,164 @@ fs_visitor::dead_code_eliminate()
return progress;
}
+struct dead_code_hash_key
+{
+ int vgrf;
+ int reg_offset;
+};
+
+static bool
+dead_code_hash_compare(const void *a, const void *b)
+{
+ return memcmp(a, b, sizeof(struct dead_code_hash_key)) == 0;
+}
+
+static void
+clear_dead_code_hash(struct hash_table *ht)
+{
+ struct hash_entry *entry;
+
+ hash_table_foreach(ht, entry) {
+ _mesa_hash_table_remove(ht, entry);
+ }
+}
+
+static void
+insert_dead_code_hash(struct hash_table *ht,
+ int vgrf, int reg_offset, fs_inst *inst)
+{
+ /* We don't bother freeing keys, because they'll be GCed with the ht. */
+ struct dead_code_hash_key *key = ralloc(ht, struct dead_code_hash_key);
+
+ key->vgrf = vgrf;
+ key->reg_offset = reg_offset;
+
+ _mesa_hash_table_insert(ht, _mesa_hash_data(key, sizeof(*key)), key, inst);
+}
+
+static struct hash_entry *
+get_dead_code_hash_entry(struct hash_table *ht, int vgrf, int reg_offset)
+{
+ struct dead_code_hash_key key;
+
+ key.vgrf = vgrf;
+ key.reg_offset = reg_offset;
+
+ return _mesa_hash_table_search(ht, _mesa_hash_data(&key, sizeof(key)), &key);
+}
+
+static void
+remove_dead_code_hash(struct hash_table *ht,
+ int vgrf, int reg_offset)
+{
+ struct hash_entry *entry = get_dead_code_hash_entry(ht, vgrf, reg_offset);
+ if (!entry)
+ return;
+
+ _mesa_hash_table_remove(ht, entry);
+}
+
+/**
+ * Walks basic blocks, removing any regs that are written but not read before
+ * being redefined.
+ *
+ * The dead_code_eliminate() function implements a global dead code
+ * elimination, but it only handles the removing the last write to a register
+ * if it's never read. This one can handle intermediate writes, but only
+ * within a basic block.
+ */
+bool
+fs_visitor::dead_code_eliminate_local()
+{
+ struct hash_table *ht;
+ bool progress = false;
+
+ ht = _mesa_hash_table_create(mem_ctx, dead_code_hash_compare);
+
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ /* At a basic block, empty the HT since we don't understand dataflow
+ * here.
+ */
+ if (inst->is_control_flow()) {
+ clear_dead_code_hash(ht);
+ continue;
+ }
+
+ /* Clear the HT of any instructions that got read. */
+ for (int i = 0; i < 3; i++) {
+ fs_reg src = inst->src[i];
+ if (src.file != GRF)
+ continue;
+
+ int read = 1;
+ if (inst->is_send_from_grf())
+ read = virtual_grf_sizes[src.reg] - src.reg_offset;
+
+ for (int reg_offset = src.reg_offset;
+ reg_offset < src.reg_offset + read;
+ reg_offset++) {
+ remove_dead_code_hash(ht, src.reg, reg_offset);
+ }
+ }
+
+ /* Add any update of a GRF to the HT, removing a previous write if it's
+ * wasn't read.
+ */
+ if (inst->dst.file == GRF) {
+ if (inst->regs_written > 1) {
+ /* We don't know how to trim channels from an instruction's
+ * writes, so we can't incrementally remove unread channels from
+ * it. Just remove whatever it overwrites from the table
+ */
+ for (int i = 0; i < inst->regs_written; i++) {
+ remove_dead_code_hash(ht,
+ inst->dst.reg,
+ inst->dst.reg_offset + i);
+ }
+ } else {
+ struct hash_entry *entry =
+ get_dead_code_hash_entry(ht, inst->dst.reg,
+ inst->dst.reg_offset);
+
+ if (inst->is_partial_write()) {
+ /* For a partial write, we can't remove any previous dead code
+ * candidate, since we're just modifying their result, but we can
+ * be dead code eliminiated ourselves.
+ */
+ if (entry) {
+ entry->data = inst;
+ } else {
+ insert_dead_code_hash(ht, inst->dst.reg, inst->dst.reg_offset,
+ inst);
+ }
+ } else {
+ if (entry) {
+ /* We're completely updating a channel, and there was a
+ * previous write to the channel that wasn't read. Kill it!
+ */
+ fs_inst *inst = (fs_inst *)entry->data;
+ inst->remove();
+ progress = true;
+ _mesa_hash_table_remove(ht, entry);
+ }
+
+ insert_dead_code_hash(ht, inst->dst.reg, inst->dst.reg_offset,
+ inst);
+ }
+ }
+ }
+ }
+
+ _mesa_hash_table_destroy(ht, NULL);
+
+ if (progress)
+ live_intervals_valid = false;
+
+ return progress;
+}
+
/**
* Implements a second type of register coalescing: This one checks if
* the two regs involved in a raw move don't interfere, in which case
@@ -2805,6 +2964,7 @@ fs_visitor::run()
progress = opt_cse() || progress;
progress = opt_copy_propagate() || progress;
progress = dead_code_eliminate() || progress;
+ progress = dead_code_eliminate_local() || progress;
progress = register_coalesce() || progress;
progress = register_coalesce_2() || progress;
progress = compute_to_mrf() || progress;
diff --git a/src/mesa/drivers/dri/i965/brw_fs.h b/src/mesa/drivers/dri/i965/brw_fs.h
index f0901e7..83b4b71 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.h
+++ b/src/mesa/drivers/dri/i965/brw_fs.h
@@ -329,6 +329,7 @@ public:
bool register_coalesce_2();
bool compute_to_mrf();
bool dead_code_eliminate();
+ bool dead_code_eliminate_local();
bool remove_dead_constants();
bool remove_duplicate_mrf_writes();
bool virtual_grf_interferes(int a, int b);
--
1.7.10.4
More information about the mesa-dev
mailing list