-/* Midgard IR only knows vector ALU types, but we sometimes need to actually
- * use scalar ALU instructions, for functional or performance reasons. To do
- * this, we just demote vector ALU payloads to scalar. */
-
-static int
-component_from_mask(unsigned mask)
-{
- for (int c = 0; c < 4; ++c) {
- if (mask & (3 << (2 * c)))
- return c;
- }
-
- assert(0);
- return 0;
-}
-
-static bool
-is_single_component_mask(unsigned mask)
-{
- int components = 0;
-
- for (int c = 0; c < 4; ++c)
- if (mask & (3 << (2 * c)))
- components++;
-
- return components == 1;
-}
-
-/* Create a mask of accessed components from a swizzle to figure out vector
- * dependencies */
-
-static unsigned
-swizzle_to_access_mask(unsigned swizzle)
-{
- unsigned component_mask = 0;
-
- for (int i = 0; i < 4; ++i) {
- unsigned c = (swizzle >> (2 * i)) & 3;
- component_mask |= (1 << c);
- }
-
- return component_mask;
-}
-
-static unsigned
-vector_to_scalar_source(unsigned u, bool is_int)
-{
- midgard_vector_alu_src v;
- memcpy(&v, &u, sizeof(v));
-
- /* TODO: Integers */
-
- midgard_scalar_alu_src s = {
- .full = !v.half,
- .component = (v.swizzle & 3) << 1
- };
-
- if (is_int) {
- /* TODO */
- } else {
- s.abs = v.mod & MIDGARD_FLOAT_MOD_ABS;
- s.negate = v.mod & MIDGARD_FLOAT_MOD_NEG;
- }
-
- unsigned o;
- memcpy(&o, &s, sizeof(s));
-
- return o & ((1 << 6) - 1);
-}
-
-static midgard_scalar_alu
-vector_to_scalar_alu(midgard_vector_alu v, midgard_instruction *ins)
-{
- bool is_int = midgard_is_integer_op(v.op);
-
- /* The output component is from the mask */
- midgard_scalar_alu s = {
- .op = v.op,
- .src1 = vector_to_scalar_source(v.src1, is_int),
- .src2 = vector_to_scalar_source(v.src2, is_int),
- .unknown = 0,
- .outmod = v.outmod,
- .output_full = 1, /* TODO: Half */
- .output_component = component_from_mask(v.mask) << 1,
- };
-
- /* Inline constant is passed along rather than trying to extract it
- * from v */
-
- if (ins->ssa_args.inline_constant) {
- uint16_t imm = 0;
- int lower_11 = ins->inline_constant & ((1 << 12) - 1);
- imm |= (lower_11 >> 9) & 3;
- imm |= (lower_11 >> 6) & 4;
- imm |= (lower_11 >> 2) & 0x38;
- imm |= (lower_11 & 63) << 6;
-
- s.src2 = imm;
- }
-
- return s;
-}
-
-/* Midgard prefetches instruction types, so during emission we need to
- * lookahead too. Unless this is the last instruction, in which we return 1. Or
- * if this is the second to last and the last is an ALU, then it's also 1... */
-
-#define IS_ALU(tag) (tag == TAG_ALU_4 || tag == TAG_ALU_8 || \
- tag == TAG_ALU_12 || tag == TAG_ALU_16)
-
-#define EMIT_AND_COUNT(type, val) util_dynarray_append(emission, type, val); \
- bytes_emitted += sizeof(type)
-
-static void
-emit_binary_vector_instruction(midgard_instruction *ains,
- uint16_t *register_words, int *register_words_count,
- uint64_t *body_words, size_t *body_size, int *body_words_count,
- size_t *bytes_emitted)
-{
- memcpy(®ister_words[(*register_words_count)++], &ains->registers, sizeof(ains->registers));
- *bytes_emitted += sizeof(midgard_reg_info);
-
- body_size[*body_words_count] = sizeof(midgard_vector_alu);
- memcpy(&body_words[(*body_words_count)++], &ains->alu, sizeof(ains->alu));
- *bytes_emitted += sizeof(midgard_vector_alu);
-}
-
-/* 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)
-{
