* stage constant folding, where, witht he help of the AST, operator
* usages can be constant folded. Then there is the constant folding
* in the IR for things like eliding if statements, can occur.
- *
+ *
* This file is thus, split into two parts.
*/
/*
* Implementation of basic vector math for vec3_t, for trivial constant
* folding.
- *
+ *
* TODO: gcc/clang hinting for autovectorization
*/
static GMQCC_INLINE vec3_t vec3_add(vec3_t a, vec3_t b) {
return out;
}
+static GMQCC_INLINE vec3_t vec3_or(vec3_t a, vec3_t b) {
+ vec3_t out;
+ out.x = (qcfloat_t)(((qcint_t)a.x) | ((qcint_t)b.x));
+ out.y = (qcfloat_t)(((qcint_t)a.y) | ((qcint_t)b.y));
+ out.z = (qcfloat_t)(((qcint_t)a.z) | ((qcint_t)b.z));
+ return out;
+}
+
+static GMQCC_INLINE vec3_t vec3_orvf(vec3_t a, qcfloat_t b) {
+ vec3_t out;
+ out.x = (qcfloat_t)(((qcint_t)a.x) | ((qcint_t)b));
+ out.y = (qcfloat_t)(((qcint_t)a.y) | ((qcint_t)b));
+ out.z = (qcfloat_t)(((qcint_t)a.z) | ((qcint_t)b));
+ return out;
+}
+
+static GMQCC_INLINE vec3_t vec3_and(vec3_t a, vec3_t b) {
+ vec3_t out;
+ out.x = (qcfloat_t)(((qcint_t)a.x) & ((qcint_t)b.x));
+ out.y = (qcfloat_t)(((qcint_t)a.y) & ((qcint_t)b.y));
+ out.z = (qcfloat_t)(((qcint_t)a.z) & ((qcint_t)b.z));
+ return out;
+}
+
+static GMQCC_INLINE vec3_t vec3_andvf(vec3_t a, qcfloat_t b) {
+ vec3_t out;
+ out.x = (qcfloat_t)(((qcint_t)a.x) & ((qcint_t)b));
+ out.y = (qcfloat_t)(((qcint_t)a.y) & ((qcint_t)b));
+ out.z = (qcfloat_t)(((qcint_t)a.z) & ((qcint_t)b));
+ return out;
+}
+
static GMQCC_INLINE vec3_t vec3_xor(vec3_t a, vec3_t b) {
vec3_t out;
out.x = (qcfloat_t)(((qcint_t)a.x) ^ ((qcint_t)b.x));
return out;
}
+static GMQCC_INLINE vec3_t vec3_not(vec3_t a) {
+ vec3_t out;
+ out.x = (qcfloat_t)(~((qcint_t)a.x));
+ out.y = (qcfloat_t)(~((qcint_t)a.y));
+ out.z = (qcfloat_t)(~((qcint_t)a.z));
+ return out;
+}
+
static GMQCC_INLINE qcfloat_t vec3_mulvv(vec3_t a, vec3_t b) {
return (a.x * b.x + a.y * b.y + a.z * b.z);
}
return !!v->constval.vfloat;
case TYPE_INTEGER:
return !!v->constval.vint;
- case TYPE_VECTOR:
+ case TYPE_VECTOR:
if (OPTS_FLAG(CORRECT_LOGIC))
return vec3_pbool(v->constval.vvec);
return !!(v->constval.vvec.x);
(void)fold_constgen_float (fold, -1.0f);
(void)fold_constgen_vector(fold, vec3_create(0.0f, 0.0f, 0.0f));
+ (void)fold_constgen_vector(fold, vec3_create(-1.0f, -1.0f, -1.0f));
return fold;
}
/*
* vector-component constant folding works by matching the component sets
* to eliminate expensive operations on whole-vectors (3 components at runtime).
- * to achive this effect in a clean manner this function generalizes the
+ * to achive this effect in a clean manner this function generalizes the
* values through the use of a set paramater, which is used as an indexing method
* for creating the elided ast binary expression.
