2 * Copyright (C) 2012, 2013
5 * Permission is hereby granted, free of charge, to any person obtaining a copy of
6 * this software and associated documentation files (the "Software"), to deal in
7 * the Software without restriction, including without limitation the rights to
8 * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
9 * of the Software, and to permit persons to whom the Software is furnished to do
10 * so, subject to the following conditions:
12 * The above copyright notice and this permission notice shall be included in all
13 * copies or substantial portions of the Software.
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
29 #define FOLD_STRING_UNTRANSLATE_HTSIZE 1024
30 #define FOLD_STRING_DOTRANSLATE_HTSIZE 1024
33 * There is two stages to constant folding in GMQCC: there is the parse
34 * stage constant folding, where, witht he help of the AST, operator
35 * usages can be constant folded. Then there is the constant folding
36 * in the IR for things like eliding if statements, can occur.
38 * This file is thus, split into two parts.
41 #define isfloat(X) (((ast_expression*)(X))->vtype == TYPE_FLOAT)
42 #define isvector(X) (((ast_expression*)(X))->vtype == TYPE_VECTOR)
43 #define isstring(X) (((ast_expression*)(X))->vtype == TYPE_STRING)
44 #define isfloats(X,Y) (isfloat (X) && isfloat (Y))
47 * Implementation of basic vector math for vec3_t, for trivial constant
50 * TODO: gcc/clang hinting for autovectorization
52 static GMQCC_INLINE vec3_t vec3_add(vec3_t a, vec3_t b) {
60 static GMQCC_INLINE vec3_t vec3_sub(vec3_t a, vec3_t b) {
68 static GMQCC_INLINE vec3_t vec3_neg(vec3_t a) {
76 static GMQCC_INLINE vec3_t vec3_xor(vec3_t a, vec3_t b) {
78 out.x = (qcfloat_t)(((qcint_t)a.x) ^ ((qcint_t)b.x));
79 out.y = (qcfloat_t)(((qcint_t)a.y) ^ ((qcint_t)b.y));
80 out.z = (qcfloat_t)(((qcint_t)a.z) ^ ((qcint_t)b.z));
84 static GMQCC_INLINE vec3_t vec3_xorvf(vec3_t a, qcfloat_t b) {
86 out.x = (qcfloat_t)(((qcint_t)a.x) ^ ((qcint_t)b));
87 out.y = (qcfloat_t)(((qcint_t)a.y) ^ ((qcint_t)b));
88 out.z = (qcfloat_t)(((qcint_t)a.z) ^ ((qcint_t)b));
92 static GMQCC_INLINE qcfloat_t vec3_mulvv(vec3_t a, vec3_t b) {
93 return (a.x * b.x + a.y * b.y + a.z * b.z);
96 static GMQCC_INLINE vec3_t vec3_mulvf(vec3_t a, qcfloat_t b) {
104 static GMQCC_INLINE bool vec3_cmp(vec3_t a, vec3_t b) {
110 static GMQCC_INLINE vec3_t vec3_create(float x, float y, float z) {
118 static GMQCC_INLINE qcfloat_t vec3_notf(vec3_t a) {
119 return (!a.x && !a.y && !a.z);
122 static GMQCC_INLINE bool vec3_pbool(vec3_t a) {
123 return (a.x && a.y && a.z);
126 static lex_ctx_t fold_ctx(fold_t *fold) {
128 if (fold->parser->lex)
129 return parser_ctx(fold->parser);
131 memset(&ctx, 0, sizeof(ctx));
135 static GMQCC_INLINE bool fold_immediate_true(fold_t *fold, ast_value *v) {
136 switch (v->expression.vtype) {
138 return !!v->constval.vfloat;
140 return !!v->constval.vint;
142 if (OPTS_FLAG(CORRECT_LOGIC))
143 return vec3_pbool(v->constval.vvec);
144 return !!(v->constval.vvec.x);
146 if (!v->constval.vstring)
148 if (OPTS_FLAG(TRUE_EMPTY_STRINGS))
150 return !!v->constval.vstring[0];
152 compile_error(fold_ctx(fold), "internal error: fold_immediate_true on invalid type");
155 return !!v->constval.vfunc;
158 /* Handy macros to determine if an ast_value can be constant folded. */
159 #define fold_can_1(X) \
160 (ast_istype(((ast_expression*)(X)), ast_value) && (X)->hasvalue && ((X)->cvq == CV_CONST) && \
161 ((ast_expression*)(X))->vtype != TYPE_FUNCTION)
163 #define fold_can_2(X, Y) (fold_can_1(X) && fold_can_1(Y))
165 #define fold_immvalue_float(E) ((E)->constval.vfloat)
166 #define fold_immvalue_vector(E) ((E)->constval.vvec)
167 #define fold_immvalue_string(E) ((E)->constval.vstring)
