static ir_node *equivalent_node_Block(ir_node *n)
{
ir_node *oldn = n;
- int n_preds = get_Block_n_cfgpreds(n);
+ int n_preds;
- /* The Block constructor does not call optimize, but mature_immBlock
- calls the optimization. */
+ /* don't optimize dead blocks */
+ if (is_Block_dead(n))
+ return n;
+
+ n_preds = get_Block_n_cfgpreds(n);
+
+ /* The Block constructor does not call optimize, but mature_immBlock()
+ calls the optimization. */
assert(get_Block_matured(n));
/* Straightening: a single entry Block following a single exit Block
/**
* Eor is commutative and has neutral 0.
*/
-#define equivalent_node_Eor equivalent_node_neutral_zero
+static ir_node *equivalent_node_Eor(ir_node *n)
+{
+ ir_node *oldn = n;
+ ir_node *a;
+ ir_node *b;
+
+ n = equivalent_node_neutral_zero(n);
+ if (n != oldn) return n;
+
+ a = get_Eor_left(n);
+ b = get_Eor_right(n);
+
+ if (is_Eor(a)) {
+ ir_node *aa = get_Eor_left(a);
+ ir_node *ab = get_Eor_right(a);
+
+ if (aa == b) {
+ /* (a ^ b) ^ a -> b */
+ n = ab;
+ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
+ return n;
+ } else if (ab == b) {
+ /* (a ^ b) ^ b -> a */
+ n = aa;
+ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
+ return n;
+ }
+ }
+ if (is_Eor(b)) {
+ ir_node *ba = get_Eor_left(b);
+ ir_node *bb = get_Eor_right(b);
+
+ if (ba == a) {
+ /* a ^ (a ^ b) -> b */
+ n = bb;
+ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
+ return n;
+ } else if (bb == a) {
+ /* a ^ (b ^ a) -> b */
+ n = ba;
+ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
+ return n;
+ }
+ }
+
+ return n;
+}
/*
* Optimize a - 0 and (a - x) + x (for modes with wrap-around).
/**
- * Optimize an "idempotent unary op", ie op(op(n)) = n.
+ * Optimize an "self-inverse unary op", ie op(op(n)) = n.
*
* @todo
* -(-a) == a, but might overflow two times.
set_Conv_op(n, get_Conv_op(a));
return n;
}
- /* else both are strict conv, second is superflous */
+ /* else both are strict conv, second is superfluous */
} else if(is_Proj(a)) {
ir_node *pred = get_Proj_pred(a);
if(is_Load(pred)) {
int i, n_preds;
ir_node *oldn = n;
- ir_node *block = NULL; /* to shutup gcc */
+ ir_node *block;
ir_node *first_val = NULL; /* to shutup gcc */
if (!get_opt_normalize()) return n;
n_preds = get_Phi_n_preds(n);
block = get_nodes_block(n);
- /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
- assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
if ((is_Block_dead(block)) || /* Control dead */
(block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
return new_Bad(); /* in the Start Block. */
* themselves.
