*
*/
-/* Create a mask of accessed components from a swizzle to figure out vector
- * dependencies */
+/* We create the dependency graph with per-component granularity */
-static unsigned
-swizzle_to_access_mask(unsigned swizzle)
-{
- unsigned component_mask = 0;
+#define COMPONENT_COUNT 8
- for (int i = 0; i < 4; ++i) {
- unsigned c = (swizzle >> (2 * i)) & 3;
- component_mask |= (1 << c);
- }
+static void
+add_dependency(struct util_dynarray *table, unsigned index, unsigned mask, midgard_instruction **instructions, unsigned child)
+{
+ for (unsigned i = 0; i < COMPONENT_COUNT; ++i) {
+ if (!(mask & (1 << i)))
+ continue;
- return component_mask;
-}
+ struct util_dynarray *parents = &table[(COMPONENT_COUNT * index) + i];
-/* Does the mask cover more than a scalar? */
+ util_dynarray_foreach(parents, unsigned, parent) {
+ BITSET_WORD *dependents = instructions[*parent]->dependents;
-static bool
-is_single_component_mask(unsigned mask)
-{
- int components = 0;
+ /* Already have the dependency */
+ if (BITSET_TEST(dependents, child))
+ continue;
- for (int c = 0; c < 8; ++c) {
- if (mask & (1 << c))
- components++;
+ BITSET_SET(dependents, child);
+ instructions[child]->nr_dependencies++;
+ }
}
-
- return components == 1;
}
-/* Checks for an SSA data hazard between two adjacent instructions, keeping in
- * mind that we are a vector architecture and we can write to different
- * components simultaneously */
-
-static bool
-can_run_concurrent_ssa(midgard_instruction *first, midgard_instruction *second)
+static void
+mark_access(struct util_dynarray *table, unsigned index, unsigned mask, unsigned parent)
{
- /* Writeout has its own rules anyway */
- if (first->compact_branch || second->compact_branch)
- return true;
-
- /* Each instruction reads some registers and writes to a register. See
- * where the first writes */
-
- int source = first->dest;
- int source_mask = first->mask;
-
- /* As long as the second doesn't read from the first, we're okay */
- for (unsigned i = 0; i < ARRAY_SIZE(second->src); ++i) {
- if (second->src[i] != source)
+ for (unsigned i = 0; i < COMPONENT_COUNT; ++i) {
+ if (!(mask & (1 << i)))
continue;
- if (first->type != TAG_ALU_4)
- return false;
-
- /* Figure out which components we just read from */
-
- int q = (i == 0) ? second->alu.src1 : second->alu.src2;
- midgard_vector_alu_src *m = (midgard_vector_alu_src *) &q;
-
- /* Check if there are components in common, and fail if so */
- if (swizzle_to_access_mask(m->swizzle) & source_mask)
- return false;
+ util_dynarray_append(&table[(COMPONENT_COUNT * index) + i], unsigned, parent);
}
-
- /* Otherwise, it's safe in that regard. Another data hazard is both
- * writing to the same place, of course */
-
- if (second->dest == source) {
- /* ...but only if the components overlap */
-
- if (second->mask & source_mask)
- return false;
- }
-
- /* ...That's it */
- return true;
}
-static bool
-midgard_has_hazard(
- midgard_instruction **segment, unsigned segment_size,
- midgard_instruction *ains)
+static void
+mir_create_dependency_graph(midgard_instruction **instructions, unsigned count, unsigned node_count)
{
- for (int s = 0; s < segment_size; ++s)
- if (!can_run_concurrent_ssa(segment[s], ains))
- return true;
+ size_t sz = node_count * COMPONENT_COUNT;
- return false;
+ struct util_dynarray *last_read = calloc(sizeof(struct util_dynarray), sz);
+ struct util_dynarray *last_write = calloc(sizeof(struct util_dynarray), sz);
+ for (unsigned i = 0; i < sz; ++i) {
+ util_dynarray_init(&last_read[i], NULL);
+ util_dynarray_init(&last_write[i], NULL);
+ }
-}
-
-/* Fragment writeout (of r0) is allowed when:
- *
- * - All components of r0 are written in the bundle
- * - No components of r0 are written in VLUT
- * - Non-pipelined dependencies of r0 are not written in the bundle
- *
- * This function checks if these requirements are satisfied given the content
- * of a scheduled bundle.
