more c++ migration for ast/ir/code; reached a working condition here
[xonotic/gmqcc.git] / ir.cpp
1 #include <stdlib.h>
2 #include <string.h>
3
4 #include "gmqcc.h"
5 #include "ir.h"
6
7 /***********************************************************************
8  * Type sizes used at multiple points in the IR codegen
9  */
10
11 const char *type_name[TYPE_COUNT] = {
12     "void",
13     "string",
14     "float",
15     "vector",
16     "entity",
17     "field",
18     "function",
19     "pointer",
20     "integer",
21     "variant",
22     "struct",
23     "union",
24     "array",
25
26     "nil",
27     "<no-expression>"
28 };
29
30 static size_t type_sizeof_[TYPE_COUNT] = {
31     1, /* TYPE_VOID     */
32     1, /* TYPE_STRING   */
33     1, /* TYPE_FLOAT    */
34     3, /* TYPE_VECTOR   */
35     1, /* TYPE_ENTITY   */
36     1, /* TYPE_FIELD    */
37     1, /* TYPE_FUNCTION */
38     1, /* TYPE_POINTER  */
39     1, /* TYPE_INTEGER  */
40     3, /* TYPE_VARIANT  */
41     0, /* TYPE_STRUCT   */
42     0, /* TYPE_UNION    */
43     0, /* TYPE_ARRAY    */
44     0, /* TYPE_NIL      */
45     0, /* TYPE_NOESPR   */
46 };
47
48 const uint16_t type_store_instr[TYPE_COUNT] = {
49     INSTR_STORE_F, /* should use I when having integer support */
50     INSTR_STORE_S,
51     INSTR_STORE_F,
52     INSTR_STORE_V,
53     INSTR_STORE_ENT,
54     INSTR_STORE_FLD,
55     INSTR_STORE_FNC,
56     INSTR_STORE_ENT, /* should use I */
57 #if 0
58     INSTR_STORE_I, /* integer type */
59 #else
60     INSTR_STORE_F,
61 #endif
62
63     INSTR_STORE_V, /* variant, should never be accessed */
64
65     VINSTR_END, /* struct */
66     VINSTR_END, /* union  */
67     VINSTR_END, /* array  */
68     VINSTR_END, /* nil    */
69     VINSTR_END, /* noexpr */
70 };
71
72 const uint16_t field_store_instr[TYPE_COUNT] = {
73     INSTR_STORE_FLD,
74     INSTR_STORE_FLD,
75     INSTR_STORE_FLD,
76     INSTR_STORE_V,
77     INSTR_STORE_FLD,
78     INSTR_STORE_FLD,
79     INSTR_STORE_FLD,
80     INSTR_STORE_FLD,
81 #if 0
82     INSTR_STORE_FLD, /* integer type */
83 #else
84     INSTR_STORE_FLD,
85 #endif
86
87     INSTR_STORE_V, /* variant, should never be accessed */
88
89     VINSTR_END, /* struct */
90     VINSTR_END, /* union  */
91     VINSTR_END, /* array  */
92     VINSTR_END, /* nil    */
93     VINSTR_END, /* noexpr */
94 };
95
96 const uint16_t type_storep_instr[TYPE_COUNT] = {
97     INSTR_STOREP_F, /* should use I when having integer support */
98     INSTR_STOREP_S,
99     INSTR_STOREP_F,
100     INSTR_STOREP_V,
101     INSTR_STOREP_ENT,
102     INSTR_STOREP_FLD,
103     INSTR_STOREP_FNC,
104     INSTR_STOREP_ENT, /* should use I */
105 #if 0
106     INSTR_STOREP_ENT, /* integer type */
107 #else
108     INSTR_STOREP_F,
109 #endif
110
111     INSTR_STOREP_V, /* variant, should never be accessed */
112
113     VINSTR_END, /* struct */
114     VINSTR_END, /* union  */
115     VINSTR_END, /* array  */
116     VINSTR_END, /* nil    */
117     VINSTR_END, /* noexpr */
118 };
119
120 const uint16_t type_eq_instr[TYPE_COUNT] = {
121     INSTR_EQ_F, /* should use I when having integer support */
122     INSTR_EQ_S,
123     INSTR_EQ_F,
124     INSTR_EQ_V,
125     INSTR_EQ_E,
126     INSTR_EQ_E, /* FLD has no comparison */
127     INSTR_EQ_FNC,
128     INSTR_EQ_E, /* should use I */
129 #if 0
130     INSTR_EQ_I,
131 #else
132     INSTR_EQ_F,
133 #endif
134
135     INSTR_EQ_V, /* variant, should never be accessed */
136
137     VINSTR_END, /* struct */
138     VINSTR_END, /* union  */
139     VINSTR_END, /* array  */
140     VINSTR_END, /* nil    */
141     VINSTR_END, /* noexpr */
142 };
143
144 const uint16_t type_ne_instr[TYPE_COUNT] = {
145     INSTR_NE_F, /* should use I when having integer support */
146     INSTR_NE_S,
147     INSTR_NE_F,
148     INSTR_NE_V,
149     INSTR_NE_E,
150     INSTR_NE_E, /* FLD has no comparison */
151     INSTR_NE_FNC,
152     INSTR_NE_E, /* should use I */
153 #if 0
154     INSTR_NE_I,
155 #else
156     INSTR_NE_F,
157 #endif
158
159     INSTR_NE_V, /* variant, should never be accessed */
160
161     VINSTR_END, /* struct */
162     VINSTR_END, /* union  */
163     VINSTR_END, /* array  */
164     VINSTR_END, /* nil    */
165     VINSTR_END, /* noexpr */
166 };
167
168 const uint16_t type_not_instr[TYPE_COUNT] = {
169     INSTR_NOT_F, /* should use I when having integer support */
170     VINSTR_END,  /* not to be used, depends on string related -f flags */
171     INSTR_NOT_F,
172     INSTR_NOT_V,
173     INSTR_NOT_ENT,
174     INSTR_NOT_ENT,
175     INSTR_NOT_FNC,
176     INSTR_NOT_ENT, /* should use I */
177 #if 0
178     INSTR_NOT_I, /* integer type */
179 #else
180     INSTR_NOT_F,
181 #endif
182
183     INSTR_NOT_V, /* variant, should never be accessed */
184
185     VINSTR_END, /* struct */
186     VINSTR_END, /* union  */
187     VINSTR_END, /* array  */
188     VINSTR_END, /* nil    */
189     VINSTR_END, /* noexpr */
190 };
191
192 /* protos */
193 static void            ir_value_dump(ir_value*, int (*oprintf)(const char*,...));
194
195 static ir_value*       ir_gen_extparam_proto(ir_builder *ir);
196 static void            ir_gen_extparam      (ir_builder *ir);
197
198 static void            ir_function_dump(ir_function*, char *ind, int (*oprintf)(const char*,...));
199
200 static ir_value*       ir_block_create_general_instr(ir_block *self, lex_ctx_t, const char *label,
201                                                      int op, ir_value *a, ir_value *b, qc_type outype);
202 static bool GMQCC_WARN ir_block_create_store(ir_block*, lex_ctx_t, ir_value *target, ir_value *what);
203 static void            ir_block_dump(ir_block*, char *ind, int (*oprintf)(const char*,...));
204
205 static bool            ir_instr_op(ir_instr*, int op, ir_value *value, bool writing);
206 static void            ir_instr_dump(ir_instr* in, char *ind, int (*oprintf)(const char*,...));
207 /* error functions */
208
209 static void irerror(lex_ctx_t ctx, const char *msg, ...)
210 {
211     va_list ap;
212     va_start(ap, msg);
213     con_cvprintmsg(ctx, LVL_ERROR, "internal error", msg, ap);
214     va_end(ap);
215 }
216
217 static bool GMQCC_WARN irwarning(lex_ctx_t ctx, int warntype, const char *fmt, ...)
218 {
219     bool    r;
220     va_list ap;
221     va_start(ap, fmt);
222     r = vcompile_warning(ctx, warntype, fmt, ap);
223     va_end(ap);
224     return r;
225 }
226
227 /***********************************************************************
228  * Vector utility functions
229  */
230
231 static bool GMQCC_WARN vec_ir_value_find(std::vector<ir_value *> &vec, const ir_value *what, size_t *idx)
232 {
233     for (auto &it : vec) {
234         if (it != what)
235             continue;
236         if (idx)
237             *idx = &it - &vec[0];
238         return true;
239     }
240     return false;
241 }
242
243 static bool GMQCC_WARN vec_ir_block_find(ir_block **vec, ir_block *what, size_t *idx)
244 {
245     size_t i;
246     size_t len = vec_size(vec);
247     for (i = 0; i < len; ++i) {
248         if (vec[i] == what) {
249             if (idx) *idx = i;
250             return true;
251         }
252     }
253     return false;
254 }
255
256 static bool GMQCC_WARN vec_ir_instr_find(std::vector<ir_instr *> &vec, ir_instr *what, size_t *idx)
257 {
258     for (auto &it : vec) {
259         if (it != what)
260             continue;
261         if (idx)
262             *idx = &it - &vec[0];
263         return true;
264     }
265     return false;
266 }
267
268 /***********************************************************************
269  * IR Builder
270  */
271
272 static void ir_block_delete_quick(ir_block* self);
273 static void ir_instr_delete_quick(ir_instr *self);
274 static void ir_function_delete_quick(ir_function *self);
275
276 void* ir_builder::operator new(std::size_t bytes)
277 {
278     return mem_a(bytes);
279 }
280
281 void ir_builder::operator delete(void *ptr)
282 {
283     mem_d(ptr);
284 }
285
286 ir_builder::ir_builder(const std::string& modulename)
287 : name(modulename),
288   code(new code_t)
289 {
290     htglobals   = util_htnew(IR_HT_SIZE);
291     htfields    = util_htnew(IR_HT_SIZE);
292     htfunctions = util_htnew(IR_HT_SIZE);
293
294     nil = new ir_value("nil", store_value, TYPE_NIL);
295     nil->cvq = CV_CONST;
296
297     for (size_t i = 0; i != IR_MAX_VINSTR_TEMPS; ++i) {
298         /* we write to them, but they're not supposed to be used outside the IR, so
299          * let's not allow the generation of ir_instrs which use these.
300          * So it's a constant noexpr.
301          */
302         vinstr_temp[i] = new ir_value("vinstr_temp", store_value, TYPE_NOEXPR);
303         vinstr_temp[i]->cvq = CV_CONST;
304     }
305 }
306
307 ir_builder::~ir_builder()
308 {
309     util_htdel(htglobals);
310     util_htdel(htfields);
311     util_htdel(htfunctions);
312     for (auto& f : functions)
313         ir_function_delete_quick(f.release());
314     functions.clear(); // delete them now before deleting the rest:
315
316     delete nil;
317
318     for (size_t i = 0; i != IR_MAX_VINSTR_TEMPS; ++i) {
319         delete vinstr_temp[i];
320     }
321
322     extparams.clear();
323     extparam_protos.clear();
324 }
325
326 static ir_function* ir_builder_get_function(ir_builder *self, const char *name)
327 {
328     return (ir_function*)util_htget(self->htfunctions, name);
329 }
330
331 ir_function* ir_builder_create_function(ir_builder *self, const std::string& name, qc_type outtype)
332 {
333     ir_function *fn = ir_builder_get_function(self, name.c_str());
334     if (fn) {
335         return nullptr;
336     }
337
338     fn = new ir_function(self, outtype);
339     fn->name = name;
340     self->functions.emplace_back(fn);
341     util_htset(self->htfunctions, name.c_str(), fn);
342
343     fn->value = ir_builder_create_global(self, fn->name, TYPE_FUNCTION);
344     if (!fn->value) {
345         delete fn;
346         return nullptr;
347     }
348
349     fn->value->hasvalue = true;
350     fn->value->outtype = outtype;
351     fn->value->constval.vfunc = fn;
352     fn->value->context = fn->context;
353
354     return fn;
355 }
356
357 static ir_value* ir_builder_get_global(ir_builder *self, const char *name)
358 {
359     return (ir_value*)util_htget(self->htglobals, name);
360 }
361
362 ir_value* ir_builder_create_global(ir_builder *self, const std::string& name, qc_type vtype)
363 {
364     ir_value *ve;
365
366     if (name[0] != '#')
367     {
368         ve = ir_builder_get_global(self, name.c_str());
369         if (ve) {
370             return nullptr;
371         }
372     }
373
374     ve = new ir_value(std::string(name), store_global, vtype);
375     self->globals.emplace_back(ve);
376     util_htset(self->htglobals, name.c_str(), ve);
377     return ve;
378 }
379
380 ir_value* ir_builder_get_va_count(ir_builder *self)
381 {
382     if (self->reserved_va_count)
383         return self->reserved_va_count;
384     return (self->reserved_va_count = ir_builder_create_global(self, "reserved:va_count", TYPE_FLOAT));
385 }
386
387 static ir_value* ir_builder_get_field(ir_builder *self, const char *name)
388 {
389     return (ir_value*)util_htget(self->htfields, name);
390 }
391
392
393 ir_value* ir_builder_create_field(ir_builder *self, const std::string& name, qc_type vtype)
394 {
395     ir_value *ve = ir_builder_get_field(self, name.c_str());
396     if (ve) {
397         return nullptr;
398     }
399
400     ve = new ir_value(std::string(name), store_global, TYPE_FIELD);
401     ve->fieldtype = vtype;
402     self->fields.emplace_back(ve);
403     util_htset(self->htfields, name.c_str(), ve);
404     return ve;
405 }
406
407 /***********************************************************************
408  *IR Function
409  */
410
411 static bool ir_function_naive_phi(ir_function*);
412 static void ir_function_enumerate(ir_function*);
413 static bool ir_function_calculate_liferanges(ir_function*);
414 static bool ir_function_allocate_locals(ir_function*);
415
416 void* ir_function::operator new(std::size_t bytes)
417 {
418     return mem_a(bytes);
419 }
420
421 void ir_function::operator delete(void *ptr)
422 {
423     mem_d(ptr);
424 }
425
426 ir_function::ir_function(ir_builder* owner_, qc_type outtype_)
427 : owner(owner_),
428   name("<@unnamed>"),
429   outtype(outtype_)
430 {
431     owner = owner;
432     context.file = "<@no context>";
433     context.line = 0;
434     outtype = outtype;
435 }
436
437 ir_function::~ir_function()
438 {
439 }
440
441 static void ir_function_delete_quick(ir_function *self)
442 {
443     for (auto& b : self->blocks)
444         ir_block_delete_quick(b.release());
445     delete self;
446 }
447
448 static void ir_function_collect_value(ir_function *self, ir_value *v)
449 {
450     self->values.emplace_back(v);
451 }
452
453 ir_block* ir_function_create_block(lex_ctx_t ctx, ir_function *self, const char *label)
454 {
455     ir_block* bn = new ir_block(self, label ? std::string(label) : std::string());
456     bn->context = ctx;
457     self->blocks.emplace_back(bn);
458
459     if ((self->flags & IR_FLAG_BLOCK_COVERAGE) && self->owner->coverage_func)
460         (void)ir_block_create_call(bn, ctx, nullptr, self->owner->coverage_func, false);
461
462     return bn;
463 }
464
465 static bool instr_is_operation(uint16_t op)
466 {
467     return ( (op >= INSTR_MUL_F  && op <= INSTR_GT) ||
468              (op >= INSTR_LOAD_F && op <= INSTR_LOAD_FNC) ||
469              (op == INSTR_ADDRESS) ||
470              (op >= INSTR_NOT_F  && op <= INSTR_NOT_FNC) ||
471              (op >= INSTR_AND    && op <= INSTR_BITOR) ||
472              (op >= INSTR_CALL0  && op <= INSTR_CALL8) ||
473              (op >= VINSTR_BITAND_V && op <= VINSTR_NEG_V) );
474 }
475
476 static bool ir_function_pass_peephole(ir_function *self)
477 {
478     for (auto& bp : self->blocks) {
479         ir_block *block = bp.get();
480         for (size_t i = 0; i < vec_size(block->instr); ++i) {
481             ir_instr *inst;
482             inst = block->instr[i];
483
484             if (i >= 1 &&
485                 (inst->opcode >= INSTR_STORE_F &&
486                  inst->opcode <= INSTR_STORE_FNC))
487             {
488                 ir_instr *store;
489                 ir_instr *oper;
490                 ir_value *value;
491
492                 store = inst;
493
494                 oper  = block->instr[i-1];
495                 if (!instr_is_operation(oper->opcode))
496                     continue;
497
498                 /* Don't change semantics of MUL_VF in engines where these may not alias. */
499                 if (OPTS_FLAG(LEGACY_VECTOR_MATHS)) {
500                     if (oper->opcode == INSTR_MUL_VF && oper->_ops[2]->memberof == oper->_ops[1])
501                         continue;
502                     if (oper->opcode == INSTR_MUL_FV && oper->_ops[1]->memberof == oper->_ops[2])
503                         continue;
504                 }
505
506                 value = oper->_ops[0];
507
508                 /* only do it for SSA values */
509                 if (value->store != store_value)
510                     continue;
511
512                 /* don't optimize out the temp if it's used later again */
513                 if (value->reads.size() != 1)
514                     continue;
515
