namespace aco {
-#ifndef NDEBUG
-void perfwarn(bool cond, const char *msg, Instruction *instr)
+static void aco_log(Program *program, enum radv_compiler_debug_level level,
+ const char *prefix, const char *file, unsigned line,
+ const char *fmt, va_list args)
{
- if (cond) {
- fprintf(stderr, "ACO performance warning: %s\n", msg);
- if (instr) {
- fprintf(stderr, "instruction: ");
- aco_print_instr(instr, stderr);
- fprintf(stderr, "\n");
- }
+ char *msg;
- if (debug_flags & DEBUG_PERFWARN)
- exit(1);
- }
+ msg = ralloc_strdup(NULL, prefix);
+
+ ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
+ ralloc_asprintf_append(&msg, " ");
+ ralloc_vasprintf_append(&msg, fmt, args);
+
+ if (program->debug.func)
+ program->debug.func(program->debug.private_data, level, msg);
+
+ fprintf(stderr, "%s\n", msg);
+
+ ralloc_free(msg);
}
-#endif
-bool instr_can_access_subdword(aco_ptr<Instruction>& instr)
+void _aco_perfwarn(Program *program, const char *file, unsigned line,
+ const char *fmt, ...)
{
- return instr->isSDWA() || instr->format == Format::PSEUDO;
+ va_list args;
+
+ va_start(args, fmt);
+ aco_log(program, RADV_COMPILER_DEBUG_LEVEL_PERFWARN,
+ "ACO PERFWARN:\n", file, line, fmt, args);
+ va_end(args);
}
-void validate(Program* program, FILE * output)
+void _aco_err(Program *program, const char *file, unsigned line,
+ const char *fmt, ...)
{
- if (!(debug_flags & DEBUG_VALIDATE))
- return;
+ va_list args;
+ va_start(args, fmt);
+ aco_log(program, RADV_COMPILER_DEBUG_LEVEL_ERROR,
+ "ACO ERROR:\n", file, line, fmt, args);
+ va_end(args);
+}
+
+bool validate_ir(Program* program)
+{
bool is_valid = true;
- auto check = [&output, &is_valid](bool check, const char * msg, aco::Instruction * instr) -> void {
+ auto check = [&program, &is_valid](bool check, const char * msg, aco::Instruction * instr) -> void {
if (!check) {
- fprintf(output, "%s: ", msg);
- aco_print_instr(instr, output);
- fprintf(output, "\n");
+ char *out;
+ size_t outsize;
+ FILE *memf = open_memstream(&out, &outsize);
+
+ fprintf(memf, "%s: ", msg);
+ aco_print_instr(instr, memf);
+ fclose(memf);
+
+ aco_err(program, out);
+ free(out);
+
is_valid = false;
}
};
- auto check_block = [&output, &is_valid](bool check, const char * msg, aco::Block * block) -> void {
+
+ auto check_block = [&program, &is_valid](bool check, const char * msg, aco::Block * block) -> void {
if (!check) {
- fprintf(output, "%s: BB%u\n", msg, block->index);
+ aco_err(program, "%s: BB%u", msg, block->index);
is_valid = false;
}
};
instr->opcode != aco_opcode::v_fmac_f16,
"SDWA can't be used with this opcode", instr.get());
}
+
+ for (unsigned i = 0; i < MIN2(instr->operands.size(), 2); i++) {
+ if (instr->operands[i].hasRegClass() && instr->operands[i].regClass().is_subdword())
+ check((sdwa->sel[i] & sdwa_asuint) == (sdwa_isra | instr->operands[i].bytes()), "Unexpected SDWA sel for sub-dword operand", instr.get());
+ }
+ if (instr->definitions[0].regClass().is_subdword())
+ check((sdwa->dst_sel & sdwa_asuint) == (sdwa_isra | instr->definitions[0].bytes()), "Unexpected SDWA sel for sub-dword definition", instr.get());
}
/* check opsel */
if (instr->isVOP3()) {
VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(instr.get());
check(vop3->opsel == 0 || program->chip_class >= GFX9, "Opsel is only supported on GFX9+", instr.get());
