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))
45 #define isvectors(X,Y) (isvector (X) && isvector(Y))
48 * Implementation of basic vector math for vec3_t, for trivial constant
51 * TODO: gcc/clang hinting for autovectorization
53 static GMQCC_INLINE vec3_t vec3_add(vec3_t a, vec3_t b) {
61 static GMQCC_INLINE vec3_t vec3_sub(vec3_t a, vec3_t b) {
69 static GMQCC_INLINE vec3_t vec3_neg(vec3_t a) {
77 static GMQCC_INLINE vec3_t vec3_xor(vec3_t a, vec3_t b) {
79 out.x = (qcfloat_t)((qcint_t)a.x ^ (qcint_t)b.x);
80 out.y = (qcfloat_t)((qcint_t)a.y ^ (qcint_t)b.y);
81 out.z = (qcfloat_t)((qcint_t)a.z ^ (qcint_t)b.z);
85 static GMQCC_INLINE vec3_t vec3_xorvf(vec3_t a, qcfloat_t b) {
87 out.x = (qcfloat_t)((qcint_t)a.x ^ (qcint_t)b);
88 out.y = (qcfloat_t)((qcint_t)a.y ^ (qcint_t)b);
89 out.z = (qcfloat_t)((qcint_t)a.z ^ (qcint_t)b);
93 static GMQCC_INLINE qcfloat_t vec3_mulvv(vec3_t a, vec3_t b) {
94 return (a.x * b.x + a.y * b.y + a.z * b.z);
97 static GMQCC_INLINE vec3_t vec3_mulvf(vec3_t a, qcfloat_t b) {
105 static GMQCC_INLINE bool vec3_cmp(vec3_t a, vec3_t b) {
111 static GMQCC_INLINE vec3_t vec3_create(float x, float y, float z) {
119 static GMQCC_INLINE qcfloat_t vec3_notf(vec3_t a) {
120 return (!a.x && !a.y && !a.z);
123 static GMQCC_INLINE bool vec3_pbool(vec3_t a) {
124 return (a.x && a.y && a.z);
127 static GMQCC_INLINE bool fold_can_1(const ast_value *val) {
128 return (ast_istype((ast_expression*)val, ast_value) && val->hasvalue && val->cvq == CV_CONST && ((ast_expression*)val)->vtype != TYPE_FUNCTION);
131 static GMQCC_INLINE bool fold_can_2(const ast_value *v1, const ast_value *v2) {
132 return fold_can_1(v1) && fold_can_1(v2);
135 static lex_ctx_t fold_ctx(fold_t *fold) {
137 if (fold->parser->lex)
138 return parser_ctx(fold->parser);
140 memset(&ctx, 0, sizeof(ctx));
144 static GMQCC_INLINE bool fold_immediate_true(fold_t *fold, ast_value *v) {
145 switch (v->expression.vtype) {
147 return !!v->constval.vfloat;
149 return !!v->constval.vint;
151 if (OPTS_FLAG(CORRECT_LOGIC))
152 return vec3_pbool(v->constval.vvec);
153 return !!v->constval.vvec.x;
155 if (!v->constval.vstring)
157 if (OPTS_FLAG(TRUE_EMPTY_STRINGS))
159 return !!v->constval.vstring[0];
161 compile_error(fold_ctx(fold), "internal error: fold_immediate_true on invalid type");
164 return !!v->constval.vfunc;
167 #define fold_immvalue_float(E) ((E)->constval.vfloat)
168 #define fold_immvalue_vector(E) ((E)->constval.vvec)
169 #define fold_immvalue_string(E) ((E)->constval.vstring)
171 fold_t *fold_init(parser_t *parser) {
172 fold_t *fold = (fold_t*)mem_a(sizeof(fold_t));
173 fold->parser = parser;
174 fold->imm_float = NULL;
175 fold->imm_vector = NULL;
176 fold->imm_string = NULL;
177 fold->imm_string_untranslate = util_htnew(FOLD_STRING_UNTRANSLATE_HTSIZE);
178 fold->imm_string_dotranslate = util_htnew(FOLD_STRING_DOTRANSLATE_HTSIZE);
181 * prime the tables with common constant values at constant
184 (void)fold_constgen_float (fold, 0.0f);
185 (void)fold_constgen_float (fold, 1.0f);
186 (void)fold_constgen_float (fold, -1.0f);
188 (void)fold_constgen_vector(fold, vec3_create(0.0f, 0.0f, 0.0f));