- /* Each instruction reads some registers and writes to a register. See
- * where the first writes */
-
- /* Figure out where exactly we wrote to */
- int source = first->ssa_args.dest;
- int source_mask = first->type == TAG_ALU_4 ? squeeze_writemask(first->alu.mask) : 0xF;
-
- /* As long as the second doesn't read from the first, we're okay */
- if (second->ssa_args.src0 == source) {
- if (first->type == TAG_ALU_4) {
- /* Figure out which components we just read from */
-
- int q = second->alu.src1;
- 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;
- } else
- return false;
-
- }
-
- if (second->ssa_args.src1 == source)
- return false;
-
- /* Otherwise, it's safe in that regard. Another data hazard is both
- * writing to the same place, of course */
-
- if (second->ssa_args.dest == source) {
- /* ...but only if the components overlap */
- int dest_mask = second->type == TAG_ALU_4 ? squeeze_writemask(second->alu.mask) : 0xF;
-
- if (dest_mask & source_mask)
- return false;
- }
-
- /* ...That's it */
- return true;
-}
-
-static bool
-midgard_has_hazard(
- midgard_instruction **segment, unsigned segment_size,
- midgard_instruction *ains)
-{
- for (int s = 0; s < segment_size; ++s)
- if (!can_run_concurrent_ssa(segment[s], ains))
- return true;
-
- return false;
-
-
-}
-
-/* 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
- * */
-
-static midgard_bundle
-schedule_bundle(compiler_context *ctx, midgard_block *block, midgard_instruction *ins, int *skip)
-{
- int instructions_emitted = 0, instructions_consumed = -1;
- midgard_bundle bundle = { 0 };
-
- uint8_t tag = ins->type;
-
- /* Default to the instruction's tag */
- bundle.tag = tag;
-
- switch (ins->type) {
- case TAG_ALU_4: {
- uint32_t control = 0;
- size_t bytes_emitted = sizeof(control);
-
- /* TODO: Constant combining */
- int index = 0, last_unit = 0;
-
- /* Previous instructions, for the purpose of parallelism */
- midgard_instruction *segment[4] = {0};
- int segment_size = 0;
-
- instructions_emitted = -1;
- midgard_instruction *pins = ins;
-
- for (;;) {
- midgard_instruction *ains = pins;
-
- /* Advance instruction pointer */
- if (index) {
- ains = mir_next_op(pins);
- pins = ains;
- }
-
- /* 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;
-
- /* TODO: Promotion of scalars to vectors */
- int vector = ((!is_single_component_mask(ains->alu.mask)) || ((units & UNITS_SCALAR) == 0)) && (units & UNITS_ANY_VECTOR);
-
- if (!vector)
- assert(units & UNITS_SCALAR);
-
- if (vector) {
- if (last_unit >= UNIT_VADD) {
- if (units & UNIT_VLUT)
- unit = UNIT_VLUT;
- else
- break;
- } else {
- if ((units & UNIT_VMUL) && !(control & 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_SADD) && !(control & UNIT_SADD) && !midgard_has_hazard(segment, segment_size, ains))
- unit = UNIT_SADD;
- else if (units & UNIT_SMUL)
- unit = ((units & UNIT_VMUL) && !(control & UNIT_VMUL)) ? UNIT_VMUL : UNIT_SMUL;
- else if ((units & UNIT_VADD) && !(control & UNIT_VADD))
- unit = UNIT_VADD;
- else
- break;
- }
- }
-
- assert(unit & units);
- }
-
- /* Late unit check, this time for encoding (not parallelism) */
- if (unit <= last_unit) break;
-
- /* Clear the segment */
- if (last_unit < UNIT_VADD && unit >= UNIT_VADD)
- segment_size = 0;
-
- if (midgard_has_hazard(segment, segment_size, ains))