- check((vop3->opsel & ~(0x10 | ((1 << instr->operands.size()) - 1))) == 0, "Unused bits in opsel must be zeroed out", instr.get());
+
+ for (unsigned i = 0; i < 3; i++) {
+ if (i >= instr->operands.size() ||
+ (instr->operands[i].hasRegClass() && instr->operands[i].regClass().is_subdword() && !instr->operands[i].isFixed()))
+ check((vop3->opsel & (1 << i)) == 0, "Unexpected opsel for operand", instr.get());
+ }
+ if (instr->definitions[0].regClass().is_subdword() && !instr->definitions[0].isFixed())
+ check((vop3->opsel & (1 << 3)) == 0, "Unexpected opsel for sub-dword definition", instr.get());
}
/* check for undefs */
/* check subdword definitions */
for (unsigned i = 0; i < instr->definitions.size(); i++) {
if (instr->definitions[i].regClass().is_subdword())
- check(instr_can_access_subdword(instr) || instr->definitions[i].bytes() <= 4, "Only SDWA and Pseudo instructions can write subdword registers larger than 4 bytes", instr.get());
+ check(instr->format == Format::PSEUDO || instr->definitions[i].bytes() <= 4, "Only Pseudo instructions can write subdword registers larger than 4 bytes", instr.get());
}
if (instr->isSALU() || instr->isVALU()) {
if (instr->isSDWA())
scalar_mask = program->chip_class >= GFX9 ? 0x7 : 0x4;
- check(instr->definitions[0].getTemp().type() == RegType::vgpr ||
- (int) instr->format & (int) Format::VOPC ||
- instr->opcode == aco_opcode::v_readfirstlane_b32 ||
- instr->opcode == aco_opcode::v_readlane_b32 ||
- instr->opcode == aco_opcode::v_readlane_b32_e64,
- "Wrong Definition type for VALU instruction", instr.get());
+ if ((int) instr->format & (int) Format::VOPC ||
+ instr->opcode == aco_opcode::v_readfirstlane_b32 ||
+ instr->opcode == aco_opcode::v_readlane_b32 ||
+ instr->opcode == aco_opcode::v_readlane_b32_e64) {
+ check(instr->definitions[0].getTemp().type() == RegType::sgpr,
+ "Wrong Definition type for VALU instruction", instr.get());
+ } else {
+ check(instr->definitions[0].getTemp().type() == RegType::vgpr,
+ "Wrong Definition type for VALU instruction", instr.get());
+ }
+
unsigned num_sgprs = 0;
unsigned sgpr[] = {0, 0};
for (unsigned i = 0; i < instr->operands.size(); i++)
Operand op = instr->operands[i];
if (instr->opcode == aco_opcode::v_readfirstlane_b32 ||
instr->opcode == aco_opcode::v_readlane_b32 ||
- instr->opcode == aco_opcode::v_readlane_b32_e64 ||
- instr->opcode == aco_opcode::v_writelane_b32 ||
+ instr->opcode == aco_opcode::v_readlane_b32_e64) {
+ check(i != 1 ||
+ (op.isTemp() && op.regClass().type() == RegType::sgpr) ||
+ op.isConstant(),
+ "Must be a SGPR or a constant", instr.get());
+ check(i == 1 ||
+ (op.isTemp() && op.regClass().type() == RegType::vgpr && op.bytes() <= 4),
+ "Wrong Operand type for VALU instruction", instr.get());
+ continue;
+ }
+
+ if (instr->opcode == aco_opcode::v_writelane_b32 ||
instr->opcode == aco_opcode::v_writelane_b32_e64) {
- check(!op.isLiteral(), "No literal allowed on VALU instruction", instr.get());
- check(i == 1 || (op.isTemp() && op.regClass() == v1), "Wrong Operand type for VALU instruction", instr.get());
+ check(i != 2 ||
+ (op.isTemp() && op.regClass().type() == RegType::vgpr && op.bytes() <= 4),
+ "Wrong Operand type for VALU instruction", instr.get());
+ check(i == 2 ||
+ (op.isTemp() && op.regClass().type() == RegType::sgpr) ||
+ op.isConstant(),
+ "Must be a SGPR or a constant", instr.get());
continue;
}
if (op.isTemp() && instr->operands[i].regClass().type() == RegType::sgpr) {