193 bool fold_generate(fold_t *fold, ir_builder *ir) {
194 /* generate globals for immediate folded values */
198 for (i = 0; i < vec_size(fold->imm_float); ++i)
199 if (!ast_global_codegen ((cur = fold->imm_float[i]), ir, false)) goto err;
200 for (i = 0; i < vec_size(fold->imm_vector); ++i)
201 if (!ast_global_codegen((cur = fold->imm_vector[i]), ir, false)) goto err;
202 for (i = 0; i < vec_size(fold->imm_string); ++i)
203 if (!ast_global_codegen((cur = fold->imm_string[i]), ir, false)) goto err;
208 con_out("failed to generate global %s\n", cur->name);
209 ir_builder_delete(ir);
213 void fold_cleanup(fold_t *fold) {
216 for (i = 0; i < vec_size(fold->imm_float); ++i) ast_delete(fold->imm_float[i]);
217 for (i = 0; i < vec_size(fold->imm_vector); ++i) ast_delete(fold->imm_vector[i]);
218 for (i = 0; i < vec_size(fold->imm_string); ++i) ast_delete(fold->imm_string[i]);
220 vec_free(fold->imm_float);
221 vec_free(fold->imm_vector);
222 vec_free(fold->imm_string);
224 util_htdel(fold->imm_string_untranslate);
225 util_htdel(fold->imm_string_dotranslate);
230 ast_expression *fold_constgen_float(fold_t *fold, qcfloat_t value) {
231 ast_value *out = NULL;
234 for (i = 0; i < vec_size(fold->imm_float); i++) {
235 if (fold->imm_float[i]->constval.vfloat == value)
236 return (ast_expression*)fold->imm_float[i];
239 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_FLOAT);
241 out->hasvalue = true;
242 out->constval.vfloat = value;
244 vec_push(fold->imm_float, out);
246 return (ast_expression*)out;
249 ast_expression *fold_constgen_vector(fold_t *fold, vec3_t value) {
253 for (i = 0; i < vec_size(fold->imm_vector); i++) {
254 if (vec3_cmp(fold->imm_vector[i]->constval.vvec, value))
255 return (ast_expression*)fold->imm_vector[i];
258 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_VECTOR);
260 out->hasvalue = true;
261 out->constval.vvec = value;
263 vec_push(fold->imm_vector, out);
265 return (ast_expression*)out;
268 ast_expression *fold_constgen_string(fold_t *fold, const char *str, bool translate) {
269 hash_table_t *table = (translate) ? fold->imm_string_untranslate : fold->imm_string_dotranslate;
270 ast_value *out = NULL;
271 size_t hash = util_hthash(table, str);
273 if ((out = (ast_value*)util_htgeth(table, str, hash)))
274 return (ast_expression*)out;
278 util_snprintf(name, sizeof(name), "dotranslate_%lu", (unsigned long)(fold->parser->translated++));
279 out = ast_value_new(parser_ctx(fold->parser), name, TYPE_STRING);
280 out->expression.flags |= AST_FLAG_INCLUDE_DEF; /* def needs to be included for translatables */
282 out = ast_value_new(fold_ctx(fold), "#IMMEDIATE", TYPE_STRING);
285 out->hasvalue = true;
287 out->constval.vstring = parser_strdup(str);
289 vec_push(fold->imm_string, out);
290 util_htseth(table, str, hash, out);
292 return (ast_expression*)out;
296 static GMQCC_INLINE ast_expression *fold_op_mul_vec(fold_t *fold, vec3_t vec, ast_value *sel, const char *set) {
298 * vector-component constant folding works by matching the component sets
299 * to eliminate expensive operations on whole-vectors (3 components at runtime).
