* 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 Muchnick's Advanced Compiler Design and Implementation, section
+ * 12.5 (p356).
+ */
+
+#define ACP_HASH_SIZE 16
+
+#include "util/bitset.h"
#include "brw_fs.h"
-#include "brw_fs_cfg.h"
+#include "brw_cfg.h"
namespace { /* avoid conflict with opt_copy_propagation_elements */
struct acp_entry : public exec_node {
fs_reg dst;
fs_reg src;
+ uint8_t regs_written;
+ enum opcode opcode;
+ bool saturate;
};
+
+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;
+ foreach_block (block, cfg) {
+ for (int i = 0; i < ACP_HASH_SIZE; i++) {
+ num_acp += out_acp[block->num][i].length();
+ }
+ }
+
+ acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
+
+ bitset_words = BITSET_WORDS(num_acp);
+
+ int next_acp = 0;
+ foreach_block (block, cfg) {
+ bd[block->num].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[block->num].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[block->num].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
+ bd[block->num].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
+
+ for (int i = 0; i < ACP_HASH_SIZE; i++) {
+ foreach_in_list(acp_entry, entry, &out_acp[block->num][i]) {
+ 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[block->num].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. */
+ foreach_block (block, cfg) {
+ foreach_inst_in_block(fs_inst, inst, block) {
+ 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[block->num].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).
+ */
+ foreach_block (block, cfg) {
+ if (block->parents.is_empty()) {
+ for (int i = 0; i < bitset_words; i++) {
+ bd[block->num].livein[i] = 0u;
+ bd[block->num].liveout[i] = bd[block->num].copy[i];
+ }
+ } else {
+ for (int i = 0; i < bitset_words; i++) {
+ bd[block->num].liveout[i] = 0u;
+ bd[block->num].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. */
+ foreach_block (block, cfg) {
+ if (block->parents.is_empty())
+ continue;
+
+ for (int i = 0; i < bitset_words; i++) {
+ const BITSET_WORD old_liveout = bd[block->num].liveout[i];
+
+ bd[block->num].liveout[i] =
+ bd[block->num].copy[i] | (bd[block->num].livein[i] &
+ ~bd[block->num].kill[i]);
+
+ if (old_liveout != bd[block->num].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.
+ */
+ foreach_block (block, cfg) {
+ if (block->parents.is_empty())
+ continue;
+
+ for (int i = 0; i < bitset_words; i++) {
+ const BITSET_WORD old_livein = bd[block->num].livein[i];
+
+ bd[block->num].livein[i] = ~0u;
+ foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
+ bblock_t *parent = parent_link->block;
+ bd[block->num].livein[i] &= bd[parent->num].liveout[i];
+ }
+
+ if (old_livein != bd[block->num].livein[i])
+ progress = true;
+ }
+ }
+ } while (progress);
+}
+
+void
+fs_copy_prop_dataflow::dump_block_data() const
+{
+ foreach_block (block, cfg) {
+ fprintf(stderr, "Block %d [%d, %d] (parents ", block->num,
+ block->start_ip, block->end_ip);
+ foreach_list_typed(bblock_link, link, link, &block->parents) {
+ bblock_t *parent = link->block;
+ fprintf(stderr, "%d ", parent->num);
+ }
+ fprintf(stderr, "):\n");
+ fprintf(stderr, " livein = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ fprintf(stderr, "%08x", bd[block->num].livein[i]);
+ fprintf(stderr, ", liveout = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ fprintf(stderr, "%08x", bd[block->num].liveout[i]);
+ fprintf(stderr, ",\n copy = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ fprintf(stderr, "%08x", bd[block->num].copy[i]);
+ fprintf(stderr, ", kill = 0x");
+ for (int i = 0; i < bitset_words; i++)
+ fprintf(stderr, "%08x", bd[block->num].kill[i]);
+ fprintf(stderr, "\n");
+ }
+}
+
+static bool
+is_logic_op(enum opcode opcode)
+{
+ return (opcode == BRW_OPCODE_AND ||
+ opcode == BRW_OPCODE_OR ||
+ opcode == BRW_OPCODE_XOR ||
+ opcode == BRW_OPCODE_NOT);
}
bool
fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
{
+ if (inst->src[arg].file != GRF)
+ return false;
+
if (entry->src.file == IMM)
return false;
+ assert(entry->src.file == GRF || entry->src.file == UNIFORM);
- if (inst->src[arg].file != entry->dst.file ||
- inst->src[arg].reg != entry->dst.reg ||
- inst->src[arg].reg_offset != entry->dst.reg_offset) {
+ if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
+ inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
+ return false;
+
+ assert(entry->dst.file == GRF);
+ if (inst->src[arg].reg != entry->dst.reg)
+ return false;
+
+ /* Bail if inst is reading a range that isn't contained in the range
+ * that entry is writing.