*
* Consider 'n 0 0' where y, and z need to be tested for 0, and x is
- * used as the value in a binary operation generating an INSTR_MUL instruction
+ * used as the value in a binary operation generating an INSTR_MUL instruction,
* to acomplish the indexing of the correct component value we use set[0], set[1], set[2]
* as x, y, z, where the values of those operations return 'x', 'y', 'z'. Because
* of how ASCII works we can easily deliniate:
* vec.z is the same as set[2]-'x' for when set[2] is 'z', 'z'-'x' results in a
* literal value of 2, using this 2, we know that taking the address of vec->x (float)
* and indxing it with this literal will yeild the immediate address of that component
- *
+ *
* Of course more work needs to be done to generate the correct index for the ast_member_new
* call, which is no problem: set[0]-'x' suffices that job.
*/
}
static GMQCC_INLINE ast_expression *fold_op_bor(fold_t *fold, ast_value *a, ast_value *b) {
- if (fold_can_2(a, b))
- return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) | ((qcint_t)fold_immvalue_float(b))));
+ if (isfloat(a)) {
+ if (fold_can_2(a, b))
+ return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) | ((qcint_t)fold_immvalue_float(b))));
+ } else {
+ if (isvector(b)) {
+ if (fold_can_2(a, b))
+ return fold_constgen_vector(fold, vec3_or(fold_immvalue_vector(a), fold_immvalue_vector(b)));
+ } else {
+ if (fold_can_2(a, b))
+ return fold_constgen_vector(fold, vec3_orvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
+ }
+ }
return NULL;
}
static GMQCC_INLINE ast_expression *fold_op_band(fold_t *fold, ast_value *a, ast_value *b) {
- if (fold_can_2(a, b))
- return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) & ((qcint_t)fold_immvalue_float(b))));
+ if (isfloat(a)) {
+ if (fold_can_2(a, b))
+ return fold_constgen_float(fold, (qcfloat_t)(((qcint_t)fold_immvalue_float(a)) & ((qcint_t)fold_immvalue_float(b))));
+ } else {
+ if (isvector(b)) {
+ if (fold_can_2(a, b))
+ return fold_constgen_vector(fold, vec3_and(fold_immvalue_vector(a), fold_immvalue_vector(b)));
+ } else {
+ if (fold_can_2(a, b))
+ return fold_constgen_vector(fold, vec3_andvf(fold_immvalue_vector(a), fold_immvalue_float(b)));
+ }
+ }
return NULL;
}
return NULL;
}
-static GMQCC_INLINE ast_expression *fold_op_andor(fold_t *fold, ast_value *a, ast_value *b, float or) {
+static GMQCC_INLINE ast_expression *fold_op_andor(fold_t *fold, ast_value *a, ast_value *b, float expr) {
if (fold_can_2(a, b)) {
if (OPTS_FLAG(PERL_LOGIC)) {
if (fold_immediate_true(fold, a))
return (ast_expression*)b;
} else {
return fold_constgen_float (
- fold,
- ((or) ? (fold_immediate_true(fold, a) || fold_immediate_true(fold, b))
- : (fold_immediate_true(fold, a) && fold_immediate_true(fold, b)))
+ fold,
+ ((expr) ? (fold_immediate_true(fold, a) || fold_immediate_true(fold, b))
+ : (fold_immediate_true(fold, a) && fold_immediate_true(fold, b)))
? 1
: 0
);
}
static GMQCC_INLINE ast_expression *fold_op_bnot(fold_t *fold, ast_value *a) {
- if (fold_can_1(a))
- return fold_constgen_float(fold, ~((qcint_t)fold_immvalue_float(a)));
+ if (isfloat(a)) {
+ if (fold_can_1(a))
+ return fold_constgen_float(fold, ~((qcint_t)fold_immvalue_float(a)));
+ } else {
+ if (isvector(a)) {
+ if (fold_can_1(a))
+ return fold_constgen_vector(fold, vec3_not(fold_immvalue_vector(a)));
+ }
+ }
return NULL;
}
ast_expression *fold_op(fold_t *fold, const oper_info *info, ast_expression **opexprs) {
- ast_value *a = (ast_value*)opexprs[0];
- ast_value *b = (ast_value*)opexprs[1];
- ast_value *c = (ast_value*)opexprs[2];
+ ast_value *a = (ast_value*)opexprs[0];
+ ast_value *b = (ast_value*)opexprs[1];
+ ast_value *c = (ast_value*)opexprs[2];
+ ast_expression *e = NULL;
/* can a fold operation be applied to this operator usage? */
if (!info->folds)
case 2: if(!b) return NULL;
case 1:
if(!a) {
- compile_error(fold_ctx(fold), "interal error: fold_op no operands to fold\n");
+ compile_error(fold_ctx(fold), "internal error: fold_op no operands to fold\n");
return NULL;
}
}
+ /*
+ * we could use a boolean and default case but ironically gcc produces
+ * invalid broken assembly from that operation. clang/tcc get it right,
+ * but interestingly ignore compiling this to a jump-table when I do that,
+ * this happens to be the most efficent method, since you have per-level
+ * granularity on the pointer check happening only for the case you check
+ * it in. Opposed to the default method which would involve a boolean and
+ * pointer check after wards.
