};
/* protos */
-static void ir_gen_extparam(ir_builder *ir);
+static ir_value* ir_gen_extparam_proto(ir_builder *ir);
+static void ir_gen_extparam (ir_builder *ir);
/* error functions */
* Vector utility functions
*/
-bool GMQCC_WARN vec_ir_value_find(ir_value **vec, ir_value *what, size_t *idx)
+bool GMQCC_WARN vec_ir_value_find(ir_value **vec, const ir_value *what, size_t *idx)
{
size_t i;
size_t len = vec_size(vec);
self->functions = NULL;
self->globals = NULL;
self->fields = NULL;
- self->extparams = NULL;
self->filenames = NULL;
self->filestrings = NULL;
self->htglobals = util_htnew(IR_HT_SIZE);
self->htfields = util_htnew(IR_HT_SIZE);
self->htfunctions = util_htnew(IR_HT_SIZE);
- self->max_locals = 0;
+ self->extparams = NULL;
+ self->extparam_protos = NULL;
+
+ self->first_common_globaltemp = 0;
+ self->max_globaltemps = 0;
+ self->first_common_local = 0;
+ self->max_locals = 0;
self->str_immediate = 0;
self->name = NULL;
self->code_function_def = -1;
self->allocated_locals = 0;
+ self->globaltemps = 0;
self->run_id = 0;
return self;
bool ir_function_finalize(ir_function *self)
{
+ size_t i;
+
if (self->builtin)
return true;
if (!ir_function_naive_phi(self))
return false;
+ for (i = 0; i < vec_size(self->locals); ++i) {
+ ir_value *v = self->locals[i];
+ if (v->vtype == TYPE_VECTOR ||
+ (v->vtype == TYPE_FIELD && v->outtype == TYPE_VECTOR))
+ {
+ ir_value_vector_member(v, 0);
+ ir_value_vector_member(v, 1);
+ ir_value_vector_member(v, 2);
+ }
+ }
+ for (i = 0; i < vec_size(self->values); ++i) {
+ ir_value *v = self->values[i];
+ if (v->vtype == TYPE_VECTOR ||
+ (v->vtype == TYPE_FIELD && v->outtype == TYPE_VECTOR))
+ {
+ ir_value_vector_member(v, 0);
+ ir_value_vector_member(v, 1);
+ ir_value_vector_member(v, 2);
+ }
+ }
+
ir_function_enumerate(self);
if (!ir_function_calculate_liferanges(self))
}
ve = ir_value_var(name, (param ? store_param : store_local), vtype);
+ if (param)
+ ve->locked = true;
vec_push(self->locals, ve);
return ve;
}
self->memberof = NULL;
self->unique_life = false;
+ self->locked = false;
+ self->callparam = false;
self->life = NULL;
return self;
bool *unique;
} function_allocator;
-static bool function_allocator_alloc(function_allocator *alloc, const ir_value *var)
+static bool function_allocator_alloc(function_allocator *alloc, ir_value *var)
{
ir_value *slot;
size_t vsize = ir_value_sizeof(var);
+ var->code.local = vec_size(alloc->locals);
+
slot = ir_value_var("reg", store_global, var->vtype);
if (!slot)
return false;
return false;
}
+static bool ir_function_allocator_assign(ir_function *self, function_allocator *alloc, ir_value *v)
+{
+ size_t a;
+ ir_value *slot;
+
+ for (a = 0; a < vec_size(alloc->locals); ++a)
+ {
+ /* if it's reserved for a unique liferange: skip */
+ if (alloc->unique[a])
+ continue;
+
+ slot = alloc->locals[a];
+
+ /* never resize parameters
+ * will be required later when overlapping temps + locals
+ */
+ if (a < vec_size(self->params) &&
+ alloc->sizes[a] < ir_value_sizeof(v))
+ {
+ continue;
+ }
+
+ if (ir_values_overlap(v, slot))
+ continue;
+
+ if (!ir_value_life_merge_into(slot, v))
+ return false;
+
+ /* adjust size for this slot */
+ if (alloc->sizes[a] < ir_value_sizeof(v))
+ alloc->sizes[a] = ir_value_sizeof(v);
+
+ v->code.local = a;
+ return true;
+ }
+ if (a >= vec_size(alloc->locals)) {
+ if (!function_allocator_alloc(alloc, v))
+ return false;
+ }
+ return true;
+}
+
bool ir_function_allocate_locals(ir_function *self)
{
- size_t i, a;
+ size_t i;
bool retval = true;