- break;
-
- /* We're good to go -- emit the instruction */
- ains->unit = unit;
-
- segment[segment_size++] = ains;
-
- /* Only one set of embedded constants per
- * bundle possible; if we have more, we must
- * break the chain early, unfortunately */
-
- if (ains->has_constants) {
- if (bundle.has_embedded_constants) {
- /* The blend constant needs to be
- * alone, since it conflicts with
- * everything by definition*/
-
- if (ains->has_blend_constant || bundle.has_blend_constant)
- break;
-
- /* ...but if there are already
- * constants but these are the
- * *same* constants, we let it
- * through */
-
- if (memcmp(bundle.constants, ains->constants, sizeof(bundle.constants)))
- break;
- } else {
- bundle.has_embedded_constants = true;
- memcpy(bundle.constants, ains->constants, sizeof(bundle.constants));
-
- /* If this is a blend shader special constant, track it for patching */
- bundle.has_blend_constant |= ains->has_blend_constant;
- }
- }
-
- if (ains->unit & UNITS_ANY_VECTOR) {
- emit_binary_vector_instruction(ains, bundle.register_words,
- &bundle.register_words_count, bundle.body_words,
- bundle.body_size, &bundle.body_words_count, &bytes_emitted);
- } else if (ains->compact_branch) {
- /* All of r0 has to be written out
- * along with the branch writeout.
- * (slow!) */
-
- if (ains->writeout) {
- if (index == 0) {
- midgard_instruction ins = v_fmov(0, blank_alu_src, SSA_FIXED_REGISTER(0));
- ins.unit = UNIT_VMUL;
-
- control |= ins.unit;
-
- emit_binary_vector_instruction(&ins, bundle.register_words,
- &bundle.register_words_count, bundle.body_words,
- bundle.body_size, &bundle.body_words_count, &bytes_emitted);
- } else {
- /* Analyse the group to see if r0 is written in full, on-time, without hanging dependencies*/
- bool written_late = false;
- bool components[4] = { 0 };
- uint16_t register_dep_mask = 0;
- uint16_t written_mask = 0;
-
- midgard_instruction *qins = ins;
- for (int t = 0; t < index; ++t) {
- if (qins->registers.out_reg != 0) {
- /* Mark down writes */
-
- written_mask |= (1 << qins->registers.out_reg);
- } else {
- /* Mark down the register dependencies for errata check */
-
- if (qins->registers.src1_reg < 16)
- register_dep_mask |= (1 << qins->registers.src1_reg);
-
- if (qins->registers.src2_reg < 16)
- register_dep_mask |= (1 << qins->registers.src2_reg);
-
- int mask = qins->alu.mask;
-
- for (int c = 0; c < 4; ++c)
- if (mask & (0x3 << (2 * c)))
- components[c] = true;
-
- /* ..but if the writeout is too late, we have to break up anyway... for some reason */
-
- if (qins->unit == UNIT_VLUT)
- written_late = true;
- }
-
- /* Advance instruction pointer */
- qins = mir_next_op(qins);
- }
-
-
- /* ERRATA (?): In a bundle ending in a fragment writeout, the register dependencies of r0 cannot be written within this bundle (discovered in -bshading:shading=phong) */
- if (register_dep_mask & written_mask) {
- DBG("ERRATA WORKAROUND: Breakup for writeout dependency masks %X vs %X (common %X)\n", register_dep_mask, written_mask, register_dep_mask & written_mask);
- break;
- }
-
- if (written_late)
- break;
-
- /* If even a single component is not written, break it up (conservative check). */
- bool breakup = false;
-
- for (int c = 0; c < 4; ++c)
- if (!components[c])