- */
-
-static bool
-can_writeout_fragment(compiler_context *ctx, midgard_instruction **bundle, unsigned count, unsigned node_count, unsigned r0)
-{
- /* First scan for which components of r0 are written out. Initially
- * none are written */
-
- uint8_t r0_written_mask = 0x0;
-
- /* Simultaneously we scan for the set of dependencies */
+ /* Initialize dependency graph */
+ for (unsigned i = 0; i < count; ++i) {
+ instructions[i]->dependents =
+ calloc(BITSET_WORDS(count), sizeof(BITSET_WORD));
- size_t sz = sizeof(BITSET_WORD) * BITSET_WORDS(node_count);
- BITSET_WORD *dependencies = calloc(1, sz);
- memset(dependencies, 0, sz);
+ instructions[i]->nr_dependencies = 0;
+ }
- bool success = false;
+ /* Populate dependency graph */
+ for (signed i = count - 1; i >= 0; --i) {
+ if (instructions[i]->compact_branch)
+ continue;
- for (unsigned i = 0; i < count; ++i) {
- midgard_instruction *ins = bundle[i];
+ unsigned dest = instructions[i]->dest;
+ unsigned mask = instructions[i]->mask;
- if (ins->dest != r0)
- continue;
+ mir_foreach_src((*instructions), s) {
+ unsigned src = instructions[i]->src[s];
- /* Record written out mask */
- r0_written_mask |= ins->mask;
+ if (src < node_count) {
+ unsigned readmask = mir_mask_of_read_components(instructions[i], src);
+ add_dependency(last_write, src, readmask, instructions, i);
+ }
+ }
- /* Record dependencies, but only if they won't become pipeline
- * registers. We know we can't be live after this, because
- * we're writeout at the very end of the shader. So check if
- * they were written before us. */
+ if (dest < node_count) {
+ add_dependency(last_read, dest, mask, instructions, i);
+ add_dependency(last_write, dest, mask, instructions, i);
+ mark_access(last_write, dest, mask, i);
+ }
- unsigned src0 = ins->src[0];
- unsigned src1 = ins->src[1];
+ mir_foreach_src((*instructions), s) {
+ unsigned src = instructions[i]->src[s];
- if (!mir_is_written_before(ctx, bundle[0], src0))
- src0 = ~0;
+ if (src < node_count) {
+ unsigned readmask = mir_mask_of_read_components(instructions[i], src);
+ mark_access(last_read, src, readmask, i);
+ }
+ }
+ }
- if (!mir_is_written_before(ctx, bundle[0], src1))
- src1 = ~0;
+ /* If there is a branch, all instructions depend on it, as interblock
+ * execution must be purely in-order */
- if (src0 < node_count)
- BITSET_SET(dependencies, src0);
+ if (instructions[count - 1]->compact_branch) {
+ BITSET_WORD *dependents = instructions[count - 1]->dependents;
- if (src1 < node_count)
- BITSET_SET(dependencies, src1);
+ for (signed i = count - 2; i >= 0; --i) {
+ if (BITSET_TEST(dependents, i))
+ continue;
- /* Requirement 2 */
- if (ins->unit == UNIT_VLUT)
- goto done;
+ BITSET_SET(dependents, i);
+ instructions[i]->nr_dependencies++;
+ }
}
- /* Requirement 1 */
- if ((r0_written_mask & 0xF) != 0xF)
- goto done;
+ /* Free the intermediate structures */
+ for (unsigned i = 0; i < sz; ++i) {
+ util_dynarray_fini(&last_read[i]);
+ util_dynarray_fini(&last_write[i]);
+ }
+}
- /* Requirement 3 */
+/* Does the mask cover more than a scalar? */
- for (unsigned i = 0; i < count; ++i) {
- unsigned dest = bundle[i]->dest;
+static bool
+is_single_component_mask(unsigned mask)
+{
+ int components = 0;
- if (dest < node_count && BITSET_TEST(dependencies, dest))
- goto done;
+ for (int c = 0; c < 8; ++c) {
+ if (mask & (1 << c))
+ components++;
}
- /* Otherwise, we're good to go */
- success = true;
-
-done:
- free(dependencies);
- return success;
+ return components == 1;
}
/* Helpers for scheudling */
static bool
mir_is_scalar(midgard_instruction *ains)
{
- /* Does the op support scalar units? */
- if (!(alu_opcode_props[ains->alu.op].props & UNITS_SCALAR))
- return false;
-
/* Do we try to use it as a vector op? */
if (!is_single_component_mask(ains->mask))
return false;
return sizeof(midgard_reg_info) + sizeof(midgard_scalar_alu);
}
-/* Schedules, but does not emit, a single basic block. After scheduling, the
- * final tag and size of the block are known, which are necessary for branching
- * */
+/* We would like to flatten the linked list of midgard_instructions in a bundle
+ * to an array of pointers on the heap for easy indexing */
-static midgard_bundle
-schedule_bundle(compiler_context *ctx, midgard_block *block, midgard_instruction *ins, int *skip)
+static midgard_instruction **
+flatten_mir(midgard_block *block, unsigned *len)
{
- int instructions_emitted = 0, packed_idx = 0;
- midgard_bundle bundle = { 0 };
+ *len = list_length(&block->instructions);
- midgard_instruction *scheduled[5] = { NULL };
+ if (!(*len))
+ return NULL;
- uint8_t tag = ins->type;
+ midgard_instruction **instructions =
+ calloc(sizeof(midgard_instruction *), *len);
- /* Default to the instruction's tag */
- bundle.tag = tag;
+ unsigned i = 0;
- switch (ins->type) {
- case TAG_ALU_4: {
- uint32_t control = 0;
- size_t bytes_emitted = sizeof(control);
+ mir_foreach_instr_in_block(block, ins)
+ instructions[i++] = ins;
- /* TODO: Constant combining */
- int index = 0, last_unit = 0;
+ return instructions;
+}
- /* Previous instructions, for the purpose of parallelism */
- midgard_instruction *segment[4] = {0};
- int segment_size = 0;
+/* The worklist is the set of instructions that can be scheduled now; that is,
+ * the set of instructions with no remaining dependencies */
- instructions_emitted = -1;
- midgard_instruction *pins = ins;
+static void
+mir_initialize_worklist(BITSET_WORD *worklist, midgard_instruction **instructions, unsigned count)
+{
+ for (unsigned i = 0; i < count; ++i) {
+ if (instructions[i]->nr_dependencies == 0)
+ BITSET_SET(worklist, i);
+ }
+}
- unsigned constant_count = 0;
+/* Update the worklist after an instruction terminates. Remove its edges from
+ * the graph and if that causes any node to have no dependencies, add it to the
+ * worklist */
- for (;;) {
- midgard_instruction *ains = pins;
+static void
+mir_update_worklist(
+ BITSET_WORD *worklist, unsigned count,
+ midgard_instruction **instructions, midgard_instruction *done)
+{
+ /* Sanity check: if no instruction terminated, there is nothing to do.