size_t pos;
+ bool opt_gt = OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS);
- ir_value *slot;
- const ir_value *v;
+ ir_value *v;
- function_allocator alloc;
+ function_allocator lockalloc, globalloc;
if (!vec_size(self->locals) && !vec_size(self->values))
return true;
- alloc.locals = NULL;
- alloc.sizes = NULL;
- alloc.positions = NULL;
- alloc.unique = NULL;
+ globalloc.locals = NULL;
+ globalloc.sizes = NULL;
+ globalloc.positions = NULL;
+ globalloc.unique = NULL;
+ lockalloc.locals = NULL;
+ lockalloc.sizes = NULL;
+ lockalloc.positions = NULL;
+ lockalloc.unique = NULL;
for (i = 0; i < vec_size(self->locals); ++i)
{
- if (!OPTS_OPTIMIZATION(OPTIM_LOCAL_TEMPS))
- self->locals[i]->unique_life = true;
- if (!function_allocator_alloc(&alloc, self->locals[i]))
+ v = self->locals[i];
+ if (!OPTS_OPTIMIZATION(OPTIM_LOCAL_TEMPS)) {
+ v->locked = true;
+ v->unique_life = true;
+ }
+ else if (i >= vec_size(self->params))
+ break;
+ else
+ v->locked = true; /* lock parameters locals */
+ if (!function_allocator_alloc((v->locked || !opt_gt ? &lockalloc : &globalloc), self->locals[i]))
+ goto error;
+ }
+ for (; i < vec_size(self->locals); ++i)
+ {
+ v = self->locals[i];
+ if (!vec_size(v->life))
+ continue;
+ if (!ir_function_allocator_assign(self, (v->locked || !opt_gt ? &lockalloc : &globalloc), v))
goto error;
}
if (!vec_size(v->life))
continue;
- for (a = 0; a < vec_size(alloc.locals); ++a)
- {
- /* if it's reserved for a unique liferange: skip */
- if (alloc.unique[a])
- continue;
-
- slot = alloc.locals[a];
-
- /* never resize parameters
- * will be required later when overlapping temps + locals
- */
- if (a < vec_size(self->params) &&
- alloc.sizes[a] < ir_value_sizeof(v))
+ /* CALL optimization:
+ * If the value is a parameter-temp: 1 write, 1 read from a CALL
+ * and it's not "locked", write it to the OFS_PARM directly.
+ */
+ if (OPTS_OPTIMIZATION(OPTIM_CALL_STORES) && !v->locked && !v->unique_life) {
+ if (vec_size(v->reads) == 1 && vec_size(v->writes) == 1 &&
+ (v->reads[0]->opcode == VINSTR_NRCALL ||
+ (v->reads[0]->opcode >= INSTR_CALL0 && v->reads[0]->opcode <= INSTR_CALL8)
+ )
+ )
{
+ size_t param;
+ ir_instr *call = v->reads[0];
+ if (!vec_ir_value_find(call->params, v, ¶m)) {
+ irerror(call->context, "internal error: unlocked parameter %s not found", v->name);
+ goto error;
+ }
+
+ ++opts_optimizationcount[OPTIM_CALL_STORES];
+ v->callparam = true;
+ if (param < 8)
+ ir_value_code_setaddr(v, OFS_PARM0 + 3*param);
+ else {
+ ir_value *ep;
+ param -= 8;
+ if (vec_size(self->owner->extparam_protos) <= param)
+ ep = ir_gen_extparam_proto(self->owner);
+ else
+ ep = self->owner->extparam_protos[param];
+ ir_instr_op(v->writes[0], 0, ep, true);
+ call->params[param+8] = ep;
+ }
continue;
}
-
- if (ir_values_overlap(v, slot))
+ if (vec_size(v->writes) == 1 && v->writes[0]->opcode == INSTR_CALL0)
+ {
+ v->store = store_return;
+ ++opts_optimizationcount[OPTIM_CALL_STORES];
continue;
-
- if (!ir_value_life_merge_into(slot, v))
- goto error;
-
- /* adjust size for this slot */
- if (alloc.sizes[a] < ir_value_sizeof(v))
- alloc.sizes[a] = ir_value_sizeof(v);
-
- self->values[i]->code.local = a;
- break;
- }
- if (a >= vec_size(alloc.locals)) {
- self->values[i]->code.local = vec_size(alloc.locals);
- if (!function_allocator_alloc(&alloc, v))
- goto error;
+ }
}
+
+ if (!ir_function_allocator_assign(self, (v->locked || !opt_gt ? &lockalloc : &globalloc), v))
+ goto error;
}
- if (!alloc.sizes) {
+ if (!lockalloc.sizes && !globalloc.sizes) {