+ */
+ #define fold_op_case(ARGS, ARGS_OPID, OP, ARGS_FOLD) \
+ case opid##ARGS ARGS_OPID: \
+ if ((e = fold_op_##OP ARGS_FOLD)) { \
+ ++opts_optimizationcount[OPTIM_CONST_FOLD]; \
+ } \
+ return e
+
switch(info->id) {
- case opid2('-','P'): return fold_op_neg (fold, a);
- case opid2('!','P'): return fold_op_not (fold, a);
- case opid1('+'): return fold_op_add (fold, a, b);
- case opid1('-'): return fold_op_sub (fold, a, b);
- case opid1('*'): return fold_op_mul (fold, a, b);
- case opid1('/'): return fold_op_div (fold, a, b);
- case opid1('%'): return fold_op_mod (fold, a, b);
- case opid1('|'): return fold_op_bor (fold, a, b);
- case opid1('&'): return fold_op_band (fold, a, b);
- case opid1('^'): return fold_op_xor (fold, a, b);
- case opid2('<','<'): return fold_op_lshift (fold, a, b);
- case opid2('>','>'): return fold_op_rshift (fold, a, b);
- case opid2('|','|'): return fold_op_andor (fold, a, b, true);
- case opid2('&','&'): return fold_op_andor (fold, a, b, false);
- case opid2('?',':'): return fold_op_tern (fold, a, b, c);
- case opid2('*','*'): return fold_op_exp (fold, a, b);
- case opid3('<','=','>'): return fold_op_lteqgt (fold, a, b);
- case opid2('!','='): return fold_op_cmp (fold, a, b, true);
- case opid2('=','='): return fold_op_cmp (fold, a, b, false);
- case opid2('~','P'): return fold_op_bnot (fold, a);
+ fold_op_case(2, ('-', 'P'), neg, (fold, a));
+ fold_op_case(2, ('!', 'P'), not, (fold, a));
+ fold_op_case(1, ('+'), add, (fold, a, b));
+ fold_op_case(1, ('-'), sub, (fold, a, b));
+ fold_op_case(1, ('*'), mul, (fold, a, b));
+ fold_op_case(1, ('/'), div, (fold, a, b));
+ fold_op_case(1, ('%'), mod, (fold, a, b));
+ fold_op_case(1, ('|'), bor, (fold, a, b));
+ fold_op_case(1, ('&'), band, (fold, a, b));
+ fold_op_case(1, ('^'), xor, (fold, a, b));
+ fold_op_case(2, ('<', '<'), lshift, (fold, a, b));
+ fold_op_case(2, ('>', '>'), rshift, (fold, a, b));
+ fold_op_case(2, ('|', '|'), andor, (fold, a, b, true));
+ fold_op_case(2, ('&', '&'), andor, (fold, a, b, false));
+ fold_op_case(2, ('?', ':'), tern, (fold, a, b, c));
+ fold_op_case(2, ('*', '*'), exp, (fold, a, b));
+ fold_op_case(3, ('<','=','>'), lteqgt, (fold, a, b));
+ fold_op_case(2, ('!', '='), cmp, (fold, a, b, true));
+ fold_op_case(2, ('=', '='), cmp, (fold, a, b, false));
+ fold_op_case(2, ('~', 'P'), bnot, (fold, a));
}
- compile_error(fold_ctx(fold), "internal error: attempted to constant for unsupported operator");
+ #undef fold_op_case
+ compile_error(fold_ctx(fold), "internal error: attempted to constant-fold for unsupported operator");
+ return NULL;
+}
+
+#define expect(X) \
+ do { \
+ if (vec_size(params) != (X)) { \
+ compile_error( \
+ fold_ctx(fold), \
+ "internal error: attempted to constant-fold with invalid paramaters for intrinsic `%s`", \
