* IN THE SOFTWARE.
*/
+/** @file brw_fs_copy_propagation.cpp
+ *
+ * Support for global copy propagation in two passes: A local pass that does
+ * intra-block copy (and constant) propagation, and a global pass that uses
+ * dataflow analysis on the copies available at the end of each block to re-do
+ * local copy propagation with more copies available.
+ *
+ * See Muchnik's Advanced Compiler Design and Implementation, section
+ * 12.5 (p356).
+ */
+
+#define ACP_HASH_SIZE 16
+
+#include "main/bitset.h"
#include "brw_fs.h"
#include "brw_cfg.h"
fs_reg dst;
fs_reg src;
};
+
+struct block_data {
+ /**
+ * Which entries in the fs_copy_prop_dataflow acp table are live at the
+ * start of this block. This is the useful output of the analysis, since
+ * it lets us plug those into the local copy propagation on the second
+ * pass.
+ */
+ BITSET_WORD *livein;
+
+ /**
+ * Which entries in the fs_copy_prop_dataflow acp table are live at the end
+ * of this block. This is done in initial setup from the per-block acps
+ * returned by the first local copy prop pass.
+ */
+ BITSET_WORD *liveout;
+
+ /**
+ * Which entries in the fs_copy_prop_dataflow acp table are generated by
+ * instructions in this block which reach the end of the block without
+ * being killed.
+ */
+ BITSET_WORD *copy;
+
+ /**
+ * Which entries in the fs_copy_prop_dataflow acp table are killed over the
+ * course of this block.
+ */
+ BITSET_WORD *kill;
+};
+
+class fs_copy_prop_dataflow
+{
+public:
+ fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
+ exec_list *out_acp[ACP_HASH_SIZE]);
+
+ void setup_initial_values();
+ void run();
+
+ void dump_block_data() const;
+
+ void *mem_ctx;
+ cfg_t *cfg;
+
+ acp_entry **acp;
+ int num_acp;
+ int bitset_words;
+
+ struct block_data *bd;
+};
+} /* anonymous namespace */
+
+fs_copy_prop_dataflow::fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
+ exec_list *out_acp[ACP_HASH_SIZE])
+ : mem_ctx(mem_ctx), cfg(cfg)
+{
+ bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
+
+ num_acp = 0;
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ for (int i = 0; i < ACP_HASH_SIZE; i++) {
+ foreach_list(entry_node, &out_acp[b][i]) {
+ num_acp++;
+ }
+ }
+ }
+
+ acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
+
+ bitset_words = BITSET_WORDS(num_acp);
+
+ int next_acp = 0;
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ bd[b].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[b].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[b].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[b].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
+
+ for (int i = 0; i < ACP_HASH_SIZE; i++) {
+ foreach_list(entry_node, &out_acp[b][i]) {
+ acp_entry *entry = (acp_entry *)entry_node;
+
+ acp[next_acp] = entry;
+
+ /* opt_copy_propagate_local populates out_acp with copies created
+ * in a block which are still live at the end of the block. This
+ * is exactly what we want in the COPY set.
+ */
+ BITSET_SET(bd[b].copy, next_acp);
+
+ next_acp++;
+ }
+ }
+ }
+
+ assert(next_acp == num_acp);
+
+ setup_initial_values();
+ run();
+}
+
+/**
+ * Set up initial values for each of the data flow sets, prior to running
+ * the fixed-point algorithm.
+ */
+void
+fs_copy_prop_dataflow::setup_initial_values()
+{
+ /* Initialize the COPY and KILL sets. */
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ bblock_t *block = cfg->blocks[b];
+
+ for (fs_inst *inst = (fs_inst *)block->start;
+ inst != block->end->next;
+ inst = (fs_inst *)inst->next) {
+ if (inst->dst.file != GRF)
+ continue;
+
+ /* Mark ACP entries which are killed by this instruction. */
+ for (int i = 0; i < num_acp; i++) {
+ if (inst->overwrites_reg(acp[i]->dst) ||
+ inst->overwrites_reg(acp[i]->src)) {
+ BITSET_SET(bd[b].kill, i);
+ }
+ }
+ }
+ }
+
+ /* Populate the initial values for the livein and liveout sets. For the
+ * block at the start of the program, livein = 0 and liveout = copy.
+ * For the others, set liveout to 0 (the empty set) and livein to ~0
+ * (the universal set).