516                 /* The very next store must use this value */
517                 if (value->reads[0] != store)
518                     continue;
519
520                 /* And of course the store must _read_ from it, so it's in
521                  * OP 1 */
522                 if (store->_ops[1] != value)
523                     continue;
524
525                 ++opts_optimizationcount[OPTIM_PEEPHOLE];
526                 (void)!ir_instr_op(oper, 0, store->_ops[0], true);
527
528                 vec_remove(block->instr, i, 1);
529                 delete store;
530             }
531             else if (inst->opcode == VINSTR_COND)
532             {
533                 /* COND on a value resulting from a NOT could
534                  * remove the NOT and swap its operands
535                  */
536                 while (true) {
537                     ir_block *tmp;
538                     size_t    inotid;
539                     ir_instr *inot;
540                     ir_value *value;
541                     value = inst->_ops[0];
542
543                     if (value->store != store_value || value->reads.size() != 1 || value->reads[0] != inst)
544                         break;
545
546                     inot = value->writes[0];
547                     if (inot->_ops[0] != value ||
548                         inot->opcode < INSTR_NOT_F ||
549                         inot->opcode > INSTR_NOT_FNC ||
550                         inot->opcode == INSTR_NOT_V || /* can't do these */
551                         inot->opcode == INSTR_NOT_S)
552                     {
553                         break;
554                     }
555
556                     /* count */
557                     ++opts_optimizationcount[OPTIM_PEEPHOLE];
558                     /* change operand */
559                     (void)!ir_instr_op(inst, 0, inot->_ops[1], false);
560                     /* remove NOT */
561                     tmp = inot->owner;
562                     for (inotid = 0; inotid < vec_size(tmp->instr); ++inotid) {
563                         if (tmp->instr[inotid] == inot)
564                             break;
565                     }
566                     if (inotid >= vec_size(tmp->instr)) {
567                         compile_error(inst->context, "sanity-check failed: failed to find instruction to optimize out");
568                         return false;
569                     }
570                     vec_remove(tmp->instr, inotid, 1);
571                     delete inot;
572                     /* swap ontrue/onfalse */
573                     tmp = inst->bops[0];
574                     inst->bops[0] = inst->bops[1];
575                     inst->bops[1] = tmp;
576                 }
577                 continue;
578             }
579         }
580     }
581
582     return true;
583 }
584
585 static bool ir_function_pass_tailrecursion(ir_function *self)
586 {
587     size_t p;
588
589     for (auto& bp : self->blocks) {
590         ir_block *block = bp.get();
591
592         ir_value *funcval;
593         ir_instr *ret, *call, *store = nullptr;
594
595         if (!block->final || vec_size(block->instr) < 2)
596             continue;
597
598         ret = block->instr[vec_size(block->instr)-1];
599         if (ret->opcode != INSTR_DONE && ret->opcode != INSTR_RETURN)
600             continue;
601
602         call = block->instr[vec_size(block->instr)-2];
603         if (call->opcode >= INSTR_STORE_F && call->opcode <= INSTR_STORE_FNC) {
604             /* account for the unoptimized
605              * CALL
606              * STORE %return, %tmp
607              * RETURN %tmp
608              * version
609              */
610             if (vec_size(block->instr) < 3)
611                 continue;
612
613             store = call;
614             call = block->instr[vec_size(block->instr)-3];
615         }
616
617         if (call->opcode < INSTR_CALL0 || call->opcode > INSTR_CALL8)
618             continue;
619
620         if (store) {
621             /* optimize out the STORE */
622             if (ret->_ops[0]   &&
623                 ret->_ops[0]   == store->_ops[0] &&
624                 store->_ops[1] == call->_ops[0])
625             {
626                 ++opts_optimizationcount[OPTIM_PEEPHOLE];
627                 call->_ops[0] = store->_ops[0];
628                 vec_remove(block->instr, vec_size(block->instr) - 2, 1);
629                 delete store;
630             }
631             else
632                 continue;
633         }
634
635         if (!call->_ops[0])
636             continue;
637
638         funcval = call->_ops[1];
639         if (!funcval)
640             continue;
641         if (funcval->vtype != TYPE_FUNCTION || funcval->constval.vfunc != self)
642             continue;
643
644         /* now we have a CALL and a RET, check if it's a tailcall */
645         if (ret->_ops[0] && call->_ops[0] != ret->_ops[0])
646             continue;
647
648         ++opts_optimizationcount[OPTIM_TAIL_RECURSION];
649         vec_shrinkby(block->instr, 2);
650
651         block->final = false; /* open it back up */
652
653         /* emite parameter-stores */
654         for (p = 0; p < call->params.size(); ++p) {
655             /* assert(call->params_count <= self->locals_count); */
656             if (!ir_block_create_store(block, call->context, self->locals[p].get(), call->params[p])) {
657                 irerror(call->context, "failed to create tailcall store instruction for parameter %i", (int)p);
658                 return false;
659             }
660         }
661         if (!ir_block_create_jump(block, call->context, self->blocks[0].get())) {
662             irerror(call->context, "failed to create tailcall jump");
663             return false;
664         }
665
666         delete call;
667         delete ret;
668     }
669
670     return true;
671 }
672
673 bool ir_function_finalize(ir_function *self)
674 {
675     if (self->builtin)
676         return true;
677
678     if (OPTS_OPTIMIZATION(OPTIM_PEEPHOLE)) {
679         if (!ir_function_pass_peephole(self)) {
680             irerror(self->context, "generic optimization pass broke something in `%s`", self->name.c_str());
681             return false;
682         }
683     }
684
685     if (OPTS_OPTIMIZATION(OPTIM_TAIL_RECURSION)) {
686         if (!ir_function_pass_tailrecursion(self)) {
687             irerror(self->context, "tail-recursion optimization pass broke something in `%s`", self->name.c_str());
688             return false;
689         }
690     }
691
692     if (!ir_function_naive_phi(self)) {
693         irerror(self->context, "internal error: ir_function_naive_phi failed");
694         return false;
695     }
696
697     for (auto& lp : self->locals) {
698         ir_value *v = lp.get();
699         if (v->vtype == TYPE_VECTOR ||
700             (v->vtype == TYPE_FIELD && v->outtype == TYPE_VECTOR))
701         {
702             ir_value_vector_member(v, 0);
703             ir_value_vector_member(v, 1);
704             ir_value_vector_member(v, 2);
705         }
706     }
707     for (auto& vp : self->values) {
708         ir_value *v = vp.get();
709         if (v->vtype == TYPE_VECTOR ||
710             (v->vtype == TYPE_FIELD && v->outtype == TYPE_VECTOR))
711         {
712             ir_value_vector_member(v, 0);
713             ir_value_vector_member(v, 1);
714             ir_value_vector_member(v, 2);
715         }
716     }
717
718     ir_function_enumerate(self);
719
720     if (!ir_function_calculate_liferanges(self))
721         return false;
722     if (!ir_function_allocate_locals(self))
723         return false;
724     return true;
725 }
726
727 ir_value* ir_function_create_local(ir_function *self, const std::string& name, qc_type vtype, bool param)
728 {
729     ir_value *ve;
730
731     if (param &&
732         !self->locals.empty() &&
733         self->locals.back()->store != store_param)
734     {
735         irerror(self->context, "cannot add parameters after adding locals");
736         return nullptr;
737     }
738
739     ve = new ir_value(std::string(name), (param ? store_param : store_local), vtype);
740     if (param)
741         ve->locked = true;
742     self->locals.emplace_back(ve);
743     return ve;
744 }
745
746 /***********************************************************************
747  *IR Block
748  */
749
750 void* ir_block::operator new(std::size_t bytes) {
751   return mem_a(bytes);
752 }
753
754 void ir_block::operator delete(void *data) {
755     mem_d(data);
756 }
757
758 ir_block::ir_block(ir_function* owner, const std::string& name)
759 : owner(owner),
760   label(name)
761 {
762     context.file = "<@no context>";
763     context.line = 0;
764 }
765
766 ir_block::~ir_block()
767 {
768     for (size_t i = 0; i != vec_size(instr); ++i)
769         delete instr[i];
770     vec_free(instr);
771     vec_free(entries);
772     vec_free(exits);
773 }
774
775 static void ir_block_delete_quick(ir_block* self)
776 {
777     size_t i;
778     for (i = 0; i != vec_size(self->instr); ++i)
779         ir_instr_delete_quick(self->instr[i]);
780     vec_free(self->instr);
781     delete self;
782 }
783
784 /***********************************************************************
785  *IR Instructions
786  */
787
788 void* ir_instr::operator new(std::size_t bytes) {
789   return mem_a(bytes);
790 }
791
792 void ir_instr::operator delete(void *data) {
793     mem_d(data);
794 }
795
796 ir_instr::ir_instr(lex_ctx_t ctx, ir_block* owner_, int op)
797 : opcode(op),
798   context(ctx),
799   owner(owner_)
800 {
801 }
802
803 ir_instr::~ir_instr()
804 {
805     // The following calls can only delete from
806     // vectors, we still want to delete this instruction
807     // so ignore the return value. Since with the warn_unused_result attribute
808     // gcc doesn't care about an explicit: (void)foo(); to ignore the result,
809     // I have to improvise here and use if(foo());
810     for (auto &it : phi) {
811         size_t idx;
812         if (vec_ir_instr_find(it.value->writes, this, &idx))
813             it.value->writes.erase(it.value->writes.begin() + idx);
814         if (vec_ir_instr_find(it.value->reads, this, &idx))
815             it.value->reads.erase(it.value->reads.begin() + idx);
816     }
817     for (auto &it : params) {
818         size_t idx;
819         if (vec_ir_instr_find(it->writes, this, &idx))
820             it->writes.erase(it->writes.begin() + idx);
821         if (vec_ir_instr_find(it->reads, this, &idx))
822             it->reads.erase(it->reads.begin() + idx);
823     }
824     (void)!ir_instr_op(this, 0, nullptr, false);
825     (void)!ir_instr_op(this, 1, nullptr, false);
826     (void)!ir_instr_op(this, 2, nullptr, false);
827 }
828
829 static void ir_instr_delete_quick(ir_instr *self)
830 {
831     self->phi.clear();
832     self->params.clear();
833     delete self;
834 }
835
836 static bool ir_instr_op(ir_instr *self, int op, ir_value *v, bool writing)
837 {
838     if (v && v->vtype == TYPE_NOEXPR) {
839         irerror(self->context, "tried to use a NOEXPR value");
840         return false;
841     }
842
843     if (self->_ops[op]) {
844         size_t idx;
845         if (writing && vec_ir_instr_find(self->_ops[op]->writes, self, &idx))
846             self->_ops[op]->writes.erase(self->_ops[op]->writes.begin() + idx);
847         else if (vec_ir_instr_find(self->_ops[op]->reads, self, &idx))
848             self->_ops[op]->reads.erase(self->_ops[op]->reads.begin() + idx);
849     }
850     if (v) {
851         if (writing)
852             v->writes.push_back(self);
853         else
854             v->reads.push_back(self);
855     }
856     self->_ops[op] = v;
857     return true;
858 }
859
860 /***********************************************************************
861  *IR Value
862  */
863
864 static void ir_value_code_setaddr(ir_value *self, int32_t gaddr)
865 {
866     self->code.globaladdr = gaddr;
867     if (self->members[0]) self->members[0]->code.globaladdr = gaddr;
868     if (self->members[1]) self->members[1]->code.globaladdr = gaddr;
869     if (self->members[2]) self->members[2]->code.globaladdr = gaddr;
870 }
871
872 static int32_t ir_value_code_addr(const ir_value *self)
873 {
874     if (self->store == store_return)
875         return OFS_RETURN + self->code.addroffset;
876     return self->code.globaladdr + self->code.addroffset;
877 }
878
879 void* ir_value::operator new(std::size_t bytes) {
880   return mem_a(bytes);
881 }
882
883 void ir_value::operator delete(void *data) {
884     mem_d(data);
885 }
886
887 ir_value::ir_value(std::string&& name_, store_type store_, qc_type vtype_)
888 : name(move(name_)),
889   vtype(vtype_),
890   store(store_)
891 {
892     fieldtype = TYPE_VOID;
893     outtype = TYPE_VOID;
894     flags = 0;
895
896     cvq          = CV_NONE;
897     hasvalue     = false;
898     context.file = "<@no context>";
899     context.line = 0;
900
901     memset(&constval, 0, sizeof(constval));
902     memset(&code,     0, sizeof(code));
903
904     members[0] = nullptr;
905     members[1] = nullptr;
906     members[2] = nullptr;
907     memberof = nullptr;
908
909     unique_life = false;
910     locked = false;
911     callparam  = false;
912 }
913
914 ir_value::~ir_value()
915 {
916     size_t i;
917     if (hasvalue) {
918         if (vtype == TYPE_STRING)
919             mem_d((void*)constval.vstring);
920     }
921     if (!(flags & IR_FLAG_SPLIT_VECTOR)) {
922         for (i = 0; i < 3; ++i) {
923             if (members[i])
924                 delete members[i];
925         }
926     }
927 }
928
929
930 /*  helper function */
931 static ir_value* ir_builder_imm_float(ir_builder *self, float value, bool add_to_list) {
932     ir_value *v = new ir_value("#IMMEDIATE", store_global, TYPE_FLOAT);
933     v->flags |= IR_FLAG_ERASABLE;
934     v->hasvalue = true;
935     v->cvq = CV_CONST;
936     v->constval.vfloat = value;
937
938     self->globals.emplace_back(v);
939     if (add_to_list)
940         self->const_floats.emplace_back(v);
941     return v;
942 }
943
944 ir_value* ir_value_vector_member(ir_value *self, unsigned int member)
945 {
946     std::string name;
947     ir_value *m;
948     if (member >= 3)
949         return nullptr;
950
951     if (self->members[member])
952         return self->members[member];
953
954     if (!self->name.empty()) {
955         char member_name[3] = { '_', char('x' + member), 0 };
956         name = self->name + member_name;
957     }
958
959     if (self->vtype == TYPE_VECTOR)
960     {
961         m = new ir_value(move(name), self->store, TYPE_FLOAT);
962         if (!m)
963             return nullptr;
964         m->context = self->context;
965
966         self->members[member] = m;
967         m->code.addroffset = member;
968     }
969     else if (self->vtype == TYPE_FIELD)
970     {
971         if (self->fieldtype != TYPE_VECTOR)
972             return nullptr;
973         m = new ir_value(move(name), self->store, TYPE_FIELD);
974         if (!m)
975             return nullptr;
976         m->fieldtype = TYPE_FLOAT;
977         m->context = self->context;
978
979         self->members[member] = m;
980         m->code.addroffset = member;
981     }
982     else
983     {
984         irerror(self->context, "invalid member access on %s", self->name.c_str());