+ */
+ if (inst->src[arg].reg_offset < entry->dst.reg_offset ||
+ (inst->src[arg].reg_offset * 32 + inst->src[arg].subreg_offset +
+ inst->regs_read(arg) * inst->src[arg].stride * 32) >
+ (entry->dst.reg_offset + entry->regs_written) * 32)
return false;
- }
- /* See resolve_ud_negate() and comment in brw_fs_emit.cpp. */
- if (inst->conditional_mod &&
- inst->src[arg].type == BRW_REGISTER_TYPE_UD &&
+ /* we can't generally copy-propagate UD negations because we
+ * can end up accessing the resulting values as signed integers
+ * instead. See also resolve_ud_negate() and comment in
+ * fs_generator::generate_code.
+ */
+ if (inst->src[arg].type == BRW_REGISTER_TYPE_UD &&
entry->src.negate)
return false;
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.is_contiguous()) &&
+ !inst->can_do_source_mods(brw))
+ return false;
+
+ if (has_source_modifiers &&
+ inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
return false;
+ /* Bail if the result of composing both strides would exceed the
+ * hardware limit.
+ */
+ if (entry->src.stride * inst->src[arg].stride > 4)
+ return false;
+
+ /* Bail if the result of composing both strides cannot be expressed
+ * as another stride. This avoids, for example, trying to transform
+ * this:
+ *
+ * MOV (8) rX<1>UD rY<0;1,0>UD
+ * FOO (8) ... rX<8;8,1>UW
+ *
+ * into this:
+ *
+ * FOO (8) ... rY<0;1,0>UW
+ *
+ * Which would have different semantics.
+ */
+ if (entry->src.stride != 1 &&
+ (inst->src[arg].stride *
+ type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
+ return false;
+
+ if (has_source_modifiers && entry->dst.type != inst->src[arg].type)
+ return false;
+
+ if (brw->gen >= 8 && (entry->src.negate || entry->src.abs) &&
+ is_logic_op(inst->opcode)) {
+ return false;
+ }
+
+ if (entry->saturate) {
+ switch(inst->opcode) {
+ case BRW_OPCODE_SEL:
+ if (inst->src[1].file != IMM ||
+ inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
+ inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
+ return false;
+ }
+ break;
+ default:
+ 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;
+ inst->src[arg].stride *= entry->src.stride;
+ inst->saturate = inst->saturate || entry->saturate;
+
+ switch (entry->src.file) {
+ case UNIFORM:
+ assert(entry->src.width == 1);
+ case BAD_FILE:
+ case HW_REG:
+ inst->src[arg].width = entry->src.width;
+ inst->src[arg].reg_offset = entry->src.reg_offset;
+ inst->src[arg].subreg_offset = entry->src.subreg_offset;
+ break;
+ case GRF:
+ {
+ assert(entry->src.width % inst->src[arg].width == 0);
+ /* In this case, we'll just leave the width alone. The source
+ * register could have different widths depending on how it is
+ * being used. For instance, if only half of the register was
+ * used then we want to preserve that and continue to only use
+ * half.