switch (instr->format) {
case Format::PSEUDO: {
+ bool is_subdword = false;
+ bool has_const_sgpr = false;
+ bool has_literal = false;
+ for (Definition def : instr->definitions)
+ is_subdword |= def.regClass().is_subdword();
+ for (unsigned i = 0; i < instr->operands.size(); i++) {
+ if (instr->opcode == aco_opcode::p_extract_vector && i == 1)
+ continue;
+ Operand op = instr->operands[i];
+ is_subdword |= op.hasRegClass() && op.regClass().is_subdword();
+ has_const_sgpr |= op.isConstant() || (op.hasRegClass() && op.regClass().type() == RegType::sgpr);
+ has_literal |= op.isLiteral();
+ }
+
+ check(!is_subdword || !has_const_sgpr || program->chip_class >= GFX9,
+ "Sub-dword pseudo instructions can only take constants or SGPRs on GFX9+", instr.get());
+ check(!is_subdword || !has_literal, "Sub-dword pseudo instructions cannot take literals", instr.get());
+
if (instr->opcode == aco_opcode::p_create_vector) {
unsigned size = 0;
for (const Operand& op : instr->operands) {
}
} else if (instr->opcode == aco_opcode::p_phi) {
check(instr->operands.size() == block.logical_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
- check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->definitions[0].getTemp().regClass() == program->lane_mask, "Logical Phi Definition must be vgpr or divergent boolean", instr.get());
+ check(instr->definitions[0].getTemp().type() == RegType::vgpr, "Logical Phi Definition must be vgpr", instr.get());
} else if (instr->opcode == aco_opcode::p_linear_phi) {
for (const Operand& op : instr->operands)
check(!op.isTemp() || op.getTemp().is_linear(), "Wrong Operand type", instr.get());
}
case Format::SMEM: {
if (instr->operands.size() >= 1)
- check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::sgpr, "SMEM operands must be sgpr", instr.get());
+ check((instr->operands[0].isFixed() && !instr->operands[0].isConstant()) ||
+ (instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::sgpr), "SMEM operands must be sgpr", instr.get());
if (instr->operands.size() >= 2)
check(instr->operands[1].isConstant() || (instr->operands[1].isTemp() && instr->operands[1].regClass().type() == RegType::sgpr),
"SMEM offset must be constant or sgpr", instr.get());
}
}
- assert(is_valid);
+ return is_valid;
}
/* RA validation */
PhysReg reg;
};
-bool ra_fail(FILE *output, Location loc, Location loc2, const char *fmt, ...) {
+bool ra_fail(Program *program, Location loc, Location loc2, const char *fmt, ...) {
va_list args;
va_start(args, fmt);
char msg[1024];
vsprintf(msg, fmt, args);
va_end(args);
- fprintf(stderr, "RA error found at instruction in BB%d:\n", loc.block->index);
+ char *out;
+ size_t outsize;
+ FILE *memf = open_memstream(&out, &outsize);
+
+ fprintf(memf, "RA error found at instruction in BB%d:\n", loc.block->index);
if (loc.instr) {
- aco_print_instr(loc.instr, stderr);
- fprintf(stderr, "\n%s", msg);
+ aco_print_instr(loc.instr, memf);
+ fprintf(memf, "\n%s", msg);
} else {
- fprintf(stderr, "%s", msg);
+ fprintf(memf, "%s", msg);
}
if (loc2.block) {
- fprintf(stderr, " in BB%d:\n", loc2.block->index);
- aco_print_instr(loc2.instr, stderr);
+ fprintf(memf, " in BB%d:\n", loc2.block->index);
+ aco_print_instr(loc2.instr, memf);
}
- fprintf(stderr, "\n\n");
+ fprintf(memf, "\n\n");
+ fclose(memf);
+
+ aco_err(program, out);
+ free(out);
return true;
}