985         return nullptr;
986     }
987
988     m->memberof = self;
989     return m;
990 }
991
992 static GMQCC_INLINE size_t ir_value_sizeof(const ir_value *self)
993 {
994     if (self->vtype == TYPE_FIELD && self->fieldtype == TYPE_VECTOR)
995         return type_sizeof_[TYPE_VECTOR];
996     return type_sizeof_[self->vtype];
997 }
998
999 static ir_value* ir_value_out(ir_function *owner, const char *name, store_type storetype, qc_type vtype)
1000 {
1001     ir_value *v = new ir_value(name ? std::string(name) : std::string(), storetype, vtype);
1002     if (!v)
1003         return nullptr;
1004     ir_function_collect_value(owner, v);
1005     return v;
1006 }
1007
1008 bool ir_value_set_float(ir_value *self, float f)
1009 {
1010     if (self->vtype != TYPE_FLOAT)
1011         return false;
1012     self->constval.vfloat = f;
1013     self->hasvalue = true;
1014     return true;
1015 }
1016
1017 bool ir_value_set_func(ir_value *self, int f)
1018 {
1019     if (self->vtype != TYPE_FUNCTION)
1020         return false;
1021     self->constval.vint = f;
1022     self->hasvalue = true;
1023     return true;
1024 }
1025
1026 bool ir_value_set_vector(ir_value *self, vec3_t v)
1027 {
1028     if (self->vtype != TYPE_VECTOR)
1029         return false;
1030     self->constval.vvec = v;
1031     self->hasvalue = true;
1032     return true;
1033 }
1034
1035 bool ir_value_set_field(ir_value *self, ir_value *fld)
1036 {
1037     if (self->vtype != TYPE_FIELD)
1038         return false;
1039     self->constval.vpointer = fld;
1040     self->hasvalue = true;
1041     return true;
1042 }
1043
1044 bool ir_value_set_string(ir_value *self, const char *str)
1045 {
1046     if (self->vtype != TYPE_STRING)
1047         return false;
1048     self->constval.vstring = util_strdupe(str);
1049     self->hasvalue = true;
1050     return true;
1051 }
1052
1053 #if 0
1054 bool ir_value_set_int(ir_value *self, int i)
1055 {
1056     if (self->vtype != TYPE_INTEGER)
1057         return false;
1058     self->constval.vint = i;
1059     self->hasvalue = true;
1060     return true;
1061 }
1062 #endif
1063
1064 bool ir_value_lives(ir_value *self, size_t at)
1065 {
1066     for (auto& l : self->life) {
1067         if (l.start <= at && at <= l.end)
1068             return true;
1069         if (l.start > at) /* since it's ordered */
1070             return false;
1071     }
1072     return false;
1073 }
1074
1075 static bool ir_value_life_insert(ir_value *self, size_t idx, ir_life_entry_t e)
1076 {
1077     self->life.insert(self->life.begin() + idx, e);
1078     return true;
1079 }
1080
1081 static bool ir_value_life_merge(ir_value *self, size_t s)
1082 {
1083     size_t i;
1084     const size_t vs = self->life.size();
1085     ir_life_entry_t *life_found = nullptr;
1086     ir_life_entry_t *before = nullptr;
1087     ir_life_entry_t new_entry;
1088
1089     /* Find the first range >= s */
1090     for (i = 0; i < vs; ++i)
1091     {
1092         before = life_found;
1093         life_found = &self->life[i];
1094         if (life_found->start > s)
1095             break;
1096     }
1097     /* nothing found? append */
1098     if (i == vs) {
1099         ir_life_entry_t e;
1100         if (life_found && life_found->end+1 == s)
1101         {
1102             /* previous life range can be merged in */
1103             life_found->end++;
1104             return true;
1105         }
1106         if (life_found && life_found->end >= s)
1107             return false;
1108         e.start = e.end = s;
1109         self->life.emplace_back(e);
1110         return true;
1111     }
1112     /* found */
1113     if (before)
1114     {
1115         if (before->end + 1 == s &&
1116             life_found->start - 1 == s)
1117         {
1118             /* merge */
1119             before->end = life_found->end;
1120             self->life.erase(self->life.begin()+i);
1121             return true;
1122         }
1123         if (before->end + 1 == s)
1124         {
1125             /* extend before */
1126             before->end++;
1127             return true;
1128         }
1129         /* already contained */
1130         if (before->end >= s)
1131             return false;
1132     }
1133     /* extend */
1134     if (life_found->start - 1 == s)
1135     {
1136         life_found->start--;
1137         return true;
1138     }
1139     /* insert a new entry */
1140     new_entry.start = new_entry.end = s;
1141     return ir_value_life_insert(self, i, new_entry);
1142 }
1143
1144 static bool ir_value_life_merge_into(ir_value *self, const ir_value *other)
1145 {
1146     size_t i, myi;
1147
1148     if (other->life.empty())
1149         return true;
1150
1151     if (self->life.empty()) {
1152         self->life = other->life;
1153         return true;
1154     }
1155
1156     myi = 0;
1157     for (i = 0; i < other->life.size(); ++i)
1158     {
1159         const ir_life_entry_t &otherlife = other->life[i];
1160         while (true)
1161         {
1162             ir_life_entry_t *entry = &self->life[myi];
1163
1164             if (otherlife.end+1 < entry->start)
1165             {
1166                 /* adding an interval before entry */
1167                 if (!ir_value_life_insert(self, myi, otherlife))
1168                     return false;
1169                 ++myi;
1170                 break;
1171             }
1172
1173             if (otherlife.start <  entry->start &&
1174                 otherlife.end+1 >= entry->start)
1175             {
1176                 /* starts earlier and overlaps */
1177                 entry->start = otherlife.start;
1178             }
1179
1180             if (otherlife.end   >  entry->end &&
1181                 otherlife.start <= entry->end+1)
1182             {
1183                 /* ends later and overlaps */
1184                 entry->end = otherlife.end;
1185             }
1186
1187             /* see if our change combines it with the next ranges */
1188             while (myi+1 < self->life.size() &&
1189                    entry->end+1 >= self->life[1+myi].start)
1190             {
1191                 /* overlaps with (myi+1) */
1192                 if (entry->end < self->life[1+myi].end)
1193                     entry->end = self->life[1+myi].end;
1194                 self->life.erase(self->life.begin() + (myi + 1));
1195                 entry = &self->life[myi];
1196             }
1197
1198             /* see if we're after the entry */
1199             if (otherlife.start > entry->end)
1200             {
1201                 ++myi;
1202                 /* append if we're at the end */
1203                 if (myi >= self->life.size()) {
1204                     self->life.emplace_back(otherlife);
1205                     break;
1206                 }
1207                 /* otherweise check the next range */
1208                 continue;
1209             }
1210             break;
1211         }
1212     }
1213     return true;
1214 }
1215
1216 static bool ir_values_overlap(const ir_value *a, const ir_value *b)
1217 {
1218     /* For any life entry in A see if it overlaps with
1219      * any life entry in B.
1220      * Note that the life entries are orderes, so we can make a
1221      * more efficient algorithm there than naively translating the
1222      * statement above.
1223      */
1224
1225     const ir_life_entry_t *la, *lb, *enda, *endb;
1226
1227     /* first of all, if either has no life range, they cannot clash */
1228     if (a->life.empty() || b->life.empty())
1229         return false;
1230
1231     la = &a->life.front();
1232     lb = &b->life.front();
1233     enda = &a->life.back() + 1;
1234     endb = &b->life.back() + 1;
1235     while (true)
1236     {
1237         /* check if the entries overlap, for that,
1238          * both must start before the other one ends.
1239          */
1240         if (la->start < lb->end &&
1241             lb->start < la->end)
1242         {
1243             return true;
1244         }
1245
1246         /* entries are ordered
1247          * one entry is earlier than the other
1248          * that earlier entry will be moved forward
1249          */
1250         if (la->start < lb->start)
1251         {
1252             /* order: A B, move A forward
1253              * check if we hit the end with A
1254              */
1255             if (++la == enda)
1256                 break;
1257         }
1258         else /* if (lb->start < la->start)  actually <= */
1259         {
1260             /* order: B A, move B forward
1261              * check if we hit the end with B
1262              */
1263             if (++lb == endb)
1264                 break;
1265         }
1266     }
1267     return false;
1268 }
1269
1270 /***********************************************************************
1271  *IR main operations
1272  */
1273
1274 static bool ir_check_unreachable(ir_block *self)
1275 {
1276     /* The IR should never have to deal with unreachable code */
1277     if (!self->final/* || OPTS_FLAG(ALLOW_UNREACHABLE_CODE)*/)
1278         return true;
1279     irerror(self->context, "unreachable statement (%s)", self->label.c_str());
1280     return false;
1281 }
1282
1283 bool ir_block_create_store_op(ir_block *self, lex_ctx_t ctx, int op, ir_value *target, ir_value *what)
1284 {
1285     ir_instr *in;
1286     if (!ir_check_unreachable(self))
1287         return false;
1288
1289     if (target->store == store_value &&
1290         (op < INSTR_STOREP_F || op > INSTR_STOREP_FNC))
1291     {
1292         irerror(self->context, "cannot store to an SSA value");
1293         irerror(self->context, "trying to store: %s <- %s", target->name.c_str(), what->name.c_str());
1294         irerror(self->context, "instruction: %s", util_instr_str[op]);
1295         return false;
1296     }
1297
1298     in = new ir_instr(ctx, self, op);
1299     if (!in)
1300         return false;
1301
1302     if (!ir_instr_op(in, 0, target, (op < INSTR_STOREP_F || op > INSTR_STOREP_FNC)) ||
1303         !ir_instr_op(in, 1, what, false))
1304     {
1305         delete in;
1306         return false;
1307     }
1308     vec_push(self->instr, in);
1309     return true;
1310 }
1311
1312 bool ir_block_create_state_op(ir_block *self, lex_ctx_t ctx, ir_value *frame, ir_value *think)
1313 {
1314     ir_instr *in;
1315     if (!ir_check_unreachable(self))
1316         return false;
1317
1318     in = new ir_instr(ctx, self, INSTR_STATE);
1319     if (!in)
1320         return false;
1321
1322     if (!ir_instr_op(in, 0, frame, false) ||
1323         !ir_instr_op(in, 1, think, false))
1324     {
1325         delete in;
1326         return false;
1327     }
1328     vec_push(self->instr, in);
1329     return true;
1330 }
1331
1332 static bool ir_block_create_store(ir_block *self, lex_ctx_t ctx, ir_value *target, ir_value *what)
1333 {
1334     int op = 0;
1335     qc_type vtype;
1336     if (target->vtype == TYPE_VARIANT)
1337         vtype = what->vtype;
1338     else
1339         vtype = target->vtype;
1340
1341 #if 0
1342     if      (vtype == TYPE_FLOAT   && what->vtype == TYPE_INTEGER)
1343         op = INSTR_CONV_ITOF;
1344     else if (vtype == TYPE_INTEGER && what->vtype == TYPE_FLOAT)
1345         op = INSTR_CONV_FTOI;
1346 #endif
1347         op = type_store_instr[vtype];
1348
1349     if (OPTS_FLAG(ADJUST_VECTOR_FIELDS)) {
1350         if (op == INSTR_STORE_FLD && what->fieldtype == TYPE_VECTOR)
1351             op = INSTR_STORE_V;
1352     }
1353
1354     return ir_block_create_store_op(self, ctx, op, target, what);
1355 }
1356
1357 bool ir_block_create_storep(ir_block *self, lex_ctx_t ctx, ir_value *target, ir_value *what)
1358 {
1359     int op = 0;
1360     qc_type vtype;
1361
1362     if (target->vtype != TYPE_POINTER)
1363         return false;
1364
1365     /* storing using pointer - target is a pointer, type must be
1366      * inferred from source
1367      */
1368     vtype = what->vtype;
1369
1370     op = type_storep_instr[vtype];
1371     if (OPTS_FLAG(ADJUST_VECTOR_FIELDS)) {
1372         if (op == INSTR_STOREP_FLD && what->fieldtype == TYPE_VECTOR)
1373             op = INSTR_STOREP_V;
1374     }
1375
1376     return ir_block_create_store_op(self, ctx, op, target, what);
1377 }
1378
1379 bool ir_block_create_return(ir_block *self, lex_ctx_t ctx, ir_value *v)
1380 {
1381     ir_instr *in;
1382     if (!ir_check_unreachable(self))
1383         return false;
1384
1385     self->final = true;
1386
1387     self->is_return = true;
1388     in = new ir_instr(ctx, self, INSTR_RETURN);
1389     if (!in)
1390         return false;
1391
1392     if (v && !ir_instr_op(in, 0, v, false)) {
1393         delete in;
1394         return false;
1395     }
1396
1397     vec_push(self->instr, in);
1398     return true;
1399 }
1400
1401 bool ir_block_create_if(ir_block *self, lex_ctx_t ctx, ir_value *v,
1402                         ir_block *ontrue, ir_block *onfalse)
1403 {
1404     ir_instr *in;
1405     if (!ir_check_unreachable(self))
1406         return false;
1407     self->final = true;
1408     /*in = new ir_instr(ctx, self, (v->vtype == TYPE_STRING ? INSTR_IF_S : INSTR_IF_F));*/
1409     in = new ir_instr(ctx, self, VINSTR_COND);
1410     if (!in)
1411         return false;
1412
1413     if (!ir_instr_op(in, 0, v, false)) {
1414         delete in;
1415         return false;
1416     }
1417
1418     in->bops[0] = ontrue;
1419     in->bops[1] = onfalse;
1420
1421     vec_push(self->instr, in);
1422
1423     vec_push(self->exits, ontrue);
1424     vec_push(self->exits, onfalse);
1425     vec_push(ontrue->entries,  self);
1426     vec_push(onfalse->entries, self);
1427     return true;
1428 }
1429
1430 bool ir_block_create_jump(ir_block *self, lex_ctx_t ctx, ir_block *to)
1431 {
1432     ir_instr *in;
1433     if (!ir_check_unreachable(self))
1434         return false;
1435     self->final = true;
1436     in = new ir_instr(ctx, self, VINSTR_JUMP);
1437     if (!in)
1438         return false;
1439
1440     in->bops[0] = to;
1441     vec_push(self->instr, in);
1442
1443     vec_push(self->exits, to);
1444     vec_push(to->entries, self);
1445     return true;
1446 }
1447
1448 bool ir_block_create_goto(ir_block *self, lex_ctx_t ctx, ir_block *to)
1449 {
1450     self->owner->flags |= IR_FLAG_HAS_GOTO;
1451     return ir_block_create_jump(self, ctx, to);
1452 }
1453
1454 ir_instr* ir_block_create_phi(ir_block *self, lex_ctx_t ctx, const char *label, qc_type ot)
1455 {
1456     ir_value *out;
1457     ir_instr *in;
1458     if (!ir_check_unreachable(self))
1459         return nullptr;
1460     in = new ir_instr(ctx, self, VINSTR_PHI);
1461     if (!in)
1462         return nullptr;
1463     out = ir_value_out(self->owner, label, store_value, ot);
1464     if (!out) {
1465         delete in;
1466         return nullptr;
1467     }
1468     if (!ir_instr_op(in, 0, out, true)) {
1469         delete in;
1470         return nullptr;
1471     }
1472     vec_push(self->instr, in);
1473     return in;
1474 }
1475
1476 ir_value* ir_phi_value(ir_instr *self)
1477 {
1478     return self->_ops[0];
1479 }
1480
1481 void ir_phi_add(ir_instr* self, ir_block *b, ir_value *v)
1482 {
1483     ir_phi_entry_t pe;
1484
1485     if (!vec_ir_block_find(self->owner->entries, b, nullptr)) {
1486         // Must not be possible to cause this, otherwise the AST
1487         // is doing something wrong.