+ *
+ * Also, we have to deal with mapping parts of vgrfs to other
+ * parts of vgrfs so we have to do some reg_offset magic.
+ */
+
+ /* Compute the offset of inst->src[arg] relative to inst->dst */
+ assert(entry->dst.subreg_offset == 0);
+ int rel_offset = inst->src[arg].reg_offset - entry->dst.reg_offset;
+ int rel_suboffset = inst->src[arg].subreg_offset;
+
+ /* Compute the final register offset (in bytes) */
+ int offset = entry->src.reg_offset * 32 + entry->src.subreg_offset;
+ offset += rel_offset * 32 + rel_suboffset;
+ inst->src[arg].reg_offset = offset / 32;
+ inst->src[arg].subreg_offset = offset % 32;
+ }
+ break;
+ default:
+ unreachable("Invalid register file");
+ break;
+ }
if (!inst->src[arg].abs) {
inst->src[arg].abs = entry->src.abs;
if (entry->src.file != IMM)
return false;
+ if (entry->saturate)
+ return false;
- for (int i = 2; i >= 0; i--) {
- if (inst->src[i].file != entry->dst.file ||
- inst->src[i].reg != entry->dst.reg ||
- inst->src[i].reg_offset != entry->dst.reg_offset)
+ for (int i = inst->sources - 1; i >= 0; i--) {
+ if (inst->src[i].file != GRF)
continue;
- /* Don't bother with cases that should have been taken care of by the
- * GLSL compiler's constant folding pass.
+ assert(entry->dst.file == GRF);
+ if (inst->src[i].reg != entry->dst.reg)
+ continue;
+
+ /* Bail if inst is reading a range that isn't contained in the range
+ * that entry is writing.
*/
- if (inst->src[i].negate || inst->src[i].abs)
+ if (inst->src[i].reg_offset < entry->dst.reg_offset ||
+ (inst->src[i].reg_offset * 32 + inst->src[i].subreg_offset +
+ inst->regs_read(i) * inst->src[i].stride * 32) >
+ (entry->dst.reg_offset + entry->regs_written) * 32)
continue;
+ fs_reg val = entry->src;
+ val.effective_width = inst->src[i].effective_width;
+ val.type = inst->src[i].type;
+
+ if (inst->src[i].abs) {
+ if ((brw->gen >= 8 && is_logic_op(inst->opcode)) ||
+ !brw_abs_immediate(val.type, &val.fixed_hw_reg)) {
+ continue;
+ }
+ }
+
+ if (inst->src[i].negate) {
+ if ((brw->gen >= 8 && is_logic_op(inst->opcode)) ||
+ !brw_negate_immediate(val.type, &val.fixed_hw_reg)) {
+ continue;
+ }
+ }
+
switch (inst->opcode) {
case BRW_OPCODE_MOV:
- inst->src[i] = entry->src;
+ case SHADER_OPCODE_LOAD_PAYLOAD:
+ inst->src[i] = val;
progress = true;
break;
+ case SHADER_OPCODE_POW:
+ case SHADER_OPCODE_INT_QUOTIENT:
+ case SHADER_OPCODE_INT_REMAINDER:
+ if (brw->gen < 8)
+ break;
+ /* fallthrough */
+ case BRW_OPCODE_BFI1:
+ case BRW_OPCODE_ASR:
+ case BRW_OPCODE_SHL:
+ case BRW_OPCODE_SHR:
+ case BRW_OPCODE_SUBB:
+ if (i == 1) {
+ inst->src[i] = val;
+ progress = true;
+ }
+ break;
+
+ case BRW_OPCODE_MACH:
case BRW_OPCODE_MUL:
case BRW_OPCODE_ADD:
+ case BRW_OPCODE_OR:
+ case BRW_OPCODE_AND:
+ case BRW_OPCODE_XOR:
+ case BRW_OPCODE_ADDC:
if (i == 1) {
- inst->src[i] = entry->src;
+ inst->src[i] = val;
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;
inst->src[0] = inst->src[1];
- inst->src[1] = entry->src;
+ inst->src[1] = val;
progress = true;
}
break;
case BRW_OPCODE_CMP:
case BRW_OPCODE_IF:
if (i == 1) {
- inst->src[i] = entry->src;
+ inst->src[i] = val;
progress = true;
} else if (i == 0 && inst->src[1].file != IMM) {
- uint32_t new_cmod;
+ enum brw_conditional_mod new_cmod;
new_cmod = brw_swap_cmod(inst->conditional_mod);
- if (new_cmod != ~0u) {
+ if (new_cmod != BRW_CONDITIONAL_NONE) {
/* Fit this constant in by swapping the operands and
* flipping the test
*/
inst->src[0] = inst->src[1];
- inst->src[1] = entry->src;
+ inst->src[1] = val;
inst->conditional_mod = new_cmod;
progress = true;
}
case BRW_OPCODE_SEL:
if (i == 1) {
- inst->src[i] = entry->src;
+ inst->src[i] = val;
progress = true;
} else if (i == 0 && inst->src[1].file != IMM) {
inst->src[0] = inst->src[1];
- inst->src[1] = entry->src;
+ inst->src[1] = val;
/* If this was predicated, flipping operands means
* we also need to flip the predicate.
* anyway.
*/
assert(i == 0);
- if (inst->src[0].imm.f != 0.0f) {
+ if (inst->src[0].fixed_hw_reg.dw1.f != 0.0f) {
inst->opcode = BRW_OPCODE_MOV;
- inst->src[0] = entry->src;
- inst->src[0].imm.f = 1.0f / inst->src[0].imm.f;
+ inst->src[0] = val;
+ inst->src[0].fixed_hw_reg.dw1.f = 1.0f / inst->src[0].fixed_hw_reg.dw1.f;
progress = true;
}
break;
- case FS_OPCODE_PULL_CONSTANT_LOAD:
- inst->src[i] = entry->src;
+ case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
+ inst->src[i] = val;
+ progress = true;
+ break;
+
+ case BRW_OPCODE_MAD:
+ case BRW_OPCODE_LRP:
+ inst->src[i] = val;
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).
- */
+static bool
+can_propagate_from(fs_inst *inst)
+{
+ return (inst->opcode == BRW_OPCODE_MOV &&
+ inst->dst.file == GRF &&
+ ((inst->src[0].file == GRF &&
+ (inst->src[0].reg != inst->dst.reg ||
+ inst->src[0].reg_offset != inst->dst.reg_offset)) ||
+ inst->src[0].file == UNIFORM ||
+ inst->src[0].file == IMM) &&
+ inst->src[0].type == inst->dst.type &&
+ !inst->is_partial_write());
+}
/* Walks a basic block and does copy propagation on it using the acp
* list.