+bool validate_subdword_operand(chip_class chip, const aco_ptr<Instruction>& instr, unsigned index)
+{
+ Operand op = instr->operands[index];
+ unsigned byte = op.physReg().byte();
+
+ if (instr->opcode == aco_opcode::p_as_uniform)
+ return byte == 0;
+ if (instr->format == Format::PSEUDO && chip >= GFX8)
+ return true;
+ if (instr->isSDWA() && (static_cast<SDWA_instruction *>(instr.get())->sel[index] & sdwa_asuint) == (sdwa_isra | op.bytes()))
+ return true;
+ if (byte == 2 && can_use_opsel(chip, instr->opcode, index, 1))
+ return true;
+
+ switch (instr->opcode) {
+ case aco_opcode::v_cvt_f32_ubyte1:
+ if (byte == 1)
+ return true;
+ break;
+ case aco_opcode::v_cvt_f32_ubyte2:
+ if (byte == 2)
+ return true;
+ break;
+ case aco_opcode::v_cvt_f32_ubyte3:
+ if (byte == 3)
+ return true;
+ break;
+ case aco_opcode::ds_write_b8_d16_hi:
+ case aco_opcode::ds_write_b16_d16_hi:
+ if (byte == 2 && index == 1)
+ return true;
+ break;
+ case aco_opcode::buffer_store_byte_d16_hi:
+ case aco_opcode::buffer_store_short_d16_hi:
+ if (byte == 2 && index == 3)
+ return true;
+ break;
+ case aco_opcode::flat_store_byte_d16_hi:
+ case aco_opcode::flat_store_short_d16_hi:
+ case aco_opcode::scratch_store_byte_d16_hi:
+ case aco_opcode::scratch_store_short_d16_hi:
+ case aco_opcode::global_store_byte_d16_hi:
+ case aco_opcode::global_store_short_d16_hi:
+ if (byte == 2 && index == 2)
+ return true;
+ default:
+ break;
+ }
+
+ return byte == 0;
+}
+
+bool validate_subdword_definition(chip_class chip, const aco_ptr<Instruction>& instr)
+{
+ Definition def = instr->definitions[0];
+ unsigned byte = def.physReg().byte();
+
+ if (instr->format == Format::PSEUDO && chip >= GFX8)
+ return true;
+ if (instr->isSDWA() && static_cast<SDWA_instruction *>(instr.get())->dst_sel == (sdwa_isra | def.bytes()))
+ return true;
+ if (byte == 2 && can_use_opsel(chip, instr->opcode, -1, 1))
+ return true;
+
+ switch (instr->opcode) {
+ case aco_opcode::buffer_load_ubyte_d16_hi:
+ case aco_opcode::buffer_load_short_d16_hi:
+ case aco_opcode::flat_load_ubyte_d16_hi:
+ case aco_opcode::flat_load_short_d16_hi:
+ case aco_opcode::scratch_load_ubyte_d16_hi:
+ case aco_opcode::scratch_load_short_d16_hi:
+ case aco_opcode::global_load_ubyte_d16_hi:
+ case aco_opcode::global_load_short_d16_hi:
+ case aco_opcode::ds_read_u8_d16_hi:
+ case aco_opcode::ds_read_u16_d16_hi:
+ return byte == 2;
+ default:
+ break;
+ }
+
+ return byte == 0;
+}
+
+unsigned get_subdword_bytes_written(Program *program, const aco_ptr<Instruction>& instr, unsigned index)
+{
+ chip_class chip = program->chip_class;
+ Definition def = instr->definitions[index];
+
+ if (instr->format == Format::PSEUDO)
+ return chip >= GFX8 ? def.bytes() : def.size() * 4u;
+ if (instr->isSDWA() && static_cast<SDWA_instruction *>(instr.get())->dst_sel == (sdwa_isra | def.bytes()))
+ return def.bytes();
+
+ switch (instr->opcode) {
+ case aco_opcode::buffer_load_ubyte_d16:
+ case aco_opcode::buffer_load_short_d16:
+ case aco_opcode::flat_load_ubyte_d16:
+ case aco_opcode::flat_load_short_d16:
+ case aco_opcode::scratch_load_ubyte_d16:
+ case aco_opcode::scratch_load_short_d16:
+ case aco_opcode::global_load_ubyte_d16:
+ case aco_opcode::global_load_short_d16:
+ case aco_opcode::ds_read_u8_d16:
+ case aco_opcode::ds_read_u16_d16:
+ case aco_opcode::buffer_load_ubyte_d16_hi:
+ case aco_opcode::buffer_load_short_d16_hi:
+ case aco_opcode::flat_load_ubyte_d16_hi:
+ case aco_opcode::flat_load_short_d16_hi:
+ case aco_opcode::scratch_load_ubyte_d16_hi:
+ case aco_opcode::scratch_load_short_d16_hi:
+ case aco_opcode::global_load_ubyte_d16_hi:
+ case aco_opcode::global_load_short_d16_hi:
+ case aco_opcode::ds_read_u8_d16_hi:
+ case aco_opcode::ds_read_u16_d16_hi:
+ return program->sram_ecc_enabled ? 4 : 2;
+ case aco_opcode::v_mad_f16:
+ case aco_opcode::v_mad_u16:
+ case aco_opcode::v_mad_i16:
+ case aco_opcode::v_fma_f16:
+ case aco_opcode::v_div_fixup_f16:
+ case aco_opcode::v_interp_p2_f16:
+ if (chip >= GFX9)
+ return 2;
+ default:
+ break;
+ }
+
+ return MAX2(chip >= GFX10 ? def.bytes() : 4, instr_info.definition_size[(int)instr->opcode] / 8u);
+}
+
} /* end namespace */
-bool validate_ra(Program *program, const struct radv_nir_compiler_options *options, FILE *output) {
+bool validate_ra(Program *program, const struct radv_nir_compiler_options *options) {
if (!(debug_flags & DEBUG_VALIDATE_RA))
return false;
if (!op.isTemp())
continue;
if (!op.isFixed())
- err |= ra_fail(output, loc, Location(), "Operand %d is not assigned a register", i);
+ err |= ra_fail(program, loc, Location(), "Operand %d is not assigned a register", i);
if (assignments.count(op.tempId()) && assignments[op.tempId()].reg != op.physReg())
- err |= ra_fail(output, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an inconsistent register assignment with instruction", i);
+ err |= ra_fail(program, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an inconsistent register assignment with instruction", i);
if ((op.getTemp().type() == RegType::vgpr && op.physReg().reg_b + op.bytes() > (256 + program->config->num_vgprs) * 4) ||
(op.getTemp().type() == RegType::sgpr && op.physReg() + op.size() > program->config->num_sgprs && op.physReg() < program->sgpr_limit))
- err |= ra_fail(output, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an out-of-bounds register assignment", i);
+ err |= ra_fail(program, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an out-of-bounds register assignment", i);
if (op.physReg() == vcc && !program->needs_vcc)
- err |= ra_fail(output, loc, Location(), "Operand %d fixed to vcc but needs_vcc=false", i);
- if (!instr_can_access_subdword(instr) && op.regClass().is_subdword() && op.physReg().byte())
- err |= ra_fail(output, loc, assignments.at(op.tempId()).firstloc, "Operand %d must be aligned to a full register", i);
+ err |= ra_fail(program, loc, Location(), "Operand %d fixed to vcc but needs_vcc=false", i);
+ if (op.regClass().is_subdword() && !validate_subdword_operand(program->chip_class, instr, i))
+ err |= ra_fail(program, loc, Location(), "Operand %d not aligned correctly", i);
if (!assignments[op.tempId()].firstloc.block)
assignments[op.tempId()].firstloc = loc;
if (!assignments[op.tempId()].defloc.block)
if (!def.isTemp())
continue;
if (!def.isFixed())
- err |= ra_fail(output, loc, Location(), "Definition %d is not assigned a register", i);
+ err |= ra_fail(program, loc, Location(), "Definition %d is not assigned a register", i);
if (assignments[def.tempId()].defloc.block)
- err |= ra_fail(output, loc, assignments.at(def.tempId()).defloc, "Temporary %%%d also defined by instruction", def.tempId());
+ err |= ra_fail(program, loc, assignments.at(def.tempId()).defloc, "Temporary %%%d also defined by instruction", def.tempId());