goto cleanup;
}
+ vec_push(lockalloc.positions, 0);
+ vec_push(globalloc.positions, 0);
/* Adjust slot positions based on sizes */
- vec_push(alloc.positions, 0);
-
- if (vec_size(alloc.sizes))
- pos = alloc.positions[0] + alloc.sizes[0];
- else
- pos = 0;
- for (i = 1; i < vec_size(alloc.sizes); ++i)
- {
- pos = alloc.positions[i-1] + alloc.sizes[i-1];
- vec_push(alloc.positions, pos);
+ if (lockalloc.sizes) {
+ pos = (vec_size(lockalloc.sizes) ? (lockalloc.positions[0] + lockalloc.sizes[0]) : 0);
+ for (i = 1; i < vec_size(lockalloc.sizes); ++i)
+ {
+ pos = lockalloc.positions[i-1] + lockalloc.sizes[i-1];
+ vec_push(lockalloc.positions, pos);
+ }
+ self->allocated_locals = pos + vec_last(lockalloc.sizes);
+ }
+ if (globalloc.sizes) {
+ pos = (vec_size(globalloc.sizes) ? (globalloc.positions[0] + globalloc.sizes[0]) : 0);
+ for (i = 1; i < vec_size(globalloc.sizes); ++i)
+ {
+ pos = globalloc.positions[i-1] + globalloc.sizes[i-1];
+ vec_push(globalloc.positions, pos);
+ }
+ self->globaltemps = pos + vec_last(globalloc.sizes);
}
-
- self->allocated_locals = pos + vec_last(alloc.sizes);
/* Locals need to know their new position */
for (i = 0; i < vec_size(self->locals); ++i) {
- self->locals[i]->code.local = alloc.positions[i];
+ v = self->locals[i];
+ if (i >= vec_size(self->params) && !vec_size(v->life))
+ continue;
+ if (v->locked || !opt_gt)
+ v->code.local = lockalloc.positions[v->code.local];
+ else
+ v->code.local = globalloc.positions[v->code.local];
}
/* Take over the actual slot positions on values */
for (i = 0; i < vec_size(self->values); ++i) {
- self->values[i]->code.local = alloc.positions[self->values[i]->code.local];
+ v = self->values[i];
+ if (!vec_size(v->life))
+ continue;
+ if (v->locked || !opt_gt)
+ v->code.local = lockalloc.positions[v->code.local];
+ else
+ v->code.local = globalloc.positions[v->code.local];
}
goto cleanup;
error:
retval = false;
cleanup:
- for (i = 0; i < vec_size(alloc.locals); ++i)
- ir_value_delete(alloc.locals[i]);
- vec_free(alloc.unique);
- vec_free(alloc.locals);
- vec_free(alloc.sizes);
- vec_free(alloc.positions);
+ for (i = 0; i < vec_size(lockalloc.locals); ++i)
+ ir_value_delete(lockalloc.locals[i]);
+ for (i = 0; i < vec_size(globalloc.locals); ++i)
+ ir_value_delete(globalloc.locals[i]);
+ vec_free(globalloc.unique);
+ vec_free(globalloc.locals);
+ vec_free(globalloc.sizes);
+ vec_free(globalloc.positions);
+ vec_free(lockalloc.unique);
+ vec_free(lockalloc.locals);
+ vec_free(lockalloc.sizes);
+ vec_free(lockalloc.positions);
return retval;
}
for (i = 0; i != vec_size(self->living); ++i)
{
tempbool = ir_value_life_merge(self->living[i], eid);
- /* debug
- if (tempbool)
- irerror(self->context, "block_living_add_instr() value instruction added %s: %i", self->living[i]->_name, (int)eid);
- */
changed = changed || tempbool;
}
return changed;
}
+static bool ir_block_living_lock(ir_block *self)
+{
+ size_t i;
+ bool changed = false;
+ for (i = 0; i != vec_size(self->living); ++i)
+ {
+ if (!self->living[i]->locked)
+ changed = true;
+ self->living[i]->locked = true;
+ }
+ return changed;
+}
+
static bool ir_block_life_prop_previous(ir_block* self, ir_block *prev, bool *changed)
{
size_t i;
}
}
+ /* on a call, all these values must be "locked" */
+ if (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8) {
+ if (ir_block_living_lock(self))
+ *changed = true;
+ }
/* call params are read operands too */
for (p = 0; p < vec_size(instr->params); ++p)
{
*
* Breaking conventions is annoying...