169 fold_t *fold_init(parser_t *parser) {
170 fold_t *fold = (fold_t*)mem_a(sizeof(fold_t));
171 fold->parser = parser;
172 fold->imm_float = NULL;
173 fold->imm_vector = NULL;
174 fold->imm_string = NULL;
175 fold->imm_string_untranslate = util_htnew(FOLD_STRING_UNTRANSLATE_HTSIZE);
176 fold->imm_string_dotranslate = util_htnew(FOLD_STRING_DOTRANSLATE_HTSIZE);
179 * prime the tables with common constant values at constant
182 (void)fold_constgen_float (fold, 0.0f);
183 (void)fold_constgen_float (fold, 1.0f);
184 (void)fold_constgen_float (fold, -1.0f);
186 (void)fold_constgen_vector(fold, vec3_create(0.0f, 0.0f, 0.0f));
191 bool fold_generate(fold_t *fold, ir_builder *ir) {
192 /* generate globals for immediate folded values */
196 for (i = 0; i < vec_size(fold->imm_float); ++i)
197 if (!ast_global_codegen ((cur = fold->imm_float[i]), ir, false)) goto err;
198 for (i = 0; i < vec_size(fold->imm_vector); ++i)
199 if (!ast_global_codegen((cur = fold->imm_vector[i]), ir, false)) goto err;
200 for (i = 0; i < vec_size(fold->imm_string); ++i)
201 if (!ast_global_codegen((cur = fold->imm_string[i]), ir, false)) goto err;
206 con_out("failed to generate global %s\n", cur->name);
207 ir_builder_delete(ir);
211 void fold_cleanup(fold_t *fold) {
214 for (i = 0; i < vec_size(fold->imm_float); ++i) ast_delete(fold->imm_float[i]);
215 for (i = 0; i < vec_size(fold->imm_vector); ++i) ast_delete(fold->imm_vector[i]);
216 for (i = 0; i < vec_size(fold->imm_string); ++i) ast_delete(fold->imm_string[i]);
218 vec_free(fold->imm_float);
219 vec_free(fold->imm_vector);
220 vec_free(fold->imm_string);
222 util_htdel(fold->imm_string_untranslate);
223 util_htdel(fold->imm_string_dotranslate);
228 ast_expression *fold_constgen_float(fold_t *fold, qcfloat_t value) {
229 ast_value *out = NULL;
232 for (i = 0; i < vec_size(fold->imm_float); i++) {
233 if (fold->imm_float[i]->constval.vfloat == value)
234 return (ast_expression*)fold->imm_float[i];
237 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_FLOAT);
239 out->hasvalue = true;
240 out->constval.vfloat = value;
242 vec_push(fold->imm_float, out);
244 return (ast_expression*)out;
247 ast_expression *fold_constgen_vector(fold_t *fold, vec3_t value) {
251 for (i = 0; i < vec_size(fold->imm_vector); i++) {
252 if (vec3_cmp(fold->imm_vector[i]->constval.vvec, value))
253 return (ast_expression*)fold->imm_vector[i];
256 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_VECTOR);
258 out->hasvalue = true;
259 out->constval.vvec = value;
261 vec_push(fold->imm_vector, out);
263 return (ast_expression*)out;
266 ast_expression *fold_constgen_string(fold_t *fold, const char *str, bool translate) {
267 hash_table_t *table = (translate) ? fold->imm_string_untranslate : fold->imm_string_dotranslate;
268 ast_value *out = NULL;
269 size_t hash = util_hthash(table, str);
271 if ((out = (ast_value*)util_htgeth(table, str, hash)))
272 return (ast_expression*)out;
276 util_snprintf(name, sizeof(name), "dotranslate_%lu", (unsigned long)(fold->parser->translated++));
277 out = ast_value_new(parser_ctx(fold->parser), name, TYPE_STRING);
278 out->expression.flags |= AST_FLAG_INCLUDE_DEF; /* def needs to be included for translatables */
280 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_STRING);
283 out->hasvalue = true;
285 out->constval.vstring = parser_strdup(str);
287 vec_push(fold->imm_string, out);
288 util_htseth(table, str, hash, out);
290 return (ast_expression*)out;
294 static GMQCC_INLINE ast_expression *fold_op_mul_vec(fold_t *fold, vec3_t vec, ast_value *sel, const char *set) {
296 * vector-component constant folding works by matching the component sets
297 * to eliminate expensive operations on whole-vectors (3 components at runtime).