- breakup = true;
-
- if (breakup)
- break;
-
- /* Otherwise, we're free to proceed */
- }
- }
-
- if (ains->unit == ALU_ENAB_BRANCH) {
- bundle.body_size[bundle.body_words_count] = sizeof(midgard_branch_extended);
- memcpy(&bundle.body_words[bundle.body_words_count++], &ains->branch_extended, sizeof(midgard_branch_extended));
- bytes_emitted += sizeof(midgard_branch_extended);
- } else {
- bundle.body_size[bundle.body_words_count] = sizeof(ains->br_compact);
- memcpy(&bundle.body_words[bundle.body_words_count++], &ains->br_compact, sizeof(ains->br_compact));
- bytes_emitted += sizeof(ains->br_compact);
- }
- } else {
- memcpy(&bundle.register_words[bundle.register_words_count++], &ains->registers, sizeof(ains->registers));
- bytes_emitted += sizeof(midgard_reg_info);
-
- bundle.body_size[bundle.body_words_count] = sizeof(midgard_scalar_alu);
- bundle.body_words_count++;
- bytes_emitted += sizeof(midgard_scalar_alu);
- }
-
- /* Defer marking until after writing to allow for break */
- control |= ains->unit;
- last_unit = ains->unit;
- ++instructions_emitted;
- ++index;
- }
-
- /* Bubble up the number of instructions for skipping */
- instructions_consumed = index - 1;
-
- int padding = 0;
-
- /* 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 | control;
-
- break;
- }
-
- 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.
- */
-
- midgard_instruction *next_op = mir_next_op(ins);
-
- if ((struct list_head *) next_op != &block->instructions && next_op->type == TAG_LOAD_STORE_4) {
- /* As the two operate concurrently, make sure
- * they are not dependent */
-
- if (can_run_concurrent_ssa(ins, next_op) || true) {
- /* Skip ahead, since it's redundant with the pair */
- instructions_consumed = 1 + (instructions_emitted++);
- }
- }
-
- break;
- }
-
- default:
- /* Texture ops default to single-op-per-bundle scheduling */
- break;
- }
-
- /* Copy the instructions into the bundle */
- bundle.instruction_count = instructions_emitted + 1;
-
- int used_idx = 0;
-
- midgard_instruction *uins = ins;
- for (int i = 0; used_idx < bundle.instruction_count; ++i) {
- bundle.instructions[used_idx++] = *uins;
- uins = mir_next_op(uins);
- }
-
- *skip = (instructions_consumed == -1) ? instructions_emitted : instructions_consumed;
-
- return bundle;
-}
-
-static int
-quadword_size(int tag)
-{
- switch (tag) {
- case TAG_ALU_4:
- return 1;
-
- case TAG_ALU_8:
- return 2;
-
- case TAG_ALU_12:
- return 3;
-
- case TAG_ALU_16:
- return 4;
-
- case TAG_LOAD_STORE_4:
- return 1;
-
- case TAG_TEXTURE_4:
- return 1;
-
- default:
- assert(0);
- return 0;
- }
-}
-
-/* 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);
-
- block->quadword_count = 0;
-
- 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);
-
- if (bundle.has_blend_constant) {
- /* TODO: Multiblock? */
- int quadwords_within_block = block->quadword_count + quadword_size(bundle.tag) - 1;
- ctx->blend_constant_offset = quadwords_within_block * 0x10;
- }
-
- while(skip--)
- ins = mir_next_op(ins);
-
- block->quadword_count += quadword_size(bundle.tag);
- }
-
- block->is_scheduled = true;
-}
-
-static void
-schedule_program(compiler_context *ctx)
-{
- /* We run RA prior to scheduling */
- struct ra_graph *g = allocate_registers(ctx);
- install_registers(ctx, g);
-