*/
static ir_node *equivalent_node_Sync(ir_node *n) {
- int i, n_preds;
-
- ir_node *oldn = n;
- ir_node *first_val = NULL; /* to shutup gcc */
-
- if (!get_opt_normalize()) return n;
+ int arity = get_Sync_n_preds(n);
+ int i;
- n_preds = get_Sync_n_preds(n);
+ for (i = 0; i < arity;) {
+ ir_node *pred = get_Sync_pred(n, i);
+ int j;
- /* Find first non-self-referencing input */
- for (i = 0; i < n_preds; ++i) {
- first_val = get_Sync_pred(n, i);
- if ((first_val != n) /* not self pointer */ &&
- (! is_Bad(first_val))
- ) { /* value not dead */
- break; /* then found first value. */
+ /* Remove Bad predecessors */
+ if (is_Bad(pred)) {
+ del_Sync_n(n, i);
+ --arity;
+ continue;
}
- }
-
- if (i >= n_preds)
- /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
- return new_Bad();
- /* search the rest of inputs, determine if any of these
- are non-self-referencing */
- while (++i < n_preds) {
- ir_node *scnd_val = get_Sync_pred(n, i);
- if ((scnd_val != n) &&
- (scnd_val != first_val) &&
- (! is_Bad(scnd_val))
- )
- break;
+ /* Remove duplicate predecessors */
+ for (j = 0;; ++j) {
+ if (j >= i) {
+ ++i;
+ break;
+ }
+ if (get_Sync_pred(n, j) == pred) {
+ del_Sync_n(n, i);
+ --arity;
+ break;
+ }
+ }
}
- if (i >= n_preds) {
- /* Fold, if no multiple distinct non-self-referencing inputs */
- n = first_val;
- DBG_OPT_SYNC(oldn, n);
- }
+ if (arity == 0) return new_Bad();
+ if (arity == 1) return get_Sync_pred(n, 0);
return n;
} /* equivalent_node_Sync */
else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
ir_node *cmp = get_Proj_pred(sel);
long proj_nr = get_Proj_proj(sel);
- ir_node *b = get_Mux_false(n);
- ir_node *a = get_Mux_true(n);
+ ir_node *f = get_Mux_false(n);
+ ir_node *t = get_Mux_true(n);
/*
- * Note: normalization puts the constant on the right site,
- * so we check only one case.
- *
* Note further that these optimization work even for floating point
* with NaN's because -NaN == NaN.
* However, if +0 and -0 is handled differently, we cannot use the first one.
*/
- if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) {
- ir_node *cmp_r = get_Cmp_right(cmp);
- if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
- /* Mux(a CMP 0, X, a) */
- if (is_Minus(b) && get_Minus_op(b) == a) {
- /* Mux(a CMP 0, -a, a) */
- if (proj_nr == pn_Cmp_Eq) {
- /* Mux(a == 0, -a, a) ==> -a */
- n = b;
+ if (is_Cmp(cmp)) {
+ ir_node *const cmp_l = get_Cmp_left(cmp);
+ ir_node *const cmp_r = get_Cmp_right(cmp);
+
+ switch (proj_nr) {
+ case pn_Cmp_Eq:
+ if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
+ (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
+ n = f;
DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
- } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
- /* Mux(a != 0, -a, a) ==> a */
- n = a;
+ return n;
+ }
+ break;
+
+ case pn_Cmp_Lg:
+ case pn_Cmp_Ne:
+ if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
+ (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
+ n = t;
DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
+ return n;
}
- } else if (is_Const(b) && is_Const_null(b)) {
- /* Mux(a CMP 0, 0, a) */
- if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
- /* Mux(a != 0, 0, a) ==> a */
- n = a;
+ break;
+ }
+
+ /*
+ * Note: normalization puts the constant on the right side,
+ * so we check only one case.