1488         irerror(self->context, "Invalid entry block for PHI");
1489         exit(EXIT_FAILURE);
1490     }
1491
1492     pe.value = v;
1493     pe.from = b;
1494     v->reads.push_back(self);
1495     self->phi.push_back(pe);
1496 }
1497
1498 /* call related code */
1499 ir_instr* ir_block_create_call(ir_block *self, lex_ctx_t ctx, const char *label, ir_value *func, bool noreturn)
1500 {
1501     ir_value *out;
1502     ir_instr *in;
1503     if (!ir_check_unreachable(self))
1504         return nullptr;
1505     in = new ir_instr(ctx, self, (noreturn ? VINSTR_NRCALL : INSTR_CALL0));
1506     if (!in)
1507         return nullptr;
1508     if (noreturn) {
1509         self->final = true;
1510         self->is_return = true;
1511     }
1512     out = ir_value_out(self->owner, label, (func->outtype == TYPE_VOID) ? store_return : store_value, func->outtype);
1513     if (!out) {
1514         delete in;
1515         return nullptr;
1516     }
1517     if (!ir_instr_op(in, 0, out, true) ||
1518         !ir_instr_op(in, 1, func, false))
1519     {
1520         delete in;
1521         return nullptr;
1522     }
1523     vec_push(self->instr, in);
1524     /*
1525     if (noreturn) {
1526         if (!ir_block_create_return(self, ctx, nullptr)) {
1527             compile_error(ctx, "internal error: failed to generate dummy-return instruction");
1528             delete in;
1529             return nullptr;
1530         }
1531     }
1532     */
1533     return in;
1534 }
1535
1536 ir_value* ir_call_value(ir_instr *self)
1537 {
1538     return self->_ops[0];
1539 }
1540
1541 void ir_call_param(ir_instr* self, ir_value *v)
1542 {
1543     self->params.push_back(v);
1544     v->reads.push_back(self);
1545 }
1546
1547 /* binary op related code */
1548
1549 ir_value* ir_block_create_binop(ir_block *self, lex_ctx_t ctx,
1550                                 const char *label, int opcode,
1551                                 ir_value *left, ir_value *right)
1552 {
1553     qc_type ot = TYPE_VOID;
1554     switch (opcode) {
1555         case INSTR_ADD_F:
1556         case INSTR_SUB_F:
1557         case INSTR_DIV_F:
1558         case INSTR_MUL_F:
1559         case INSTR_MUL_V:
1560         case INSTR_AND:
1561         case INSTR_OR:
1562 #if 0
1563         case INSTR_AND_I:
1564         case INSTR_AND_IF:
1565         case INSTR_AND_FI:
1566         case INSTR_OR_I:
1567         case INSTR_OR_IF:
1568         case INSTR_OR_FI:
1569 #endif
1570         case INSTR_BITAND:
1571         case INSTR_BITOR:
1572         case VINSTR_BITXOR:
1573 #if 0
1574         case INSTR_SUB_S: /* -- offset of string as float */
1575         case INSTR_MUL_IF:
1576         case INSTR_MUL_FI:
1577         case INSTR_DIV_IF:
1578         case INSTR_DIV_FI:
1579         case INSTR_BITOR_IF:
1580         case INSTR_BITOR_FI:
1581         case INSTR_BITAND_FI:
1582         case INSTR_BITAND_IF:
1583         case INSTR_EQ_I:
1584         case INSTR_NE_I:
1585 #endif
1586             ot = TYPE_FLOAT;
1587             break;
1588 #if 0
1589         case INSTR_ADD_I:
1590         case INSTR_ADD_IF:
1591         case INSTR_ADD_FI:
1592         case INSTR_SUB_I:
1593         case INSTR_SUB_FI:
1594         case INSTR_SUB_IF:
1595         case INSTR_MUL_I:
1596         case INSTR_DIV_I:
1597         case INSTR_BITAND_I:
1598         case INSTR_BITOR_I:
1599         case INSTR_XOR_I:
1600         case INSTR_RSHIFT_I:
1601         case INSTR_LSHIFT_I:
1602             ot = TYPE_INTEGER;
1603             break;
1604 #endif
1605         case INSTR_ADD_V:
1606         case INSTR_SUB_V:
1607         case INSTR_MUL_VF:
1608         case INSTR_MUL_FV:
1609         case VINSTR_BITAND_V:
1610         case VINSTR_BITOR_V:
1611         case VINSTR_BITXOR_V:
1612         case VINSTR_BITAND_VF:
1613         case VINSTR_BITOR_VF:
1614         case VINSTR_BITXOR_VF:
1615         case VINSTR_CROSS:
1616 #if 0
1617         case INSTR_DIV_VF:
1618         case INSTR_MUL_IV:
1619         case INSTR_MUL_VI:
1620 #endif
1621             ot = TYPE_VECTOR;
1622             break;
1623 #if 0
1624         case INSTR_ADD_SF:
1625             ot = TYPE_POINTER;
1626             break;
1627 #endif
1628     /*
1629      * after the following default case, the value of opcode can never
1630      * be 1, 2, 3, 4, 5, 6, 7, 8, 9, 62, 63, 64, 65
1631      */
1632         default:
1633             /* ranges: */
1634             /* boolean operations result in floats */
1635
1636             /*
1637              * opcode >= 10 takes true branch opcode is at least 10
1638              * opcode <= 23 takes false branch opcode is at least 24
1639              */
1640             if (opcode >= INSTR_EQ_F && opcode <= INSTR_GT)
1641                 ot = TYPE_FLOAT;
1642
1643             /*
1644              * At condition "opcode <= 23", the value of "opcode" must be
1645              * at least 24.
1646              * At condition "opcode <= 23", the value of "opcode" cannot be
1647              * equal to any of {1, 2, 3, 4, 5, 6, 7, 8, 9, 62, 63, 64, 65}.
1648              * The condition "opcode <= 23" cannot be true.
1649              *
1650              * Thus ot=2 (TYPE_FLOAT) can never be true
1651              */
1652 #if 0
1653             else if (opcode >= INSTR_LE && opcode <= INSTR_GT)
1654                 ot = TYPE_FLOAT;
1655             else if (opcode >= INSTR_LE_I && opcode <= INSTR_EQ_FI)
1656                 ot = TYPE_FLOAT;
1657 #endif
1658             break;
1659     };
1660     if (ot == TYPE_VOID) {
1661         /* The AST or parser were supposed to check this! */
1662         return nullptr;
1663     }
1664
1665     return ir_block_create_general_instr(self, ctx, label, opcode, left, right, ot);
1666 }
1667
1668 ir_value* ir_block_create_unary(ir_block *self, lex_ctx_t ctx,
1669                                 const char *label, int opcode,
1670                                 ir_value *operand)
1671 {
1672     qc_type ot = TYPE_FLOAT;
1673     switch (opcode) {
1674         case INSTR_NOT_F:
1675         case INSTR_NOT_V:
1676         case INSTR_NOT_S:
1677         case INSTR_NOT_ENT:
1678         case INSTR_NOT_FNC: /*
1679         case INSTR_NOT_I:   */
1680             ot = TYPE_FLOAT;
1681             break;
1682
1683         /*
1684          * Negation for virtual instructions is emulated with 0-value. Thankfully
1685          * the operand for 0 already exists so we just source it from here.
1686          */
1687         case VINSTR_NEG_F:
1688             return ir_block_create_general_instr(self, ctx, label, INSTR_SUB_F, nullptr, operand, ot);
1689         case VINSTR_NEG_V:
1690             return ir_block_create_general_instr(self, ctx, label, INSTR_SUB_V, nullptr, operand, TYPE_VECTOR);
1691
1692         default:
1693             ot = operand->vtype;
1694             break;
1695     };
1696     if (ot == TYPE_VOID) {
1697         /* The AST or parser were supposed to check this! */
1698         return nullptr;
1699     }
1700
1701     /* let's use the general instruction creator and pass nullptr for OPB */
1702     return ir_block_create_general_instr(self, ctx, label, opcode, operand, nullptr, ot);
1703 }
1704
1705 static ir_value* ir_block_create_general_instr(ir_block *self, lex_ctx_t ctx, const char *label,
1706                                         int op, ir_value *a, ir_value *b, qc_type outype)
1707 {
1708     ir_instr *instr;
1709     ir_value *out;
1710
1711     out = ir_value_out(self->owner, label, store_value, outype);
1712     if (!out)
1713         return nullptr;
1714
1715     instr = new ir_instr(ctx, self, op);
1716     if (!instr) {
1717         return nullptr;
1718     }
1719
1720     if (!ir_instr_op(instr, 0, out, true) ||
1721         !ir_instr_op(instr, 1, a, false) ||
1722         !ir_instr_op(instr, 2, b, false) )
1723     {
1724         goto on_error;
1725     }
1726
1727     vec_push(self->instr, instr);
1728
1729     return out;
1730 on_error:
1731     delete instr;
1732     return nullptr;
1733 }
1734
1735 ir_value* ir_block_create_fieldaddress(ir_block *self, lex_ctx_t ctx, const char *label, ir_value *ent, ir_value *field)
1736 {
1737     ir_value *v;
1738
1739     /* Support for various pointer types todo if so desired */
1740     if (ent->vtype != TYPE_ENTITY)
1741         return nullptr;
1742
1743     if (field->vtype != TYPE_FIELD)
1744         return nullptr;
1745
1746     v = ir_block_create_general_instr(self, ctx, label, INSTR_ADDRESS, ent, field, TYPE_POINTER);
1747     v->fieldtype = field->fieldtype;
1748     return v;
1749 }
1750
1751 ir_value* ir_block_create_load_from_ent(ir_block *self, lex_ctx_t ctx, const char *label, ir_value *ent, ir_value *field, qc_type outype)
1752 {
1753     int op;
1754     if (ent->vtype != TYPE_ENTITY)
1755         return nullptr;
1756
1757     /* at some point we could redirect for TYPE_POINTER... but that could lead to carelessness */
1758     if (field->vtype != TYPE_FIELD)
1759         return nullptr;
1760
1761     switch (outype)
1762     {
1763         case TYPE_FLOAT:    op = INSTR_LOAD_F;   break;
1764         case TYPE_VECTOR:   op = INSTR_LOAD_V;   break;
1765         case TYPE_STRING:   op = INSTR_LOAD_S;   break;
1766         case TYPE_FIELD:    op = INSTR_LOAD_FLD; break;
1767         case TYPE_ENTITY:   op = INSTR_LOAD_ENT; break;
1768         case TYPE_FUNCTION: op = INSTR_LOAD_FNC; break;
1769 #if 0
1770         case TYPE_POINTER: op = INSTR_LOAD_I;   break;
1771         case TYPE_INTEGER: op = INSTR_LOAD_I;   break;
1772 #endif
1773         default:
1774             irerror(self->context, "invalid type for ir_block_create_load_from_ent: %s", type_name[outype]);
1775             return nullptr;
1776     }
1777
1778     return ir_block_create_general_instr(self, ctx, label, op, ent, field, outype);
1779 }
1780
1781 /* PHI resolving breaks the SSA, and must thus be the last
1782  * step before life-range calculation.
1783  */
1784
1785 static bool ir_block_naive_phi(ir_block *self);
1786 bool ir_function_naive_phi(ir_function *self)
1787 {
1788     for (auto& b : self->blocks)
1789         if (!ir_block_naive_phi(b.get()))
1790             return false;
1791     return true;
1792 }
1793
1794 static bool ir_block_naive_phi(ir_block *self)
1795 {
1796     size_t i;
1797     /* FIXME: optionally, create_phi can add the phis
1798      * to a list so we don't need to loop through blocks
1799      * - anyway: "don't optimize YET"
1800      */
1801     for (i = 0; i < vec_size(self->instr); ++i)
1802     {
1803         ir_instr *instr = self->instr[i];
1804         if (instr->opcode != VINSTR_PHI)
1805             continue;
1806
1807         vec_remove(self->instr, i, 1);
1808         --i; /* NOTE: i+1 below */
1809
1810         for (auto &it : instr->phi) {
1811             ir_value *v = it.value;
1812             ir_block *b = it.from;
1813             if (v->store == store_value && v->reads.size() == 1 && v->writes.size() == 1) {
1814                 /* replace the value */
1815                 if (!ir_instr_op(v->writes[0], 0, instr->_ops[0], true))
1816                     return false;
1817             } else {
1818                 /* force a move instruction */
1819                 ir_instr *prevjump = vec_last(b->instr);
1820                 vec_pop(b->instr);
1821                 b->final = false;
1822                 instr->_ops[0]->store = store_global;
1823                 if (!ir_block_create_store(b, instr->context, instr->_ops[0], v))
1824                     return false;
1825                 instr->_ops[0]->store = store_value;
1826                 vec_push(b->instr, prevjump);
1827                 b->final = true;
1828             }
1829         }
1830         delete instr;
1831     }
1832     return true;
1833 }
1834
1835 /***********************************************************************
1836  *IR Temp allocation code
1837  * Propagating value life ranges by walking through the function backwards
1838  * until no more changes are made.
1839  * In theory this should happen once more than once for every nested loop
1840  * level.
1841  * Though this implementation might run an additional time for if nests.
1842  */
1843
1844 /* Enumerate instructions used by value's life-ranges
1845  */
1846 static void ir_block_enumerate(ir_block *self, size_t *_eid)
1847 {
1848     size_t i;
1849     size_t eid = *_eid;
1850     for (i = 0; i < vec_size(self->instr); ++i)
1851     {
1852         self->instr[i]->eid = eid++;
1853     }
1854     *_eid = eid;
1855 }
1856
1857 /* Enumerate blocks and instructions.
1858  * The block-enumeration is unordered!
1859  * We do not really use the block enumreation, however
1860  * the instruction enumeration is important for life-ranges.
1861  */
1862 void ir_function_enumerate(ir_function *self)
1863 {
1864     size_t instruction_id = 0;
1865     size_t block_eid = 0;
1866     for (auto& block : self->blocks)
1867     {
1868         /* each block now gets an additional "entry" instruction id
1869          * we can use to avoid point-life issues
1870          */
1871         block->entry_id = instruction_id;
1872         block->eid      = block_eid;
1873         ++instruction_id;
1874         ++block_eid;
1875
1876         ir_block_enumerate(block.get(), &instruction_id);
1877     }
1878 }
1879
1880 /* Local-value allocator
1881  * After finishing creating the liferange of all values used in a function
1882  * we can allocate their global-positions.
1883  * This is the counterpart to register-allocation in register machines.
1884  */
1885 struct function_allocator {
1886     ir_value **locals;
1887     size_t *sizes;
1888     size_t *positions;
1889     bool *unique;
1890 };
1891
1892 static bool function_allocator_alloc(function_allocator *alloc, ir_value *var)
1893 {
1894     ir_value *slot;
1895     size_t vsize = ir_value_sizeof(var);
1896
1897     var->code.local = vec_size(alloc->locals);
1898
1899     slot = new ir_value("reg", store_global, var->vtype);
1900     if (!slot)
1901         return false;
1902
1903     if (!ir_value_life_merge_into(slot, var))
1904         goto localerror;
1905
1906     vec_push(alloc->locals, slot);
1907     vec_push(alloc->sizes, vsize);
1908     vec_push(alloc->unique, var->unique_life);
1909
1910     return true;
1911
1912 localerror:
1913     delete slot;
1914     return false;
1915 }
1916
1917 static bool ir_function_allocator_assign(ir_function *self, function_allocator *alloc, ir_value *v)
1918 {
1919     size_t a;
1920     ir_value *slot;
1921
1922     if (v->unique_life)
1923         return function_allocator_alloc(alloc, v);
1924
1925     for (a = 0; a < vec_size(alloc->locals); ++a)
1926     {
1927         /* if it's reserved for a unique liferange: skip */
1928         if (alloc->unique[a])
1929             continue;
1930
1931         slot = alloc->locals[a];
1932
1933         /* never resize parameters
1934          * will be required later when overlapping temps + locals
1935          */
1936         if (a < vec_size(self->params) &&
1937             alloc->sizes[a] < ir_value_sizeof(v))
1938         {
1939             continue;
1940         }
1941
1942         if (ir_values_overlap(v, slot))
1943             continue;
1944
1945         if (!ir_value_life_merge_into(slot, v))
1946             return false;
1947
1948         /* adjust size for this slot */
1949         if (alloc->sizes[a] < ir_value_sizeof(v))
1950             alloc->sizes[a] = ir_value_sizeof(v);
1951
1952         v->code.local = a;
1953         return true;
1954     }
1955     if (a >= vec_size(alloc->locals)) {
1956         if (!function_allocator_alloc(alloc, v))
1957             return false;
1958     }
1959     return true;
1960 }
1961
1962 bool ir_function_allocate_locals(ir_function *self)
1963 {
1964     bool   retval = true;
1965     size_t pos;
1966     bool   opt_gt = OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS);
1967
1968     function_allocator lockalloc, globalloc;
1969
1970     if (self->locals.empty() && self->values.empty())
1971         return true;
1972
1973     globalloc.locals    = nullptr;
1974     globalloc.sizes     = nullptr;
1975     globalloc.positions = nullptr;
1976     globalloc.unique    = nullptr;
1977     lockalloc.locals    = nullptr;
1978     lockalloc.sizes     = nullptr;
1979     lockalloc.positions = nullptr;
1980     lockalloc.unique    = nullptr;
1981
1982     size_t i;
1983     for (i = 0; i < self->locals.size(); ++i)
1984     {
1985         ir_value *v = self->locals[i].get();
1986         if ((self->flags & IR_FLAG_MASK_NO_LOCAL_TEMPS) || !OPTS_OPTIMIZATION(OPTIM_LOCAL_TEMPS)) {
1987             v->locked      = true;
1988             v->unique_life = true;
1989         }
1990         else if (i >= vec_size(self->params))
1991             break;
1992         else
1993             v->locked = true; /* lock parameters locals */
1994         if (!function_allocator_alloc((v->locked || !opt_gt ? &lockalloc : &globalloc), v))
1995             goto error;
1996     }
1997     for (; i < self->locals.size(); ++i)
1998     {
1999         ir_value *v = self->locals[i].get();
2000         if (v->life.empty())
2001             continue;
2002         if (!ir_function_allocator_assign(self, (v->locked || !opt_gt ? &lockalloc : &globalloc), v))
2003             goto error;
2004     }
2005
2006     /* Allocate a slot for any value that still exists */
2007     for (i = 0; i < self->values.size(); ++i)
2008     {
2009         ir_value *v = self->values[i].get();
2010
2011         if (v->life.empty())
2012             continue;
2013
2014         /* CALL optimization:
2015          * If the value is a parameter-temp: 1 write, 1 read from a CALL
2016          * and it's not "locked", write it to the OFS_PARM directly.