- mir_foreach_block(ctx, block) {
- schedule_block(ctx, block);
- }
-}
-
-/* After everything is scheduled, emit whole bundles at a time */
-
-static void
-emit_binary_bundle(compiler_context *ctx, midgard_bundle *bundle, struct util_dynarray *emission, int next_tag)
-{
- int lookahead = next_tag << 4;
-
- switch (bundle->tag) {
- case TAG_ALU_4:
- case TAG_ALU_8:
- case TAG_ALU_12:
- case TAG_ALU_16: {
- /* Actually emit each component */
- util_dynarray_append(emission, uint32_t, bundle->control | lookahead);
-
- for (int i = 0; i < bundle->register_words_count; ++i)
- util_dynarray_append(emission, uint16_t, bundle->register_words[i]);
-
- /* Emit body words based on the instructions bundled */
- for (int i = 0; i < bundle->instruction_count; ++i) {
- midgard_instruction *ins = &bundle->instructions[i];
-
- if (ins->unit & UNITS_ANY_VECTOR) {
- memcpy(util_dynarray_grow(emission, sizeof(midgard_vector_alu)), &ins->alu, sizeof(midgard_vector_alu));
- } else if (ins->compact_branch) {
- /* Dummy move, XXX DRY */
- if ((i == 0) && ins->writeout) {
- midgard_instruction ins = v_fmov(0, blank_alu_src, SSA_FIXED_REGISTER(0));
- memcpy(util_dynarray_grow(emission, sizeof(midgard_vector_alu)), &ins.alu, sizeof(midgard_vector_alu));
- }
-
- if (ins->unit == ALU_ENAB_BR_COMPACT) {
- memcpy(util_dynarray_grow(emission, sizeof(ins->br_compact)), &ins->br_compact, sizeof(ins->br_compact));
- } else {
- memcpy(util_dynarray_grow(emission, sizeof(ins->branch_extended)), &ins->branch_extended, sizeof(ins->branch_extended));
- }
- } else {
- /* Scalar */
- midgard_scalar_alu scalarised = vector_to_scalar_alu(ins->alu, ins);
- memcpy(util_dynarray_grow(emission, sizeof(scalarised)), &scalarised, sizeof(scalarised));
- }
- }
-
- /* Emit padding (all zero) */
- memset(util_dynarray_grow(emission, bundle->padding), 0, bundle->padding);
-
- /* Tack on constants */
-
- if (bundle->has_embedded_constants) {
- util_dynarray_append(emission, float, bundle->constants[0]);
- util_dynarray_append(emission, float, bundle->constants[1]);
- util_dynarray_append(emission, float, bundle->constants[2]);
- util_dynarray_append(emission, float, bundle->constants[3]);
- }
-
- break;
- }
-
- case TAG_LOAD_STORE_4: {
- /* One or two composing instructions */
-
- uint64_t current64, next64 = LDST_NOP;
-
- memcpy(¤t64, &bundle->instructions[0].load_store, sizeof(current64));
-
- if (bundle->instruction_count == 2)
- memcpy(&next64, &bundle->instructions[1].load_store, sizeof(next64));
-
- midgard_load_store instruction = {
- .type = bundle->tag,
- .next_type = next_tag,
- .word1 = current64,
- .word2 = next64
- };
-
- util_dynarray_append(emission, midgard_load_store, instruction);
-
- break;
- }
-
- case TAG_TEXTURE_4: {
- /* Texture instructions are easy, since there is no
- * pipelining nor VLIW to worry about. We may need to set the .last flag */
-
- midgard_instruction *ins = &bundle->instructions[0];
-
- ins->texture.type = TAG_TEXTURE_4;
- ins->texture.next_type = next_tag;
-
- ctx->texture_op_count--;
-
- if (!ctx->texture_op_count) {
- ins->texture.cont = 0;
- ins->texture.last = 1;
- }
-
- util_dynarray_append(emission, midgard_texture_word, ins->texture);
- break;
- }
-
- default:
- DBG("Unknown midgard instruction type\n");
- assert(0);
- break;
- }
-}
-