X-Git-Url: https://git.xonotic.org/?p=xonotic%2Fgmqcc.git;a=blobdiff_plain;f=ir.c;h=e9a73528d23ec27763e71a798ef6cf93bf1dc779;hp=3467d321057045264f4358ae15fa40ecd79fa1fe;hb=d4483bfda63dac509144741c1f0e11d68d989b66;hpb=2534076963894e7f43ba0366566184f2cd21fe57

diff --git a/ir.c b/ir.c
index 3467d32..e9a7352 100644
--- a/ir.c
+++ b/ir.c
@@ -45,7 +45,7 @@ const char *type_name[TYPE_COUNT] = {
     "array"
 };
 
-size_t type_sizeof[TYPE_COUNT] = {
+size_t type_sizeof_[TYPE_COUNT] = {
     1, /* TYPE_VOID     */
     1, /* TYPE_STRING   */
     1, /* TYPE_FLOAT    */
@@ -194,7 +194,8 @@ uint16_t type_not_instr[TYPE_COUNT] = {
 };
 
 /* 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 */
 
@@ -209,18 +210,18 @@ static void irerror(lex_ctx ctx, const char *msg, ...)
 static bool irwarning(lex_ctx ctx, int warntype, const char *fmt, ...)
 {
     bool    r;
-	va_list ap;
-	va_start(ap, fmt);
-	r = vcompile_warning(ctx, warntype, fmt, ap);
-	va_end(ap);
-	return r;
+    va_list ap;
+    va_start(ap, fmt);
+    r = vcompile_warning(ctx, warntype, fmt, ap);
+    va_end(ap);
+    return r;
 }
 
 /***********************************************************************
  * 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);
@@ -278,13 +279,20 @@ ir_builder* ir_builder_new(const char *modulename)
     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->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;
     if (!ir_builder_set_name(self, modulename)) {
@@ -433,6 +441,8 @@ ir_function* ir_function_new(ir_builder* owner, int outtype)
         mem_d(self);
         return NULL;
     }
+    self->flags = 0;
+
     self->owner = owner;
     self->context.file = "<@no context>";
     self->context.line = 0;
@@ -447,6 +457,7 @@ ir_function* ir_function_new(ir_builder* owner, int outtype)
 
     self->code_function_def = -1;
     self->allocated_locals = 0;
+    self->globaltemps      = 0;
 
     self->run_id = 0;
     return self;
@@ -527,7 +538,8 @@ static bool instr_is_operation(uint16_t op)
              (op >= INSTR_LOAD_F && op <= INSTR_LOAD_FNC) ||
              (op == INSTR_ADDRESS) ||
              (op >= INSTR_NOT_F  && op <= INSTR_NOT_FNC) ||
-             (op >= INSTR_AND    && op <= INSTR_BITOR) );
+             (op >= INSTR_AND    && op <= INSTR_BITOR) ||
+             (op >= INSTR_CALL0  && op <= INSTR_CALL8) );
 }
 
 bool ir_function_pass_peephole(ir_function *self)
@@ -575,7 +587,7 @@ bool ir_function_pass_peephole(ir_function *self)
                 if (store->_ops[1] != value)
                     continue;
 
-                ++optimization_count[OPTIM_PEEPHOLE];
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
                 (void)!ir_instr_op(oper, 0, store->_ops[0], true);
 
                 vec_remove(block->instr, i, 1);
@@ -604,13 +616,14 @@ bool ir_function_pass_peephole(ir_function *self)
                     if (inot->_ops[0] != value ||
                         inot->opcode < INSTR_NOT_F ||
                         inot->opcode > INSTR_NOT_FNC ||
-                        inot->opcode == INSTR_NOT_V) /* can't do this one */
+                        inot->opcode == INSTR_NOT_V || /* can't do these */
+                        inot->opcode == INSTR_NOT_S)
                     {
                         break;
                     }
 
                     /* count */
-                    ++optimization_count[OPTIM_PEEPHOLE];
+                    ++opts_optimizationcount[OPTIM_PEEPHOLE];
                     /* change operand */
                     (void)!ir_instr_op(inst, 0, inot->_ops[1], false);
                     /* remove NOT */
@@ -638,7 +651,7 @@ bool ir_function_pass_peephole(ir_function *self)
     return true;
 }
 
-bool ir_function_pass_tailcall(ir_function *self)
+bool ir_function_pass_tailrecursion(ir_function *self)
 {
     size_t b, p;
 
@@ -678,7 +691,7 @@ bool ir_function_pass_tailcall(ir_function *self)
                 ret->_ops[0]   == store->_ops[0] &&
                 store->_ops[1] == call->_ops[0])
             {
-                ++optimization_count[OPTIM_PEEPHOLE];
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
                 call->_ops[0] = store->_ops[0];
                 vec_remove(block->instr, vec_size(block->instr) - 2, 1);
                 ir_instr_delete(store);
@@ -700,7 +713,7 @@ bool ir_function_pass_tailcall(ir_function *self)
         if (ret->_ops[0] && call->_ops[0] != ret->_ops[0])
             continue;
 
-        ++optimization_count[OPTIM_TAIL_RECURSION];
+        ++opts_optimizationcount[OPTIM_TAIL_RECURSION];
         vec_shrinkby(block->instr, 2);
 
         block->final = false; /* open it back up */
@@ -727,6 +740,8 @@ bool ir_function_pass_tailcall(ir_function *self)
 
 bool ir_function_finalize(ir_function *self)
 {
+    size_t i;
+
     if (self->builtin)
         return true;
 
@@ -738,8 +753,8 @@ bool ir_function_finalize(ir_function *self)
     }
 
     if (OPTS_OPTIMIZATION(OPTIM_TAIL_RECURSION)) {
-        if (!ir_function_pass_tailcall(self)) {
-            irerror(self->context, "tailcall optimization pass broke something in `%s`", self->name);
+        if (!ir_function_pass_tailrecursion(self)) {
+            irerror(self->context, "tail-recursion optimization pass broke something in `%s`", self->name);
             return false;
         }
     }
@@ -747,6 +762,27 @@ bool ir_function_finalize(ir_function *self)
     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))
@@ -768,6 +804,8 @@ ir_value* ir_function_create_local(ir_function *self, const char *name, int vtyp
     }
 
     ve = ir_value_var(name, (param ? store_param : store_local), vtype);
+    if (param)
+        ve->locked = true;
     vec_push(self->locals, ve);
     return ve;
 }
@@ -981,6 +1019,8 @@ ir_value* ir_value_var(const char *name, int storetype, int vtype)
     self->memberof = NULL;
 
     self->unique_life = false;
+    self->locked      = false;
+    self->callparam   = false;
 
     self->life = NULL;
     return self;
@@ -988,6 +1028,8 @@ ir_value* ir_value_var(const char *name, int storetype, int vtype)
 
 ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
 {
+    char     *name;
+    size_t    len;
     ir_value *m;
     if (member >= 3)
         return NULL;
@@ -995,9 +1037,22 @@ ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
     if (self->members[member])
         return self->members[member];
 
+    if (self->name) {
+        len = strlen(self->name);
+        name = (char*)mem_a(len + 3);
+        memcpy(name, self->name, len);
+        name[len+0] = '_';
+        name[len+1] = 'x' + member;
+        name[len+2] = '\0';
+    }
+    else
+        name = NULL;
+
     if (self->vtype == TYPE_VECTOR)
     {
-        m = ir_value_var(self->name, self->store, TYPE_FLOAT);
+        m = ir_value_var(name, self->store, TYPE_FLOAT);
+        if (name)
+            mem_d(name);
         if (!m)
             return NULL;
         m->context = self->context;
@@ -1009,7 +1064,9 @@ ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
     {
         if (self->fieldtype != TYPE_VECTOR)
             return NULL;
-        m = ir_value_var(self->name, self->store, TYPE_FIELD);
+        m = ir_value_var(name, self->store, TYPE_FIELD);
+        if (name)
+            mem_d(name);
         if (!m)
             return NULL;
         m->fieldtype = TYPE_FLOAT;
@@ -1028,6 +1085,13 @@ ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
     return m;
 }
 
+static GMQCC_INLINE size_t ir_value_sizeof(const ir_value *self)
+{
+    if (self->vtype == TYPE_FIELD && self->fieldtype == TYPE_VECTOR)
+        return type_sizeof_[TYPE_VECTOR];
+    return type_sizeof_[self->vtype];
+}
+
 ir_value* ir_value_out(ir_function *owner, const char *name, int storetype, int vtype)
 {
     ir_value *v = ir_value_var(name, storetype, vtype);
@@ -1105,7 +1169,7 @@ static char *ir_strdup(const char *str)
 {
     if (str && !*str) {
         /* actually dup empty strings */
-        char *out = mem_a(1);
+        char *out = (char*)mem_a(1);
         *out = 0;
         return out;
     }
@@ -1350,15 +1414,19 @@ bool ir_values_overlap(const ir_value *a, const ir_value *b)
  *IR main operations
  */
 
+static bool ir_check_unreachable(ir_block *self)
+{
+    /* The IR should never have to deal with unreachable code */
+    if (!self->final/* || OPTS_FLAG(ALLOW_UNREACHABLE_CODE)*/)
+        return true;
+    irerror(self->context, "unreachable statement (%s)", self->label);
+    return false;
+}
+
 bool ir_block_create_store_op(ir_block *self, lex_ctx ctx, int op, ir_value *target, ir_value *what)
 {
     ir_instr *in;
-    if (self->final) {
-        irerror(self->context, "unreachable statement (%s)", self->label);
-        return false;
-    }
-    in = ir_instr_new(ctx, self, op);
-    if (!in)
+    if (!ir_check_unreachable(self))
         return false;
 
     if (target->store == store_value &&
@@ -1370,9 +1438,14 @@ bool ir_block_create_store_op(ir_block *self, lex_ctx ctx, int op, ir_value *tar
         return false;
     }
 
-    if (!ir_instr_op(in, 0, target, true) ||
+    in = ir_instr_new(ctx, self, op);
+    if (!in)
+        return false;
+
+    if (!ir_instr_op(in, 0, target, (op < INSTR_STOREP_F || op > INSTR_STOREP_FNC)) ||
         !ir_instr_op(in, 1, what, false))
     {
+        ir_instr_delete(in);
         return false;
     }
     vec_push(self->instr, in);
@@ -1429,18 +1502,18 @@ bool ir_block_create_storep(ir_block *self, lex_ctx ctx, ir_value *target, ir_va
 bool ir_block_create_return(ir_block *self, lex_ctx ctx, ir_value *v)
 {
     ir_instr *in;
-    if (self->final) {
-        irerror(self->context, "unreachable statement (%s)", self->label);
+    if (!ir_check_unreachable(self))
         return false;
-    }
     self->final = true;
     self->is_return = true;
     in = ir_instr_new(ctx, self, INSTR_RETURN);
     if (!in)
         return false;
 
-    if (v && !ir_instr_op(in, 0, v, false))
+    if (v && !ir_instr_op(in, 0, v, false)) {
+        ir_instr_delete(in);
         return false;
+    }
 
     vec_push(self->instr, in);
     return true;
@@ -1450,10 +1523,8 @@ bool ir_block_create_if(ir_block *self, lex_ctx ctx, ir_value *v,
                         ir_block *ontrue, ir_block *onfalse)
 {
     ir_instr *in;
-    if (self->final) {
-        irerror(self->context, "unreachable statement (%s)", self->label);
+    if (!ir_check_unreachable(self))
         return false;
-    }
     self->final = true;
     /*in = ir_instr_new(ctx, self, (v->vtype == TYPE_STRING ? INSTR_IF_S : INSTR_IF_F));*/
     in = ir_instr_new(ctx, self, VINSTR_COND);
@@ -1480,10 +1551,8 @@ bool ir_block_create_if(ir_block *self, lex_ctx ctx, ir_value *v,
 bool ir_block_create_jump(ir_block *self, lex_ctx ctx, ir_block *to)
 {
     ir_instr *in;
-    if (self->final) {
-        irerror(self->context, "unreachable statement (%s)", self->label);
+    if (!ir_check_unreachable(self))
         return false;
-    }
     self->final = true;
     in = ir_instr_new(ctx, self, VINSTR_JUMP);
     if (!in)
@@ -1499,28 +1568,16 @@ bool ir_block_create_jump(ir_block *self, lex_ctx ctx, ir_block *to)
 
 bool ir_block_create_goto(ir_block *self, lex_ctx ctx, ir_block *to)
 {
-    ir_instr *in;
-    if (self->final) {
-        irerror(self->context, "unreachable statement (%s)", self->label);
-        return false;
-    }
-    self->final = true;
-    in = ir_instr_new(ctx, self, INSTR_GOTO);
-    if (!in)
-        return false;
-
-    in->bops[0] = to;
-    vec_push(self->instr, in);
-
-    vec_push(self->exits, to);
-    vec_push(to->entries, self);
-    return true;
+    self->owner->flags |= IR_FLAG_HAS_GOTO;
+    return ir_block_create_jump(self, ctx, to);
 }
 
 ir_instr* ir_block_create_phi(ir_block *self, lex_ctx ctx, const char *label, int ot)
 {
     ir_value *out;
     ir_instr *in;
+    if (!ir_check_unreachable(self))
+        return NULL;
     in = ir_instr_new(ctx, self, VINSTR_PHI);
     if (!in)
         return NULL;
@@ -1562,13 +1619,19 @@ void ir_phi_add(ir_instr* self, ir_block *b, ir_value *v)
 }
 