298 * to achive this effect in a clean manner this function generalizes the
299 * values through the use of a set paramater, which is used as an indexing method
300 * for creating the elided ast binary expression.
302 * Consider 'n 0 0' where y, and z need to be tested for 0, and x is
303 * used as the value in a binary operation generating an INSTR_MUL instruction
304 * to acomplish the indexing of the correct component value we use set[0], set[1], set[2]
305 * as x, y, z, where the values of those operations return 'x', 'y', 'z'. Because
306 * of how ASCII works we can easily deliniate:
307 * vec.z is the same as set[2]-'x' for when set[2] is 'z', 'z'-'x' results in a
308 * literal value of 2, using this 2, we know that taking the address of vec->x (float)
309 * and indxing it with this literal will yeild the immediate address of that component
311 * Of course more work needs to be done to generate the correct index for the ast_member_new
312 * call, which is no problem: set[0]-'x' suffices that job.
314 qcfloat_t x = (&vec.x)[set[0]-'x'];
315 qcfloat_t y = (&vec.x)[set[1]-'x'];
316 qcfloat_t z = (&vec.x)[set[2]-'x'];
320 ++opts_optimizationcount[OPTIM_VECTOR_COMPONENTS];
321 out = (ast_expression*)ast_member_new(fold_ctx(fold), (ast_expression*)sel, set[0]-'x', NULL);
322 out->node.keep = false;
323 ((ast_member*)out)->rvalue = true;
325 return (ast_expression*)ast_binary_new(fold_ctx(fold), INSTR_MUL_F, fold_constgen_float(fold, x), out);
331 static GMQCC_INLINE ast_expression *fold_op_neg(fold_t *fold, ast_value *a) {
334 return fold_constgen_float(fold, -fold_immvalue_float(a));
335 } else if (isvector(a)) {
337 return fold_constgen_vector(fold, vec3_neg(fold_immvalue_vector(a)));
342 static GMQCC_INLINE ast_expression *fold_op_not(fold_t *fold, ast_value *a) {
345 return fold_constgen_float(fold, !fold_immvalue_float(a));
346 } else if (isvector(a)) {
348 return fold_constgen_float(fold, vec3_notf(fold_immvalue_vector(a)));
349 } else if (isstring(a)) {
351 if (OPTS_FLAG(TRUE_EMPTY_STRINGS))
352 return fold_constgen_float(fold, !fold_immvalue_string(a));
354 return fold_constgen_float(fold, !fold_immvalue_string(a) || !*fold_immvalue_string(a));
360 static GMQCC_INLINE ast_expression *fold_op_add(fold_t *fold, ast_value *a, ast_value *b) {
362 if (fold_can_2(a, b))
363 return fold_constgen_float(fold, fold_immvalue_float(a) + fold_immvalue_float(b));
364 } else if (isvector(a)) {
365 if (fold_can_2(a, b))
366 return fold_constgen_vector(fold, vec3_add(fold_immvalue_vector(a), fold_immvalue_vector(b)));
371 static GMQCC_INLINE ast_expression *fold_op_sub(fold_t *fold, ast_value *a, ast_value *b) {
373 if (fold_can_2(a, b))
374 return fold_constgen_float(fold, fold_immvalue_float(a) - fold_immvalue_float(b));
375 } else if (isvector(a)) {
376 if (fold_can_2(a, b))
377 return fold_constgen_vector(fold, vec3_sub(fold_immvalue_vector(a), fold_immvalue_vector(b)));
382 static GMQCC_INLINE ast_expression *fold_op_mul(fold_t *fold, ast_value *a, ast_value *b) {
385 if (fold_can_2(a, b))
386 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(b), fold_immvalue_float(a)));
388 if (fold_can_2(a, b))
389 return fold_constgen_float(fold, fold_immvalue_float(a) * fold_immvalue_float(b));
391 } else if (isvector(a)) {
393 if (fold_can_2(a, b))
394 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
396 if (fold_can_2(a, b)) {
397 return fold_constgen_float(fold, vec3_mulvv(fold_immvalue_vector(a), fold_immvalue_vector(b)));
398 } else if (OPTS_OPTIMIZATION(OPTIM_VECTOR_COMPONENTS) && fold_can_1(a)) {
400 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "xyz"))) return out;
401 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "yxz"))) return out;
402 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "zxy"))) return out;
403 } else if (OPTS_OPTIMIZATION(OPTIM_VECTOR_COMPONENTS) && fold_can_1(b)) {