2017          */
2018         if (OPTS_OPTIMIZATION(OPTIM_CALL_STORES) && !v->locked && !v->unique_life) {
2019             if (v->reads.size() == 1 && v->writes.size() == 1 &&
2020                 (v->reads[0]->opcode == VINSTR_NRCALL ||
2021                  (v->reads[0]->opcode >= INSTR_CALL0 && v->reads[0]->opcode <= INSTR_CALL8)
2022                 )
2023                )
2024             {
2025                 size_t param;
2026                 ir_instr *call = v->reads[0];
2027                 if (!vec_ir_value_find(call->params, v, &param)) {
2028                     irerror(call->context, "internal error: unlocked parameter %s not found", v->name.c_str());
2029                     goto error;
2030                 }
2031                 ++opts_optimizationcount[OPTIM_CALL_STORES];
2032                 v->callparam = true;
2033                 if (param < 8)
2034                     ir_value_code_setaddr(v, OFS_PARM0 + 3*param);
2035                 else {
2036                     size_t nprotos = self->owner->extparam_protos.size();
2037                     ir_value *ep;
2038                     param -= 8;
2039                     if (nprotos > param)
2040                         ep = self->owner->extparam_protos[param].get();
2041                     else
2042                     {
2043                         ep = ir_gen_extparam_proto(self->owner);
2044                         while (++nprotos <= param)
2045                             ep = ir_gen_extparam_proto(self->owner);
2046                     }
2047                     ir_instr_op(v->writes[0], 0, ep, true);
2048                     call->params[param+8] = ep;
2049                 }
2050                 continue;
2051             }
2052             if (v->writes.size() == 1 && v->writes[0]->opcode == INSTR_CALL0) {
2053                 v->store = store_return;
2054                 if (v->members[0]) v->members[0]->store = store_return;
2055                 if (v->members[1]) v->members[1]->store = store_return;
2056                 if (v->members[2]) v->members[2]->store = store_return;
2057                 ++opts_optimizationcount[OPTIM_CALL_STORES];
2058                 continue;
2059             }
2060         }
2061
2062         if (!ir_function_allocator_assign(self, (v->locked || !opt_gt ? &lockalloc : &globalloc), v))
2063             goto error;
2064     }
2065
2066     if (!lockalloc.sizes && !globalloc.sizes) {
2067         goto cleanup;
2068     }
2069     vec_push(lockalloc.positions, 0);
2070     vec_push(globalloc.positions, 0);
2071
2072     /* Adjust slot positions based on sizes */
2073     if (lockalloc.sizes) {
2074         pos = (vec_size(lockalloc.sizes) ? lockalloc.positions[0] : 0);
2075         for (i = 1; i < vec_size(lockalloc.sizes); ++i)
2076         {
2077             pos = lockalloc.positions[i-1] + lockalloc.sizes[i-1];
2078             vec_push(lockalloc.positions, pos);
2079         }
2080         self->allocated_locals = pos + vec_last(lockalloc.sizes);
2081     }
2082     if (globalloc.sizes) {
2083         pos = (vec_size(globalloc.sizes) ? globalloc.positions[0] : 0);
2084         for (i = 1; i < vec_size(globalloc.sizes); ++i)
2085         {
2086             pos = globalloc.positions[i-1] + globalloc.sizes[i-1];
2087             vec_push(globalloc.positions, pos);
2088         }
2089         self->globaltemps = pos + vec_last(globalloc.sizes);
2090     }
2091
2092     /* Locals need to know their new position */
2093     for (auto& local : self->locals) {
2094         if (local->locked || !opt_gt)
2095             local->code.local = lockalloc.positions[local->code.local];
2096         else
2097             local->code.local = globalloc.positions[local->code.local];
2098     }
2099     /* Take over the actual slot positions on values */
2100     for (auto& value : self->values) {
2101         if (value->locked || !opt_gt)
2102             value->code.local = lockalloc.positions[value->code.local];
2103         else
2104             value->code.local = globalloc.positions[value->code.local];
2105     }
2106
2107     goto cleanup;
2108
2109 error:
2110     retval = false;
2111 cleanup:
2112     for (i = 0; i < vec_size(lockalloc.locals); ++i)
2113         delete lockalloc.locals[i];
2114     for (i = 0; i < vec_size(globalloc.locals); ++i)
2115         delete globalloc.locals[i];
2116     vec_free(globalloc.unique);
2117     vec_free(globalloc.locals);
2118     vec_free(globalloc.sizes);
2119     vec_free(globalloc.positions);
2120     vec_free(lockalloc.unique);
2121     vec_free(lockalloc.locals);
2122     vec_free(lockalloc.sizes);
2123     vec_free(lockalloc.positions);
2124     return retval;
2125 }
2126
2127 /* Get information about which operand
2128  * is read from, or written to.
2129  */
2130 static void ir_op_read_write(int op, size_t *read, size_t *write)
2131 {
2132     switch (op)
2133     {
2134     case VINSTR_JUMP:
2135     case INSTR_GOTO:
2136         *write = 0;
2137         *read = 0;
2138         break;
2139     case INSTR_IF:
2140     case INSTR_IFNOT:
2141 #if 0
2142     case INSTR_IF_S:
2143     case INSTR_IFNOT_S:
2144 #endif
2145     case INSTR_RETURN:
2146     case VINSTR_COND:
2147         *write = 0;
2148         *read = 1;
2149         break;
2150     case INSTR_STOREP_F:
2151     case INSTR_STOREP_V:
2152     case INSTR_STOREP_S:
2153     case INSTR_STOREP_ENT:
2154     case INSTR_STOREP_FLD:
2155     case INSTR_STOREP_FNC:
2156         *write = 0;
2157         *read  = 7;
2158         break;
2159     default:
2160         *write = 1;
2161         *read = 6;
2162         break;
2163     };
2164 }
2165
2166 static bool ir_block_living_add_instr(ir_block *self, size_t eid) {
2167     bool changed = false;
2168     for (auto &it : self->living)
2169         if (ir_value_life_merge(it, eid))
2170             changed = true;
2171     return changed;
2172 }
2173
2174 static bool ir_block_living_lock(ir_block *self) {
2175     bool changed = false;
2176     for (auto &it : self->living) {
2177         if (it->locked)
2178             continue;
2179         it->locked = true;
2180         changed = true;
2181     }
2182     return changed;
2183 }
2184
2185 static bool ir_block_life_propagate(ir_block *self, bool *changed)
2186 {
2187     ir_instr *instr;
2188     ir_value *value;
2189     size_t i, o, p, mem;
2190     // bitmasks which operands are read from or written to
2191     size_t read, write;
2192
2193     self->living.clear();
2194
2195     p = vec_size(self->exits);
2196     for (i = 0; i < p; ++i) {
2197         ir_block *prev = self->exits[i];
2198         for (auto &it : prev->living)
2199             if (!vec_ir_value_find(self->living, it, nullptr))
2200                 self->living.push_back(it);
2201     }
2202
2203     i = vec_size(self->instr);
2204     while (i)
2205     { --i;
2206         instr = self->instr[i];
2207
2208         /* See which operands are read and write operands */
2209         ir_op_read_write(instr->opcode, &read, &write);
2210
2211         /* Go through the 3 main operands
2212          * writes first, then reads
2213          */
2214         for (o = 0; o < 3; ++o)
2215         {
2216             if (!instr->_ops[o]) /* no such operand */
2217                 continue;
2218
2219             value = instr->_ops[o];
2220
2221             /* We only care about locals */
2222             /* we also calculate parameter liferanges so that locals
2223              * can take up parameter slots */
2224             if (value->store != store_value &&
2225                 value->store != store_local &&
2226                 value->store != store_param)
2227                 continue;
2228
2229             /* write operands */
2230             /* When we write to a local, we consider it "dead" for the
2231              * remaining upper part of the function, since in SSA a value
2232              * can only be written once (== created)
2233              */
2234             if (write & (1<<o))
2235             {
2236                 size_t idx;
2237                 bool in_living = vec_ir_value_find(self->living, value, &idx);
2238                 if (!in_living)
2239                 {
2240                     /* If the value isn't alive it hasn't been read before... */
2241                     /* TODO: See if the warning can be emitted during parsing or AST processing
2242                      * otherwise have warning printed here.
2243                      * IF printing a warning here: include filecontext_t,
2244                      * and make sure it's only printed once
2245                      * since this function is run multiple times.
2246                      */
2247                     /* con_err( "Value only written %s\n", value->name); */
2248                     if (ir_value_life_merge(value, instr->eid))
2249                         *changed = true;
2250                 } else {
2251                     /* since 'living' won't contain it
2252                      * anymore, merge the value, since
2253                      * (A) doesn't.
2254                      */
2255                     if (ir_value_life_merge(value, instr->eid))
2256                         *changed = true;
2257                     // Then remove
2258                     self->living.erase(self->living.begin() + idx);
2259                 }
2260                 /* Removing a vector removes all members */
2261                 for (mem = 0; mem < 3; ++mem) {
2262                     if (value->members[mem] && vec_ir_value_find(self->living, value->members[mem], &idx)) {
2263                         if (ir_value_life_merge(value->members[mem], instr->eid))
2264                             *changed = true;
2265                         self->living.erase(self->living.begin() + idx);
2266                     }
2267                 }
2268                 /* Removing the last member removes the vector */
2269                 if (value->memberof) {
2270                     value = value->memberof;
2271                     for (mem = 0; mem < 3; ++mem) {
2272                         if (value->members[mem] && vec_ir_value_find(self->living, value->members[mem], nullptr))
2273                             break;
2274                     }
2275                     if (mem == 3 && vec_ir_value_find(self->living, value, &idx)) {
2276                         if (ir_value_life_merge(value, instr->eid))
2277                             *changed = true;
2278                         self->living.erase(self->living.begin() + idx);
2279                     }
2280                 }
2281             }
2282         }
2283
2284         /* These operations need a special case as they can break when using
2285          * same source and destination operand otherwise, as the engine may
2286          * read the source multiple times. */
2287         if (instr->opcode == INSTR_MUL_VF ||
2288             instr->opcode == VINSTR_BITAND_VF ||
2289             instr->opcode == VINSTR_BITOR_VF ||
2290             instr->opcode == VINSTR_BITXOR ||
2291             instr->opcode == VINSTR_BITXOR_VF ||
2292             instr->opcode == VINSTR_BITXOR_V ||
2293             instr->opcode == VINSTR_CROSS)
2294         {
2295             value = instr->_ops[2];
2296             /* the float source will get an additional lifetime */
2297             if (ir_value_life_merge(value, instr->eid+1))
2298                 *changed = true;
2299             if (value->memberof && ir_value_life_merge(value->memberof, instr->eid+1))
2300                 *changed = true;
2301         }
2302
2303         if (instr->opcode == INSTR_MUL_FV ||
2304             instr->opcode == INSTR_LOAD_V ||
2305             instr->opcode == VINSTR_BITXOR ||
2306             instr->opcode == VINSTR_BITXOR_VF ||
2307             instr->opcode == VINSTR_BITXOR_V ||
2308             instr->opcode == VINSTR_CROSS)
2309         {
2310             value = instr->_ops[1];
2311             /* the float source will get an additional lifetime */
2312             if (ir_value_life_merge(value, instr->eid+1))
2313                 *changed = true;
2314             if (value->memberof && ir_value_life_merge(value->memberof, instr->eid+1))
2315                 *changed = true;
2316         }
2317
2318         for (o = 0; o < 3; ++o)
2319         {
2320             if (!instr->_ops[o]) /* no such operand */
2321                 continue;
2322
2323             value = instr->_ops[o];
2324
2325             /* We only care about locals */
2326             /* we also calculate parameter liferanges so that locals
2327              * can take up parameter slots */
2328             if (value->store != store_value &&
2329                 value->store != store_local &&
2330                 value->store != store_param)
2331                 continue;
2332
2333             /* read operands */
2334             if (read & (1<<o))
2335             {
2336                 if (!vec_ir_value_find(self->living, value, nullptr))
2337                     self->living.push_back(value);
2338                 /* reading adds the full vector */
2339                 if (value->memberof && !vec_ir_value_find(self->living, value->memberof, nullptr))
2340                     self->living.push_back(value->memberof);
2341                 for (mem = 0; mem < 3; ++mem) {
2342                     if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], nullptr))
2343                         self->living.push_back(value->members[mem]);
2344                 }
2345             }
2346         }
2347         /* PHI operands are always read operands */
2348         for (auto &it : instr->phi) {
2349             value = it.value;
2350             if (!vec_ir_value_find(self->living, value, nullptr))
2351                 self->living.push_back(value);
2352             /* reading adds the full vector */
2353             if (value->memberof && !vec_ir_value_find(self->living, value->memberof, nullptr))
2354                 self->living.push_back(value->memberof);
2355             for (mem = 0; mem < 3; ++mem) {
2356                 if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], nullptr))
2357                     self->living.push_back(value->members[mem]);
2358             }
2359         }
2360
2361         /* on a call, all these values must be "locked" */
2362         if (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8) {
2363             if (ir_block_living_lock(self))
2364                 *changed = true;
2365         }
2366         /* call params are read operands too */
2367         for (auto &it : instr->params) {
2368             value = it;
2369             if (!vec_ir_value_find(self->living, value, nullptr))
2370                 self->living.push_back(value);
2371             /* reading adds the full vector */
2372             if (value->memberof && !vec_ir_value_find(self->living, value->memberof, nullptr))
2373                 self->living.push_back(value->memberof);
2374             for (mem = 0; mem < 3; ++mem) {
2375                 if (value->members[mem] && !vec_ir_value_find(self->living, value->members[mem], nullptr))
2376                     self->living.push_back(value->members[mem]);
2377             }
2378         }
2379
2380         /* (A) */
2381         if (ir_block_living_add_instr(self, instr->eid))
2382             *changed = true;
2383     }
2384     /* the "entry" instruction ID */
2385     if (ir_block_living_add_instr(self, self->entry_id))
2386         *changed = true;
2387
2388     return true;
2389 }
2390
2391 bool ir_function_calculate_liferanges(ir_function *self)
2392 {
2393     /* parameters live at 0 */
2394     for (size_t i = 0; i < vec_size(self->params); ++i)
2395         if (!ir_value_life_merge(self->locals[i].get(), 0))
2396             compile_error(self->context, "internal error: failed value-life merging");
2397
2398     bool changed;
2399     do {
2400         self->run_id++;
2401         changed = false;
2402         for (auto i = self->blocks.rbegin(); i != self->blocks.rend(); ++i)
2403             ir_block_life_propagate(i->get(), &changed);
2404     } while (changed);
2405
2406     if (self->blocks.size()) {
2407         ir_block *block = self->blocks[0].get();
2408         for (auto &it : block->living) {
2409             ir_value *v = it;
2410             if (v->store != store_local)
2411                 continue;
2412             if (v->vtype == TYPE_VECTOR)
2413                 continue;
2414             self->flags |= IR_FLAG_HAS_UNINITIALIZED;
2415             /* find the instruction reading from it */
2416             size_t s = 0;
2417             for (; s < v->reads.size(); ++s) {
2418                 if (v->reads[s]->eid == v->life[0].end)
2419                     break;
2420             }
2421             if (s < v->reads.size()) {
2422                 if (irwarning(v->context, WARN_USED_UNINITIALIZED,
2423                               "variable `%s` may be used uninitialized in this function\n"
2424                               " -> %s:%i",
2425                               v->name.c_str(),
2426                               v->reads[s]->context.file, v->reads[s]->context.line)
2427                    )
2428                 {
2429                     return false;
2430                 }
2431                 continue;
2432             }
2433             if (v->memberof) {
2434                 ir_value *vec = v->memberof;
2435                 for (s = 0; s < vec->reads.size(); ++s) {
2436                     if (vec->reads[s]->eid == v->life[0].end)
2437                         break;
2438                 }
2439                 if (s < vec->reads.size()) {
2440                     if (irwarning(v->context, WARN_USED_UNINITIALIZED,
2441                                   "variable `%s` may be used uninitialized in this function\n"
2442                                   " -> %s:%i",
2443                                   v->name.c_str(),
2444                                   vec->reads[s]->context.file, vec->reads[s]->context.line)
2445                        )
2446                     {
2447                         return false;
2448                     }
2449                     continue;
2450                 }
2451             }
2452             if (irwarning(v->context, WARN_USED_UNINITIALIZED,
2453                           "variable `%s` may be used uninitialized in this function", v->name.c_str()))
2454             {
2455                 return false;
2456             }
2457         }
2458     }
2459     return true;
2460 }
2461
2462 /***********************************************************************
2463  *IR Code-Generation
2464  *
2465  * Since the IR has the convention of putting 'write' operands
2466  * at the beginning, we have to rotate the operands of instructions
2467  * properly in order to generate valid QCVM code.
2468  *
2469  * Having destinations at a fixed position is more convenient. In QC
2470  * this is *mostly* OPC,  but FTE adds at least 2 instructions which
2471  * read from from OPA,  and store to OPB rather than OPC.   Which is
2472  * partially the reason why the implementation of these instructions
2473  * in darkplaces has been delayed for so long.
2474  *
2475  * Breaking conventions is annoying...