 /* call related code */
-ir_instr* ir_block_create_call(ir_block *self, lex_ctx ctx, const char *label, ir_value *func)
+ir_instr* ir_block_create_call(ir_block *self, lex_ctx ctx, const char *label, ir_value *func, bool noreturn)
 {
     ir_value *out;
     ir_instr *in;
-    in = ir_instr_new(ctx, self, INSTR_CALL0);
+    if (!ir_check_unreachable(self))
+        return NULL;
+    in = ir_instr_new(ctx, self, (noreturn ? VINSTR_NRCALL : INSTR_CALL0));
     if (!in)
         return NULL;
+    if (noreturn) {
+        self->final = true;
+        self->is_return = true;
+    }
     out = ir_value_out(self->owner, label, (func->outtype == TYPE_VOID) ? store_return : store_value, func->outtype);
     if (!out) {
         ir_instr_delete(in);
@@ -1582,6 +1645,15 @@ ir_instr* ir_block_create_call(ir_block *self, lex_ctx ctx, const char *label, i
         return NULL;
     }
     vec_push(self->instr, in);
+    /*
+    if (noreturn) {
+        if (!ir_block_create_return(self, ctx, NULL)) {
+            compile_error(ctx, "internal error: failed to generate dummy-return instruction");
+            ir_instr_delete(in);
+            return NULL;
+        }
+    }
+    */
     return in;
 }
 
@@ -1801,175 +1873,6 @@ ir_value* ir_block_create_load_from_ent(ir_block *self, lex_ctx ctx, const char
     return ir_block_create_general_instr(self, ctx, label, op, ent, field, outype);
 }
 
-ir_value* ir_block_create_add(ir_block *self, lex_ctx ctx,
-                              const char *label,
-                              ir_value *left, ir_value *right)
-{
-    int op = 0;
-    int l = left->vtype;
-    int r = right->vtype;
-    if (l == r) {
-        switch (l) {
-            default:
-                irerror(self->context, "invalid type for ir_block_create_add: %s", type_name[l]);
-                return NULL;
-            case TYPE_FLOAT:
-                op = INSTR_ADD_F;
-                break;
-#if 0
-            case TYPE_INTEGER:
-                op = INSTR_ADD_I;
-                break;
-#endif
-            case TYPE_VECTOR:
-                op = INSTR_ADD_V;
-                break;
-        }
-    } else {
-#if 0
-        if ( (l == TYPE_FLOAT && r == TYPE_INTEGER) )
-            op = INSTR_ADD_FI;
-        else if ( (l == TYPE_INTEGER && r == TYPE_FLOAT) )
-            op = INSTR_ADD_IF;
-        else
-#endif
-        {
-            irerror(self->context, "invalid type for ir_block_create_add: %s", type_name[l]);
-            return NULL;
-        }
-    }
-    return ir_block_create_binop(self, ctx, label, op, left, right);
-}
-
-ir_value* ir_block_create_sub(ir_block *self, lex_ctx ctx,
-                              const char *label,
-                              ir_value *left, ir_value *right)
-{
-    int op = 0;
-    int l = left->vtype;
-    int r = right->vtype;
-    if (l == r) {
-
-        switch (l) {
-            default:
-                irerror(self->context, "invalid type for ir_block_create_sub: %s", type_name[l]);
-                return NULL;
-            case TYPE_FLOAT:
-                op = INSTR_SUB_F;
-                break;
-#if 0
-            case TYPE_INTEGER:
-                op = INSTR_SUB_I;
-                break;
-#endif
-            case TYPE_VECTOR:
-                op = INSTR_SUB_V;
-                break;
-        }
-    } else {
-#if 0
-        if ( (l == TYPE_FLOAT && r == TYPE_INTEGER) )
-            op = INSTR_SUB_FI;
-        else if ( (l == TYPE_INTEGER && r == TYPE_FLOAT) )
-            op = INSTR_SUB_IF;
-        else
-#endif
-        {
-            irerror(self->context, "invalid type for ir_block_create_sub: %s", type_name[l]);
-            return NULL;
-        }
-    }
-    return ir_block_create_binop(self, ctx, label, op, left, right);
-}
-
-ir_value* ir_block_create_mul(ir_block *self, lex_ctx ctx,
-                              const char *label,
-                              ir_value *left, ir_value *right)
-{
-    int op = 0;
-    int l = left->vtype;
-    int r = right->vtype;
-    if (l == r) {
-
-        switch (l) {
-            default:
-                irerror(self->context, "invalid type for ir_block_create_mul: %s", type_name[l]);
-                return NULL;
-            case TYPE_FLOAT:
-                op = INSTR_MUL_F;
-                break;
-#if 0
-            case TYPE_INTEGER:
-                op = INSTR_MUL_I;
-                break;
-#endif
-            case TYPE_VECTOR:
-                op = INSTR_MUL_V;
-                break;
-        }
-    } else {
-        if ( (l == TYPE_VECTOR && r == TYPE_FLOAT) )
-            op = INSTR_MUL_VF;
-        else if ( (l == TYPE_FLOAT && r == TYPE_VECTOR) )
-            op = INSTR_MUL_FV;
-#if 0
-        else if ( (l == TYPE_VECTOR && r == TYPE_INTEGER) )
-            op = INSTR_MUL_VI;
-        else if ( (l == TYPE_INTEGER && r == TYPE_VECTOR) )
-            op = INSTR_MUL_IV;
-        else if ( (l == TYPE_FLOAT && r == TYPE_INTEGER) )
-            op = INSTR_MUL_FI;
-        else if ( (l == TYPE_INTEGER && r == TYPE_FLOAT) )
-            op = INSTR_MUL_IF;
-#endif
-        else {
-            irerror(self->context, "invalid type for ir_block_create_mul: %s", type_name[l]);
-            return NULL;
-        }
-    }
-    return ir_block_create_binop(self, ctx, label, op, left, right);
-}
-
-ir_value* ir_block_create_div(ir_block *self, lex_ctx ctx,
-                              const char *label,
-                              ir_value *left, ir_value *right)
-{
-    int op = 0;
-    int l = left->vtype;
-    int r = right->vtype;
-    if (l == r) {
-
-        switch (l) {
-            default:
-                irerror(self->context, "invalid type for ir_block_create_div: %s", type_name[l]);
-                return NULL;
-            case TYPE_FLOAT:
-                op = INSTR_DIV_F;
-                break;
-#if 0
-            case TYPE_INTEGER:
-                op = INSTR_DIV_I;
-                break;
-#endif
-        }
-    } else {
-#if 0
-        if ( (l == TYPE_VECTOR && r == TYPE_FLOAT) )
-            op = INSTR_DIV_VF;
-        else if ( (l == TYPE_FLOAT && r == TYPE_INTEGER) )
-            op = INSTR_DIV_FI;
-        else if ( (l == TYPE_INTEGER && r == TYPE_FLOAT) )
-            op = INSTR_DIV_IF;
-        else
-#endif
-        {
-            irerror(self->context, "invalid type for ir_block_create_div: %s", type_name[l]);
-            return NULL;
-        }
-    }
-    return ir_block_create_binop(self, ctx, label, op, left, right);
-}
-
 /* PHI resolving breaks the SSA, and must thus be the last
  * step before life-range calculation.
  */
@@ -2137,7 +2040,7 @@ static void ir_block_enumerate(ir_block *self, size_t *_eid)
 void ir_function_enumerate(ir_function *self)
 {
     size_t i;
-    size_t instruction_id = 0;
+    size_t instruction_id = 1;
     for (i = 0; i < vec_size(self->blocks); ++i)
     {
         self->blocks[i]->eid = i;
@@ -2149,9 +2052,13 @@ void ir_function_enumerate(ir_function *self)
 static bool ir_block_life_propagate(ir_block *b, ir_block *prev, bool *changed);
 bool ir_function_calculate_liferanges(ir_function *self)
 {
-    size_t i;
+    size_t i, s;
     bool changed;
 