+ intrin \
+ ); \
+ return NULL; \
+ } \
+ } while (0)
+
+ast_expression *fold_intrin(fold_t *fold, const char *intrin, ast_expression **params) {
+ if (!fold) return NULL;
+ if (!intrin) return NULL;
+
+ if (!strcmp(intrin, "__builtin_exp")) {
+ expect(1);
+ ++opts_optimizationcount[OPTIM_CONST_FOLD];
+ return fold_constgen_float(fold, exp(fold_immvalue_float((ast_value*)params[0])));
+ }
+
+ if (!strcmp(intrin, "__builtin_mod")) {
+ expect(2);
+ ++opts_optimizationcount[OPTIM_CONST_FOLD];
+ return fold_constgen_float(
+ fold,
+ fmodf(
+ fold_immvalue_float((ast_value*)params[0]),
+ fold_immvalue_float((ast_value*)params[1])
+ )
+ );
+ }
+
+ if (!strcmp(intrin, "__builtin_pow")) {
+ expect(2);
+ ++opts_optimizationcount[OPTIM_CONST_FOLD];
+ return fold_constgen_float(
+ fold,
+ powf(
+ fold_immvalue_float((ast_value*)params[0]),
+ fold_immvalue_float((ast_value*)params[1])
+ )
+ );
+ }
+
+ if (!strcmp(intrin, "__builtin_isnan")) {
+ expect(1);
+ ++opts_optimizationcount[OPTIM_CONST_FOLD];
+ return fold_constgen_float(fold, isnan(fold_immvalue_float((ast_value*)params[0])) != 0.0f);
+ }
+
+ if (!strcmp(intrin, "__builtin_fabs")) {
+ expect(1);
+ ++opts_optimizationcount[OPTIM_CONST_FOLD];
+ return fold_constgen_float(fold, fabs(fold_immvalue_float((ast_value*)params[0])));
+ }
+
return NULL;
}
* expressions, which is the only supported thing so far. We undefine the
* testing macros here because an ir_value is differant than an ast_value.
*/
+#undef expect
#undef isfloat
#undef isstring
#undef isvector
ast_expression_codegen *cgen;
ir_block *elide;
ir_value *dummy;
- bool istrue = (fold_immvalue_float(condval) == 1.0f && branch->on_true);
+ bool istrue = (fold_immvalue_float(condval) != 0.0f && branch->on_true);
bool isfalse = (fold_immvalue_float(condval) == 0.0f && branch->on_false);
ast_expression *path = (istrue) ? branch->on_true :
(isfalse) ? branch->on_false : NULL;
- if (!path)
- return false;
+ if (!path) {
+ /*
+ * no path to take implies that the evaluation is if(0) and there
+ * is no else block. so eliminate all the code.
+ */
+ ++opts_optimizationcount[OPTIM_CONST_FOLD_DCE];
+ return true;
+ }
+
if (!(elide = ir_function_create_block(ast_ctx(branch), func->ir_func, ast_function_label(func, ((istrue) ? "ontrue" : "onfalse")))))
return false;
if (!(*(cgen = path->codegen))((ast_expression*)path, func, false, &dummy))
if (!ir_block_create_jump(func->curblock, ast_ctx(branch), elide))
return false;
/*
- * now the branch has been eliminates, and the correct block for the constant evaluation
+ * now the branch has been eliminated and the correct block for the constant evaluation
* is expanded into the current block for the function.
*/
func->curblock = elide;