405 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "xyz"))) return out;
406 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "yxz"))) return out;
407 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "zxy"))) return out;
414 static GMQCC_INLINE ast_expression *fold_op_div(fold_t *fold, ast_value *a, ast_value *b) {
416 if (fold_can_2(a, b))
417 return fold_constgen_float(fold, fold_immvalue_float(a) / fold_immvalue_float(b));
418 } else if (isvector(a)) {
419 if (fold_can_2(a, b))
420 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(a), 1.0f / fold_immvalue_float(b)));
422 return (ast_expression*)ast_binary_new(
427 ? (ast_expression*)fold_constgen_float(fold, 1.0f / fold_immvalue_float(b))
428 : (ast_expression*)ast_binary_new(
431 (ast_expression*)fold->imm_float[1],
440 static GMQCC_INLINE ast_expression *fold_op_mod(fold_t *fold, ast_value *a, ast_value *b) {
441 if (fold_can_2(a, b))
442 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) % ((qcint_t)fold_immvalue_float(b))));
446 static GMQCC_INLINE ast_expression *fold_op_bor(fold_t *fold, ast_value *a, ast_value *b) {
447 if (fold_can_2(a, b))
448 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) | ((qcint_t)fold_immvalue_float(b))));
452 static GMQCC_INLINE ast_expression *fold_op_band(fold_t *fold, ast_value *a, ast_value *b) {
453 if (fold_can_2(a, b))
454 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) & ((qcint_t)fold_immvalue_float(b))));
458 static GMQCC_INLINE ast_expression *fold_op_xor(fold_t *fold, ast_value *a, ast_value *b) {
460 if (fold_can_2(a, b))
461 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) ^ ((qcint_t)fold_immvalue_float(b))));
464 if (fold_can_2(a, b))
465 return fold_constgen_vector(fold, vec3_xor(fold_immvalue_vector(a), fold_immvalue_vector(b)));
467 if (fold_can_2(a, b))
468 return fold_constgen_vector(fold, vec3_xorvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
474 static GMQCC_INLINE ast_expression *fold_op_lshift(fold_t *fold, ast_value *a, ast_value *b) {
475 if (fold_can_2(a, b) && isfloats(a, b))
476 return fold_constgen_float(fold, (qcfloat_t)((qcuint_t)(fold_immvalue_float(a)) << (qcuint_t)(fold_immvalue_float(b))));
480 static GMQCC_INLINE ast_expression *fold_op_rshift(fold_t *fold, ast_value *a, ast_value *b) {
481 if (fold_can_2(a, b) && isfloats(a, b))
482 return fold_constgen_float(fold, (qcfloat_t)((qcuint_t)(fold_immvalue_float(a)) >> (qcuint_t)(fold_immvalue_float(b))));
486 static GMQCC_INLINE ast_expression *fold_op_andor(fold_t *fold, ast_value *a, ast_value *b, float or) {
487 if (fold_can_2(a, b)) {
488 if (OPTS_FLAG(PERL_LOGIC)) {
489 if (fold_immediate_true(fold, a))
490 return (ast_expression*)b;
492 return fold_constgen_float (
494 ((or) ? (fold_immediate_true(fold, a) || fold_immediate_true(fold, b))
495 : (fold_immediate_true(fold, a) && fold_immediate_true(fold, b)))
504 static GMQCC_INLINE ast_expression *fold_op_tern(fold_t *fold, ast_value *a, ast_value *b, ast_value *c) {
506 return fold_immediate_true(fold, a)
508 : (ast_expression*)c;
513 static GMQCC_INLINE ast_expression *fold_op_exp(fold_t *fold, ast_value *a, ast_value *b) {
514 if (fold_can_2(a, b))
515 return fold_constgen_float(fold, (qcfloat_t)powf(fold_immvalue_float(a), fold_immvalue_float(b)));
519 static GMQCC_INLINE ast_expression *fold_op_lteqgt(fold_t *fold, ast_value *a, ast_value *b) {
520 if (fold_can_2(a,b)) {
521 if (fold_immvalue_float(a) < fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[2];
522 if (fold_immvalue_float(a) == fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[0];
523 if (fold_immvalue_float(a) > fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[1];
528 static GMQCC_INLINE ast_expression *fold_op_cmp(fold_t *fold, ast_value *a, ast_value *b, bool ne) {
529 if (fold_can_2(a, b)) {