+ * If the instruction that terminated had dependencies, that makes no
+ * sense and means we messed up the worklist. Finally, as the purpose
+ * of this routine is to update dependents, we abort early if there are
+ * no dependents defined. */
- /* Advance instruction pointer */
- if (index) {
- ains = mir_next_op(pins);
- pins = ains;
- }
+ if (!done)
+ return;
- /* Out-of-work condition */
- if ((struct list_head *) ains == &block->instructions)
- break;
-
- /* Ensure that the chain can continue */
- if (ains->type != TAG_ALU_4) break;
-
- /* If there's already something in the bundle and we
- * have weird scheduler constraints, break now */
- if (ains->precede_break && index) break;
-
- /* According to the presentation "The ARM
- * Mali-T880 Mobile GPU" from HotChips 27,
- * there are two pipeline stages. Branching
- * position determined experimentally. Lines
- * are executed in parallel:
- *
- * [ VMUL ] [ SADD ]
- * [ VADD ] [ SMUL ] [ LUT ] [ BRANCH ]
- *
- * Verify that there are no ordering dependencies here.
- *
- * TODO: Allow for parallelism!!!
- */
-
- /* Pick a unit for it if it doesn't force a particular unit */
-
- int unit = ains->unit;
-
- if (!unit) {
- int op = ains->alu.op;
- int units = alu_opcode_props[op].props;
- bool scalar = mir_is_scalar(ains);
-
- if (!scalar) {
- if (last_unit >= UNIT_VADD) {
- if (units & UNIT_VLUT)
- unit = UNIT_VLUT;
- else
- break;
- } else {
- if ((units & UNIT_VMUL) && last_unit < UNIT_VMUL)
- unit = UNIT_VMUL;
- else if ((units & UNIT_VADD) && !(control & UNIT_VADD))
- unit = UNIT_VADD;
- else if (units & UNIT_VLUT)
- unit = UNIT_VLUT;
- else
- break;
- }
- } else {
- if (last_unit >= UNIT_VADD) {
- if ((units & UNIT_SMUL) && !(control & UNIT_SMUL))
- unit = UNIT_SMUL;
- else if (units & UNIT_VLUT)
- unit = UNIT_VLUT;
- else
- break;
- } else {
- if ((units & UNIT_VMUL) && (last_unit < UNIT_VMUL))
- unit = UNIT_VMUL;
- else if ((units & UNIT_SADD) && !(control & UNIT_SADD) && !midgard_has_hazard(segment, segment_size, ains))
- unit = UNIT_SADD;
- else if (units & UNIT_VADD)
- unit = UNIT_VADD;
- else if (units & UNIT_SMUL)
- unit = UNIT_SMUL;
- else if (units & UNIT_VLUT)
- unit = UNIT_VLUT;
- else
- break;
- }
- }
+ assert(done->nr_dependencies == 0);
- assert(unit & units);
- }
+ if (!done->dependents)
+ return;
- /* Late unit check, this time for encoding (not parallelism) */
- if (unit <= last_unit) break;
+ /* We have an instruction with dependents. Iterate each dependent to
+ * remove one dependency (`done`), adding dependents to the worklist
+ * where possible. */
- /* Clear the segment */
- if (last_unit < UNIT_VADD && unit >= UNIT_VADD)
- segment_size = 0;
+ unsigned i;
+ BITSET_WORD tmp;
+ BITSET_FOREACH_SET(i, tmp, done->dependents, count) {
+ assert(instructions[i]->nr_dependencies);
- if (midgard_has_hazard(segment, segment_size, ains))
- break;
+ if (!(--instructions[i]->nr_dependencies))
+ BITSET_SET(worklist, i);
+ }
- /* We're good to go -- emit the instruction */
- ains->unit = unit;
+ free(done->dependents);
+}
- segment[segment_size++] = ains;
+/* While scheduling, we need to choose instructions satisfying certain
+ * criteria. As we schedule backwards, we choose the *last* instruction in the
+ * worklist to simulate in-order scheduling. Chosen instructions must satisfy a
+ * given predicate. */
- /* We try to reuse constants if possible, by adjusting
- * the swizzle */
+struct midgard_predicate {
+ /* TAG or ~0 for dont-care */
+ unsigned tag;
- if (ains->has_blend_constant) {
- /* Everything conflicts with the blend constant */
- if (bundle.has_embedded_constants)
- break;
+ /* True if we want to pop off the chosen instruction */
+ bool destructive;
- bundle.has_blend_constant = 1;
- bundle.has_embedded_constants = 1;
- } else if (ains->has_constants && ains->alu.reg_mode == midgard_reg_mode_16) {
- /* TODO: DRY with the analysis pass */
+ /* For ALU, choose only this unit */
+ unsigned unit;
- if (bundle.has_blend_constant)
- break;
+ /* State for bundle constants. constants is the actual constants
+ * for the bundle. constant_count is the number of bytes (up to
+ * 16) currently in use for constants. When picking in destructive
+ * mode, the constants array will be updated, and the instruction
+ * will be adjusted to index into the constants array */
- if (constant_count)
- break;
+ uint8_t *constants;
+ unsigned constant_count;
+ bool blend_constant;
- /* TODO: Fix packing XXX */
- uint16_t *bundles = (uint16_t *) bundle.constants;
- uint32_t *constants = (uint32_t *) ains->constants;
+ /* Exclude this destination (if not ~0) */
+ unsigned exclude;
- /* Copy them wholesale */
- for (unsigned i = 0; i < 4; ++i)
- bundles[i] = constants[i];
+ /* Don't schedule instructions consuming conditionals (since we already
