]> git.xonotic.org Git - xonotic/gmqcc.git/blobdiff - ir.c
Remove old tempalloc code; disable -Olocaltemps for now
[xonotic/gmqcc.git] / ir.c
diff --git a/ir.c b/ir.c
index f1ffc7fa4fbaa5e0ca265fd0ce8ed43953620c0b..84832fedf5a5aa7dc3bc4aec35a7dffa2b6a9a8f 100644 (file)
--- 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    */
@@ -208,20 +208,12 @@ static void irerror(lex_ctx ctx, const char *msg, ...)
 
 static bool irwarning(lex_ctx ctx, int warntype, const char *fmt, ...)
 {
-       va_list ap;
-       int lvl = LVL_WARNING;
-
-    if (warntype && !OPTS_WARN(warntype))
-        return false;
-
-    if (opts_werror)
-           lvl = LVL_ERROR;
-
-       va_start(ap, fmt);
-    con_vprintmsg(lvl, ctx.file, ctx.line, "warning", fmt, ap);
-       va_end(ap);
-
-       return opts_werror;
+    bool    r;
+    va_list ap;
+    va_start(ap, fmt);
+    r = vcompile_warning(ctx, warntype, fmt, ap);
+    va_end(ap);
+    return r;
 }
 
 /***********************************************************************
@@ -529,6 +521,123 @@ ir_block* ir_function_create_block(lex_ctx ctx, ir_function *self, const char *l
     return bn;
 }
 
+static bool instr_is_operation(uint16_t op)
+{
+    return ( (op >= INSTR_MUL_F  && op <= INSTR_GT) ||
+             (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) );
+}
+
+bool ir_function_pass_peephole(ir_function *self)
+{
+    size_t b;
+
+    for (b = 0; b < vec_size(self->blocks); ++b) {
+        size_t    i;
+        ir_block *block = self->blocks[b];
+
+        for (i = 0; i < vec_size(block->instr); ++i) {
+            ir_instr *inst;
+            inst = block->instr[i];
+
+            if (i >= 1 &&
+                (inst->opcode >= INSTR_STORE_F &&
+                 inst->opcode <= INSTR_STORE_FNC))
+            {
+                ir_instr *store;
+                ir_instr *oper;
+                ir_value *value;
+
+                store = inst;
+
+                oper  = block->instr[i-1];
+                if (!instr_is_operation(oper->opcode))
+                    continue;
+
+                value = oper->_ops[0];
+
+                /* only do it for SSA values */
+                if (value->store != store_value)
+                    continue;
+
+                /* don't optimize out the temp if it's used later again */
+                if (vec_size(value->reads) != 1)
+                    continue;
+
+                /* The very next store must use this value */
+                if (value->reads[0] != store)
+                    continue;
+
+                /* And of course the store must _read_ from it, so it's in
+                 * OP 1 */
+                if (store->_ops[1] != value)
+                    continue;
+
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
+                (void)!ir_instr_op(oper, 0, store->_ops[0], true);
+
+                vec_remove(block->instr, i, 1);
+                ir_instr_delete(store);
+            }
+            else if (inst->opcode == VINSTR_COND)
+            {
+                /* COND on a value resulting from a NOT could
+                 * remove the NOT and swap its operands
+                 */
+                while (true) {
+                    ir_block *tmp;
+                    size_t    inotid;
+                    ir_instr *inot;
+                    ir_value *value;
+                    value = inst->_ops[0];
+
+                    if (value->store != store_value ||
+                        vec_size(value->reads) != 1 ||
+                        value->reads[0] != inst)
+                    {
+                        break;
+                    }
+
+                    inot = value->writes[0];
+                    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 */
+                    {
+                        break;
+                    }
+
+                    /* count */
+                    ++opts_optimizationcount[OPTIM_PEEPHOLE];
+                    /* change operand */
+                    (void)!ir_instr_op(inst, 0, inot->_ops[1], false);
+                    /* remove NOT */
+                    tmp = inot->owner;
+                    for (inotid = 0; inotid < vec_size(tmp->instr); ++inotid) {
+                        if (tmp->instr[inotid] == inot)
+                            break;
+                    }
+                    if (inotid >= vec_size(tmp->instr)) {
+                        compile_error(inst->context, "sanity-check failed: failed to find instruction to optimize out");
+                        return false;
+                    }
+                    vec_remove(tmp->instr, inotid, 1);
+                    ir_instr_delete(inot);
+                    /* swap ontrue/onfalse */
+                    tmp = inst->bops[0];
+                    inst->bops[0] = inst->bops[1];
+                    inst->bops[1] = tmp;
+                }
+                continue;
+            }
+        }
+    }
+
+    return true;
+}
+
 bool ir_function_pass_tailcall(ir_function *self)
 {
     size_t b, p;
@@ -569,7 +678,7 @@ bool ir_function_pass_tailcall(ir_function *self)
                 ret->_ops[0]   == store->_ops[0] &&
                 store->_ops[1] == call->_ops[0])
             {
-                ++optimization_count[OPTIM_MINOR];
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
                 call->_ops[0] = store->_ops[0];
                 vec_remove(block->instr, vec_size(block->instr) - 2, 1);
                 ir_instr_delete(store);
@@ -591,7 +700,7 @@ 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 */
@@ -621,6 +730,13 @@ bool ir_function_finalize(ir_function *self)
     if (self->builtin)
         return true;
 