2476  */
2477 static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal);
2478
2479 static bool gen_global_field(code_t *code, ir_value *global)
2480 {
2481     if (global->hasvalue)
2482     {
2483         ir_value *fld = global->constval.vpointer;
2484         if (!fld) {
2485             irerror(global->context, "Invalid field constant with no field: %s", global->name.c_str());
2486             return false;
2487         }
2488
2489         /* copy the field's value */
2490         ir_value_code_setaddr(global, code->globals.size());
2491         code->globals.push_back(fld->code.fieldaddr);
2492         if (global->fieldtype == TYPE_VECTOR) {
2493             code->globals.push_back(fld->code.fieldaddr+1);
2494             code->globals.push_back(fld->code.fieldaddr+2);
2495         }
2496     }
2497     else
2498     {
2499         ir_value_code_setaddr(global, code->globals.size());
2500         code->globals.push_back(0);
2501         if (global->fieldtype == TYPE_VECTOR) {
2502             code->globals.push_back(0);
2503             code->globals.push_back(0);
2504         }
2505     }
2506     if (global->code.globaladdr < 0)
2507         return false;
2508     return true;
2509 }
2510
2511 static bool gen_global_pointer(code_t *code, ir_value *global)
2512 {
2513     if (global->hasvalue)
2514     {
2515         ir_value *target = global->constval.vpointer;
2516         if (!target) {
2517             irerror(global->context, "Invalid pointer constant: %s", global->name.c_str());
2518             /* nullptr pointers are pointing to the nullptr constant, which also
2519              * sits at address 0, but still has an ir_value for itself.
2520              */
2521             return false;
2522         }
2523
2524         /* Here, relocations ARE possible - in fteqcc-enhanced-qc:
2525          * void() foo; <- proto
2526          * void() *fooptr = &foo;
2527          * void() foo = { code }
2528          */
2529         if (!target->code.globaladdr) {
2530             /* FIXME: Check for the constant nullptr ir_value!
2531              * because then code.globaladdr being 0 is valid.
2532              */
2533             irerror(global->context, "FIXME: Relocation support");
2534             return false;
2535         }
2536
2537         ir_value_code_setaddr(global, code->globals.size());
2538         code->globals.push_back(target->code.globaladdr);
2539     }
2540     else
2541     {
2542         ir_value_code_setaddr(global, code->globals.size());
2543         code->globals.push_back(0);
2544     }
2545     if (global->code.globaladdr < 0)
2546         return false;
2547     return true;
2548 }
2549
2550 static bool gen_blocks_recursive(code_t *code, ir_function *func, ir_block *block)
2551 {
2552     prog_section_statement_t stmt;
2553     ir_instr *instr;
2554     ir_block *target;
2555     ir_block *ontrue;
2556     ir_block *onfalse;
2557     size_t    stidx;
2558     size_t    i;
2559     int       j;
2560
2561     block->generated = true;
2562     block->code_start = code->statements.size();
2563     for (i = 0; i < vec_size(block->instr); ++i)
2564     {
2565         instr = block->instr[i];
2566
2567         if (instr->opcode == VINSTR_PHI) {
2568             irerror(block->context, "cannot generate virtual instruction (phi)");
2569             return false;
2570         }
2571
2572         if (instr->opcode == VINSTR_JUMP) {
2573             target = instr->bops[0];
2574             /* for uncoditional jumps, if the target hasn't been generated
2575              * yet, we generate them right here.
2576              */
2577             if (!target->generated)
2578                 return gen_blocks_recursive(code, func, target);
2579
2580             /* otherwise we generate a jump instruction */
2581             stmt.opcode = INSTR_GOTO;
2582             stmt.o1.s1 = target->code_start - code->statements.size();
2583             stmt.o2.s1 = 0;
2584             stmt.o3.s1 = 0;
2585             if (stmt.o1.s1 != 1)
2586                 code_push_statement(code, &stmt, instr->context);
2587
2588             /* no further instructions can be in this block */
2589             return true;
2590         }
2591
2592         if (instr->opcode == VINSTR_BITXOR) {
2593             stmt.opcode = INSTR_BITOR;
2594             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2595             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2596             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2597             code_push_statement(code, &stmt, instr->context);
2598             stmt.opcode = INSTR_BITAND;
2599             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2600             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2601             stmt.o3.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]);
2602             code_push_statement(code, &stmt, instr->context);
2603             stmt.opcode = INSTR_SUB_F;
2604             stmt.o1.s1 = ir_value_code_addr(instr->_ops[0]);
2605             stmt.o2.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]);
2606             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2607             code_push_statement(code, &stmt, instr->context);
2608
2609             /* instruction generated */
2610             continue;
2611         }
2612
2613         if (instr->opcode == VINSTR_BITAND_V) {
2614             stmt.opcode = INSTR_BITAND;
2615             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2616             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2617             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2618             code_push_statement(code, &stmt, instr->context);
2619             ++stmt.o1.s1;
2620             ++stmt.o2.s1;
2621             ++stmt.o3.s1;
2622             code_push_statement(code, &stmt, instr->context);
2623             ++stmt.o1.s1;
2624             ++stmt.o2.s1;
2625             ++stmt.o3.s1;
2626             code_push_statement(code, &stmt, instr->context);
2627
2628             /* instruction generated */
2629             continue;
2630         }
2631
2632         if (instr->opcode == VINSTR_BITOR_V) {
2633             stmt.opcode = INSTR_BITOR;
2634             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2635             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2636             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2637             code_push_statement(code, &stmt, instr->context);
2638             ++stmt.o1.s1;
2639             ++stmt.o2.s1;
2640             ++stmt.o3.s1;
2641             code_push_statement(code, &stmt, instr->context);
2642             ++stmt.o1.s1;
2643             ++stmt.o2.s1;
2644             ++stmt.o3.s1;
2645             code_push_statement(code, &stmt, instr->context);
2646
2647             /* instruction generated */
2648             continue;
2649         }
2650
2651         if (instr->opcode == VINSTR_BITXOR_V) {
2652             for (j = 0; j < 3; ++j) {
2653                 stmt.opcode = INSTR_BITOR;
2654                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + j;
2655                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]) + j;
2656                 stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]) + j;
2657                 code_push_statement(code, &stmt, instr->context);
2658                 stmt.opcode = INSTR_BITAND;
2659                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + j;
2660                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]) + j;
2661                 stmt.o3.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]) + j;
2662                 code_push_statement(code, &stmt, instr->context);
2663             }
2664             stmt.opcode = INSTR_SUB_V;
2665             stmt.o1.s1 = ir_value_code_addr(instr->_ops[0]);
2666             stmt.o2.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]);
2667             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2668             code_push_statement(code, &stmt, instr->context);
2669
2670             /* instruction generated */
2671             continue;
2672         }
2673
2674         if (instr->opcode == VINSTR_BITAND_VF) {
2675             stmt.opcode = INSTR_BITAND;
2676             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2677             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2678             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2679             code_push_statement(code, &stmt, instr->context);
2680             ++stmt.o1.s1;
2681             ++stmt.o3.s1;
2682             code_push_statement(code, &stmt, instr->context);
2683             ++stmt.o1.s1;
2684             ++stmt.o3.s1;
2685             code_push_statement(code, &stmt, instr->context);
2686
2687             /* instruction generated */
2688             continue;
2689         }
2690
2691         if (instr->opcode == VINSTR_BITOR_VF) {
2692             stmt.opcode = INSTR_BITOR;
2693             stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]);
2694             stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2695             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2696             code_push_statement(code, &stmt, instr->context);
2697             ++stmt.o1.s1;
2698             ++stmt.o3.s1;
2699             code_push_statement(code, &stmt, instr->context);
2700             ++stmt.o1.s1;
2701             ++stmt.o3.s1;
2702             code_push_statement(code, &stmt, instr->context);
2703
2704             /* instruction generated */
2705             continue;
2706         }
2707
2708         if (instr->opcode == VINSTR_BITXOR_VF) {
2709             for (j = 0; j < 3; ++j) {
2710                 stmt.opcode = INSTR_BITOR;
2711                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + j;
2712                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2713                 stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]) + j;
2714                 code_push_statement(code, &stmt, instr->context);
2715                 stmt.opcode = INSTR_BITAND;
2716                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + j;
2717                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]);
2718                 stmt.o3.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]) + j;
2719                 code_push_statement(code, &stmt, instr->context);
2720             }
2721             stmt.opcode = INSTR_SUB_V;
2722             stmt.o1.s1 = ir_value_code_addr(instr->_ops[0]);
2723             stmt.o2.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]);
2724             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2725             code_push_statement(code, &stmt, instr->context);
2726
2727             /* instruction generated */
2728             continue;
2729         }
2730
2731         if (instr->opcode == VINSTR_CROSS) {
2732             stmt.opcode = INSTR_MUL_F;
2733             for (j = 0; j < 3; ++j) {
2734                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + (j + 1) % 3;
2735                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]) + (j + 2) % 3;
2736                 stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]) + j;
2737                 code_push_statement(code, &stmt, instr->context);
2738                 stmt.o1.s1 = ir_value_code_addr(instr->_ops[1]) + (j + 2) % 3;
2739                 stmt.o2.s1 = ir_value_code_addr(instr->_ops[2]) + (j + 1) % 3;
2740                 stmt.o3.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]) + j;
2741                 code_push_statement(code, &stmt, instr->context);
2742             }
2743             stmt.opcode = INSTR_SUB_V;
2744             stmt.o1.s1 = ir_value_code_addr(instr->_ops[0]);
2745             stmt.o2.s1 = ir_value_code_addr(func->owner->vinstr_temp[0]);
2746             stmt.o3.s1 = ir_value_code_addr(instr->_ops[0]);
2747             code_push_statement(code, &stmt, instr->context);
2748
2749             /* instruction generated */
2750             continue;
2751         }
2752
2753         if (instr->opcode == VINSTR_COND) {
2754             ontrue  = instr->bops[0];
2755             onfalse = instr->bops[1];
2756             /* TODO: have the AST signal which block should
2757              * come first: eg. optimize IFs without ELSE...
2758              */
2759
2760             stmt.o1.u1 = ir_value_code_addr(instr->_ops[0]);
2761             stmt.o2.u1 = 0;
2762             stmt.o3.s1 = 0;
2763
2764             if (ontrue->generated) {
2765                 stmt.opcode = INSTR_IF;
2766                 stmt.o2.s1 = ontrue->code_start - code->statements.size();
2767                 if (stmt.o2.s1 != 1)
2768                     code_push_statement(code, &stmt, instr->context);
2769             }
2770             if (onfalse->generated) {
2771                 stmt.opcode = INSTR_IFNOT;
2772                 stmt.o2.s1 = onfalse->code_start - code->statements.size();
2773                 if (stmt.o2.s1 != 1)
2774                     code_push_statement(code, &stmt, instr->context);
2775             }
2776             if (!ontrue->generated) {
2777                 if (onfalse->generated)
2778                     return gen_blocks_recursive(code, func, ontrue);
2779             }
2780             if (!onfalse->generated) {
2781                 if (ontrue->generated)
2782                     return gen_blocks_recursive(code, func, onfalse);
2783             }
2784             /* neither ontrue nor onfalse exist */
2785             stmt.opcode = INSTR_IFNOT;
2786             if (!instr->likely) {
2787                 /* Honor the likelyhood hint */
2788                 ir_block *tmp = onfalse;
2789                 stmt.opcode = INSTR_IF;
2790                 onfalse = ontrue;
2791                 ontrue = tmp;
2792             }
2793             stidx = code->statements.size();
2794             code_push_statement(code, &stmt, instr->context);
2795             /* on false we jump, so add ontrue-path */
2796             if (!gen_blocks_recursive(code, func, ontrue))
2797                 return false;
2798             /* fixup the jump address */
2799             code->statements[stidx].o2.s1 = code->statements.size() - stidx;
2800             /* generate onfalse path */
2801             if (onfalse->generated) {
2802                 /* fixup the jump address */
2803                 code->statements[stidx].o2.s1 = onfalse->code_start - stidx;
2804                 if (stidx+2 == code->statements.size() && code->statements[stidx].o2.s1 == 1) {
2805                     code->statements[stidx] = code->statements[stidx+1];
2806                     if (code->statements[stidx].o1.s1 < 0)
2807                         code->statements[stidx].o1.s1++;
2808                     code_pop_statement(code);
2809                 }
2810                 stmt.opcode = code->statements.back().opcode;
2811                 if (stmt.opcode == INSTR_GOTO ||
2812                     stmt.opcode == INSTR_IF ||
2813                     stmt.opcode == INSTR_IFNOT ||
2814                     stmt.opcode == INSTR_RETURN ||
2815                     stmt.opcode == INSTR_DONE)
2816                 {
2817                     /* no use jumping from here */
2818                     return true;
2819                 }
2820                 /* may have been generated in the previous recursive call */
2821                 stmt.opcode = INSTR_GOTO;
2822                 stmt.o1.s1 = onfalse->code_start - code->statements.size();
2823                 stmt.o2.s1 = 0;
2824                 stmt.o3.s1 = 0;
2825                 if (stmt.o1.s1 != 1)
2826                     code_push_statement(code, &stmt, instr->context);
2827                 return true;
2828             }
2829             else if (stidx+2 == code->statements.size() && code->statements[stidx].o2.s1 == 1) {
2830                 code->statements[stidx] = code->statements[stidx+1];
2831                 if (code->statements[stidx].o1.s1 < 0)
2832                     code->statements[stidx].o1.s1++;
2833                 code_pop_statement(code);
2834             }
2835             /* if not, generate now */
2836             return gen_blocks_recursive(code, func, onfalse);
2837         }
2838
2839         if ( (instr->opcode >= INSTR_CALL0 && instr->opcode <= INSTR_CALL8)
2840            || instr->opcode == VINSTR_NRCALL)
2841         {
2842             size_t p, first;
2843             ir_value *retvalue;
2844
2845             first = instr->params.size();
2846             if (first > 8)
2847                 first = 8;
2848             for (p = 0; p < first; ++p)
2849             {
2850                 ir_value *param = instr->params[p];
2851                 if (param->callparam)
2852                     continue;
2853
2854                 stmt.opcode = INSTR_STORE_F;
2855                 stmt.o3.u1 = 0;
2856
2857                 if (param->vtype == TYPE_FIELD)
2858                     stmt.opcode = field_store_instr[param->fieldtype];
2859                 else if (param->vtype == TYPE_NIL)
2860                     stmt.opcode = INSTR_STORE_V;
2861                 else
2862                     stmt.opcode = type_store_instr[param->vtype];
2863                 stmt.o1.u1 = ir_value_code_addr(param);
2864                 stmt.o2.u1 = OFS_PARM0 + 3 * p;
2865
2866                 if (param->vtype == TYPE_VECTOR && (param->flags & IR_FLAG_SPLIT_VECTOR)) {
2867                     /* fetch 3 separate floats */
2868                     stmt.opcode = INSTR_STORE_F;
2869                     stmt.o1.u1 = ir_value_code_addr(param->members[0]);
2870                     code_push_statement(code, &stmt, instr->context);
2871                     stmt.o2.u1++;
2872                     stmt.o1.u1 = ir_value_code_addr(param->members[1]);
2873                     code_push_statement(code, &stmt, instr->context);
2874                     stmt.o2.u1++;
2875                     stmt.o1.u1 = ir_value_code_addr(param->members[2]);
2876                     code_push_statement(code, &stmt, instr->context);
2877                 }
2878                 else
2879                     code_push_statement(code, &stmt, instr->context);
2880             }
2881             /* Now handle extparams */
2882             first = instr->params.size();
2883             for (; p < first; ++p)
2884             {
2885                 ir_builder *ir = func->owner;
2886                 ir_value *param = instr->params[p];
2887                 ir_value *targetparam;
2888
2889                 if (param->callparam)
2890                     continue;
2891
2892                 if (p-8 >= ir->extparams.size())
2893                     ir_gen_extparam(ir);
2894
2895                 targetparam = ir->extparams[p-8];
2896
2897                 stmt.opcode = INSTR_STORE_F;
2898                 stmt.o3.u1 = 0;
2899
2900                 if (param->vtype == TYPE_FIELD)
2901                     stmt.opcode = field_store_instr[param->fieldtype];
2902                 else if (param->vtype == TYPE_NIL)
2903                     stmt.opcode = INSTR_STORE_V;
2904                 else
2905                     stmt.opcode = type_store_instr[param->vtype];
2906                 stmt.o1.u1 = ir_value_code_addr(param);
2907                 stmt.o2.u1 = ir_value_code_addr(targetparam);
2908                 if (param->vtype == TYPE_VECTOR && (param->flags & IR_FLAG_SPLIT_VECTOR)) {
2909                     /* fetch 3 separate floats */
2910                     stmt.opcode = INSTR_STORE_F;
2911                     stmt.o1.u1 = ir_value_code_addr(param->members[0]);
2912                     code_push_statement(code, &stmt, instr->context);
2913                     stmt.o2.u1++;
2914                     stmt.o1.u1 = ir_value_code_addr(param->members[1]);
2915                     code_push_statement(code, &stmt, instr->context);
2916                     stmt.o2.u1++;
2917                     stmt.o1.u1 = ir_value_code_addr(param->members[2]);
2918                     code_push_statement(code, &stmt, instr->context);
2919                 }
2920                 else
2921                     code_push_statement(code, &stmt, instr->context);
2922             }
2923
2924             stmt.opcode = INSTR_CALL0 + instr->params.size();
2925             if (stmt.opcode > INSTR_CALL8)
2926                 stmt.opcode = INSTR_CALL8;
2927             stmt.o1.u1 = ir_value_code_addr(instr->_ops[1]);
2928             stmt.o2.u1 = 0;
2929             stmt.o3.u1 = 0;
2930             code_push_statement(code, &stmt, instr->context);
2931
2932             retvalue = instr->_ops[0];
2933             if (retvalue && retvalue->store != store_return &&
2934                 (retvalue->store == store_global || retvalue->life.size()))
2935             {
2936                 /* not to be kept in OFS_RETURN */
2937                 if (retvalue->vtype == TYPE_FIELD && OPTS_FLAG(ADJUST_VECTOR_FIELDS))
2938                     stmt.opcode = field_store_instr[retvalue->fieldtype];
2939                 else
2940                     stmt.opcode = type_store_instr[retvalue->vtype];
2941                 stmt.o1.u1 = OFS_RETURN;
2942                 stmt.o2.u1 = ir_value_code_addr(retvalue);
2943                 stmt.o3.u1 = 0;
2944                 code_push_statement(code, &stmt, instr->context);
2945             }
2946             continue;
2947         }
2948
2949         if (instr->opcode == INSTR_STATE) {
2950             stmt.opcode = instr->opcode;
2951             if (instr->_ops[0])
2952                 stmt.o1.u1 = ir_value_code_addr(instr->_ops[0]);
2953             if (instr->_ops[1])
2954                 stmt.o2.u1 = ir_value_code_addr(instr->_ops[1]);
2955             stmt.o3.u1 = 0;
2956             code_push_statement(code, &stmt, instr->context);
2957             continue;
2958         }
2959
2960         stmt.opcode = instr->opcode;
2961         stmt.o1.u1 = 0;
2962         stmt.o2.u1 = 0;
2963         stmt.o3.u1 = 0;
2964
2965         /* This is the general order of operands */
2966         if (instr->_ops[0])
2967             stmt.o3.u1 = ir_value_code_addr(instr->_ops[0]);
2968
2969         if (instr->_ops[1])
2970             stmt.o1.u1 = ir_value_code_addr(instr->_ops[1]);
2971
2972         if (instr->_ops[2])
2973             stmt.o2.u1 = ir_value_code_addr(instr->_ops[2]);
2974
2975         if (stmt.opcode == INSTR_RETURN || stmt.opcode == INSTR_DONE)
2976         {
2977             stmt.o1.u1 = stmt.o3.u1;
2978             stmt.o3.u1 = 0;
2979         }
2980         else if ((stmt.opcode >= INSTR_STORE_F &&
2981                   stmt.opcode <= INSTR_STORE_FNC) ||
2982                  (stmt.opcode >= INSTR_STOREP_F &&
2983                   stmt.opcode <= INSTR_STOREP_FNC))
2984         {
2985             /* 2-operand instructions with A -> B */
2986             stmt.o2.u1 = stmt.o3.u1;
2987             stmt.o3.u1 = 0;
2988
2989             /* tiny optimization, don't output
2990              * STORE a, a
2991              */
2992             if (stmt.o2.u1 == stmt.o1.u1 &&
2993                 OPTS_OPTIMIZATION(OPTIM_PEEPHOLE))
2994             {
2995                 ++opts_optimizationcount[OPTIM_PEEPHOLE];
2996                 continue;
2997             }
2998         }
2999         code_push_statement(code, &stmt, instr->context);
3000     }
3001     return true;
3002 }
3003
3004 static bool gen_function_code(code_t *code, ir_function *self)
3005 {
3006     ir_block *block;
3007     prog_section_statement_t stmt, *retst;
3008
3009     /* Starting from entry point, we generate blocks "as they come"
3010      * for now. Dead blocks will not be translated obviously.