+ * scheduled one). Excludes conditional branches and csel */
+ bool no_cond;
+};
- bundle.has_embedded_constants = true;
- constant_count = 4;
- } else if (ains->has_constants) {
- /* By definition, blend constants conflict with
- * everything, so if there are already
- * constants we break the bundle *now* */
+/* For an instruction that can fit, adjust it to fit and update the constants
+ * array, in destructive mode. Returns whether the fitting was successful. */
- if (bundle.has_blend_constant)
- break;
+static bool
+mir_adjust_constants(midgard_instruction *ins,
+ struct midgard_predicate *pred,
+ bool destructive)
+{
+ /* Blend constants dominate */
+ if (ins->has_blend_constant) {
+ if (pred->constant_count)
+ return false;
+ else if (destructive) {
+ pred->blend_constant = true;
+ pred->constant_count = 16;
+ return true;
+ }
+ }
- /* For anything but blend constants, we can do
- * proper analysis, however */
+ /* No constant, nothing to adjust */
+ if (!ins->has_constants)
+ return true;
- /* TODO: Mask by which are used */
- uint32_t *constants = (uint32_t *) ains->constants;
- uint32_t *bundles = (uint32_t *) bundle.constants;
+ /* TODO: Deduplicate; permit multiple constants within a bundle */
- uint32_t indices[4] = { 0 };
- bool break_bundle = false;
+ if (destructive && !pred->constant_count) {
+ if (ins->alu.reg_mode == midgard_reg_mode_16) {
+ /* TODO: Fix packing XXX */
+ uint16_t *bundles = (uint16_t *) pred->constants;
+ uint32_t *constants = (uint32_t *) ins->constants;
- for (unsigned i = 0; i < 4; ++i) {
- uint32_t cons = constants[i];
- bool constant_found = false;
+ /* Copy them wholesale */
+ for (unsigned i = 0; i < 4; ++i)
+ bundles[i] = constants[i];
+ } else {
+ memcpy(pred->constants, ins->constants, 16);
+ }
- /* Search for the constant */
- for (unsigned j = 0; j < constant_count; ++j) {
- if (bundles[j] != cons)
- continue;
+ pred->constant_count = 16;
+ return true;
+ }
- /* We found it, reuse */
- indices[i] = j;
- constant_found = true;
- break;
- }
+ return !pred->constant_count;
+}
- if (constant_found)
- continue;
+static midgard_instruction *
+mir_choose_instruction(
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned count,
+ struct midgard_predicate *predicate)
+{
+ /* Parse the predicate */
+ unsigned tag = predicate->tag;
+ bool alu = tag == TAG_ALU_4;
+ unsigned unit = predicate->unit;
+ bool branch = alu && (unit == ALU_ENAB_BR_COMPACT);
+ bool scalar = (unit != ~0) && (unit & UNITS_SCALAR);
+ bool no_cond = predicate->no_cond;
+
+ /* Iterate to find the best instruction satisfying the predicate */
+ unsigned i;
+ BITSET_WORD tmp;
+
+ signed best_index = -1;
+ bool best_conditional = false;
+
+ /* Enforce a simple metric limiting distance to keep down register
+ * pressure. TOOD: replace with liveness tracking for much better
+ * results */
+
+ unsigned max_active = 0;
+ unsigned max_distance = 6;
+
+ BITSET_FOREACH_SET(i, tmp, worklist, count) {
+ max_active = MAX2(max_active, i);
+ }
- /* We didn't find it, so allocate it */
- unsigned idx = constant_count++;
+ BITSET_FOREACH_SET(i, tmp, worklist, count) {
+ if ((max_active - i) >= max_distance)
+ continue;
- if (idx >= 4) {
- /* Uh-oh, out of space */
- break_bundle = true;
- break;
- }
+ if (tag != ~0 && instructions[i]->type != tag)
+ continue;
- /* We have space, copy it in! */
- bundles[idx] = cons;
- indices[i] = idx;
- }
+ if (predicate->exclude != ~0 && instructions[i]->dest == predicate->exclude)
+ continue;
- if (break_bundle)
- break;
+ if (alu && !branch && !(alu_opcode_props[instructions[i]->alu.op].props & unit))
+ continue;
- /* Cool, we have it in. So use indices as a
- * swizzle */
+ if (branch && !instructions[i]->compact_branch)
+ continue;
- unsigned swizzle = SWIZZLE_FROM_ARRAY(indices);
- unsigned r_constant = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
+ if (alu && scalar && !mir_is_scalar(instructions[i]))
+ continue;
- if (ains->src[0] == r_constant)
- ains->alu.src1 = vector_alu_apply_swizzle(ains->alu.src1, swizzle);
+ if (alu && !mir_adjust_constants(instructions[i], predicate, false))
+ continue;
- if (ains->src[1] == r_constant)
- ains->alu.src2 = vector_alu_apply_swizzle(ains->alu.src2, swizzle);
+ bool conditional = alu && !branch && OP_IS_CSEL(instructions[i]->alu.op);
+ conditional |= (branch && !instructions[i]->prepacked_branch && instructions[i]->branch.conditional);
- bundle.has_embedded_constants = true;
- }
+ if (conditional && no_cond)
+ continue;
- if (ains->compact_branch) {
- /* All of r0 has to be written out along with
- * the branch writeout */
-
- if (ains->writeout && !can_writeout_fragment(ctx, scheduled, index, ctx->temp_count, ains->src[0])) {
- /* We only work on full moves