*/
-static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal, bool defs_only);
+static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal);
static bool gen_global_field(ir_value *global)
{
if ( (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8)
|| instr->opcode == VINSTR_NRCALL)
{
- /* Trivial call translation:
- * copy all params to OFS_PARM*
- * if the output's storetype is not store_return,
- * add append a STORE instruction!
- *
- * NOTES on how to do it better without much trouble:
- * -) The liferanges!
- * Simply check the liferange of all parameters for
- * other CALLs. For each param with no CALL in its
- * liferange, we can store it in an OFS_PARM at
- * generation already. This would even include later
- * reuse.... probably... :)
- */
size_t p, first;
ir_value *retvalue;
for (p = 0; p < first; ++p)
{
ir_value *param = instr->params[p];
+ if (param->callparam)
+ continue;
stmt.opcode = INSTR_STORE_F;
stmt.o3.u1 = 0;
ir_value *param = instr->params[p];
ir_value *targetparam;
+ if (param->callparam)
+ continue;
+
if (p-8 >= vec_size(ir->extparams))
ir_gen_extparam(ir);
static bool gen_function_code(ir_function *self)
{
ir_block *block;
- prog_section_statement stmt;
+ prog_section_statement stmt, *retst;
/* Starting from entry point, we generate blocks "as they come"
* for now. Dead blocks will not be translated obviously.
}
/* code_write and qcvm -disasm need to know that the function ends here */
- stmt.opcode = INSTR_DONE;
- stmt.o1.u1 = 0;
- stmt.o2.u1 = 0;
- stmt.o3.u1 = 0;
- code_push_statement(&stmt, vec_last(code_linenums));
+ retst = &vec_last(code_statements);
+ if (OPTS_OPTIMIZATION(OPTIM_VOID_RETURN) &&
+ self->outtype == TYPE_VOID &&
+ retst->opcode == INSTR_RETURN &&
+ !retst->o1.u1 && !retst->o2.u1 && !retst->o3.u1)
+ {
+ retst->opcode = INSTR_DONE;
+ ++opts_optimizationcount[OPTIM_VOID_RETURN];
+ } else {
+ stmt.opcode = INSTR_DONE;
+ stmt.o1.u1 = 0;
+ stmt.o2.u1 = 0;
+ stmt.o3.u1 = 0;
+ code_push_statement(&stmt, vec_last(code_linenums));
+ }
return true;
}
return true;
}
+static ir_value* ir_gen_extparam_proto(ir_builder *ir)
+{
+ ir_value *global;
+ char name[128];
+
+ snprintf(name, sizeof(name), "EXTPARM#%i", (int)(vec_size(ir->extparam_protos)+8));
+ global = ir_value_var(name, store_global, TYPE_VECTOR);
+
+ vec_push(ir->extparam_protos, global);
+ return global;
+}
+
static void ir_gen_extparam(ir_builder *ir)
{
prog_section_def def;
ir_value *global;
- char name[128];
- snprintf(name, sizeof(name), "EXTPARM#%i", (int)(vec_size(ir->extparams)+8));
- global = ir_value_var(name, store_global, TYPE_VECTOR);
+ if (vec_size(ir->extparam_protos) < vec_size(ir->extparams)+1)
+ global = ir_gen_extparam_proto(ir);
+ else
+ global = ir->extparam_protos[vec_size(ir->extparams)];
- def.name = code_genstring(name);
+ def.name = code_genstring(global->name);
def.type = TYPE_VECTOR;
def.offset = vec_size(code_globals);
prog_section_function *def;
ir_function *irfun;
size_t i;
- uint32_t firstlocal;
+ uint32_t firstlocal, firstglobal;
irfun = global->constval.vfunc;
def = code_functions + irfun->code_function_def;
++opts_optimizationcount[OPTIM_OVERLAP_LOCALS];
}
+ firstglobal = (OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS) ? ir->first_common_globaltemp : firstlocal);
+
for (i = vec_size(code_globals); i < firstlocal + irfun->allocated_locals; ++i)
vec_push(code_globals, 0);
for (i = 0; i < vec_size(irfun->locals); ++i) {
- ir_value_code_setaddr(irfun->locals[i], firstlocal + irfun->locals[i]->code.local);
