X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Freturn.c;h=3b5338aaf2563ab8da6fbd4295104578d9d7a746;hb=46de9c58b7126770f460c97a3b95b6e0402804f8;hp=ad66d18bd41f691b25af9dc94bc4cc6a86bda8b9;hpb=c59c04ea5eb29275896b6f8ec8ef1646ed11265b;p=libfirm diff --git a/ir/opt/return.c b/ir/opt/return.c index ad66d18bd..3b5338aaf 100644 --- a/ir/opt/return.c +++ b/ir/opt/return.c @@ -33,7 +33,7 @@ /* * Normalize the Returns of a graph by creating a new End block * with One Return(Phi). - * This is the prefered input for the if-conversion. + * This is the preferred input for the if-conversion. * * In pseudocode, it means: * @@ -55,19 +55,25 @@ void normalize_one_return(ir_graph *irg) ir_node *endbl = get_irg_end_block(irg); int i, j, k, n, last_idx, n_rets, n_ret_vals = -1; unsigned char *returns; - ir_node **in, **retvals; + ir_node **in, **retvals, **endbl_in; - ir_node *block, *new_ret; + ir_node *block; /* look, if we have more than one return */ - n = get_Block_n_cfgpreds(endbl); + n = get_Block_n_cfgpreds(endbl); + if (n <= 0) { + /* The end block has no predecessors, we have an endless + loop. In that case, no returns exists. */ + return; + } + returns = alloca((n + 7) >> 3); memset(returns, 0, (n + 7) >> 3); for (n_rets = i = 0; i < n; ++i) { ir_node *node = get_Block_cfgpred(endbl, i); - if (get_irn_op(node) == op_Return) { + if (is_Return(node)) { ++n_rets; set_bit(i); @@ -81,12 +87,15 @@ void normalize_one_return(ir_graph *irg) if (n_rets <= 1) return; - in = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals)); - retvals = alloca(sizeof(*in) * n_rets * n_ret_vals); + in = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals)); + retvals = alloca(sizeof(*retvals) * n_rets * n_ret_vals); + endbl_in = alloca(sizeof(*endbl_in) * n); + last_idx = 0; for (j = i = 0; i < n; ++i) { + ir_node *ret = get_Block_cfgpred(endbl, i); + if (get_bit(i)) { - ir_node *ret = get_Block_cfgpred(endbl, i); ir_node *block = get_nodes_block(ret); /* create a new Jmp for every Ret and place the in in */ @@ -94,13 +103,12 @@ void normalize_one_return(ir_graph *irg) /* save the return values and shuffle them */ for (k = 0; k < n_ret_vals; ++k) - retvals[j + k*n_ret_vals] = get_irn_n(ret, k); + retvals[j + k*n_rets] = get_irn_n(ret, k); ++j; - - set_Block_cfgpred(endbl, i, new_r_Bad(irg)); - last_idx = i; } + else + endbl_in[last_idx++] = ret; } /* ok, create a new block with all created in's */ @@ -108,21 +116,38 @@ void normalize_one_return(ir_graph *irg) /* now create the Phi nodes */ for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) { + int k; + ir_node *first; /* the return values are already shuffled */ - in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j])); + + /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a + but NOT, if a is Unknown. Here, we known that this case can be optimize also, + so do it here */ + first = retvals[j + 0]; + for (k = 1; k < n_rets; ++k) { + if (retvals[j + k] != first) { + first = NULL; + break; + } + } + if (first) + in[i] = first; + else + in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j])); } - new_ret = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]); + endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]); - set_Block_cfgpred(endbl, last_idx, new_ret); + set_irn_in(endbl, last_idx, endbl_in); /* invalidate analysis information: * a new Block was added, so dominator, outs and loop are inconsistent, * trouts and callee-state should be still valid */ - set_irg_dom_inconsistent(irg); + set_irg_doms_inconsistent(irg); set_irg_outs_inconsistent(irg); - set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent); + set_irg_extblk_inconsistent(irg); + set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent); } /** @@ -169,7 +194,7 @@ static int can_move_ret(ir_node *ret) /* * Normalize the Returns of a graph by moving * the Returns upwards as much as possible. - * This might be prefered for code generation. + * This might be preferred for code generation. * * In pseudocode, it means: * @@ -195,12 +220,17 @@ void normalize_n_returns(ir_graph *irg) ir_node *endbl = get_irg_end_block(irg); ir_node *end; - /* first, link all returns */ + /* + * First, link all returns: + * These must be predecessors of the endblock. + * Place Returns that can be moved on list, all others + * on final. + */ n = get_Block_n_cfgpreds(endbl); for (n_finals = n_rets = i = 0; i < n; ++i) { ir_node *ret = get_Block_cfgpred(endbl, i); - if (get_irn_op(ret) == op_Return && can_move_ret(ret)) { + if (is_Return(ret) && can_move_ret(ret)) { /* * Ok, all conditions met, we can move this Return, put it * on our work list. @@ -238,15 +268,19 @@ void normalize_n_returns(ir_graph *irg) n = get_Block_n_cfgpreds(block); for (i = 0; i < n; ++i) { - ir_node *jmp = get_Block_cfgpred(block, i); - ir_node *new_bl = get_nodes_block(jmp); - ir_node *new_ret; + ir_node *jmp = get_Block_cfgpred(block, i); + ir_node *new_bl, *new_ret; + + if (get_irn_op(jmp) != op_Jmp) + continue; + + new_bl = get_nodes_block(jmp); /* create the in-array for the new Ret */ for (j = 0; j < n_ret_vals; ++j) { ir_node *pred = get_irn_n(ret, j); - in[j] = is_Phi(pred) ? get_Phi_pred(pred, i) : pred; + in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred; } new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]); @@ -270,6 +304,9 @@ void normalize_n_returns(ir_graph *irg) ++n_finals; } } + + /* remove the Jmp, we have placed a Return here */ + exchange(jmp, new_r_Bad(irg)); } /* @@ -300,14 +337,14 @@ void normalize_n_returns(ir_graph *irg) exchange(endbl, new_r_Block(irg, n_finals, in)); - /* the end block is not automatically skiped, so do it here */ + /* the end block is not automatically skipped, so do it here */ set_irg_end_block(irg, skip_Id(get_irg_end_block(irg))); /* Invalidate analysis information: - * Blocks become dead and new Eeturns were deleted, so dominator, outs and loop are inconsistent, + * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent, * trouts and callee-state should be still valid */ - set_irg_dom_inconsistent(irg); + set_irg_doms_inconsistent(irg); set_irg_outs_inconsistent(irg); set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent); }