+    if (OPTS_OPTIMIZATION(OPTIM_PEEPHOLE)) {
+        if (!ir_function_pass_peephole(self)) {
+            irerror(self->context, "generic optimization pass broke something in `%s`", self->name);
+            return false;
+        }
+    }
+
     if (OPTS_OPTIMIZATION(OPTIM_TAIL_RECURSION)) {
         if (!ir_function_pass_tailcall(self)) {
             irerror(self->context, "tailcall optimization pass broke something in `%s`", self->name);
@@ -864,12 +980,16 @@ ir_value* ir_value_var(const char *name, int storetype, int vtype)
     self->members[2] = NULL;
     self->memberof = NULL;
 
+    self->unique_life = false;
+
     self->life = NULL;
     return self;
 }
 
 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;
@@ -877,9 +997,16 @@ ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
     if (self->members[member])
         return self->members[member];
 
+    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';
     if (self->vtype == TYPE_VECTOR)
     {
-        m = ir_value_var(self->name, self->store, TYPE_FLOAT);
+        m = ir_value_var(name, self->store, TYPE_FLOAT);
+        mem_d(name);
         if (!m)
             return NULL;
         m->context = self->context;
@@ -891,7 +1018,8 @@ 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);
+        mem_d(name);
         if (!m)
             return NULL;
         m->fieldtype = TYPE_FLOAT;
@@ -910,6 +1038,13 @@ 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);
@@ -1232,15 +1367,19 @@ 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 &&
@@ -1252,9 +1391,14 @@ bool ir_block_create_store_op(ir_block *self, lex_ctx ctx, int op, ir_value *tar
         return false;
     }
 
+    in = ir_instr_new(ctx, self, op);
+    if (!in)
+        return false;
+
     if (!ir_instr_op(in, 0, target, true) ||
         !ir_instr_op(in, 1, what, false))
     {
+        ir_instr_delete(in);
         return false;
     }
     vec_push(self->instr, in);
@@ -1311,18 +1455,18 @@ 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;
@@ -1332,10 +1476,8 @@ 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);
@@ -1362,10 +1504,8 @@ 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)
@@ -1382,10 +1522,8 @@ 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);
+    if (!ir_check_unreachable(self))
         return false;
-    }
     self->final = true;
     in = ir_instr_new(ctx, self, INSTR_GOTO);
     if (!in)
@@ -1403,6 +1541,8 @@ ir_instr* ir_block_create_phi(ir_block *self, lex_ctx ctx, const char *label, in
 {
     ir_value *out;
     ir_instr *in;
+    if (!ir_check_unreachable(self))
+        return false;
     in = ir_instr_new(ctx, self, VINSTR_PHI);
     if (!in)
         return NULL;
@@ -1444,13 +1584,19 @@ 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 false;
+    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);
@@ -1464,6 +1610,15 @@ 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;
 }
 
