From 7552843d51f33d3132dd55df6d9e1eff8a39fe6e Mon Sep 17 00:00:00 2001 From: Christoph Mallon Date: Wed, 3 May 2006 08:29:19 +0000 Subject: [PATCH] Add if conversion, use with -fif-conversion [r7685] --- ir/opt/ifconv.c | 536 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 536 insertions(+) diff --git a/ir/opt/ifconv.c b/ir/opt/ifconv.c index a9b06fdb0..f31b802a0 100644 --- a/ir/opt/ifconv.c +++ b/ir/opt/ifconv.c @@ -9,6 +9,538 @@ * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. */ +#if 1 + +#include +#include +#include "cdep.h" +#include "ircons.h" +#include "ifconv.h" +#include "irdom.h" +#include "irgmod.h" +#include "irgopt.h" +#include "irgwalk.h" +#include "irtools.h" +#include "return.h" + +// debug +#include "irdump.h" +#include "irprintf.h" + + +static ir_node* walk_to_projx(ir_node* start) +{ + ir_node* pred; + ir_node* cdep; + + pred = get_nodes_block(start); + cdep = get_unique_cdep(pred); + if (cdep == NULL) return NULL; + + assert(get_irn_arity(pred) == 1); + pred = get_irn_n(pred, 0); + if (get_irn_op(pred) == op_Proj) { + assert(get_irn_mode(pred) == mode_X); + return pred; + } else { + return NULL; + } +} + + +typedef struct block_info { + ir_node* phi; + int evil; +} block_info; + + +static int can_empty_block(ir_node* block) +{ + return !((block_info*)get_irn_link(block))->evil; +} + + +/* + * Copies the DAG starting at node to the ith predecessor block of src_block + * -if the node isn't in the src_block, this is a nop and the node is returned + * -if the node is a phi in the src_block, the ith predecessor of the phi is + * returned + * otherwise returns the copy of the passed node + */ +static ir_node* copy_to(ir_node* node, ir_node* src_block, int i) +{ + ir_node* dst_block; + ir_node* copy; + int arity; + int j; + + if (get_nodes_block(node) != src_block) return node; + if (get_irn_op(node) == op_Phi) return get_irn_n(node, i); + + copy_irn_to_irg(node, current_ir_graph); + copy = get_irn_link(node); + dst_block = get_nodes_block(get_irn_n(src_block, i)); + set_nodes_block(copy, dst_block); + + ir_fprintf(stderr, "Copying node %+F to block %+F, copy is %+F\n", + node, dst_block, copy + ); + + arity = get_irn_arity(node); + for (j = 0; j < arity; ++j) { + set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i)); + ir_fprintf(stderr, "-- pred %d is %+F\n", j, get_irn_n(copy, j)); + } + return copy; +} + + +/* + * Duplicate and move the contents of ith block predecessor into its + * predecessors if the block has multiple control dependencies and only one + * successor. + * Also bail out if the block contains non-movable nodes, because later + * if-conversion would be pointless. + */ +static int fission_block(ir_node* block, int i) +{ + ir_node* pred = get_irn_n(block, i); + ir_node* pred_block; + block_info* info; + ir_node* phi; + int pred_arity; + int arity; + ir_node** ins; + int j; + + if (get_irn_op(pred) != op_Jmp) return 0; + pred_block = get_nodes_block(pred); + + if (!has_multiple_cdep(pred_block)) return 0; + if (!can_empty_block(pred_block)) return 0; + + ir_fprintf(stderr, "Fissioning block %+F\n", pred_block); + + pred_arity = get_irn_arity(pred_block); + arity = get_irn_arity(block); + info = get_irn_link(block); + ins = malloc(sizeof(*ins) * (arity + pred_arity - 1)); + for (phi = info->phi; phi != NULL; phi = get_irn_link(phi)) { + for (j = 0; j < i; ++j) ins[j] = get_irn_n(phi, j); + for (j = 0; j < pred_arity; ++j) { + ins[i + j] = copy_to(get_irn_n(phi, i), pred_block, j); + } + for (j = i + 1; j < arity; ++j) { + ins[pred_arity - 1 + j] = get_irn_n(phi, j); + } + set_irn_in(phi, arity + pred_arity - 1, ins); + } + for (j = 0; j < i; ++j) ins[j] = get_irn_n(block, j); + for (j = 0; j < pred_arity; ++j) ins[i + j] = get_irn_n(pred_block, j); + for (j = i + 1; j < arity; ++j) ins[pred_arity - 1 + j] = get_irn_n(block, j); + set_irn_in(block, arity + pred_arity - 1, ins); + free(ins); + + /* Kill all Phis in the fissioned block + * This is to make sure they're not kept alive + */ + info = get_irn_link(pred_block); + phi = info->phi; + while (phi != NULL) { + ir_node* next = get_irn_link(phi); + exchange(phi, new_Bad()); + phi = next; + } + return 1; +} + + +static void if_conv_walker(ir_node* block, void* env) +{ + ir_node* phi; + int arity; + int i; + + // Bail out, if there are no Phis at all + if (((block_info*)get_irn_link(block))->phi == NULL) return; + +restart: + arity = get_irn_arity(block); + for (i = 0; i < arity; ++i) { + if (fission_block(block, i)) goto restart; + } + //return; + + arity = get_irn_arity(block); + for (i = 0; i < arity; ++i) { + ir_node* pred; + ir_node* cond; + ir_node* projx0; + int j; + + projx0 = walk_to_projx(get_irn_n(block, i)); + if (projx0 == NULL) return; + pred = get_Proj_pred(projx0); + if (get_irn_op(pred) != op_Cond || get_irn_mode(get_Cond_selector(pred)) != mode_b) continue; + cond = pred; + + if (!can_empty_block(get_nodes_block(get_irn_n(block, i)))) { + ir_fprintf(stderr, "Cannot empty block %+F\n", + get_nodes_block(get_irn_n(block, i)) + ); + continue; + } + + for (j = i + 1; j < arity; ++j) { + ir_node* projx1; + ir_node* psi_block; + ir_node* conds[1]; + ir_node* vals[2]; + ir_node* psi; + + projx1 = walk_to_projx(get_irn_n(block, j)); + if (projx1 == NULL) continue; + pred = get_Proj_pred(projx1); + if (get_irn_op(pred) != op_Cond || get_irn_mode(get_Cond_selector(pred)) != mode_b) continue; + if (pred != cond) continue; + ir_fprintf(stderr, "Found Cond %+F with proj %+F and %+F\n", cond, projx0, projx1); + + if (!can_empty_block(get_nodes_block(get_irn_n(block, j)))) { + ir_fprintf(stderr, "Cannot empty block %+F\n", + get_nodes_block(get_irn_n(block, j)) + ); + continue; + } + + conds[0] = get_Cond_selector(cond); + + psi_block = get_nodes_block(cond); + phi = ((block_info*)get_irn_link(block))->phi; + do { + // Don't generate PsiMs + if (get_irn_mode(phi) == mode_M) { + /* Something is very fishy if to predecessors of a PhiM point into the + * block but not at the same memory node + */ + assert(get_irn_n(phi, i) == get_irn_n(phi, j)); + // fake memory Psi + psi = get_irn_n(phi, i); + ir_fprintf(stderr, "Handling memory Phi %+F\n", phi); + } else { + if (get_Proj_proj(projx0) == pn_Cond_true) { + vals[0] = get_irn_n(phi, i); + vals[1] = get_irn_n(phi, j); + } else { + vals[0] = get_irn_n(phi, j); + vals[1] = get_irn_n(phi, i); + } + psi = new_r_Psi( + current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi) + ); + ir_fprintf(stderr, "Generating %+F for %+F\n", psi, phi); + } + + if (arity == 2) { + exchange(phi, psi); + } else { + ir_node** ins = malloc(sizeof(*ins) * (arity - 1)); + int k; + int l; + + l = 0; + for (k = 0; k < i; ++k) ins[l++] = get_irn_n(phi, k); + for (++k; k < j; ++k) ins[l++] = get_irn_n(phi, k); + for (++k; k < arity; ++k) ins[l++] = get_irn_n(phi, k); + ins[l++] = psi; + assert(l == arity - 