+    /* parameters live at 0 */
+    for (i = 0; i < vec_size(self->params); ++i)
+        ir_value_life_merge(self->locals[i], 0);
+
     do {
         self->run_id++;
         changed = false;
@@ -2169,8 +2076,47 @@ bool ir_function_calculate_liferanges(ir_function *self)
         ir_block *block = self->blocks[0];
         for (i = 0; i < vec_size(block->living); ++i) {
             ir_value *v = block->living[i];
-            if (v->memberof || v->store != store_local)
+            if (v->store != store_local)
                 continue;
+            if (v->vtype == TYPE_VECTOR)
+                continue;
+            self->flags |= IR_FLAG_HAS_UNINITIALIZED;
+            /* find the instruction reading from it */
+            for (s = 0; s < vec_size(v->reads); ++s) {
+                if (v->reads[s]->eid == v->life[0].end)
+                    break;
+            }
+            if (s < vec_size(v->reads)) {
+                if (irwarning(v->context, WARN_USED_UNINITIALIZED,
+                              "variable `%s` may be used uninitialized in this function\n"
+                              " -> %s:%i",
+                              v->name,
+                              v->reads[s]->context.file, v->reads[s]->context.line)
+                   )
+                {
+                    return false;
+                }
+                continue;
+            }
+            if (v->memberof) {
+                ir_value *vec = v->memberof;
+                for (s = 0; s < vec_size(vec->reads); ++s) {
+                    if (vec->reads[s]->eid == v->life[0].end)
+                        break;
+                }
+                if (s < vec_size(vec->reads)) {
+                    if (irwarning(v->context, WARN_USED_UNINITIALIZED,
+                                  "variable `%s` may be used uninitialized in this function\n"
+                                  " -> %s:%i",
+                                  v->name,
+                                  vec->reads[s]->context.file, vec->reads[s]->context.line)
+                       )
+                    {
+                        return false;
+                    }
+                    continue;
+                }
+            }
             if (irwarning(v->context, WARN_USED_UNINITIALIZED,
                           "variable `%s` may be used uninitialized in this function", v->name))
             {
@@ -2193,10 +2139,12 @@ typedef struct {
     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 = type_sizeof[var->vtype];
+    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)
@@ -2216,30 +2164,91 @@ localerror:
     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_LOCALTEMPS))
-            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;
     }
 
@@ -2251,69 +2260,97 @@ bool ir_function_allocate_locals(ir_function *self)
         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] < type_sizeof[v->vtype])
+        /* 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, &param)) {
+                    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] < type_sizeof[v->vtype])
-                alloc.sizes[a] = type_sizeof[v->vtype];
-
-            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;
@@ -2321,11 +2358,18 @@ bool ir_function_allocate_locals(ir_function *self)
 error:
     retval = false;
 cleanup:
-    for (i = 0; i < vec_size(alloc.locals); ++i)
-        ir_value_delete(alloc.locals[i]);
-    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;
 }
 
@@ -2376,15 +2420,24 @@ static bool ir_block_living_add_instr(ir_block *self, size_t eid)
     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;
@@ -2425,7 +2478,7 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
     ir_instr *instr;
     ir_value *value;
     bool  tempbool;
-    size_t i, o, p;
+    size_t i, o, p, mem;
     /* bitmasks which operands are read from or written to */
     size_t read, write;
     char dbg_ind[16] = { '#', '0' };
@@ -2442,26 +2495,6 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
     { --i;
         instr = self->instr[i];
 
-        /* PHI operands are always read operands */
-        for (p = 0; p < vec_size(instr->phi); ++p)
-        {
-            value = instr->phi[p].value;
-            if (value->memberof)
-                value = value->memberof;
-            if (!vec_ir_value_find(self->living, value, NULL))
-                vec_push(self->living, value);
-        }
-
-        /* call params are read operands too */
-        for (p = 0; p < vec_size(instr->params); ++p)
-        {
-            value = instr->params[p];
-            if (value->memberof)
-                value = value->memberof;
-            if (!vec_ir_value_find(self->living, value, NULL))
-                vec_push(self->living, value);
-        }
-
         /* See which operands are read and write operands */
         ir_op_read_write(instr->opcode, &read, &write);
 