if ((def.getTemp().type() == RegType::vgpr && def.physReg().reg_b + def.bytes() > (256 + program->config->num_vgprs) * 4) ||
(def.getTemp().type() == RegType::sgpr && def.physReg() + def.size() > program->config->num_sgprs && def.physReg() < program->sgpr_limit))
- err |= ra_fail(output, loc, assignments.at(def.tempId()).firstloc, "Definition %d has an out-of-bounds register assignment", i);
+ err |= ra_fail(program, loc, assignments.at(def.tempId()).firstloc, "Definition %d has an out-of-bounds register assignment", i);
if (def.physReg() == vcc && !program->needs_vcc)
- err |= ra_fail(output, loc, Location(), "Definition %d fixed to vcc but needs_vcc=false", i);
- if (!instr_can_access_subdword(instr) && def.regClass().is_subdword() && def.physReg().byte())
- err |= ra_fail(output, loc, assignments.at(def.tempId()).firstloc, "Definition %d must be aligned to a full register", i);
+ err |= ra_fail(program, loc, Location(), "Definition %d fixed to vcc but needs_vcc=false", i);
+ if (def.regClass().is_subdword() && !validate_subdword_definition(program->chip_class, instr))
+ err |= ra_fail(program, loc, Location(), "Definition %d not aligned correctly", i);
if (!assignments[def.tempId()].firstloc.block)
assignments[def.tempId()].firstloc = loc;
assignments[def.tempId()].defloc = loc;
PhysReg reg = assignments.at(tmp.id()).reg;
for (unsigned i = 0; i < tmp.bytes(); i++) {
if (regs[reg.reg_b + i]) {
- err |= ra_fail(output, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
+ err |= ra_fail(program, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
}
regs[reg.reg_b + i] = tmp.id();
}
PhysReg reg = assignments.at(tmp.id()).reg;
for (unsigned i = 0; i < tmp.bytes(); i++) {
if (regs[reg.reg_b + i])
- err |= ra_fail(output, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
+ err |= ra_fail(program, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
}
live.emplace(tmp);
}
PhysReg reg = assignments.at(tmp.id()).reg;
for (unsigned j = 0; j < tmp.bytes(); j++) {
if (regs[reg.reg_b + j])
- err |= ra_fail(output, loc, assignments.at(regs[reg.reg_b + j]).defloc, "Assignment of element %d of %%%d already taken by %%%d from instruction", i, tmp.id(), regs[reg.reg_b + j]);
+ err |= ra_fail(program, loc, assignments.at(regs[reg.reg_b + j]).defloc, "Assignment of element %d of %%%d already taken by %%%d from instruction", i, tmp.id(), regs[reg.reg_b + j]);
regs[reg.reg_b + j] = tmp.id();
}
- if (def.regClass().is_subdword() && !instr_can_access_subdword(instr)) {
- for (unsigned j = tmp.bytes(); j < 4; j++)
- if (regs[reg.reg_b + j])
- err |= ra_fail(output, loc, assignments.at(regs[reg.reg_b + j]).defloc, "Assignment of element %d of %%%d overwrites the full register taken by %%%d from instruction", i, tmp.id(), regs[reg.reg_b + j]);
+ if (def.regClass().is_subdword() && def.bytes() < 4) {
+ unsigned written = get_subdword_bytes_written(program, instr, i);
+ /* If written=4, the instruction still might write the upper half. In that case, it's the lower half that isn't preserved */
+ for (unsigned j = reg.byte() & ~(written - 1); j < written; j++) {
+ unsigned written_reg = reg.reg() * 4u + j;
+ if (regs[written_reg] && regs[written_reg] != def.tempId())
+ err |= ra_fail(program, loc, assignments.at(regs[written_reg]).defloc, "Assignment of element %d of %%%d overwrites the full register taken by %%%d from instruction", i, tmp.id(), regs[written_reg]);
+ }
}
}