530 return fold_constgen_float(
532 (ne) ? (fold_immvalue_float(a) != fold_immvalue_float(b))
533 : (fold_immvalue_float(a) == fold_immvalue_float(b))
539 static GMQCC_INLINE ast_expression *fold_op_bnot(fold_t *fold, ast_value *a) {
541 return fold_constgen_float(fold, ~((qcint_t)fold_immvalue_float(a)));
545 ast_expression *fold_op(fold_t *fold, const oper_info *info, ast_expression **opexprs) {
546 ast_value *a = (ast_value*)opexprs[0];
547 ast_value *b = (ast_value*)opexprs[1];
548 ast_value *c = (ast_value*)opexprs[2];
550 /* can a fold operation be applied to this operator usage? */
554 switch(info->operands) {
555 case 3: if(!c) return NULL;
556 case 2: if(!b) return NULL;
559 compile_error(fold_ctx(fold), "interal error: fold_op no operands to fold\n");
565 case opid2('-','P'): return fold_op_neg (fold, a);
566 case opid2('!','P'): return fold_op_not (fold, a);
567 case opid1('+'): return fold_op_add (fold, a, b);
568 case opid1('-'): return fold_op_sub (fold, a, b);
569 case opid1('*'): return fold_op_mul (fold, a, b);
570 case opid1('/'): return fold_op_div (fold, a, b);
571 case opid1('%'): return fold_op_mod (fold, a, b);
572 case opid1('|'): return fold_op_bor (fold, a, b);
573 case opid1('&'): return fold_op_band (fold, a, b);
574 case opid1('^'): return fold_op_xor (fold, a, b);
575 case opid2('<','<'): return fold_op_lshift (fold, a, b);
576 case opid2('>','>'): return fold_op_rshift (fold, a, b);
577 case opid2('|','|'): return fold_op_andor (fold, a, b, true);
578 case opid2('&','&'): return fold_op_andor (fold, a, b, false);
579 case opid2('?',':'): return fold_op_tern (fold, a, b, c);
580 case opid2('*','*'): return fold_op_exp (fold, a, b);
581 case opid3('<','=','>'): return fold_op_lteqgt (fold, a, b);
582 case opid2('!','='): return fold_op_cmp (fold, a, b, true);
583 case opid2('=','='): return fold_op_cmp (fold, a, b, false);
584 case opid2('~','P'): return fold_op_bnot (fold, a);
586 compile_error(fold_ctx(fold), "internal error: attempted to constant for unsupported operator");
591 * These are all the actual constant folding methods that happen in between
592 * the AST/IR stage of the compiler , i.e eliminating branches for const
593 * expressions, which is the only supported thing so far. We undefine the
594 * testing macros here because an ir_value is differant than an ast_value.
599 #undef fold_immvalue_float
600 #undef fold_immvalue_string
601 #undef fold_immvalue_vector
605 #define isfloat(X) ((X)->vtype == TYPE_FLOAT)
606 #define isstring(X) ((X)->vtype == TYPE_STRING)
607 #define isvector(X) ((X)->vtype == TYPE_VECTOR)
608 #define fold_immvalue_float(X) ((X)->constval.vfloat)
609 #define fold_immvalue_vector(X) ((X)->constval.vvec)
610 #define fold_immvalue_string(X) ((X)->constval.vstring)
611 #define fold_can_1(X) ((X)->hasvalue && (X)->cvq == CV_CONST)
612 #define fold_can_2(X,Y) (fold_can_1(X) && fold_can_1(Y))
615 int fold_cond(ir_value *condval, ast_function *func, ast_ifthen *branch) {
616 if (isfloat(condval) && fold_can_1(condval) && OPTS_OPTIMIZATION(OPTIM_CONST_FOLD_DCE)) {
617 ast_expression_codegen *cgen;
620 bool istrue = (fold_immvalue_float(condval) == 1.0f && branch->on_true);
621 bool isfalse = (fold_immvalue_float(condval) == 0.0f && branch->on_false);
622 ast_expression *path = (istrue) ? branch->on_true :
623 (isfalse) ? branch->on_false : NULL;
626 if (!(elide = ir_function_create_block(ast_ctx(branch), func->ir_func, ast_function_label(func, ((istrue) ? "ontrue" : "onfalse")))))
628 if (!(*(cgen = path->codegen))((ast_expression*)path, func, false, &dummy))
630 if (!ir_block_create_jump(func->curblock, ast_ctx(branch), elide))
633 * now the branch has been eliminates, and the correct block for the constant evaluation
634 * is expanded into the current block for the function.
636 func->curblock = elide;
639 return -1; /* nothing done */