@@ -2478,15 +2511,15 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
             *changed = *changed || tempbool;
         }
 
-        /* Go through the 3 main operands */
+        /* Go through the 3 main operands
+         * writes first, then reads
+         */
         for (o = 0; o < 3; ++o)
         {
             if (!instr->_ops[o]) /* no such operand */
                 continue;
 
             value = instr->_ops[o];
-            if (value->memberof)
-                value = value->memberof;
 
             /* We only care about locals */
             /* we also calculate parameter liferanges so that locals
@@ -2496,13 +2529,6 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
                 value->store != store_param)
                 continue;
 
-            /* read operands */
-            if (read & (1<<o))
-            {
-                if (!vec_ir_value_find(self->living, value, NULL))
-                    vec_push(self->living, value);
-            }
-
             /* write operands */
             /* When we write to a local, we consider it "dead" for the
              * remaining upper part of the function, since in SSA a value
@@ -2521,30 +2547,107 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
                      * and make sure it's only printed once
                      * since this function is run multiple times.
                      */
-                    /* For now: debug info: */
                     /* con_err( "Value only written %s\n", value->name); */
                     tempbool = ir_value_life_merge(value, instr->eid);
                     *changed = *changed || tempbool;
-                    /*
-                    ir_instr_dump(instr, dbg_ind, printf);
-                    abort();
-                    */
                 } else {
                     /* since 'living' won't contain it
                      * anymore, merge the value, since
                      * (A) doesn't.
                      */
                     tempbool = ir_value_life_merge(value, instr->eid);
-                    /*
-                    if (tempbool)
-                        con_err( "value added id %s %i\n", value->name, (int)instr->eid);
-                    */
                     *changed = *changed || tempbool;
                     /* Then remove */
                     vec_remove(self->living, idx, 1);
                 }
+                /* Removing a vector removes all members */
+                for (mem = 0; mem < 3; ++mem) {
+                    if (value->members[mem] && vec_ir_value_find(self->living, value->members[mem], &idx)) {
+                        tempbool = ir_value_life_merge(value->members[mem], instr->eid);
+                        *changed = *changed || tempbool;
+                        vec_remove(self->living, idx, 1);
+                    }
+                }
+                /* Removing the last member removes the vector */
+                if (value->memberof) {
+                    value = value->memberof;
+                    for (mem = 0; mem < 3; ++mem) {
+                        if (value->members[mem] && vec_ir_value_find(self->living, value->members[mem], NULL))
+                            break;
+                    }
+                    if (mem == 3 && vec_ir_value_find(self->living, value, &idx)) {
+                        tempbool = ir_value_life_merge(value, instr->eid);
+                        *changed = *changed || tempbool;
+                        vec_remove(self->living, idx, 1);
+                    }
+                }
+            }
+        }
+
+        for (o = 0; o < 3; ++o)
+        {
+            if (!instr->_ops[o]) /* no such operand */
+                continue;
+
+            value = instr->_ops[o];
+
+            /* We only care about locals */
+            /* we also calculate parameter liferanges so that locals
+             * can take up parameter slots */
+            if (value->store != store_value &&
+                value->store != store_local &&
+                value->store != store_param)
+                continue;
+
+            /* read operands */
+            if (read & (1<<o))
+            {
+                if (!vec_ir_value_find(self->living, value, NULL))
+                    vec_push(self->living, value);
+                /* reading adds the full vector */
+                if (value->memberof && !vec_ir_value_find(self->living, value->memberof, NULL))
+                    vec_push(self->living, value->memberof);
+                for (mem = 0; mem < 3; ++mem) {
+                    if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], NULL))
+                        vec_push(self->living, value->members[mem]);
+                }
+            }
+        }
+        /* PHI operands are always read operands */
+        for (p = 0; p < vec_size(instr->phi); ++p)
+        {
+            value = instr->phi[p].value;
+            if (!vec_ir_value_find(self->living, value, NULL))
+                vec_push(self->living, value);
+            /* reading adds the full vector */
+            if (value->memberof && !vec_ir_value_find(self->living, value->memberof, NULL))
+                vec_push(self->living, value->memberof);
+            for (mem = 0; mem < 3; ++mem) {
+                if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], NULL))
+                    vec_push(self->living, value->members[mem]);
+            }
+        }
+
+        /* 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)
+        {
+            value = instr->params[p];
+            if (!vec_ir_value_find(self->living, value, NULL))
+                vec_push(self->living, value);
+            /* reading adds the full vector */
+            if (value->memberof && !vec_ir_value_find(self->living, value->memberof, NULL))
+                vec_push(self->living, value->memberof);
+            for (mem = 0; mem < 3; ++mem) {
+                if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], NULL))
+                    vec_push(self->living, value->members[mem]);
             }
         }
+
         /* (A) */
         tempbool = ir_block_living_add_instr(self, instr->eid);
         /*con_err( "living added values\n");*/
@@ -2691,7 +2794,8 @@ tailcall:
             stmt.o1.s1 = (target->code_start) - vec_size(code_statements);
             stmt.o2.s1 = 0;
             stmt.o3.s1 = 0;
-            code_push_statement(&stmt, instr->context.line);
+            if (stmt.o1.s1 != 1)
+                code_push_statement(&stmt, instr->context.line);
 