300 * to achive this effect in a clean manner this function generalizes the
301 * values through the use of a set paramater, which is used as an indexing method
302 * for creating the elided ast binary expression.
304 * Consider 'n 0 0' where y, and z need to be tested for 0, and x is
305 * used as the value in a binary operation generating an INSTR_MUL instruction
306 * to acomplish the indexing of the correct component value we use set[0], set[1], set[2]
307 * as x, y, z, where the values of those operations return 'x', 'y', 'z'. Because
308 * of how ASCII works we can easily deliniate:
309 * vec.z is the same as set[2]-'x' for when set[2] is 'z', 'z'-'x' results in a
310 * literal value of 2, using this 2, we know that taking the address of vec->x (float)
311 * and indxing it with this literal will yeild the immediate address of that component
313 * Of course more work needs to be done to generate the correct index for the ast_member_new
314 * call, which is no problem: set[0]-'x' suffices that job.
316 qcfloat_t x = (&vec.x)[set[0]-'x'];
317 qcfloat_t y = (&vec.x)[set[1]-'x'];
318 qcfloat_t z = (&vec.x)[set[2]-'x'];
322 ++opts_optimizationcount[OPTIM_VECTOR_COMPONENTS];
323 out = (ast_expression*)ast_member_new(fold_ctx(fold), (ast_expression*)sel, set[0]-'x', NULL);
324 out->node.keep = false;
325 ((ast_member*)out)->rvalue = true;
327 return (ast_expression*)ast_binary_new(fold_ctx(fold), INSTR_MUL_F, fold_constgen_float(fold, x), out);
333 static GMQCC_INLINE ast_expression *fold_op_neg(fold_t *fold, ast_value *a) {
336 return fold_constgen_float(fold, -fold_immvalue_float(a));
337 } else if (isvector(a)) {
339 return fold_constgen_vector(fold, vec3_neg(fold_immvalue_vector(a)));
344 static GMQCC_INLINE ast_expression *fold_op_not(fold_t *fold, ast_value *a) {
347 return fold_constgen_float(fold, !fold_immvalue_float(a));
348 } else if (isvector(a)) {
350 return fold_constgen_float(fold, vec3_notf(fold_immvalue_vector(a)));
351 } else if (isstring(a)) {
353 if (OPTS_FLAG(TRUE_EMPTY_STRINGS))
354 return fold_constgen_float(fold, !fold_immvalue_string(a));
356 return fold_constgen_float(fold, !fold_immvalue_string(a) || !*fold_immvalue_string(a));
362 static GMQCC_INLINE ast_expression *fold_op_add(fold_t *fold, ast_value *a, ast_value *b) {
364 if (fold_can_2(a, b))
365 return fold_constgen_float(fold, fold_immvalue_float(a) + fold_immvalue_float(b));
366 } else if (isvector(a)) {
367 if (fold_can_2(a, b))
368 return fold_constgen_vector(fold, vec3_add(fold_immvalue_vector(a), fold_immvalue_vector(b)));
373 static GMQCC_INLINE ast_expression *fold_op_sub(fold_t *fold, ast_value *a, ast_value *b) {
375 if (fold_can_2(a, b))
376 return fold_constgen_float(fold, fold_immvalue_float(a) - fold_immvalue_float(b));
377 } else if (isvector(a)) {
378 if (fold_can_2(a, b))
379 return fold_constgen_vector(fold, vec3_sub(fold_immvalue_vector(a), fold_immvalue_vector(b)));
384 static GMQCC_INLINE ast_expression *fold_op_mul(fold_t *fold, ast_value *a, ast_value *b) {
386 if (isfloat(b) && fold_can_2(a, b))
387 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(b), fold_immvalue_float(a)));
388 else if (fold_can_2(a, b))