+ */
+ if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
+ /* Mux(t CMP 0, X, t) */
+ if (is_Minus(f) && get_Minus_op(f) == t) {
+ /* Mux(t CMP 0, -t, t) */
+ if (proj_nr == pn_Cmp_Eq) {
+ /* Mux(t == 0, -t, t) ==> -t */
+ n = f;
DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
- } else if (proj_nr == pn_Cmp_Eq) {
- /* Mux(a == 0, 0, a) ==> 0 */
- n = b;
+ } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
+ /* Mux(t != 0, -t, t) ==> t */
+ n = t;
DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
}
}
int i, n;
if (get_nodes_block(a) != get_nodes_block(b))
- return NULL;
+ return NULL;
n = get_irn_arity(a);
NEW_ARR_A(void *, res, n);
unsigned ref_bits = get_mode_size_bits(mode);
if (is_Conv(left)) {
- ir_mode *mode = get_irn_mode(left);
- unsigned bits = get_mode_size_bits(mode);
+ ir_mode *lmode = get_irn_mode(left);
+ unsigned bits = get_mode_size_bits(lmode);
if (ref_bits == bits &&
- mode_is_int(mode) &&
- get_mode_arithmetic(mode) == irma_twos_complement) {
+ mode_is_int(lmode) &&
+ get_mode_arithmetic(lmode) == irma_twos_complement) {
ir_node *pre = get_Conv_op(left);
ir_mode *pre_mode = get_irn_mode(pre);
}
if (is_Conv(right)) {
- ir_mode *mode = get_irn_mode(right);
- unsigned bits = get_mode_size_bits(mode);
+ ir_mode *rmode = get_irn_mode(right);
+ unsigned bits = get_mode_size_bits(rmode);
if (ref_bits == bits &&
- mode_is_int(mode) &&
- get_mode_arithmetic(mode) == irma_twos_complement) {
+ mode_is_int(rmode) &&
+ get_mode_arithmetic(rmode) == irma_twos_complement) {
ir_node *pre = get_Conv_op(right);
ir_mode *pre_mode = get_irn_mode(pre);
}
}
}
+
+ /* let address arithmetic use unsigned modes */
+ if (is_Const(right)) {
+ ir_mode *rmode = get_irn_mode(right);
+
+ if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
+ /* convert a AddP(P, *s) into AddP(P, *u) */
+ ir_mode *nm = get_reference_mode_unsigned_eq(mode);
+
+ ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
+ set_binop_right(n, pre);
+ }
+ }
}
return n;
} /* transform_node_AddSub */
if (is_Const(b)) {
tarval *tv = get_Const_tarval(b);
+ int rem;
+ /*
+ * Floating point constant folding might be disabled here to
+ * prevent rounding.
+ * However, as we check for exact result, doing it is safe.
+ * Switch it on.
+ */
+ rem = tarval_enable_fp_ops(1);
tv = tarval_quo(get_mode_one(mode), tv);
+ (void)tarval_enable_fp_ops(rem);
/* Do the transformation if the result is either exact or we are not
using strict rules. */
/*
* We can replace the Abs by -x here.
- * We even could add a new Confirm here.
+ * We even could add a new Confirm here
+ * (if not twos complement)
*
* Note that -x would create a new node, so we could
* not run it in the equivalent_node() context.
(get_opt_unreachable_code())) {
/* It's a boolean Cond, branching on a boolean constant.
Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
- jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
+ ir_node *blk = get_nodes_block(n);
+ jmp = new_r_Jmp(current_ir_graph, blk);
turn_into_tuple(n, pn_Cond_max);
if (ta == tarval_b_true) {
set_Tuple_pred(n, pn_Cond_false, new_Bad());
set_Tuple_pred(n, pn_Cond_true, new_Bad());
}
/* We might generate an endless loop, so keep it alive. */
- add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
+ add_End_keepalive(get_irg_end(current_ir_graph), blk);
}
return n;
} /* transform_node_Cond */
-typedef ir_node* (*recursive_transform) (ir_node *n);
+/**
+ * Prototype of a recursive transform function
+ * for bitwise distributive transformations.