*/
bool
-fs_visitor::opt_copy_propagate_local(void *mem_ctx,
- fs_bblock *block, exec_list *acp)
+fs_visitor::opt_copy_propagate_local(void *copy_prop_ctx, bblock_t *block,
+ exec_list *acp)
{
bool progress = false;
- for (fs_inst *inst = block->start;
- inst != block->end->next;
- inst = (fs_inst *)inst->next) {
-
+ foreach_inst_in_block(fs_inst, inst, block) {
/* Try propagating into this instruction. */
- foreach_list(entry_node, acp) {
- acp_entry *entry = (acp_entry *)entry_node;
+ for (int i = 0; i < inst->sources; i++) {
+ if (inst->src[i].file != GRF)
+ continue;
- if (try_constant_propagate(inst, entry))
- progress = true;
+ foreach_in_list(acp_entry, entry, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
+ if (try_constant_propagate(inst, entry))
+ progress = true;
- for (int i = 0; i < 3; i++) {
- if (try_copy_propagate(inst, i, entry))
- progress = true;
- }
+ if (try_copy_propagate(inst, i, entry))
+ progress = true;
+ }
}
/* kill the destination from the ACP */
if (inst->dst.file == GRF) {
- foreach_list_safe(entry_node, acp) {
- acp_entry *entry = (acp_entry *)entry_node;
-
- if (inst->overwrites_reg(entry->dst) ||
- inst->overwrites_reg(entry->src)) {
+ foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
+ if (inst->overwrites_reg(entry->dst)) {
entry->remove();
}
}
+
+ /* 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_HASH_SIZE; i++) {
+ foreach_in_list_safe(acp_entry, entry, &acp[i]) {
+ if (inst->overwrites_reg(entry->src))
+ entry->remove();
+ }
+ }
}
- /* If this instruction is a raw copy, add it to the ACP. */
- if (inst->opcode == BRW_OPCODE_MOV &&
- inst->dst.file == GRF &&
- ((inst->src[0].file == GRF &&
- (inst->src[0].reg != inst->dst.reg ||
- inst->src[0].reg_offset != inst->dst.reg_offset)) ||
- inst->src[0].file == UNIFORM ||
- 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) {
- acp_entry *entry = ralloc(mem_ctx, acp_entry);
+ /* If this instruction's source could potentially be folded into the
+ * operand of another instruction, add it to the ACP.
+ */
+ if (can_propagate_from(inst)) {
+ acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
entry->dst = inst->dst;
entry->src = inst->src[0];
- acp->push_tail(entry);
+ entry->regs_written = inst->regs_written;
+ entry->opcode = inst->opcode;
+ entry->saturate = inst->saturate;
+ acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
+ } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
+ inst->dst.file == GRF) {
+ int offset = 0;
+ for (int i = 0; i < inst->sources; i++) {
+ int regs_written = ((inst->src[i].effective_width *
+ type_sz(inst->src[i].type)) + 31) / 32;
+ if (inst->src[i].file == GRF) {
+ acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
+ entry->dst = inst->dst;
+ entry->dst.reg_offset = offset;
+ entry->dst.width = inst->src[i].effective_width;
+ entry->src = inst->src[i];
+ entry->regs_written = regs_written;
+ entry->opcode = inst->opcode;
+ if (!entry->dst.equals(inst->src[i])) {
+ acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
+ } else {
+ ralloc_free(entry);
+ }
+ }
+ offset += regs_written;
+ }
}
}
fs_visitor::opt_copy_propagate()
{
bool progress = false;
- void *mem_ctx = ralloc_context(this->mem_ctx);
+ void *copy_prop_ctx = ralloc_context(NULL);
+ 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.
+ */
+ foreach_block (block, cfg) {
+ progress = opt_copy_propagate_local(copy_prop_ctx, block,
+ out_acp[block->num]) || progress;
+ }
+
+ /* Do dataflow analysis for those available copies. */
+ fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, out_acp);
- fs_cfg cfg(this);
+ /* 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.
+ */
+ foreach_block (block, cfg) {
+ exec_list in_acp[ACP_HASH_SIZE];
- for (int b = 0; b < cfg.num_blocks; b++) {
- fs_bblock *block = cfg.blocks[b];
- exec_list acp;
+ for (int i = 0; i < dataflow.num_acp; i++) {
+ if (BITSET_TEST(dataflow.bd[block->num].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, &acp) || progress;
+ progress = opt_copy_propagate_local(copy_prop_ctx, block, in_acp) || progress;
}
- ralloc_free(mem_ctx);
+ for (int i = 0; i < cfg->num_blocks; i++)
+ delete [] out_acp[i];
+ ralloc_free(copy_prop_ctx);
if (progress)
- live_intervals_valid = false;
+ invalidate_live_intervals();
return progress;
}
projects / mesa.git / blobdiff