3011      */
3012     if (self->blocks.empty()) {
3013         irerror(self->context, "Function '%s' declared without body.", self->name.c_str());
3014         return false;
3015     }
3016
3017     block = self->blocks[0].get();
3018     if (block->generated)
3019         return true;
3020
3021     if (!gen_blocks_recursive(code, self, block)) {
3022         irerror(self->context, "failed to generate blocks for '%s'", self->name.c_str());
3023         return false;
3024     }
3025
3026     /* code_write and qcvm -disasm need to know that the function ends here */
3027     retst = &code->statements.back();
3028     if (OPTS_OPTIMIZATION(OPTIM_VOID_RETURN) &&
3029         self->outtype == TYPE_VOID &&
3030         retst->opcode == INSTR_RETURN &&
3031         !retst->o1.u1 && !retst->o2.u1 && !retst->o3.u1)
3032     {
3033         retst->opcode = INSTR_DONE;
3034         ++opts_optimizationcount[OPTIM_VOID_RETURN];
3035     } else {
3036         lex_ctx_t last;
3037
3038         stmt.opcode = INSTR_DONE;
3039         stmt.o1.u1  = 0;
3040         stmt.o2.u1  = 0;
3041         stmt.o3.u1  = 0;
3042         last.line   = code->linenums.back();
3043         last.column = code->columnnums.back();
3044
3045         code_push_statement(code, &stmt, last);
3046     }
3047     return true;
3048 }
3049
3050 static qcint_t ir_builder_filestring(ir_builder *ir, const char *filename)
3051 {
3052     /* NOTE: filename pointers are copied, we never strdup them,
3053      * thus we can use pointer-comparison to find the string.
3054      */
3055     qcint_t  str;
3056
3057     for (size_t i = 0; i != ir->filenames.size(); ++i) {
3058         if (!strcmp(ir->filenames[i], filename))
3059             return i;
3060     }
3061
3062     str = code_genstring(ir->code.get(), filename);
3063     ir->filenames.push_back(filename);
3064     ir->filestrings.push_back(str);
3065     return str;
3066 }
3067
3068 static bool gen_global_function(ir_builder *ir, ir_value *global)
3069 {
3070     prog_section_function_t fun;
3071     ir_function            *irfun;
3072
3073     size_t i;
3074
3075     if (!global->hasvalue || (!global->constval.vfunc)) {
3076         irerror(global->context, "Invalid state of function-global: not constant: %s", global->name.c_str());
3077         return false;
3078     }
3079
3080     irfun = global->constval.vfunc;
3081     fun.name = global->code.name;
3082     fun.file = ir_builder_filestring(ir, global->context.file);
3083     fun.profile = 0; /* always 0 */
3084     fun.nargs = vec_size(irfun->params);
3085     if (fun.nargs > 8)
3086         fun.nargs = 8;
3087
3088     for (i = 0; i < 8; ++i) {
3089         if ((int32_t)i >= fun.nargs)
3090             fun.argsize[i] = 0;
3091         else
3092             fun.argsize[i] = type_sizeof_[irfun->params[i]];
3093     }
3094
3095     fun.firstlocal = 0;
3096     fun.locals = irfun->allocated_locals;
3097
3098     if (irfun->builtin)
3099         fun.entry = irfun->builtin+1;
3100     else {
3101         irfun->code_function_def = ir->code->functions.size();
3102         fun.entry = ir->code->statements.size();
3103     }
3104
3105     ir->code->functions.push_back(fun);
3106     return true;
3107 }
3108
3109 static ir_value* ir_gen_extparam_proto(ir_builder *ir)
3110 {
3111     char      name[128];
3112
3113     util_snprintf(name, sizeof(name), "EXTPARM#%i", (int)(ir->extparam_protos.size()));
3114     ir_value *global = new ir_value(name, store_global, TYPE_VECTOR);
3115     ir->extparam_protos.emplace_back(global);
3116
3117     return global;
3118 }
3119
3120 static void ir_gen_extparam(ir_builder *ir)
3121 {
3122     prog_section_def_t def;
3123     ir_value          *global;
3124
3125     if (ir->extparam_protos.size() < ir->extparams.size()+1)
3126         global = ir_gen_extparam_proto(ir);
3127     else
3128         global = ir->extparam_protos[ir->extparams.size()].get();
3129
3130     def.name = code_genstring(ir->code.get(), global->name.c_str());
3131     def.type = TYPE_VECTOR;
3132     def.offset = ir->code->globals.size();
3133
3134     ir->code->defs.push_back(def);
3135
3136     ir_value_code_setaddr(global, def.offset);
3137
3138     ir->code->globals.push_back(0);
3139     ir->code->globals.push_back(0);
3140     ir->code->globals.push_back(0);
3141
3142     ir->extparams.emplace_back(global);
3143 }
3144
3145 static bool gen_function_extparam_copy(code_t *code, ir_function *self)
3146 {
3147     ir_builder *ir = self->owner;
3148
3149     size_t numparams = vec_size(self->params);
3150     if (!numparams)
3151         return true;
3152
3153     prog_section_statement_t stmt;
3154     stmt.opcode = INSTR_STORE_F;
3155     stmt.o3.s1 = 0;
3156     for (size_t i = 8; i < numparams; ++i) {
3157         size_t ext = i - 8;
3158         if (ext >= ir->extparams.size())
3159             ir_gen_extparam(ir);
3160
3161         ir_value *ep = ir->extparams[ext];
3162
3163         stmt.opcode = type_store_instr[self->locals[i]->vtype];
3164         if (self->locals[i]->vtype == TYPE_FIELD &&
3165             self->locals[i]->fieldtype == TYPE_VECTOR)
3166         {
3167             stmt.opcode = INSTR_STORE_V;
3168         }
3169         stmt.o1.u1 = ir_value_code_addr(ep);
3170         stmt.o2.u1 = ir_value_code_addr(self->locals[i].get());
3171         code_push_statement(code, &stmt, self->context);
3172     }
3173
3174     return true;
3175 }
3176
3177 static bool gen_function_varargs_copy(code_t *code, ir_function *self)
3178 {
3179     size_t i, ext, numparams, maxparams;
3180
3181     ir_builder *ir = self->owner;
3182     ir_value   *ep;
3183     prog_section_statement_t stmt;
3184
3185     numparams = vec_size(self->params);
3186     if (!numparams)
3187         return true;
3188
3189     stmt.opcode = INSTR_STORE_V;
3190     stmt.o3.s1 = 0;
3191     maxparams = numparams + self->max_varargs;
3192     for (i = numparams; i < maxparams; ++i) {
3193         if (i < 8) {
3194             stmt.o1.u1 = OFS_PARM0 + 3*i;
3195             stmt.o2.u1 = ir_value_code_addr(self->locals[i].get());
3196             code_push_statement(code, &stmt, self->context);
3197             continue;
3198         }
3199         ext = i - 8;
3200         while (ext >= ir->extparams.size())
3201             ir_gen_extparam(ir);
3202
3203         ep = ir->extparams[ext];
3204
3205         stmt.o1.u1 = ir_value_code_addr(ep);
3206         stmt.o2.u1 = ir_value_code_addr(self->locals[i].get());
3207         code_push_statement(code, &stmt, self->context);
3208     }
3209
3210     return true;
3211 }
3212
3213 static bool gen_function_locals(ir_builder *ir, ir_value *global)
3214 {
3215     prog_section_function_t *def;
3216     ir_function             *irfun;
3217     uint32_t                 firstlocal, firstglobal;
3218
3219     irfun = global->constval.vfunc;
3220     def   = &ir->code->functions[0] + irfun->code_function_def;
3221
3222     if (OPTS_OPTION_BOOL(OPTION_G) ||
3223         !OPTS_OPTIMIZATION(OPTIM_OVERLAP_LOCALS)        ||
3224         (irfun->flags & IR_FLAG_MASK_NO_OVERLAP))
3225     {
3226         firstlocal = def->firstlocal = ir->code->globals.size();
3227     } else {
3228         firstlocal = def->firstlocal = ir->first_common_local;
3229         ++opts_optimizationcount[OPTIM_OVERLAP_LOCALS];
3230     }
3231
3232     firstglobal = (OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS) ? ir->first_common_globaltemp : firstlocal);
3233
3234     for (size_t i = ir->code->globals.size(); i < firstlocal + irfun->allocated_locals; ++i)
3235         ir->code->globals.push_back(0);
3236
3237     for (auto& lp : irfun->locals) {
3238         ir_value *v = lp.get();
3239         if (v->locked || !OPTS_OPTIMIZATION(OPTIM_GLOBAL_TEMPS)) {
3240             ir_value_code_setaddr(v, firstlocal + v->code.local);
3241             if (!ir_builder_gen_global(ir, v, true)) {
3242                 irerror(v->context, "failed to generate local %s", v->name.c_str());
3243                 return false;
3244             }
3245         }
3246         else
3247             ir_value_code_setaddr(v, firstglobal + v->code.local);
3248     }
3249     for (auto& vp : irfun->values) {
3250         ir_value *v = vp.get();
3251         if (v->callparam)
3252             continue;
3253         if (v->locked)
3254             ir_value_code_setaddr(v, firstlocal + v->code.local);
3255         else
3256             ir_value_code_setaddr(v, firstglobal + v->code.local);
3257     }
3258     return true;
3259 }
3260
3261 static bool gen_global_function_code(ir_builder *ir, ir_value *global)
3262 {
3263     prog_section_function_t *fundef;
3264     ir_function             *irfun;
3265
3266     (void)ir;
3267
3268     irfun = global->constval.vfunc;
3269     if (!irfun) {
3270         if (global->cvq == CV_NONE) {
3271             if (irwarning(global->context, WARN_IMPLICIT_FUNCTION_POINTER,
3272                           "function `%s` has no body and in QC implicitly becomes a function-pointer",
3273                           global->name.c_str()))
3274             {
3275                 /* Not bailing out just now. If this happens a lot you don't want to have
3276                  * to rerun gmqcc for each such function.
3277                  */
3278
3279                 /* return false; */
3280             }
3281         }
3282         /* this was a function pointer, don't generate code for those */
3283         return true;
3284     }
3285
3286     if (irfun->builtin)
3287         return true;
3288
3289     /*
3290      * If there is no definition and the thing is eraseable, we can ignore
3291      * outputting the function to begin with.
3292      */
3293     if (global->flags & IR_FLAG_ERASABLE && irfun->code_function_def < 0) {
3294         return true;
3295     }
3296
3297     if (irfun->code_function_def < 0) {
3298         irerror(irfun->context, "`%s`: IR global wasn't generated, failed to access function-def", irfun->name.c_str());
3299         return false;
3300     }
3301     fundef = &ir->code->functions[irfun->code_function_def];
3302
3303     fundef->entry = ir->code->statements.size();
3304     if (!gen_function_locals(ir, global)) {
3305         irerror(irfun->context, "Failed to generate locals for function %s", irfun->name.c_str());
3306         return false;
3307     }
3308     if (!gen_function_extparam_copy(ir->code.get(), irfun)) {
3309         irerror(irfun->context, "Failed to generate extparam-copy code for function %s", irfun->name.c_str());
3310         return false;
3311     }
3312     if (irfun->max_varargs && !gen_function_varargs_copy(ir->code.get(), irfun)) {
3313         irerror(irfun->context, "Failed to generate vararg-copy code for function %s", irfun->name.c_str());
3314         return false;
3315     }
3316     if (!gen_function_code(ir->code.get(), irfun)) {
3317         irerror(irfun->context, "Failed to generate code for function %s", irfun->name.c_str());
3318         return false;
3319     }
3320     return true;
3321 }
3322
3323 static void gen_vector_defs(code_t *code, prog_section_def_t def, const char *name)
3324 {
3325     char  *component;
3326     size_t len, i;
3327
3328     if (!name || name[0] == '#' || OPTS_FLAG(SINGLE_VECTOR_DEFS))
3329         return;
3330
3331     def.type = TYPE_FLOAT;
3332
3333     len = strlen(name);
3334
3335     component = (char*)mem_a(len+3);
3336     memcpy(component, name, len);
3337     len += 2;
3338     component[len-0] = 0;
3339     component[len-2] = '_';
3340
3341     component[len-1] = 'x';
3342
3343     for (i = 0; i < 3; ++i) {
3344         def.name = code_genstring(code, component);
3345         code->defs.push_back(def);
3346         def.offset++;
3347         component[len-1]++;
3348     }
3349
3350     mem_d(component);
3351 }
3352
3353 static void gen_vector_fields(code_t *code, prog_section_field_t fld, const char *name)
3354 {
3355     char  *component;
3356     size_t len, i;
3357
3358     if (!name || OPTS_FLAG(SINGLE_VECTOR_DEFS))
3359         return;
3360
3361     fld.type = TYPE_FLOAT;
3362
3363     len = strlen(name);
3364
3365     component = (char*)mem_a(len+3);
3366     memcpy(component, name, len);
3367     len += 2;
3368     component[len-0] = 0;
3369     component[len-2] = '_';
3370
3371     component[len-1] = 'x';
3372
3373     for (i = 0; i < 3; ++i) {
3374         fld.name = code_genstring(code, component);
3375         code->fields.push_back(fld);
3376         fld.offset++;
3377         component[len-1]++;
3378     }
3379
3380     mem_d(component);
3381 }
3382
3383 static bool ir_builder_gen_global(ir_builder *self, ir_value *global, bool islocal)
3384 {
3385     size_t             i;
3386     int32_t           *iptr;
3387     prog_section_def_t def;
3388     bool               pushdef = opts.optimizeoff;
3389
3390     /* we don't generate split-vectors */
3391     if (global->vtype == TYPE_VECTOR && (global->flags & IR_FLAG_SPLIT_VECTOR))
3392         return true;
3393
3394     def.type = global->vtype;
3395     def.offset = self->code->globals.size();
3396     def.name = 0;
3397     if (OPTS_OPTION_BOOL(OPTION_G) || !islocal)
3398     {
3399         pushdef = true;
3400
3401         /*
3402          * if we're eraseable and the function isn't referenced ignore outputting
3403          * the function.