389 return fold_constgen_float(fold, fold_immvalue_float(a) * fold_immvalue_float(b));
390 } else if (isvector(a)) {
391 if (isfloat(b) && fold_can_2(a, b)) {
392 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
394 if (fold_can_2(a, b)) {
395 return fold_constgen_float(fold, vec3_mulvv(fold_immvalue_vector(a), fold_immvalue_vector(b)));
396 } else if (OPTS_OPTIMIZATION(OPTIM_VECTOR_COMPONENTS) && fold_can_1(a)) {
398 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "xyz"))) return out;
399 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "yxz"))) return out;
400 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(a), b, "zxy"))) return out;
401 } else if (OPTS_OPTIMIZATION(OPTIM_VECTOR_COMPONENTS) && fold_can_1(b)) {
403 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "xyz"))) return out;
404 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "yxz"))) return out;
405 if ((out = fold_op_mul_vec(fold, fold_immvalue_vector(b), a, "zxy"))) return out;
412 static GMQCC_INLINE ast_expression *fold_op_div(fold_t *fold, ast_value *a, ast_value *b) {
414 if (fold_can_2(a, b))
415 return fold_constgen_float(fold, fold_immvalue_float(a) / fold_immvalue_float(b));
416 } else if (isvector(a)) {
417 if (fold_can_2(a, b))
418 return fold_constgen_vector(fold, vec3_mulvf(fold_immvalue_vector(a), 1.0f / fold_immvalue_float(b)));
419 else if (fold_can_1(b))
420 return fold_constgen_float (fold, 1.0f / fold_immvalue_float(b));
425 static GMQCC_INLINE ast_expression *fold_op_mod(fold_t *fold, ast_value *a, ast_value *b) {
426 if (fold_can_2(a, b))
427 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) % ((qcint_t)fold_immvalue_float(b))));
431 static GMQCC_INLINE ast_expression *fold_op_bor(fold_t *fold, ast_value *a, ast_value *b) {
432 if (fold_can_2(a, b))
433 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) | ((qcint_t)fold_immvalue_float(b))));
437 static GMQCC_INLINE ast_expression *fold_op_band(fold_t *fold, ast_value *a, ast_value *b) {
438 if (fold_can_2(a, b))
439 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) & ((qcint_t)fold_immvalue_float(b))));
443 static GMQCC_INLINE ast_expression *fold_op_xor(fold_t *fold, ast_value *a, ast_value *b) {
445 if (fold_can_2(a, b))
446 return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) ^ ((qcint_t)fold_immvalue_float(b))));
449 if (fold_can_2(a, b))
450 return fold_constgen_vector(fold, vec3_xor(fold_immvalue_vector(a), fold_immvalue_vector(b)));
452 if (fold_can_2(a, b))
453 return fold_constgen_vector(fold, vec3_xorvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
459 static GMQCC_INLINE ast_expression *fold_op_lshift(fold_t *fold, ast_value *a, ast_value *b) {
460 if (fold_can_2(a, b) && isfloats(a, b))
461 return fold_constgen_float(fold, (qcfloat_t)((qcuint_t)(fold_immvalue_float(a)) << (qcuint_t)(fold_immvalue_float(b))));
465 static GMQCC_INLINE ast_expression *fold_op_rshift(fold_t *fold, ast_value *a, ast_value *b) {
466 if (fold_can_2(a, b) && isfloats(a, b))
467 return fold_constgen_float(fold, (qcfloat_t)((qcuint_t)(fold_immvalue_float(a)) >> (qcuint_t)(fold_immvalue_float(b))));
471 static GMQCC_INLINE ast_expression *fold_op_andor(fold_t *fold, ast_value *a, ast_value *b, float or) {