- * at the beginning. We could
- * probably do better */
- if (index != 0)
- break;
-
- /* Inject a move */
- midgard_instruction ins = v_mov(0, blank_alu_src, SSA_FIXED_REGISTER(0));
- ins.unit = UNIT_VMUL;
- control |= ins.unit;
-
- /* TODO don't leak */
- midgard_instruction *move =
- mem_dup(&ins, sizeof(midgard_instruction));
- bytes_emitted += bytes_for_instruction(move);
- bundle.instructions[packed_idx++] = move;
- }
- }
+ /* Simulate in-order scheduling */
+ if ((signed) i < best_index)
+ continue;
+
+ best_index = i;
+ best_conditional = conditional;
+ }
- bytes_emitted += bytes_for_instruction(ains);
- /* Defer marking until after writing to allow for break */
- scheduled[index] = ains;
- control |= ains->unit;
- last_unit = ains->unit;
- ++instructions_emitted;
- ++index;
- }
+ /* Did we find anything? */
- int padding = 0;
+ if (best_index < 0)
+ return NULL;
- /* Pad ALU op to nearest word */
+ /* If we found something, remove it from the worklist */
+ assert(best_index < count);
- if (bytes_emitted & 15) {
- padding = 16 - (bytes_emitted & 15);
- bytes_emitted += padding;
- }
+ if (predicate->destructive) {
+ BITSET_CLEAR(worklist, best_index);
+
+ if (alu)
+ mir_adjust_constants(instructions[best_index], predicate, true);
+
+ /* Once we schedule a conditional, we can't again */
+ predicate->no_cond |= best_conditional;
+ }
+
+ return instructions[best_index];
+}
+
+/* Still, we don't choose instructions in a vacuum. We need a way to choose the
+ * best bundle type (ALU, load/store, texture). Nondestructive. */
+
+static unsigned
+mir_choose_bundle(
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned count)
+{
+ /* At the moment, our algorithm is very simple - use the bundle of the
+ * best instruction, regardless of what else could be scheduled
+ * alongside it. This is not optimal but it works okay for in-order */
+
+ struct midgard_predicate predicate = {
+ .tag = ~0,
+ .destructive = false,
+ .exclude = ~0
+ };
+
+ midgard_instruction *chosen = mir_choose_instruction(instructions, worklist, count, &predicate);
+
+ if (chosen)
+ return chosen->type;
+ else
+ return ~0;
+}
- /* Constants must always be quadwords */
- if (bundle.has_embedded_constants)
- bytes_emitted += 16;
+/* We want to choose an ALU instruction filling a given unit */
+static void
+mir_choose_alu(midgard_instruction **slot,
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned len,
+ struct midgard_predicate *predicate,
+ unsigned unit)
+{
+ /* Did we already schedule to this slot? */
+ if ((*slot) != NULL)
+ return;
+
+ /* Try to schedule something, if not */
+ predicate->unit = unit;
+ *slot = mir_choose_instruction(instructions, worklist, len, predicate);
+
+ /* Store unit upon scheduling */
+ if (*slot && !((*slot)->compact_branch))
+ (*slot)->unit = unit;
+}
+
+/* When we are scheduling a branch/csel, we need the consumed condition in the
+ * same block as a pipeline register. There are two options to enable this:
+ *
+ * - Move the conditional into the bundle. Preferred, but only works if the
+ * conditional is used only once and is from this block.
+ * - Copy the conditional.
+ *
+ * We search for the conditional. If it's in this block, single-use, and
+ * without embedded constants, we schedule it immediately. Otherwise, we
+ * schedule a move for it.
+ *
+ * mir_comparison_mobile is a helper to find the moveable condition.
+ */
+
+static unsigned
+mir_comparison_mobile(
+ compiler_context *ctx,
+ midgard_instruction **instructions,
+ unsigned count,
+ unsigned cond)
+{
+ if (!mir_single_use(ctx, cond))
+ return ~0;
+
+ unsigned ret = ~0;
+
+ for (unsigned i = 0; i < count; ++i) {
+ if (instructions[i]->dest != cond)
+ continue;
- /* Size ALU instruction for tag */
- bundle.tag = (TAG_ALU_4) + (bytes_emitted / 16) - 1;
- bundle.padding = padding;
- bundle.control = bundle.tag | control;
+ /* Must fit in an ALU bundle */
+ if (instructions[i]->type != TAG_ALU_4)
+ return ~0;
- break;
+ /* We'll need to rewrite to .w but that doesn't work for vector
+ * ops that don't replicate (ball/bany), so bail there */
+
+ if (GET_CHANNEL_COUNT(alu_opcode_props[instructions[i]->alu.op].props))
+ return ~0;
+
+ /* TODO: moving conditionals with constants */
+
+ if (instructions[i]->has_constants)
+ return ~0;
+
+ /* Ensure it is written only once */
+
+ if (ret != ~0)
+ return ~0;
+ else
+ ret = i;
}
- case TAG_LOAD_STORE_4: {
- /* Load store instructions have two words at once. If
- * we only have one queued up, we need to NOP pad.
- * Otherwise, we store both in succession to save space
- * and cycles -- letting them go in parallel -- skip
- * the next. The usefulness of this optimisation is
- * greatly dependent on the quality of the instruction
- * scheduler.