3404          */
3405         if (global->flags & IR_FLAG_ERASABLE && global->reads.empty()) {
3406             return true;
3407         }
3408
3409         if (OPTS_OPTIMIZATION(OPTIM_STRIP_CONSTANT_NAMES) &&
3410             !(global->flags & IR_FLAG_INCLUDE_DEF) &&
3411             (global->name[0] == '#' || global->cvq == CV_CONST))
3412         {
3413             pushdef = false;
3414         }
3415
3416         if (pushdef) {
3417             if (global->name[0] == '#') {
3418                 if (!self->str_immediate)
3419                     self->str_immediate = code_genstring(self->code.get(), "IMMEDIATE");
3420                 def.name = global->code.name = self->str_immediate;
3421             }
3422             else
3423                 def.name = global->code.name = code_genstring(self->code.get(), global->name.c_str());
3424         }
3425         else
3426             def.name   = 0;
3427         if (islocal) {
3428             def.offset = ir_value_code_addr(global);
3429             self->code->defs.push_back(def);
3430             if (global->vtype == TYPE_VECTOR)
3431                 gen_vector_defs(self->code.get(), def, global->name.c_str());
3432             else if (global->vtype == TYPE_FIELD && global->fieldtype == TYPE_VECTOR)
3433                 gen_vector_defs(self->code.get(), def, global->name.c_str());
3434             return true;
3435         }
3436     }
3437     if (islocal)
3438         return true;
3439
3440     switch (global->vtype)
3441     {
3442     case TYPE_VOID:
3443         if (0 == global->name.compare("end_sys_globals")) {
3444             // TODO: remember this point... all the defs before this one
3445             // should be checksummed and added to progdefs.h when we generate it.
3446         }
3447         else if (0 == global->name.compare("end_sys_fields")) {
3448             // TODO: same as above but for entity-fields rather than globsl
3449         }
3450         else if(irwarning(global->context, WARN_VOID_VARIABLES, "unrecognized variable of type void `%s`",
3451                           global->name.c_str()))
3452         {
3453             /* Not bailing out */
3454             /* return false; */
3455         }
3456         /* I'd argue setting it to 0 is sufficient, but maybe some depend on knowing how far
3457          * the system fields actually go? Though the engine knows this anyway...
3458          * Maybe this could be an -foption
3459          * fteqcc creates data for end_sys_* - of size 1, so let's do the same
3460          */
3461         ir_value_code_setaddr(global, self->code->globals.size());
3462         self->code->globals.push_back(0);
3463         /* Add the def */
3464         if (pushdef) self->code->defs.push_back(def);
3465         return true;
3466     case TYPE_POINTER:
3467         if (pushdef) self->code->defs.push_back(def);
3468         return gen_global_pointer(self->code.get(), global);
3469     case TYPE_FIELD:
3470         if (pushdef) {
3471             self->code->defs.push_back(def);
3472             if (global->fieldtype == TYPE_VECTOR)
3473                 gen_vector_defs(self->code.get(), def, global->name.c_str());
3474         }
3475         return gen_global_field(self->code.get(), global);
3476     case TYPE_ENTITY:
3477         /* fall through */
3478     case TYPE_FLOAT:
3479     {
3480         ir_value_code_setaddr(global, self->code->globals.size());
3481         if (global->hasvalue) {
3482             iptr = (int32_t*)&global->constval.ivec[0];
3483             self->code->globals.push_back(*iptr);
3484         } else {
3485             self->code->globals.push_back(0);
3486         }
3487         if (!islocal && global->cvq != CV_CONST)
3488             def.type |= DEF_SAVEGLOBAL;
3489         if (pushdef) self->code->defs.push_back(def);
3490
3491         return global->code.globaladdr >= 0;
3492     }
3493     case TYPE_STRING:
3494     {
3495         ir_value_code_setaddr(global, self->code->globals.size());
3496         if (global->hasvalue) {
3497             uint32_t load = code_genstring(self->code.get(), global->constval.vstring);
3498             self->code->globals.push_back(load);
3499         } else {
3500             self->code->globals.push_back(0);
3501         }
3502         if (!islocal && global->cvq != CV_CONST)
3503             def.type |= DEF_SAVEGLOBAL;
3504         if (pushdef) self->code->defs.push_back(def);
3505         return global->code.globaladdr >= 0;
3506     }
3507     case TYPE_VECTOR:
3508     {
3509         size_t d;
3510         ir_value_code_setaddr(global, self->code->globals.size());
3511         if (global->hasvalue) {
3512             iptr = (int32_t*)&global->constval.ivec[0];
3513             self->code->globals.push_back(iptr[0]);
3514             if (global->code.globaladdr < 0)
3515                 return false;
3516             for (d = 1; d < type_sizeof_[global->vtype]; ++d) {
3517                 self->code->globals.push_back(iptr[d]);
3518             }
3519         } else {
3520             self->code->globals.push_back(0);
3521             if (global->code.globaladdr < 0)
3522                 return false;
3523             for (d = 1; d < type_sizeof_[global->vtype]; ++d) {
3524                 self->code->globals.push_back(0);
3525             }
3526         }
3527         if (!islocal && global->cvq != CV_CONST)
3528             def.type |= DEF_SAVEGLOBAL;
3529
3530         if (pushdef) {
3531             self->code->defs.push_back(def);
3532             def.type &= ~DEF_SAVEGLOBAL;
3533             gen_vector_defs(self->code.get(), def, global->name.c_str());
3534         }
3535         return global->code.globaladdr >= 0;
3536     }
3537     case TYPE_FUNCTION:
3538         ir_value_code_setaddr(global, self->code->globals.size());
3539         if (!global->hasvalue) {
3540             self->code->globals.push_back(0);
3541             if (global->code.globaladdr < 0)
3542                 return false;
3543         } else {
3544             self->code->globals.push_back(self->code->functions.size());
3545             if (!gen_global_function(self, global))
3546                 return false;
3547         }
3548         if (!islocal && global->cvq != CV_CONST)
3549             def.type |= DEF_SAVEGLOBAL;
3550         if (pushdef) self->code->defs.push_back(def);
3551         return true;
3552     case TYPE_VARIANT:
3553         /* assume biggest type */
3554             ir_value_code_setaddr(global, self->code->globals.size());
3555             self->code->globals.push_back(0);
3556             for (i = 1; i < type_sizeof_[TYPE_VARIANT]; ++i)
3557                 self->code->globals.push_back(0);
3558             return true;
3559     default:
3560         /* refuse to create 'void' type or any other fancy business. */
3561         irerror(global->context, "Invalid type for global variable `%s`: %s",
3562                 global->name.c_str(), type_name[global->vtype]);
3563         return false;
3564     }
3565 }
3566
3567 static GMQCC_INLINE void ir_builder_prepare_field(code_t *code, ir_value *field)
3568 {
3569     field->code.fieldaddr = code_alloc_field(code, type_sizeof_[field->fieldtype]);
3570 }
3571
3572 static bool ir_builder_gen_field(ir_builder *self, ir_value *field)
3573 {
3574     prog_section_def_t def;
3575     prog_section_field_t fld;
3576
3577     (void)self;
3578
3579     def.type   = (uint16_t)field->vtype;
3580     def.offset = (uint16_t)self->code->globals.size();
3581
3582     /* create a global named the same as the field */
3583     if (OPTS_OPTION_U32(OPTION_STANDARD) == COMPILER_GMQCC) {
3584         /* in our standard, the global gets a dot prefix */
3585         size_t len = field->name.length();
3586         char name[1024];
3587
3588         /* we really don't want to have to allocate this, and 1024
3589          * bytes is more than enough for a variable/field name
3590          */
3591         if (len+2 >= sizeof(name)) {
3592             irerror(field->context, "invalid field name size: %u", (unsigned int)len);
3593             return false;
3594         }
3595
3596         name[0] = '.';
3597         memcpy(name+1, field->name.c_str(), len); // no strncpy - we used strlen above
3598         name[len+1] = 0;
3599
3600         def.name = code_genstring(self->code.get(), name);
3601         fld.name = def.name + 1; /* we reuse that string table entry */
3602     } else {
3603         /* in plain QC, there cannot be a global with the same name,
3604          * and so we also name the global the same.
3605          * FIXME: fteqcc should create a global as well
3606          * check if it actually uses the same name. Probably does
3607          */
3608         def.name = code_genstring(self->code.get(), field->name.c_str());
3609         fld.name = def.name;
3610     }
3611
3612     field->code.name = def.name;
3613
3614     self->code->defs.push_back(def);
3615
3616     fld.type = field->fieldtype;
3617
3618     if (fld.type == TYPE_VOID) {
3619         irerror(field->context, "field is missing a type: %s - don't know its size", field->name.c_str());
3620         return false;
3621     }
3622
3623     fld.offset = field->code.fieldaddr;
3624
3625     self->code->fields.push_back(fld);
3626
3627     ir_value_code_setaddr(field, self->code->globals.size());
3628     self->code->globals.push_back(fld.offset);
3629     if (fld.type == TYPE_VECTOR) {
3630         self->code->globals.push_back(fld.offset+1);
3631         self->code->globals.push_back(fld.offset+2);
3632     }
3633
3634     if (field->fieldtype == TYPE_VECTOR) {
3635         gen_vector_defs  (self->code.get(), def, field->name.c_str());
3636         gen_vector_fields(self->code.get(), fld, field->name.c_str());
3637     }
3638
3639     return field->code.globaladdr >= 0;
3640 }
3641
3642 static void ir_builder_collect_reusables(ir_builder *builder) {
3643     std::vector<ir_value*> reusables;
3644
3645     for (auto& gp : builder->globals) {
3646         ir_value *value = gp.get();
3647         if (value->vtype != TYPE_FLOAT || !value->hasvalue)
3648             continue;
3649         if (value->cvq == CV_CONST || (value->name.length() >= 1 && value->name[0] == '#'))
3650             reusables.emplace_back(value);
3651     }
3652     builder->const_floats = move(reusables);
3653 }
3654
3655 static void ir_builder_split_vector(ir_builder *self, ir_value *vec) {
3656     ir_value* found[3] = { nullptr, nullptr, nullptr };
3657
3658     // must not be written to
3659     if (vec->writes.size())
3660         return;
3661     // must not be trying to access individual members
3662     if (vec->members[0] || vec->members[1] || vec->members[2])
3663         return;
3664     // should be actually used otherwise it won't be generated anyway
3665     if (vec->reads.empty())
3666         return;
3667     //size_t count = vec->reads.size();
3668     //if (!count)
3669     //    return;
3670
3671     // may only be used directly as function parameters, so if we find some other instruction cancel
3672     for (ir_instr *user : vec->reads) {
3673         // we only split vectors if they're used directly as parameter to a call only!
3674         if ((user->opcode < INSTR_CALL0 || user->opcode > INSTR_CALL8) && user->opcode != VINSTR_NRCALL)
3675             return;
3676     }
3677
3678     vec->flags |= IR_FLAG_SPLIT_VECTOR;
3679
3680     // find existing floats making up the split
3681     for (ir_value *c : self->const_floats) {
3682         if (!found[0] && c->constval.vfloat == vec->constval.vvec.x)
3683             found[0] = c;
3684         if (!found[1] && c->constval.vfloat == vec->constval.vvec.y)
3685             found[1] = c;
3686         if (!found[2] && c->constval.vfloat == vec->constval.vvec.z)
3687             found[2] = c;
3688         if (found[0] && found[1] && found[2])
3689             break;
3690     }
3691
3692     // generate floats for not yet found components
3693     if (!found[0])
3694         found[0] = ir_builder_imm_float(self, vec->constval.vvec.x, true);
3695     if (!found[1]) {
3696         if (vec->constval.vvec.y == vec->constval.vvec.x)
3697             found[1] = found[0];
3698         else
3699             found[1] = ir_builder_imm_float(self, vec->constval.vvec.y, true);
3700     }
3701     if (!found[2]) {
3702         if (vec->constval.vvec.z == vec->constval.vvec.x)
3703             found[2] = found[0];
3704         else if (vec->constval.vvec.z == vec->constval.vvec.y)
3705             found[2] = found[1];
3706         else
3707             found[2] = ir_builder_imm_float(self, vec->constval.vvec.z, true);
3708     }
3709
3710     // the .members array should be safe to use here
3711     vec->members[0] = found[0];
3712     vec->members[1] = found[1];
3713     vec->members[2] = found[2];
3714
3715     // register the readers for these floats
3716     found[0]->reads.insert(found[0]->reads.end(), vec->reads.begin(), vec->reads.end());
3717     found[1]->reads.insert(found[1]->reads.end(), vec->reads.begin(), vec->reads.end());
3718     found[2]->reads.insert(found[2]->reads.end(), vec->reads.begin(), vec->reads.end());
3719 }
3720
3721 static void ir_builder_split_vectors(ir_builder *self) {
3722     // member values may be added to self->globals during this operation, but
3723     // no new vectors will be added, we need to iterate via an index as
3724     // c++ iterators would be invalidated
3725     const size_t count = self->globals.size();
3726     for (size_t i = 0; i != count; ++i) {
3727         ir_value *v = self->globals[i].get();
3728         if (v->vtype != TYPE_VECTOR || !v->name.length() || v->name[0] != '#')
3729             continue;
3730         ir_builder_split_vector(self, v);
3731     }
3732 }
3733
3734 bool ir_builder_generate(ir_builder *self, const char *filename)
3735 {
3736     prog_section_statement_t stmt;
3737     char  *lnofile = nullptr;
3738
3739     if (OPTS_FLAG(SPLIT_VECTOR_PARAMETERS)) {
3740         ir_builder_collect_reusables(self);
3741         if (!self->const_floats.empty())
3742             ir_builder_split_vectors(self);
3743     }
3744
3745     for (auto& fp : self->fields)
3746         ir_builder_prepare_field(self->code.get(), fp.get());
3747
3748     for (auto& gp : self->globals) {
3749         ir_value *global = gp.get();
3750         if (!ir_builder_gen_global(self, global, false)) {
3751             return false;
3752         }
3753         if (global->vtype == TYPE_FUNCTION) {
3754             ir_function *func = global->constval.vfunc;
3755             if (func && self->max_locals < func->allocated_locals &&
3756                 !(func->flags & IR_FLAG_MASK_NO_OVERLAP))
3757             {
3758                 self->max_locals = func->allocated_locals;
3759             }
3760             if (func && self->max_globaltemps < func->globaltemps)
3761                 self->max_globaltemps = func->globaltemps;
3762         }
3763     }
3764
3765     for (auto& fp : self->fields) {
3766         if (!ir_builder_gen_field(self, fp.get()))
3767             return false;
3768     }
3769
3770     // generate nil
3771     ir_value_code_setaddr(self->nil, self->code->globals.size());
3772     self->code->globals.push_back(0);
3773     self->code->globals.push_back(0);
3774     self->code->globals.push_back(0);
3775
3776     // generate virtual-instruction temps
3777     for (size_t i = 0; i < IR_MAX_VINSTR_TEMPS; ++i) {
3778         ir_value_code_setaddr(self->vinstr_temp[i], self->code->globals.size());
3779         self->code->globals.push_back(0);
3780         self->code->globals.push_back(0);
3781         self->code->globals.push_back(0);