             /* no further instructions can be in this block */
             return true;
@@ -2711,12 +2815,14 @@ tailcall:
             if (ontrue->generated) {
                 stmt.opcode = INSTR_IF;
                 stmt.o2.s1 = (ontrue->code_start) - vec_size(code_statements);
-                code_push_statement(&stmt, instr->context.line);
+                if (stmt.o2.s1 != 1)
+                    code_push_statement(&stmt, instr->context.line);
             }
             if (onfalse->generated) {
                 stmt.opcode = INSTR_IFNOT;
                 stmt.o2.s1 = (onfalse->code_start) - vec_size(code_statements);
-                code_push_statement(&stmt, instr->context.line);
+                if (stmt.o2.s1 != 1)
+                    code_push_statement(&stmt, instr->context.line);
             }
             if (!ontrue->generated) {
                 if (onfalse->generated) {
@@ -2750,6 +2856,12 @@ tailcall:
             if (onfalse->generated) {
                 /* fixup the jump address */
                 code_statements[stidx].o2.s1 = (onfalse->code_start) - (stidx);
+                if (code_statements[stidx].o2.s1 == 1) {
+                    code_statements[stidx] = code_statements[stidx+1];
+                    if (code_statements[stidx].o1.s1 < 0)
+                        code_statements[stidx].o1.s1++;
+                    code_pop_statement();
+                }
                 stmt.opcode = vec_last(code_statements).opcode;
                 if (stmt.opcode == INSTR_GOTO ||
                     stmt.opcode == INSTR_IF ||
@@ -2765,28 +2877,24 @@ tailcall:
                 stmt.o1.s1 = (onfalse->code_start) - vec_size(code_statements);
                 stmt.o2.s1 = 0;
                 stmt.o3.s1 = 0;
-                code_push_statement(&stmt, instr->context.line);
+                if (stmt.o1.s1 != 1)
+                    code_push_statement(&stmt, instr->context.line);
                 return true;
             }
+            else if (code_statements[stidx].o2.s1 == 1) {
+                code_statements[stidx] = code_statements[stidx+1];
+                if (code_statements[stidx].o1.s1 < 0)
+                    code_statements[stidx].o1.s1++;
+                code_pop_statement();
+            }
             /* if not, generate now */
             block = onfalse;
             goto tailcall;
         }
 
-        if (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8) {
-            /* 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... :)
-             */
+        if ( (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8)
+           || instr->opcode == VINSTR_NRCALL)
+        {
             size_t p, first;
             ir_value *retvalue;
 
@@ -2796,6 +2904,8 @@ tailcall:
             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;
@@ -2816,6 +2926,9 @@ tailcall:
                 ir_value *param = instr->params[p];
                 ir_value *targetparam;
 
+                if (param->callparam)
+                    continue;
+
                 if (p-8 >= vec_size(ir->extparams))
                     ir_gen_extparam(ir);
 
@@ -2842,11 +2955,12 @@ tailcall:
             code_push_statement(&stmt, instr->context.line);
 
             retvalue = instr->_ops[0];
-            if (retvalue && retvalue->store != store_return && vec_size(retvalue->life))
+            if (retvalue && retvalue->store != store_return &&
+                (retvalue->store == store_global || vec_size(retvalue->life)))
             {
                 /* not to be kept in OFS_RETURN */
-                if (retvalue->vtype == TYPE_FIELD)
-                    stmt.opcode = field_store_instr[retvalue->vtype];
+                if (retvalue->vtype == TYPE_FIELD && OPTS_FLAG(ADJUST_VECTOR_FIELDS))
+                    stmt.opcode = field_store_instr[retvalue->fieldtype];
                 else
                     stmt.opcode = type_store_instr[retvalue->vtype];
                 stmt.o1.u1 = OFS_RETURN;
@@ -2897,7 +3011,7 @@ tailcall:
             if (stmt.o2.u1 == stmt.o1.u1 &&
                 OPTS_OPTIMIZATION(OPTIM_PEEPHOLE))
             {
-                ++optimization_count[OPTIM_PEEPHOLE];
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
                 continue;
             }
         }
@@ -2910,7 +3024,7 @@ tailcall:
 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.
@@ -2929,12 +3043,22 @@ static bool gen_function_code(ir_function *self)
         return false;
     }
 
-    /* otherwise code_write crashes since it debug-prints functions until AINSTR_END */
-    stmt.opcode = AINSTR_END;
-    stmt.o1.u1 = 0;
-    stmt.o2.u1 = 0;
-    stmt.o3.u1 = 0;
-    code_push_statement(&stmt, vec_last(code_linenums));
+    /* code_write and qcvm -disasm need to know that the function ends here */
+    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;
 }
 
@@ -2963,9 +3087,6 @@ static bool gen_global_function(ir_builder *ir, ir_value *global)
     ir_function          *irfun;
 
     size_t i;
-#ifndef NEW_ALLOC_STRAT
-    size_t local_var_end;
-#endif
 
     if (!global->hasvalue || (!global->constval.vfunc))
     {
@@ -2986,58 +3107,11 @@ static bool gen_global_function(ir_builder *ir, ir_value *global)
         if ((int32_t)i >= fun.nargs)
             fun.argsize[i] = 0;
         else
-            fun.argsize[i] = type_sizeof[irfun->params[i]];
-    }
-
-    fun.firstlocal = vec_size(code_globals);
-
-#ifndef NEW_ALLOC_STRAT
-    local_var_end = fun.firstlocal;
-    for (i = 0; i < vec_size(irfun->locals); ++i) {
-        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;
-        }
-    }
-    if (vec_size(irfun->locals)) {
-        ir_value *last = vec_last(irfun->locals);
-        local_var_end = last->code.globaladdr;
-        if (last->vtype == TYPE_FIELD && last->fieldtype == TYPE_VECTOR)
-            local_var_end += type_sizeof[TYPE_VECTOR];
-        else
-            local_var_end += type_sizeof[last->vtype];
-    }
-    for (i = 0; i < vec_size(irfun->values); ++i)
-    {
-        /* generate code.globaladdr for ssa values */
-        ir_value *v = irfun->values[i];
-        ir_value_code_setaddr(v, local_var_end + v->code.local);
-    }
-    for (i = 0; i < irfun->allocated_locals; ++i) {
-        /* fill the locals with zeros */
-        vec_push(code_globals, 0);
+            fun.argsize[i] = type_sizeof_[irfun->params[i]];
     }
 
-    fun.locals = vec_size(code_globals) - fun.firstlocal;
-#else
-    fun.locals = irfun->allocated_locals;
-    for (i = 0; i < vec_size(irfun->locals); ++i) {
-        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;
-        }
-        ir_value_code_setaddr(irfun->locals[i], fun.firstlocal + irfun->locals[i]->code.local);
-    }
-    for (i = vec_size(code_globals) - fun.firstlocal; i < fun.locals; ++i) {
-        vec_push(code_globals, 0);
-    }
-    for (i = 0; i < vec_size(irfun->values); ++i)
-    {
-        /* generate code.globaladdr for ssa values */
-        ir_value *v = irfun->values[i];
-        ir_value_code_setaddr(v, fun.firstlocal + v->code.local);
-    }
-#endif
+    fun.firstlocal = 0;
+    fun.locals     = irfun->allocated_locals;
 
     if (irfun->builtin)
         fun.entry = irfun->builtin+1;
@@ -3050,16 +3124,29 @@ static bool gen_global_function(ir_builder *ir, ir_value *global)
     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);
 