- */
+ return ret;
+}
- midgard_instruction *next_op = mir_next_op(ins);
+/* Using the information about the moveable conditional itself, we either pop
+ * that condition off the worklist for use now, or create a move to
+ * artificially schedule instead as a fallback */
- if ((struct list_head *) next_op != &block->instructions && next_op->type == TAG_LOAD_STORE_4) {
- /* TODO: Concurrency check */
- instructions_emitted++;
+static midgard_instruction *
+mir_schedule_comparison(
+ compiler_context *ctx,
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned count,
+ unsigned cond, bool vector, unsigned swizzle,
+ midgard_instruction *user)
+{
+ /* TODO: swizzle when scheduling */
+ unsigned comp_i =
+ (!vector && (swizzle == 0)) ?
+ mir_comparison_mobile(ctx, instructions, count, cond) : ~0;
+
+ /* If we can, schedule the condition immediately */
+ if ((comp_i != ~0) && BITSET_TEST(worklist, comp_i)) {
+ assert(comp_i < count);
+ BITSET_CLEAR(worklist, comp_i);
+ return instructions[comp_i];
+ }
+
+ /* Otherwise, we insert a move */
+ midgard_vector_alu_src csel = {
+ .swizzle = swizzle
+ };
+
+ midgard_instruction mov = v_mov(cond, csel, cond);
+ mov.mask = vector ? 0xF : 0x1;
+
+ return mir_insert_instruction_before(ctx, user, mov);
+}
+
+/* Most generally, we need instructions writing to r31 in the appropriate
+ * components */
+
+static midgard_instruction *
+mir_schedule_condition(compiler_context *ctx,
+ struct midgard_predicate *predicate,
+ BITSET_WORD *worklist, unsigned count,
+ midgard_instruction **instructions,
+ midgard_instruction *last)
+{
+ /* For a branch, the condition is the only argument; for csel, third */
+ bool branch = last->compact_branch;
+ unsigned condition_index = branch ? 0 : 2;
+
+ /* csel_v is vector; otherwise, conditions are scalar */
+ bool vector = !branch && OP_IS_CSEL_V(last->alu.op);
+
+ /* Grab the conditional instruction */
+
+ midgard_instruction *cond = mir_schedule_comparison(
+ ctx, instructions, worklist, count, last->src[condition_index],
+ vector, last->cond_swizzle, last);
+
+ /* We have exclusive reign over this (possibly move) conditional
+ * instruction. We can rewrite into a pipeline conditional register */
+
+ predicate->exclude = cond->dest;
+ cond->dest = SSA_FIXED_REGISTER(31);
+
+ if (!vector) {
+ cond->mask = (1 << COMPONENT_W);
+
+ mir_foreach_src(cond, s) {
+ if (cond->src[s] == ~0)
+ continue;
+
+ mir_set_swizzle(cond, s, (mir_get_swizzle(cond, s) << (2*3)) & 0xFF);
}
+ }
+
+ /* Schedule the unit: csel is always in the latter pipeline, so a csel
+ * condition must be in the former pipeline stage (vmul/sadd),
+ * depending on scalar/vector of the instruction itself. A branch must
+ * be written from the latter pipeline stage and a branch condition is
+ * always scalar, so it is always in smul (exception: ball/bany, which
+ * will be vadd) */
+
+ if (branch)
+ cond->unit = UNIT_SMUL;
+ else
+ cond->unit = vector ? UNIT_VMUL : UNIT_SADD;
+
+ return cond;
+}
+
+/* Schedules a single bundle of the given type */
+
+static midgard_bundle
+mir_schedule_texture(
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned len)
+{
+ struct midgard_predicate predicate = {
+ .tag = TAG_TEXTURE_4,
+ .destructive = true,
+ .exclude = ~0
+ };
+
+ midgard_instruction *ins =
+ mir_choose_instruction(instructions, worklist, len, &predicate);
+
+ mir_update_worklist(worklist, len, instructions, ins);
+
+ struct midgard_bundle out = {
+ .tag = TAG_TEXTURE_4,
+ .instruction_count = 1,
+ .instructions = { ins }
+ };
+
+ return out;
+}
+
+static midgard_bundle
+mir_schedule_ldst(
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned len)
+{
+ struct midgard_predicate predicate = {
+ .tag = TAG_LOAD_STORE_4,
+ .destructive = true,
+ .exclude = ~0
+ };
+
+ /* Try to pick two load/store ops. Second not gauranteed to exist */
+
+ midgard_instruction *ins =
+ mir_choose_instruction(instructions, worklist, len, &predicate);
+
+ midgard_instruction *pair =
+ mir_choose_instruction(instructions, worklist, len, &predicate);
+
+ struct midgard_bundle out = {
+ .tag = TAG_LOAD_STORE_4,
+ .instruction_count = pair ? 2 : 1,
+ .instructions = { ins, pair }
+ };
+
+ /* We have to update the worklist atomically, since the two
+ * instructions run concurrently (TODO: verify it's not pipelined) */
+
+ mir_update_worklist(worklist, len, instructions, ins);
+ mir_update_worklist(worklist, len, instructions, pair);
+
+ return out;
+}
+
+static midgard_bundle
+mir_schedule_alu(
+ compiler_context *ctx,
+ midgard_instruction **instructions,
+ BITSET_WORD *worklist, unsigned len)
+{
+ struct midgard_bundle bundle = {};
- break;
+ unsigned bytes_emitted = sizeof(bundle.control);
+
+ struct midgard_predicate predicate = {
+ .tag = TAG_ALU_4,
+ .destructive = true,