+ */
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ bblock_t *block = cfg->blocks[b];
+ if (block->parents.is_empty()) {
+ for (int i = 0; i < bitset_words; i++) {
+ bd[b].livein[i] = 0u;
+ bd[b].liveout[i] = bd[b].copy[i];
+ }
+ } else {
+ for (int i = 0; i < bitset_words; i++) {
+ bd[b].liveout[i] = 0u;
+ bd[b].livein[i] = ~0u;
+ }
+ }
+ }
+}
+
+/**
+ * Walk the set of instructions in the block, marking which entries in the acp
+ * are killed by the block.
+ */
+void
+fs_copy_prop_dataflow::run()
+{
+ bool progress;
+
+ do {
+ progress = false;
+
+ /* Update liveout for all blocks. */
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ if (cfg->blocks[b]->parents.is_empty())
+ continue;
+
+ for (int i = 0; i < bitset_words; i++) {
+ const BITSET_WORD old_liveout = bd[b].liveout[i];
+
+ bd[b].liveout[i] =
+ bd[b].copy[i] | (bd[b].livein[i] & ~bd[b].kill[i]);
+
+ if (old_liveout != bd[b].liveout[i])
+ progress = true;
+ }
+ }
+
+ /* Update livein for all blocks. If a copy is live out of all parent
+ * blocks, it's live coming in to this block.
+ */
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ if (cfg->blocks[b]->parents.is_empty())
+ continue;
+
+ for (int i = 0; i < bitset_words; i++) {
+ const BITSET_WORD old_livein = bd[b].livein[i];
+
+ bd[b].livein[i] = ~0u;
+ foreach_list(block_node, &cfg->blocks[b]->parents) {
+ bblock_link *link = (bblock_link *)block_node;
+ bblock_t *block = link->block;
+ bd[b].livein[i] &= bd[block->block_num].liveout[i];
+ }
+
+ if (old_livein != bd[b].livein[i])
+ progress = true;
+ }
+ }
+ } while (progress);
+}
+
+void
+fs_copy_prop_dataflow::dump_block_data() const
+{
+ for (int b = 0; b < cfg->num_blocks; b++) {
+ bblock_t *block = cfg->blocks[b];
+ printf("Block %d [%d, %d] (parents ", block->block_num,
+ block->start_ip, block->end_ip);
+ foreach_list(block_node, &block->parents) {
+ bblock_t *parent = ((bblock_link *) block_node)->block;
+ printf("%d ", parent->block_num);
+ }
+ printf("):\n");
+ printf(" livein = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ printf("%08x", bd[b].livein[i]);
+ printf(", liveout = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ printf("%08x", bd[b].liveout[i]);
+ printf(",\n copy = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ printf("%08x", bd[b].copy[i]);
+ printf(", kill = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ printf("%08x", bd[b].kill[i]);
+ printf("\n");
+ }
}
bool
bool has_source_modifiers = entry->src.abs || entry->src.negate;
- if (intel->gen == 6 && inst->is_math() &&
- (has_source_modifiers || entry->src.file == UNIFORM))
+ if ((has_source_modifiers || entry->src.file == UNIFORM ||
+ entry->src.smear != -1) && !can_do_source_mods(inst))
+ return false;
+
+ if (has_source_modifiers && entry->dst.type != inst->src[arg].type)
return false;
inst->src[arg].file = entry->src.file;
inst->src[arg].reg = entry->src.reg;
inst->src[arg].reg_offset = entry->src.reg_offset;
+ if (entry->src.smear != -1)
+ inst->src[arg].smear = entry->src.smear;
if (!inst->src[arg].abs) {
inst->src[arg].abs = entry->src.abs;
progress = true;
break;
+ case BRW_OPCODE_MACH:
case BRW_OPCODE_MUL:
case BRW_OPCODE_ADD:
if (i == 1) {
progress = true;
} else if (i == 0 && inst->src[1].file != IMM) {
/* Fit this constant in by commuting the operands.
- * Exception: we can't do this for 32-bit integer MUL
+ * Exception: we can't do this for 32-bit integer MUL/MACH
* because it's asymmetric.
*/
- if (inst->opcode == BRW_OPCODE_MUL &&
+ if ((inst->opcode == BRW_OPCODE_MUL ||
+ inst->opcode == BRW_OPCODE_MACH) &&
(inst->src[1].type == BRW_REGISTER_TYPE_D ||
inst->src[1].type == BRW_REGISTER_TYPE_UD))
break;
}
break;
- case FS_OPCODE_PULL_CONSTANT_LOAD:
+ case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
inst->src[i] = entry->src;
progress = true;
break;
return progress;
}
-
-/** @file brw_fs_copy_propagation.cpp
- *
- * Support for local copy propagation by walking the list of instructions
- * and maintaining the ACP table of available copies for propagation.
