*/
#include "compiler.h"
+#include "util/u_memory.h"
+#include "midgard_ops.h"
-/* Basic dead code elimination on the MIR itself */
+/* SIMD-aware dead code elimination. Perform liveness analysis step-by-step,
+ * removing dead components. If an instruction ends up with a zero mask, the
+ * instruction in total is dead and should be removed. */
-bool
-midgard_opt_dead_code_eliminate(compiler_context *ctx, midgard_block *block)
+static bool
+can_cull_mask(compiler_context *ctx, midgard_instruction *ins)
+{
+ if (ins->dest >= ctx->temp_count)
+ return false;
+
+ if (ins->type == TAG_LOAD_STORE_4)
+ if (load_store_opcode_props[ins->load_store.op].props & LDST_SPECIAL_MASK)
+ return false;
+
+ return true;
+}
+
+static bool
+can_dce(midgard_instruction *ins)
+{
+ if (ins->mask)
+ return false;
+
+ if (ins->compact_branch)
+ return false;
+
+ if (ins->type == TAG_LOAD_STORE_4)
+ if (load_store_opcode_props[ins->load_store.op].props & LDST_SIDE_FX)
+ return false;
+
+ if (ins->type == TAG_TEXTURE_4)
+ if (ins->texture.op == TEXTURE_OP_BARRIER)
+ return false;
+
+ return true;
+}
+
+static bool
+midgard_opt_dead_code_eliminate_block(compiler_context *ctx, midgard_block *block)
{
bool progress = false;
+ uint16_t *live = mem_dup(block->base.live_out, ctx->temp_count * sizeof(uint16_t));
+
+ mir_foreach_instr_in_block_rev(block, ins) {
+ if (can_cull_mask(ctx, ins)) {
+ unsigned type_size = nir_alu_type_get_type_size(ins->dest_type);
+ unsigned round_size = type_size;
+ unsigned oldmask = ins->mask;
+
+ /* Make sure we're packable */
+ if (type_size == 16 && ins->type == TAG_LOAD_STORE_4)
+ round_size = 32;
+
+ unsigned rounded = mir_round_bytemask_up(live[ins->dest], round_size);
+ unsigned cmask = mir_from_bytemask(rounded, type_size);
+
+ ins->mask &= cmask;
+ progress |= (ins->mask != oldmask);
+ }
+
+ mir_liveness_ins_update(live, ins, ctx->temp_count);
+ }
+
mir_foreach_instr_in_block_safe(block, ins) {
- if (ins->type != TAG_ALU_4) continue;
- if (ins->compact_branch) continue;
+ if (can_dce(ins)) {
+ mir_remove_instruction(ins);
+ progress = true;
+ }
+ }
+
+ free(live);
- if (ins->ssa_args.dest >= SSA_FIXED_MINIMUM) continue;
- if (mir_is_live_after(ctx, block, ins, ins->ssa_args.dest)) continue;
+ return progress;
+}
+
+bool
+midgard_opt_dead_code_eliminate(compiler_context *ctx)
+{
+ /* We track liveness. In fact, it's ok if we assume more things are
+ * live than they actually are, that just reduces the effectiveness of
+ * this iterations lightly. And DCE has the effect of strictly reducing
+ * liveness, so we can run DCE across all blocks while only computing
+ * liveness at the beginning. */
+
+ mir_invalidate_liveness(ctx);
+ mir_compute_liveness(ctx);
+
+ bool progress = false;
- mir_remove_instruction(ins);
- progress = true;
+ mir_foreach_block(ctx, block) {
+ progress |= midgard_opt_dead_code_eliminate_block(ctx, (midgard_block *) block);
}
return progress;
mir_foreach_instr_in_block_from(block, q, mir_next_op(ins)) {
/* Check if used */
- if (mir_has_arg(q, ins->ssa_args.dest))
+ if (mir_has_arg(q, ins->dest))
break;
/* Check if overwritten */
- if (q->ssa_args.dest == ins->ssa_args.dest) {
+ if (q->dest == ins->dest) {
/* Special case to vec4; component tracking is
* harder */
return progress;
}
-
-/* An even further special case - to be run after RA runs but before
- * scheduling, eliminating moves that end up being useless even though they
- * appeared meaningful in the SSA. Part #2 of register coalescing. */
-
-void
-midgard_opt_post_move_eliminate(compiler_context *ctx, midgard_block *block, struct ra_graph *g)
-{
- mir_foreach_instr_in_block_safe(block, ins) {
- if (ins->type != TAG_ALU_4) continue;
- if (ins->compact_branch) continue;
- if (!OP_IS_MOVE(ins->alu.op)) continue;
-
- /* Check we're to the same place post-RA */
- unsigned iA = ins->ssa_args.dest;
- unsigned iB = ins->ssa_args.src[1];
-
- if ((iA < 0) || (iB < 0)) continue;
-
- unsigned A = iA >= SSA_FIXED_MINIMUM ?
- SSA_REG_FROM_FIXED(iA) :
- ra_get_node_reg(g, iA);
-
- unsigned B = iB >= SSA_FIXED_MINIMUM ?
- SSA_REG_FROM_FIXED(iB) :
- ra_get_node_reg(g, iB);
-
- if (A != B) continue;
-
- /* Check we're in the work zone. TODO: promoted
- * uniforms? */
- if (A >= 16) continue;
-
- /* Ensure there aren't side effects */
- if (mir_nontrivial_source2_mod(ins)) continue;
- if (mir_nontrivial_outmod(ins)) continue;
- if (ins->mask != 0xF) continue;
-
- /* We do want to rewrite to keep the graph sane for pipeline
- * register creation (TODO: is this the best approach?) */
- mir_rewrite_index_dst(ctx, ins->ssa_args.src[1], ins->ssa_args.dest);
-
- /* We're good to go */
- mir_remove_instruction(ins);
-
- }
-
-}