@@ -3107,6 +3194,52 @@ static bool gen_function_extparam_copy(ir_function *self)
     return true;
 }
 
+static bool gen_function_locals(ir_builder *ir, ir_value *global)
+{
+    prog_section_function *def;
+    ir_function           *irfun;
+    size_t                 i;
+    uint32_t               firstlocal, firstglobal;
+
+    irfun = global->constval.vfunc;
+    def   = code_functions + irfun->code_function_def;
+
+    if (opts.g || !OPTS_OPTIMIZATION(OPTIM_OVERLAP_LOCALS) || (irfun->flags & IR_FLAG_MASK_NO_OVERLAP))
+        firstlocal = def->firstlocal = vec_size(code_globals);
+    else {
+        firstlocal = def->firstlocal = ir->first_common_local;
+        ++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 *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];
+        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 gen_global_function_code(ir_builder *ir, ir_value *global)
 {
     prog_section_function *fundef;
@@ -3134,6 +3267,10 @@ static bool gen_global_function_code(ir_builder *ir, ir_value *global)
     fundef = &code_functions[irfun->code_function_def];
 
     fundef->entry = vec_size(code_statements);
+    if (!gen_function_locals(ir, global)) {
+        irerror(irfun->context, "Failed to generate locals for function %s", irfun->name);
+        return false;
+    }
     if (!gen_function_extparam_copy(irfun)) {
         irerror(irfun->context, "Failed to generate extparam-copy code for function %s", irfun->name);
         return false;
@@ -3145,26 +3282,105 @@ static bool gen_global_function_code(ir_builder *ir, ir_value *global)
     return true;
 }
 
+static void gen_vector_defs(prog_section_def def, const char *name)
+{
+    char  *component;
+    size_t len, i;
+
+    if (!name || name[0] == '#' || OPTS_FLAG(SINGLE_VECTOR_DEFS))
+        return;
+
+    def.type = TYPE_FLOAT;
+
+    len = strlen(name);
+
+    component = (char*)mem_a(len+3);
+    memcpy(component, name, len);
+    len += 2;
+    component[len-0] = 0;
+    component[len-2] = '_';
+
+    component[len-1] = 'x';
+
+    for (i = 0; i < 3; ++i) {
+        def.name = code_genstring(component);
+        vec_push(code_defs, def);
+        def.offset++;
+        component[len-1]++;
+    }
+}
+
+static void gen_vector_fields(prog_section_field fld, const char *name)
+{
+    char  *component;
+    size_t len, i;
+
+    if (!name || OPTS_FLAG(SINGLE_VECTOR_DEFS))
+        return;
+
+    fld.type = TYPE_FLOAT;
+
+    len = strlen(name);
+
+    component = (char*)mem_a(len+3);
+    memcpy(component, name, len);
+    len += 2;
+    component[len-0] = 0;
+    component[len-2] = '_';
+
+    component[len-1] = 'x';
+
+    for (i = 0; i < 3; ++i) {
+        fld.name = code_genstring(component);
+        vec_push(code_fields, fld);
+        fld.offset++;
+        component[len-1]++;
+    }
+}
+
 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;
+
+        if (OPTS_OPTIMIZATION(OPTIM_STRIP_CONSTANT_NAMES) &&
+            (global->name[0] == '#' || global->cvq == CV_CONST))
+        {
+            pushdef = false;
+        }
 
-    if (global->name) {
-        if (global->name[0] == '#') {
-            if (!self->str_immediate)
-                self->str_immediate = code_genstring("IMMEDIATE");
-            def.name = global->code.name = self->str_immediate;
+        if (pushdef && global->name) {
+            if (global->name[0] == '#') {
+                if (!self->str_immediate)
+                    self->str_immediate = code_genstring("IMMEDIATE");
+                def.name = global->code.name = self->str_immediate;
+            }
+            else
+                def.name = global->code.name = code_genstring(global->name);
         }
         else
-            def.name = global->code.name = code_genstring(global->name);
+            def.name   = 0;
+        if (islocal) {
+            def.offset = ir_value_code_addr(global);
+            vec_push(code_defs, def);
+            if (global->vtype == TYPE_VECTOR)
+                gen_vector_defs(def, global->name);
+            else if (global->vtype == TYPE_FIELD && global->fieldtype == TYPE_VECTOR)
+                gen_vector_defs(def, global->name);
+            return true;
+        }
     }
-    else
-        def.name   = 0;
+    if (islocal)
+        return true;
 
     switch (global->vtype)
     {
@@ -3189,13 +3405,17 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
         ir_value_code_setaddr(global, vec_size(code_globals));
         vec_push(code_globals, 0);
         /* Add the def */
-        vec_push(code_defs, def);
+        if (pushdef) vec_push(code_defs, def);
         return true;
     case TYPE_POINTER:
-        vec_push(code_defs, def);
+        if (pushdef) vec_push(code_defs, def);
         return gen_global_pointer(global);
     case TYPE_FIELD:
-        vec_push(code_defs, def);
+        if (pushdef) {
+            vec_push(code_defs, def);
+            if (global->fieldtype == TYPE_VECTOR)
+                gen_vector_defs(def, global->name);
+        }
         return gen_global_field(global);
     case TYPE_ENTITY:
         /* fall through */
@@ -3207,10 +3427,10 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
             vec_push(code_globals, *iptr);
         } else {
             vec_push(code_globals, 0);
-            if (!islocal)
-                def.type |= DEF_SAVEGLOBAL;
         }
-        vec_push(code_defs, def);
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
+        if (pushdef) vec_push(code_defs, def);
 
         return global->code.globaladdr >= 0;
     }
@@ -3221,10 +3441,10 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
             vec_push(code_globals, code_genstring(global->constval.vstring));
         } else {
             vec_push(code_globals, 0);
-            if (!islocal)
-                def.type |= DEF_SAVEGLOBAL;
         }
-        vec_push(code_defs, def);
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
+        if (pushdef) vec_push(code_defs, def);
         return global->code.globaladdr >= 0;
     }
     case TYPE_VECTOR:
@@ -3236,23 +3456,25 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
             vec_push(code_globals, iptr[0]);
             if (global->code.globaladdr < 0)
                 return false;
-            for (d = 1; d < type_sizeof[global->vtype]; ++d)
-            {
+            for (d = 1; d < type_sizeof_[global->vtype]; ++d) {
                 vec_push(code_globals, iptr[d]);
             }
         } else {
             vec_push(code_globals, 0);
             if (global->code.globaladdr < 0)
                 return false;
-            for (d = 1; d < type_sizeof[global->vtype]; ++d)
-            {
+            for (d = 1; d < type_sizeof_[global->vtype]; ++d) {
                 vec_push(code_globals, 0);
             }
-            if (!islocal)
-                def.type |= DEF_SAVEGLOBAL;
         }
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
 