@@ -2051,8 +2206,24 @@ 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->members[0] && v->members[1] && v->members[2])) {
+                /* all vector members have been accessed - only treat this as uninitialized
+                 * if any of them is also uninitialized.
+                 */
+                if (!vec_ir_value_find(block->living, v->members[0], NULL) &&
+                    !vec_ir_value_find(block->living, v->members[1], NULL) &&
+                    !vec_ir_value_find(block->living, v->members[2], NULL))
+                {
+                    continue;
+                }
+            }
+            if (v->memberof) {
+                /* A member is only uninitialized if the whole vector is also uninitialized */
+                if (!vec_ir_value_find(block->living, v->memberof, NULL))
+                    continue;
+            }
             if (irwarning(v->context, WARN_USED_UNINITIALIZED,
                           "variable `%s` may be used uninitialized in this function", v->name))
             {
@@ -2072,12 +2243,13 @@ typedef struct {
     ir_value **locals;
     size_t    *sizes;
     size_t    *positions;
+    bool      *unique;
 } function_allocator;
 
 static bool function_allocator_alloc(function_allocator *alloc, const ir_value *var)
 {
     ir_value *slot;
-    size_t vsize = type_sizeof[var->vtype];
+    size_t vsize = ir_value_sizeof(var);
 
     slot = ir_value_var("reg", store_global, var->vtype);
     if (!slot)
@@ -2088,6 +2260,7 @@ static bool function_allocator_alloc(function_allocator *alloc, const ir_value *
 
     vec_push(alloc->locals, slot);
     vec_push(alloc->sizes, vsize);
+    vec_push(alloc->unique, var->unique_life);
 
     return true;
 
@@ -2113,9 +2286,14 @@ bool ir_function_allocate_locals(ir_function *self)
     alloc.locals    = NULL;
     alloc.sizes     = NULL;
     alloc.positions = NULL;
+    alloc.unique    = NULL;
 
     for (i = 0; i < vec_size(self->locals); ++i)
     {
+#if 0
+        if (!OPTS_OPTIMIZATION(OPTIM_LOCALTEMPS))
+#endif
+            self->locals[i]->unique_life = true;
         if (!function_allocator_alloc(&alloc, self->locals[i]))
             goto error;
     }
@@ -2130,8 +2308,21 @@ bool ir_function_allocate_locals(ir_function *self)
 
         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;
 
@@ -2139,8 +2330,8 @@ bool ir_function_allocate_locals(ir_function *self)
                 goto error;
 
             /* adjust size for this slot */
-            if (alloc.sizes[a] < type_sizeof[v->vtype])
-                alloc.sizes[a] = type_sizeof[v->vtype];
+            if (alloc.sizes[a] < ir_value_sizeof(v))
+                alloc.sizes[a] = ir_value_sizeof(v);
 
             self->values[i]->code.local = a;
             break;
@@ -2171,7 +2362,11 @@ bool ir_function_allocate_locals(ir_function *self)
 
     self->allocated_locals = pos + vec_last(alloc.sizes);
 
-    /* Take over the actual slot positions */
+    /* Locals need to know their new position */
+    for (i = 0; i < vec_size(self->locals); ++i) {
+        self->locals[i]->code.local = alloc.positions[i];
+    }
+    /* 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];
     }
@@ -2183,6 +2378,7 @@ error:
 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);
@@ -2256,7 +2452,6 @@ static bool ir_block_life_prop_previous(ir_block* self, ir_block *prev, bool *ch
      * So we have to remove whatever does not exist in the previous block.
      * They will be re-added on-read, but the liferange merge won't cause
      * a change.
-     */
     for (i = 0; i < vec_size(self->living); ++i)
     {
         if (!vec_ir_value_find(prev->living, self->living[i], NULL)) {
@@ -2264,6 +2459,7 @@ static bool ir_block_life_prop_previous(ir_block* self, ir_block *prev, bool *ch
             --i;
         }
     }
+     */
 
     /* Whatever the previous block still has in its living set
      * must now be added to ours as well.
@@ -2306,8 +2502,6 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
         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);
         }
@@ -2316,8 +2510,6 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
         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);
         }
@@ -2345,8 +2537,6 @@ static bool ir_block_life_propagate(ir_block *self, ir_block *prev, bool *change
                 continue;
 
             value = instr->_ops[o];
-            if (value->memberof)
-                value = value->memberof;
 