+ .exclude = ~0,
+ .constants = (uint8_t *) bundle.constants
+ };
+
+ midgard_instruction *vmul = NULL;
+ midgard_instruction *vadd = NULL;
+ midgard_instruction *vlut = NULL;
+ midgard_instruction *smul = NULL;
+ midgard_instruction *sadd = NULL;
+ midgard_instruction *branch = NULL;
+
+ mir_choose_alu(&branch, instructions, worklist, len, &predicate, ALU_ENAB_BR_COMPACT);
+ mir_update_worklist(worklist, len, instructions, branch);
+ bool writeout = branch && branch->writeout;
+
+ if (branch && !branch->prepacked_branch && branch->branch.conditional) {
+ midgard_instruction *cond = mir_schedule_condition(ctx, &predicate, worklist, len, instructions, branch);
+
+ if (cond->unit == UNIT_VADD)
+ vadd = cond;
+ else if (cond->unit == UNIT_SMUL)
+ smul = cond;
+ else
+ unreachable("Bad condition");
}
- case TAG_TEXTURE_4: {
- /* Which tag we use depends on the shader stage */
- bool in_frag = ctx->stage == MESA_SHADER_FRAGMENT;
- bundle.tag = in_frag ? TAG_TEXTURE_4 : TAG_TEXTURE_4_VTX;
- break;
+ mir_choose_alu(&smul, instructions, worklist, len, &predicate, UNIT_SMUL);
+
+ if (!writeout)
+ mir_choose_alu(&vlut, instructions, worklist, len, &predicate, UNIT_VLUT);
+
+ mir_choose_alu(&vadd, instructions, worklist, len, &predicate, UNIT_VADD);
+
+ mir_update_worklist(worklist, len, instructions, vlut);
+ mir_update_worklist(worklist, len, instructions, vadd);
+ mir_update_worklist(worklist, len, instructions, smul);
+
+ bool vadd_csel = vadd && OP_IS_CSEL(vadd->alu.op);
+ bool smul_csel = smul && OP_IS_CSEL(smul->alu.op);
+
+ if (vadd_csel || smul_csel) {
+ midgard_instruction *ins = vadd_csel ? vadd : smul;
+ midgard_instruction *cond = mir_schedule_condition(ctx, &predicate, worklist, len, instructions, ins);
+
+ if (cond->unit == UNIT_VMUL)
+ vmul = cond;
+ else if (cond->unit == UNIT_SADD)
+ sadd = cond;
+ else
+ unreachable("Bad condition");
}
- default:
- unreachable("Unknown tag");
- break;
+ /* Stage 2, let's schedule sadd before vmul for writeout */
+ mir_choose_alu(&sadd, instructions, worklist, len, &predicate, UNIT_SADD);
+
+ /* Check if writeout reads its own register */
+ bool bad_writeout = false;
+
+ if (branch && branch->writeout) {
+ midgard_instruction *stages[] = { sadd, vadd, smul };
+ unsigned src = (branch->src[0] == ~0) ? SSA_FIXED_REGISTER(0) : branch->src[0];
+ unsigned writeout_mask = 0x0;
+
+ for (unsigned i = 0; i < ARRAY_SIZE(stages); ++i) {
+ if (!stages[i])
+ continue;
+
+ if (stages[i]->dest != src)
+ continue;
+
+ writeout_mask |= stages[i]->mask;
+ bad_writeout |= mir_has_arg(stages[i], branch->src[0]);
+ }
+
+ /* Add a move if necessary */
+ if (bad_writeout || writeout_mask != 0xF) {
+ unsigned temp = (branch->src[0] == ~0) ? SSA_FIXED_REGISTER(0) : make_compiler_temp(ctx);
+ midgard_instruction mov = v_mov(src, blank_alu_src, temp);
+ vmul = mem_dup(&mov, sizeof(midgard_instruction));
+ vmul->unit = UNIT_VMUL;
+ vmul->mask = 0xF ^ writeout_mask;
+ /* TODO: Don't leak */
+
+ /* Rewrite to use our temp */
+
+ for (unsigned i = 0; i < ARRAY_SIZE(stages); ++i) {
+ if (stages[i])
+ mir_rewrite_index_dst_single(stages[i], src, temp);
+ }
+
+ mir_rewrite_index_src_single(branch, src, temp);
+ }
}
- /* Copy the instructions into the bundle */
- bundle.instruction_count = instructions_emitted + 1 + packed_idx;
+ mir_choose_alu(&vmul, instructions, worklist, len, &predicate, UNIT_VMUL);
+
+ mir_update_worklist(worklist, len, instructions, vmul);
+ mir_update_worklist(worklist, len, instructions, sadd);
+
+ bundle.has_blend_constant = predicate.blend_constant;
+ bundle.has_embedded_constants = predicate.constant_count > 0;
+
+ unsigned padding = 0;
+
+ /* Now that we have finished scheduling, build up the bundle */
+ midgard_instruction *stages[] = { vmul, sadd, vadd, smul, vlut, branch };
- midgard_instruction *uins = ins;
- for (; packed_idx < bundle.instruction_count; ++packed_idx) {
- assert(&uins->link != &block->instructions);
- bundle.instructions[packed_idx] = uins;
- uins = mir_next_op(uins);
+ for (unsigned i = 0; i < ARRAY_SIZE(stages); ++i) {
+ if (stages[i]) {
+ bundle.control |= stages[i]->unit;
+ bytes_emitted += bytes_for_instruction(stages[i]);
+ bundle.instructions[bundle.instruction_count++] = stages[i];
+ }
}
- *skip = instructions_emitted;
+ /* Pad ALU op to nearest word */
+
+ if (bytes_emitted & 15) {
+ padding = 16 - (bytes_emitted & 15);
+ bytes_emitted += padding;
+ }
+
+ /* Constants must always be quadwords */
+ if (bundle.has_embedded_constants)
+ bytes_emitted += 16;
+
+ /* Size ALU instruction for tag */
+ bundle.tag = (TAG_ALU_4) + (bytes_emitted / 16) - 1;
+ bundle.padding = padding;
+ bundle.control |= bundle.tag;
return bundle;
}
/* Schedule a single block by iterating its instruction to create bundles.