-        vec_push(code_defs, def);
+        if (pushdef) {
+            vec_push(code_defs, def);
+            def.type &= ~DEF_SAVEGLOBAL;
+            gen_vector_defs(def, global->name);
+        }
         return global->code.globaladdr >= 0;
     }
     case TYPE_FUNCTION:
@@ -3265,16 +3487,16 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
             vec_push(code_globals, vec_size(code_functions));
             if (!gen_global_function(self, global))
                 return false;
-            if (!islocal)
-                def.type |= DEF_SAVEGLOBAL;
         }
-        vec_push(code_defs, def);
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
+        if (pushdef) vec_push(code_defs, def);
         return true;
     case TYPE_VARIANT:
         /* assume biggest type */
             ir_value_code_setaddr(global, vec_size(code_globals));
             vec_push(code_globals, 0);
-            for (i = 1; i < type_sizeof[TYPE_VARIANT]; ++i)
+            for (i = 1; i < type_sizeof_[TYPE_VARIANT]; ++i)
                 vec_push(code_globals, 0);
             return true;
     default:
@@ -3287,7 +3509,7 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
 
 static void ir_builder_prepare_field(ir_value *field)
 {
-    field->code.fieldaddr = code_alloc_field(type_sizeof[field->fieldtype]);
+    field->code.fieldaddr = code_alloc_field(type_sizeof_[field->fieldtype]);
 }
 
 static bool ir_builder_gen_field(ir_builder *self, ir_value *field)
@@ -3352,6 +3574,11 @@ static bool ir_builder_gen_field(ir_builder *self, ir_value *field)
         vec_push(code_globals, fld.offset+2);
     }
 
+    if (field->fieldtype == TYPE_VECTOR) {
+        gen_vector_defs(def, field->name);
+        gen_vector_fields(fld, field->name);
+    }
+
     return field->code.globaladdr >= 0;
 }
 
@@ -3359,7 +3586,7 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
 {
     prog_section_statement stmt;
     size_t i;
-    char   *lnofile = NULL;
+    char  *lnofile = NULL;
 
     code_init();
 
@@ -3373,6 +3600,16 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
         if (!ir_builder_gen_global(self, self->globals[i], false)) {
             return false;
         }
+        if (self->globals[i]->vtype == TYPE_FUNCTION) {
+            ir_function *func = self->globals[i]->constval.vfunc;
+            if (func && self->max_locals < func->allocated_locals &&
+                !(func->flags & IR_FLAG_MASK_NO_OVERLAP))
+            {
+                self->max_locals = func->allocated_locals;
+            }
+            if (func && self->max_globaltemps < func->globaltemps)
+                self->max_globaltemps = func->globaltemps;
+        }
     }
 
     for (i = 0; i < vec_size(self->fields); ++i)
@@ -3382,6 +3619,17 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
         }
     }
 
+    /* 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) {
+        vec_push(code_globals, 0);
+    }
+
     /* generate function code */
     for (i = 0; i < vec_size(self->globals); ++i)
     {
@@ -3397,15 +3645,15 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
         return false;
     }
 
-    /* DP errors if the last instruction is not an INSTR_DONE
-     * and for debugging purposes we add an additional AINSTR_END
-     * to the end of functions, so here it goes:
-     */
-    stmt.opcode = INSTR_DONE;
-    stmt.o1.u1 = 0;
-    stmt.o2.u1 = 0;
-    stmt.o3.u1 = 0;
-    code_push_statement(&stmt, vec_last(code_linenums));
+    /* DP errors if the last instruction is not an INSTR_DONE. */
+    if (vec_last(code_statements).opcode != INSTR_DONE)
+    {
+        stmt.opcode = INSTR_DONE;
+        stmt.o1.u1 = 0;
+        stmt.o2.u1 = 0;
+        stmt.o3.u1 = 0;
+        code_push_statement(&stmt, vec_last(code_linenums));
+    }
 
     if (opts.pp_only)
         return true;
@@ -3428,10 +3676,12 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
         memcpy(vec_add(lnofile, 5), ".lno", 5);
     }
 
-    if (lnofile)
-        con_out("writing '%s' and '%s'...\n", filename, lnofile);
-    else
-        con_out("writing '%s'\n", filename);
+    if (!opts.quiet) {
+        if (lnofile)
+            con_out("writing '%s' and '%s'...\n", filename, lnofile);
+        else
+            con_out("writing '%s'\n", filename);
+    }
     if (!code_write(filename, lnofile)) {
         vec_free(lnofile);
         return false;
@@ -3446,8 +3696,8 @@ bool ir_builder_generate(ir_builder *self, const char *filename)
 
 #define IND_BUFSZ 1024
 
-#ifdef WIN32
-# define strncat(dst, src, sz) strncat_s(dst, sz, src, _TRUNCATE)
+#ifdef _MSC_VER
+#   define strncat(dst, src, sz) strncat_s(dst, sz, src, _TRUNCATE)
 #endif
 
 const char *qc_opname(int op)
@@ -3505,13 +3755,36 @@ void ir_function_dump(ir_function *f, char *ind,
     }
     oprintf("%sliferanges:\n", ind);
     for (i = 0; i < vec_size(f->locals); ++i) {
-        size_t l;
+        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);
         }
         oprintf("\n");
+        for (m = 0; m < 3; ++m) {
+            ir_value *vm = v->members[m];
+            if (!vm)
+                continue;
+            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);
+            }
+            oprintf("\n");
+        }
     }
     for (i = 0; i < vec_size(f->values); ++i) {
         size_t l;
@@ -3581,7 +3854,7 @@ void ir_instr_dump(ir_instr *in, char *ind,
         if (in->_ops[1] || in->_ops[2])
             oprintf(" <- ");
     }
-    if (in->opcode == INSTR_CALL0) {
+    if (in->opcode == INSTR_CALL0 || in->opcode == VINSTR_NRCALL) {
         oprintf("CALL%i\t", vec_size(in->params));
     } else
         oprintf("%s\t", qc_opname(in->opcode));