             /* We only care about locals */
             /* we also calculate parameter liferanges so that locals
@@ -2551,7 +2741,8 @@ 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;
@@ -2571,12 +2762,14 @@ 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) {
@@ -2610,6 +2803,12 @@ 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 ||
@@ -2625,15 +2824,24 @@ 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) {
+        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,
@@ -2750,6 +2958,16 @@ tailcall:
             /* 2-operand instructions with A -> B */
             stmt.o2.u1 = stmt.o3.u1;
             stmt.o3.u1 = 0;
+
+            /* tiny optimization, don't output
+             * STORE a, a
+             */
+            if (stmt.o2.u1 == stmt.o1.u1 &&
+                OPTS_OPTIMIZATION(OPTIM_PEEPHOLE))
+            {
+                ++opts_optimizationcount[OPTIM_PEEPHOLE];
+                continue;
+            }
         }
 
         code_push_statement(&stmt, instr->context.line);
@@ -2779,8 +2997,8 @@ 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;
+    /* 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;
@@ -2813,7 +3031,6 @@ static bool gen_global_function(ir_builder *ir, ir_value *global)
     ir_function          *irfun;
 
     size_t i;
-    size_t local_var_end;
 
     if (!global->hasvalue || (!global->constval.vfunc))
     {
@@ -2834,35 +3051,27 @@ 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.argsize[i] = type_sizeof_[irfun->params[i]];
     }
 
     fun.firstlocal = vec_size(code_globals);
 
-    local_var_end = fun.firstlocal;
+    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;
         }
-    }
-    if (vec_size(irfun->locals)) {
-        ir_value *last = vec_last(irfun->locals);
-        local_var_end = last->code.globaladdr;
-        local_var_end += type_sizeof[last->vtype];
+        ir_value_code_setaddr(irfun->locals[i], fun.firstlocal + irfun->locals[i]->code.local);
     }
     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);
+        ir_value_code_setaddr(v, fun.firstlocal + v->code.local);
     }
-    for (i = 0; i < irfun->allocated_locals; ++i) {
-        /* fill the locals with zeros */
+    for (i = vec_size(code_globals); i < fun.firstlocal + irfun->allocated_locals; ++i)
         vec_push(code_globals, 0);
-    }
-
-    fun.locals = vec_size(code_globals) - fun.firstlocal;
 
     if (irfun->builtin)
         fun.entry = irfun->builtin+1;
@@ -2970,6 +3179,62 @@ 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 || 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;
@@ -3021,6 +3286,7 @@ static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool isloc
         return gen_global_pointer(global);
     case TYPE_FIELD:
         vec_push(code_defs, def);
+        gen_vector_defs(def, global->name);
         return gen_global_field(global);
     case TYPE_ENTITY:
         /* fall through */
@@ -3032,9 +3298,9 @@ 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;
         }
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
         vec_push(code_defs, def);
 
         return global->code.globaladdr >= 0;
@@ -3046,9 +3312,9 @@ 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;
         }
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
         vec_push(code_defs, def);
         return global->code.globaladdr >= 0;
     }
@@ -3061,23 +3327,23 @@ 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);
+        def.type &= ~DEF_SAVEGLOBAL;
+        gen_vector_defs(def, global->name);
         return global->code.globaladdr >= 0;
     }
     case TYPE_FUNCTION:
@@ -3090,16 +3356,16 @@ 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;
         }
+        if (!islocal && global->cvq != CV_CONST)
+            def.type |= DEF_SAVEGLOBAL;
         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:
@@ -3112,7 +3378,7 @@ 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)
@@ -3126,7 +3392,7 @@ static bool ir_builder_gen_field(ir_builder *self, ir_value *field)
     def.offset = (uint16_t)vec_size(code_globals);
 
     /* create a global named the same as the field */
-    if (opts_standard == COMPILER_GMQCC) {
+    if (opts.standard == COMPILER_GMQCC) {
         /* in our standard, the global gets a dot prefix */
         size_t len = strlen(field->name);
         char name[1024];
@@ -3177,6 +3443,11 @@ 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;
 }
 
@@ -3222,17 +3493,17 @@ 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)
+    if (opts.pp_only)
         return true;
 
     if (vec_size(code_statements) != vec_size(code_linenums)) {
@@ -3271,8 +3542,8 @@ 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)
@@ -3332,7 +3603,7 @@ void ir_function_dump(ir_function *f, char *ind,
     for (i = 0; i < vec_size(f->locals); ++i) {
         size_t l;
         ir_value *v = f->locals[i];
-        oprintf("%s\t%s: unique ", ind, v->name);
+        oprintf("%s\t%s: %s@%i ", ind, v->name, (v->unique_life ? "unique " : ""), (int)v->code.local);
         for (l = 0; l < vec_size(v->life); ++l) {
             oprintf("[%i,%i] ", v->life[l].start, v->life[l].end);
         }
@@ -3406,7 +3677,7 @@ 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));