- *
- * See Muchnik's Advanced Compiler Design and Implementation, section
- * 12.5 (p356).
- */
-
/* Walks a basic block and does copy propagation on it using the acp
* list.
*/
bool
-fs_visitor::opt_copy_propagate_local(void *mem_ctx, bblock_t *block)
+fs_visitor::opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
+ exec_list *acp)
{
bool progress = false;
- int acp_count = 16;
- exec_list acp[acp_count];
- for (fs_inst *inst = block->start;
+ for (fs_inst *inst = (fs_inst *)block->start;
inst != block->end->next;
inst = (fs_inst *)inst->next) {
if (inst->src[i].file != GRF)
continue;
- foreach_list(entry_node, &acp[inst->src[i].reg % acp_count]) {
+ foreach_list(entry_node, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
acp_entry *entry = (acp_entry *)entry_node;
if (try_constant_propagate(inst, entry))
/* kill the destination from the ACP */
if (inst->dst.file == GRF) {
- foreach_list_safe(entry_node, &acp[inst->dst.reg % acp_count]) {
+ foreach_list_safe(entry_node, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
acp_entry *entry = (acp_entry *)entry_node;
if (inst->overwrites_reg(entry->dst)) {
/* Oops, we only have the chaining hash based on the destination, not
* the source, so walk across the entire table.
*/
- for (int i = 0; i < acp_count; i++) {
+ for (int i = 0; i < ACP_HASH_SIZE; i++) {
foreach_list_safe(entry_node, &acp[i]) {
acp_entry *entry = (acp_entry *)entry_node;
if (inst->overwrites_reg(entry->src))
}
}
- /* If this instruction is a raw copy, add it to the ACP. */
+ /* If this instruction's source could potentially be folded into the
+ * operand of another instruction, add it to the ACP.
+ */
if (inst->opcode == BRW_OPCODE_MOV &&
inst->dst.file == GRF &&
((inst->src[0].file == GRF &&
inst->src[0].file == IMM) &&
inst->src[0].type == inst->dst.type &&
!inst->saturate &&
- !inst->predicated &&
- !inst->force_uncompressed &&
- !inst->force_sechalf &&
- inst->src[0].smear == -1) {
+ !inst->is_partial_write()) {
acp_entry *entry = ralloc(mem_ctx, acp_entry);
entry->dst = inst->dst;
entry->src = inst->src[0];
- acp[entry->dst.reg % acp_count].push_tail(entry);
+ acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
}
}
{
bool progress = false;
void *mem_ctx = ralloc_context(this->mem_ctx);
-
cfg_t cfg(this);
+ exec_list *out_acp[cfg.num_blocks];
+ for (int i = 0; i < cfg.num_blocks; i++)
+ out_acp[i] = new exec_list [ACP_HASH_SIZE];
+
+ /* First, walk through each block doing local copy propagation and getting
+ * the set of copies available at the end of the block.
+ */
+ for (int b = 0; b < cfg.num_blocks; b++) {
+ bblock_t *block = cfg.blocks[b];
+ progress = opt_copy_propagate_local(mem_ctx, block,
+ out_acp[b]) || progress;
+ }
+
+ /* Do dataflow analysis for those available copies. */
+ fs_copy_prop_dataflow dataflow(mem_ctx, &cfg, out_acp);
+
+ /* Next, re-run local copy propagation, this time with the set of copies
+ * provided by the dataflow analysis available at the start of a block.
+ */
for (int b = 0; b < cfg.num_blocks; b++) {
bblock_t *block = cfg.blocks[b];
+ exec_list in_acp[ACP_HASH_SIZE];
+
+ for (int i = 0; i < dataflow.num_acp; i++) {
+ if (BITSET_TEST(dataflow.bd[b].livein, i)) {
+ struct acp_entry *entry = dataflow.acp[i];
+ in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
+ }
+ }
- progress = opt_copy_propagate_local(mem_ctx, block) || progress;
+ progress = opt_copy_propagate_local(mem_ctx, block, in_acp) || progress;
}
+ for (int i = 0; i < cfg.num_blocks; i++)
+ delete [] out_acp[i];
ralloc_free(mem_ctx);
if (progress)