* While we go, tally about the bundle sizes to compute the block size. */
+
static void
schedule_block(compiler_context *ctx, midgard_block *block)
{
- util_dynarray_init(&block->bundles, NULL);
+ /* Copy list to dynamic array */
+ unsigned len = 0;
+ midgard_instruction **instructions = flatten_mir(block, &len);
- block->quadword_count = 0;
+ if (!len)
+ return;
- mir_foreach_instr_in_block(block, ins) {
- int skip;
- midgard_bundle bundle = schedule_bundle(ctx, block, ins, &skip);
- util_dynarray_append(&block->bundles, midgard_bundle, bundle);
+ /* Calculate dependencies and initial worklist */
+ unsigned node_count = ctx->temp_count + 1;
+ mir_create_dependency_graph(instructions, len, node_count);
- if (bundle.has_blend_constant) {
- unsigned offset = ctx->quadword_count + block->quadword_count + quadword_size(bundle.tag) - 1;
- ctx->blend_constant_offset = offset * 0x10;
- }
+ /* Allocate the worklist */
+ size_t sz = BITSET_WORDS(len) * sizeof(BITSET_WORD);
+ BITSET_WORD *worklist = calloc(sz, 1);
+ mir_initialize_worklist(worklist, instructions, len);
- while(skip--)
- ins = mir_next_op(ins);
+ struct util_dynarray bundles;
+ util_dynarray_init(&bundles, NULL);
- block->quadword_count += quadword_size(bundle.tag);
- }
+ block->quadword_count = 0;
+ unsigned blend_offset = 0;
- block->is_scheduled = true;
- ctx->quadword_count += block->quadword_count;
-}
+ for (;;) {
+ unsigned tag = mir_choose_bundle(instructions, worklist, len);
+ midgard_bundle bundle;
-/* The following passes reorder MIR instructions to enable better scheduling */
+ if (tag == TAG_TEXTURE_4)
+ bundle = mir_schedule_texture(instructions, worklist, len);
+ else if (tag == TAG_LOAD_STORE_4)
+ bundle = mir_schedule_ldst(instructions, worklist, len);
+ else if (tag == TAG_ALU_4)
+ bundle = mir_schedule_alu(ctx, instructions, worklist, len);
+ else
+ break;
-static void
-midgard_pair_load_store(compiler_context *ctx, midgard_block *block)
-{
- mir_foreach_instr_in_block_safe(block, ins) {
- if (ins->type != TAG_LOAD_STORE_4) continue;
-
- /* We've found a load/store op. Check if next is also load/store. */
- midgard_instruction *next_op = mir_next_op(ins);
- if (&next_op->link != &block->instructions) {
- if (next_op->type == TAG_LOAD_STORE_4) {
- /* If so, we're done since we're a pair */
- ins = mir_next_op(ins);
- continue;
- }
+ util_dynarray_append(&bundles, midgard_bundle, bundle);
- /* Maximum search distance to pair, to avoid register pressure disasters */
- int search_distance = 8;
+ if (bundle.has_blend_constant)
+ blend_offset = block->quadword_count;
- /* Otherwise, we have an orphaned load/store -- search for another load */
- mir_foreach_instr_in_block_from(block, c, mir_next_op(ins)) {
- /* Terminate search if necessary */
- if (!(search_distance--)) break;
+ block->quadword_count += quadword_size(bundle.tag);
+ }
- if (c->type != TAG_LOAD_STORE_4) continue;
+ /* We emitted bundles backwards; copy into the block in reverse-order */
- /* We can only reorder if there are no sources */
+ util_dynarray_init(&block->bundles, NULL);
+ util_dynarray_foreach_reverse(&bundles, midgard_bundle, bundle) {
+ util_dynarray_append(&block->bundles, midgard_bundle, *bundle);
+ }
- bool deps = false;
+ /* Blend constant was backwards as well. blend_offset if set is
+ * strictly positive, as an offset of zero would imply constants before
+ * any instructions which is invalid in Midgard */
- for (unsigned s = 0; s < ARRAY_SIZE(ins->src); ++s)
- deps |= (c->src[s] != ~0);
+ if (blend_offset)
+ ctx->blend_constant_offset = ((ctx->quadword_count + block->quadword_count) - blend_offset - 1) * 0x10;
- if (deps)
- continue;
+ block->is_scheduled = true;
+ ctx->quadword_count += block->quadword_count;
- /* We found one! Move it up to pair and remove it from the old location */
+ /* Reorder instructions to match bundled. First remove existing
+ * instructions and then recreate the list */
- mir_insert_instruction_before(ctx, ins, *c);
- mir_remove_instruction(c);
+ mir_foreach_instr_in_block_safe(block, ins) {
+ list_del(&ins->link);
+ }
- break;
- }
- }
+ mir_foreach_instr_in_block_scheduled_rev(block, ins) {
+ list_add(&ins->link, &block->instructions);
}
}
midgard_promote_uniforms(ctx, 16);
- mir_foreach_block(ctx, block) {
- midgard_pair_load_store(ctx, block);
- }
-
/* Must be lowered right before RA */
mir_squeeze_index(ctx);
mir_lower_special_reads(ctx);
+ mir_squeeze_index(ctx);
/* Lowering can introduce some dead moves */