- if (!ir_builder_gen_global(ir, irfun->locals[i], true, true)) {
- irerror(irfun->locals[i]->context, "failed to generate local %s", irfun->locals[i]->name);
- return false;
+ ir_value *v = irfun->locals[i];
+ if (v->locked || !OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS)) {
+ ir_value_code_setaddr(v, firstlocal + v->code.local);
+ if (!ir_builder_gen_global(ir, irfun->locals[i], true)) {
+ irerror(irfun->locals[i]->context, "failed to generate local %s", irfun->locals[i]->name);
+ return false;
+ }
}
+ else
+ ir_value_code_setaddr(v, firstglobal + v->code.local);
}
for (i = 0; i < vec_size(irfun->values); ++i)
{
ir_value *v = irfun->values[i];
- ir_value_code_setaddr(v, firstlocal + v->code.local);
+ if (v->callparam)
+ continue;
+ if (v->locked)
+ ir_value_code_setaddr(v, firstlocal + v->code.local);
+ else
+ ir_value_code_setaddr(v, firstglobal + v->code.local);
}
return true;
}
}
}
-static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal, bool defs_only)
+static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal)
{
size_t i;
int32_t *iptr;
prog_section_def def;
bool pushdef = false;
+ def.type = global->vtype;
+ def.offset = vec_size(code_globals);
+ def.name = 0;
if (opts.g || !islocal)
{
pushdef = true;
- def.type = global->vtype;
- def.offset = vec_size(code_globals);
if (OPTS_OPTIMIZATION(OPTIM_STRIP_CONSTANT_NAMES) &&
(global->name[0] == '#' || global->cvq == CV_CONST))
}
else
def.name = 0;
- if (defs_only) {
+ if (islocal) {
def.offset = ir_value_code_addr(global);
vec_push(code_defs, def);
if (global->vtype == TYPE_VECTOR)
return true;
}
}
- if (defs_only)
+ if (islocal)
return true;
switch (global->vtype)
for (i = 0; i < vec_size(self->globals); ++i)
{
- if (!ir_builder_gen_global(self, self->globals[i], false, false)) {
+ if (!ir_builder_gen_global(self, self->globals[i], false)) {
return false;
}
if (self->globals[i]->vtype == TYPE_FUNCTION) {
{
self->max_locals = func->allocated_locals;
}
+ if (func && self->max_globaltemps < func->globaltemps)
+ self->max_globaltemps = func->globaltemps;
}
}
}
}
+ /* generate global temps */
+ self->first_common_globaltemp = vec_size(code_globals);
+ for (i = 0; i < self->max_globaltemps; ++i) {
+ vec_push(code_globals, 0);
+ }
/* generate common locals */
self->first_common_local = vec_size(code_globals);
for (i = 0; i < self->max_locals; ++i) {
}
oprintf("%sliferanges:\n", ind);
for (i = 0; i < vec_size(f->locals); ++i) {
+ const char *attr = "";
size_t l, m;
ir_value *v = f->locals[i];
- oprintf("%s\t%s: %s@%i ", ind, v->name, (v->unique_life ? "unique " : ""), (int)v->code.local);
+ if (v->unique_life && v->locked)
+ attr = "unique,locked ";
+ else if (v->unique_life)
+ attr = "unique ";
+ else if (v->locked)
+ attr = "locked ";
+ oprintf("%s\t%s: %s@%i ", ind, v->name, attr, (int)v->code.local);
for (l = 0; l < vec_size(v->life); ++l) {
oprintf("[%i,%i] ", v->life[l].start, v->life[l].end);
}
ir_value *vm = v->members[m];
if (!vm)
continue;
- oprintf("%s\t%s: %s@%i ", ind, vm->name, (vm->unique_life ? "unique " : ""), (int)vm->code.local);
+ if (vm->unique_life && vm->locked)
+ attr = "unique,locked ";
+ else if (vm->unique_life)
+ attr = "unique ";
+ else if (vm->locked)
+ attr = "locked ";
+ oprintf("%s\t%s: %s@%i ", ind, vm->name, attr, (int)vm->code.local);
for (l = 0; l < vec_size(vm->life); ++l) {
oprintf("[%i,%i] ", vm->life[l].start, vm->life[l].end);
}
projects / xonotic / gmqcc.git / blobdiff