472 if (fold_can_2(a, b)) {
473 if (OPTS_FLAG(PERL_LOGIC)) {
474 if (fold_immediate_true(fold, a))
475 return (ast_expression*)b;
477 return fold_constgen_float (
479 ((or) ? (fold_immediate_true(fold, a) || fold_immediate_true(fold, b))
480 : (fold_immediate_true(fold, a) && fold_immediate_true(fold, b)))
489 static GMQCC_INLINE ast_expression *fold_op_tern(fold_t *fold, ast_value *a, ast_value *b, ast_value *c) {
491 return fold_immediate_true(fold, a)
493 : (ast_expression*)c;
498 static GMQCC_INLINE ast_expression *fold_op_exp(fold_t *fold, ast_value *a, ast_value *b) {
499 if (fold_can_2(a, b))
500 return fold_constgen_float(fold, (qcfloat_t)powf(fold_immvalue_float(a), fold_immvalue_float(b)));
504 static GMQCC_INLINE ast_expression *fold_op_lteqgt(fold_t *fold, ast_value *a, ast_value *b) {
505 if (fold_can_2(a,b)) {
506 if (fold_immvalue_float(a) < fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[2];
507 if (fold_immvalue_float(a) == fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[0];
508 if (fold_immvalue_float(a) > fold_immvalue_float(b)) return (ast_expression*)fold->imm_float[1];
513 static GMQCC_INLINE ast_expression *fold_op_cmp(fold_t *fold, ast_value *a, ast_value *b, bool ne) {
514 if (fold_can_2(a, b)) {
515 return fold_constgen_float(
517 (ne) ? (fold_immvalue_float(a) != fold_immvalue_float(b))
518 : (fold_immvalue_float(a) == fold_immvalue_float(b))
524 static GMQCC_INLINE ast_expression *fold_op_bnot(fold_t *fold, ast_value *a) {
526 return fold_constgen_float(fold, ~((qcint_t)fold_immvalue_float(a)));
530 ast_expression *fold_op(fold_t *fold, const oper_info *info, ast_expression **opexprs) {
531 ast_value *a = (ast_value*)opexprs[0];
532 ast_value *b = (ast_value*)opexprs[1];
533 ast_value *c = (ast_value*)opexprs[2];
535 /* can a fold operation be applied to this operator usage? */
539 switch(info->operands) {
540 case 3: if(!c) return NULL;
541 case 2: if(!b) return NULL;
544 compile_error(fold_ctx(fold), "interal error: fold_op no operands to fold\n");
550 case opid2('-', 'P'): return fold_op_neg (fold, a);
551 case opid2('!', 'P'): return fold_op_not (fold, a);
552 case opid1('+'): return fold_op_add (fold, a, b);
553 case opid1('-'): return fold_op_sub (fold, a, b);
554 case opid1('*'): return fold_op_mul (fold, a, b);
555 case opid1('/'): return fold_op_div (fold, a, b);
556 case opid1('%'): return fold_op_mod (fold, a, b);
557 case opid1('|'): return fold_op_bor (fold, a, b);
558 case opid1('&'): return fold_op_band (fold, a, b);
559 case opid1('^'): return fold_op_xor (fold, a, b);
560 case opid2('<','<'): return fold_op_lshift (fold, a, b);
561 case opid2('>','>'): return fold_op_rshift (fold, a, b);
562 case opid2('|','|'): return fold_op_andor (fold, a, b, true);
563 case opid2('&','&'): return fold_op_andor (fold, a, b, false);
564 case opid2('?',':'): return fold_op_tern (fold, a, b, c);
565 case opid2('*','*'): return fold_op_exp (fold, a, b);
566 case opid3('<','=','>'): return fold_op_lteqgt (fold, a, b);
567 case opid2('!','='): return fold_op_cmp (fold, a, b, true);
568 case opid2('=','='): return fold_op_cmp (fold, a, b, false);
569 case opid2('~','P'): return fold_op_bnot (fold, a);