+ */
+typedef ir_node* (*recursive_transform)(ir_node *n);
/**
* makes use of distributive laws for and, or, eor
break;
}
- /* remove Casts */
- if (is_Cast(left))
- left = get_Cast_op(left);
- if (is_Cast(right))
- right = get_Cast_op(right);
+ /* remove Casts of both sides */
+ left = skip_Cast(left);
+ right = skip_Cast(right);
/* Remove unnecessary conversions */
/* TODO handle constants */
}
}
- /* remove operation of both sides if possible */
+ /* remove operation on both sides if possible */
if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
/*
* The following operations are NOT safe for floating point operations, for instance
/** returns 1 if a == -b */
static int is_negated_value(ir_node *a, ir_node *b) {
- if(is_Minus(a) && get_Minus_op(a) == b)
+ if (is_Minus(a) && get_Minus_op(a) == b)
return 1;
- if(is_Minus(b) && get_Minus_op(b) == a)
+ if (is_Minus(b) && get_Minus_op(b) == a)
return 1;
- if(is_Sub(a) && is_Sub(b)) {
+ if (is_Sub(a) && is_Sub(b)) {
ir_node *a_left = get_Sub_left(a);
ir_node *a_right = get_Sub_right(a);
ir_node *b_left = get_Sub_left(b);
ir_node *b_right = get_Sub_right(b);
- if(a_left == b_right && a_right == b_left)
+ if (a_left == b_right && a_right == b_left)
return 1;
}
tarval *tv_t = get_Const_tarval(t);
if (tv_t == tarval_b_true) {
if (is_Const(f)) {
+ /* Muxb(sel, true, false) = sel */
assert(get_Const_tarval(f) == tarval_b_false);
+ DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
return sel;
} else {
- return new_rd_Or(dbg, irg, block, sel, f, mode_b);
+ /* Muxb(sel, true, x) = Or(sel, x) */
+ n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
+ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
+ return n;
}
} else {
ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
assert(tv_t == tarval_b_false);
if (is_Const(f)) {
+ /* Muxb(sel, false, true) = Not(sel) */
assert(get_Const_tarval(f) == tarval_b_true);
+ DBG_OPT_ALGSIM0(oldn, not_sel, FS_OPT_MUX_NOT_BOOL);
return not_sel;
} else {
- return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
+ /* Muxb(sel, false, x) = And(Not(sel), x) */
+ n = new_rd_And(dbg, irg, block, not_sel, f, mode_b);
+ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ANDNOT_BOOL);
+ return n;
}
}
} else if (is_Const(f)) {
tarval *tv_f = get_Const_tarval(f);
if (tv_f == tarval_b_true) {
+ /* Muxb(sel, x, true) = Or(Not(sel), x) */
ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
- return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
+ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
+ n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
+ return n;
} else {
+ /* Muxb(sel, x, false) = And(sel, x) */
assert(tv_f == tarval_b_false);
- return new_rd_And(dbg, irg, block, sel, t, mode_b);
+ n = new_rd_And(dbg, irg, block, sel, t, mode_b);
+ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
+ return n;
}
}
}
- if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
+ if (is_Proj(sel)) {
ir_node *cmp = get_Proj_pred(sel);
long pn = get_Proj_proj(sel);
ir_node *f = get_Mux_false(n);
*
* Note further that these optimization work even for floating point
* with NaN's because -NaN == NaN.
- * However, if +0 and -0 is handled differently, we cannot use the first
- * one.
+ * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
+ * transformations.
*/
if (is_Cmp(cmp)) {
ir_node *cmp_r = get_Cmp_right(cmp);
if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
- ir_node *block = get_irn_n(n, -1);
+ ir_node *block = get_nodes_block(n);
+ ir_node *cmp_l = get_Cmp_left(cmp);
- if(is_negated_value(f, t)) {
- ir_node *cmp_left = get_Cmp_left(cmp);
+ if (mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
+ /* f = -t */
- /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
- if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
- || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
+ if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
+ || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
{
+ /* Psi(a >/>= 0, a, -a) = Psi(a </<= 0, -a, a) ==> Abs(a) */
n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
- cmp_left, mode);
+ cmp_l, mode);
DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
return n;
- /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
- } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
- || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
+ } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
+ || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
{
+ /* Psi(a </<= 0, a, -a) = Psi(a >/>= 0, -a, a) ==> -Abs(a) */
n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
- cmp_left, mode);
+ cmp_l, mode);
n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
block, n, mode);
DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
return n;
}
}
+
+ if (mode_is_int(mode)) {
+ /* integer only */
+ if (pn == pn_Cmp_Eq) {
+ ir_node *tmp = t;
+ t = f;
+ f = tmp;
+ pn = pn_Cmp_Lg;
+ }
+ if (pn == pn_Cmp_Lg && is_And(cmp_l)) {
+ /* Psi((a & b) != 0, c, 0) */
+ ir_node *and_r = get_And_right(cmp_l);
+ ir_node *and_l;
+
+ if (and_r == t && f == cmp_r) {
+ if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
+ /* Psi((a & 2^C) != 0, 2^C, 0) */
+ n = cmp_l;
+ return n;
+ }
+ }
+ if (and_r == f && t == cmp_r) {
+ if (is_Const(f) && tarval_is_single_bit(get_Const_tarval(f))) {
+ /* Psi((a & 2^C) != 0, 0, 2^C) */
+ n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
+ block, cmp_l, f, mode);
+ return n;
+ }
+ }
+ if (is_Shl(and_r)) {
+ ir_node *shl_l = get_Shl_left(and_r);
+ if (is_Const(shl_l) && is_Const_one(shl_l)) {
+ if (and_r == t && f == cmp_r) {
+ /* (a & (1 << n)) != 0, (1 << n), 0) */
+ n = cmp_l;
+ return n;
+ }
+ else if (and_r == f && t == cmp_r) {
+ /* (a & (1 << n)) != 0, 0, (1 << n)) */
+ n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
+ block, cmp_l, f, mode);
+ return n;
+ }
+ }
+ }
+ and_l = get_And_left(cmp_l);
+ if (is_Shl(and_l)) {
+ ir_node *shl_l = get_Shl_left(and_l);
+ if (is_Const(shl_l) && is_Const_one(shl_l)) {
+ if (and_l == t && f == cmp_r) {
+ /* ((1 << n) & a) != 0, (1 << n), 0) */
+ n = cmp_l;
+ return n;
+ }
+ else if (and_l == f && t == cmp_r) {
+ /* ((1 << n) & a) != 0, 0, (1 << n)) */
+ n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
+ block, cmp_l, f, mode);
+ return n;
+ }
+ }
+ }
+ }
+ }
}
}
}
* of the other sync to our own inputs
*/
static ir_node *transform_node_Sync(ir_node *n) {
- int i, arity;
+ int arity = get_Sync_n_preds(n);
+ int i;
- arity = get_irn_arity(n);
- for(i = 0; i < get_irn_arity(n); /*empty*/) {
- int i2, arity2;
- ir_node *in = get_irn_n(n, i);
- if(!is_Sync(in)) {
+ for (i = 0; i < arity;) {
+ ir_node *pred = get_Sync_pred(n, i);
+ int pred_arity;
+ int j;
+
+ if (!is_Sync(pred)) {
++i;
continue;
}
- /* set sync input 0 instead of the sync */
- set_irn_n(n, i, get_irn_n(in, 0));
- /* so we check this input again for syncs */
-
- /* append all other inputs of the sync to our sync */
- arity2 = get_irn_arity(in);
- for(i2 = 1; i2 < arity2; ++i2) {
- ir_node *in_in = get_irn_n(in, i2);
- add_irn_n(n, in_in);
- /* increase arity so we also check the new inputs for syncs */
- arity++;
+ del_Sync_n(n, i);
+ --arity;
+
+ pred_arity = get_Sync_n_preds(pred);
+ for (j = 0; j < pred_arity; ++j) {
+ ir_node *pred_pred = get_Sync_pred(pred, j);
+ int k;
+
+ for (k = 0;; ++k) {
+ if (k >= arity) {
+ add_irn_n(n, pred_pred);
+ ++arity;
+ break;
+ }
+ if (get_Sync_pred(n, k) == pred_pred) break;
+ }
}
}
static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
const symconst_attr *pa = get_irn_symconst_attr(a);
const symconst_attr *pb = get_irn_symconst_attr(b);
- return (pa->num != pb->num)
+ return (pa->kind != pb->kind)
|| (pa->sym.type_p != pb->sym.type_p)
|| (pa->tp != pb->tp);
} /* node_cmp_attr_SymConst */
projects / libfirm / blobdiff