1); + set_irn_in(phi, l, ins); + + free(ins); + } + + phi = get_irn_link(phi); + } while (phi != NULL && get_irn_op(phi) == op_Phi); + + exchange(get_nodes_block(get_irn_n(block, i)), psi_block); + exchange(get_nodes_block(get_irn_n(block, j)), psi_block); + + if (arity == 2) { + ir_fprintf(stderr, "Welding block %+F to %+F\n", block, psi_block); + ((block_info*)get_irn_link(psi_block))->evil |= + ((block_info*)get_irn_link(block))->evil; + exchange(block, psi_block); + return; + } else { + ir_node** ins = malloc(sizeof(*ins) * (arity - 1)); + int k; + int l; + + l = 0; + for (k = 0; k < i; ++k) ins[l++] = get_irn_n(block, k); + for (++k; k < j; ++k) ins[l++] = get_irn_n(block, k); + for (++k; k < arity; ++k) ins[l++] = get_irn_n(block, k); + ins[l++] = new_r_Jmp(current_ir_graph, psi_block); + assert(l == arity - 1); + set_irn_in(block, l, ins); + + free(ins); + goto restart; + } + } + } +} + + +static void init_block_link(ir_node* block, void* env) +{ + block_info* bi = malloc(sizeof(*bi)); + + bi->phi = NULL; + bi->evil = 0; + set_irn_link(block, bi); +} + + +/* Daisy-chain all phis in a block + * If a non-movable node is encountered set the evil flag + */ +static void collect_phis(ir_node* node, void* env) +{ + ir_node* block; + block_info* bi; + + if (get_irn_op(node) == op_Block) return; + + block = get_nodes_block(node); + bi = get_irn_link(block); + + if (get_irn_op(node) == op_Phi) { + set_irn_link(node, bi->phi); + bi->phi = node; + } else { +#if 1 + if (get_irn_op(node) == op_Call || + get_irn_op(node) == op_Store || + get_irn_op(node) == op_Load) { + ir_fprintf(stderr, "Node %+F in block %+F is unmovable\n", node, block); + bi->evil = 1; + } +#else + if (get_irn_op(node) != op_Jmp && + get_irn_op(node) != op_Proj && + get_irn_op(node) != op_Cond && + get_irn_op(node) != op_Cmp && + !mode_is_datab(get_irn_mode(node))) { + ir_fprintf(stderr, "Node %+F in block %+F is unmovable\n", node, block); + bi->evil = 1; + } +#endif + } +} + + +/* + * Transform multiple cascaded Psis into one Psi + */ +static ir_node* fold_psi(ir_node* psi) +{ + int arity = get_Psi_n_conds(psi); + int new_arity = 0; + int i; + ir_node* n; + ir_node** conds; + ir_node** vals; + int j; + int k; + int a; + ir_node* new_psi; + + for (i = 0; i < arity; ++i) { + n = get_Psi_val(psi, i); + if (get_irn_op(n) == op_Psi) { + new_arity += get_Psi_n_conds(n) + 1; + } else { + ++new_arity; + } + } + n = get_Psi_default(psi); + if (get_irn_op(n) == op_Psi) { + new_arity += get_Psi_n_conds(n); + } + + if (arity == new_arity) return psi; // no attached Psis found + ir_fprintf(stderr, "Folding %+F from %d to %d conds\n", psi, arity, new_arity); + + conds = malloc(new_arity * sizeof(*conds)); + vals = malloc((new_arity + 1) * sizeof(*vals)); + j = 0; + for (i = 0; i < arity; ++i) { + ir_node* c = get_Psi_cond(psi, i); + + n = get_Psi_val(psi, i); + if (get_irn_op(n) == op_Psi) { + a = get_Psi_n_conds(n); + for (k = 0; k < a; ++k) { + conds[j] = new_r_And( + current_ir_graph, get_nodes_block(psi), + c, get_Psi_cond(n, k), mode_b + ); + vals[j] = get_Psi_val(n, k); + ++j; + } + conds[j] = c; + vals[j] = get_Psi_default(n); + } else { + conds[j] = c; + vals[j] = n; + } + ++j; + } + n = get_Psi_default(psi); + if (get_irn_op(n) == op_Psi) { + a = get_Psi_n_conds(n); + for (k = 0; k < a; ++k) { + conds[j] = get_Psi_cond(n, k); + vals[j] = get_Psi_val(n, k); + ++j; + } + vals[j] = get_Psi_default(n); + } else { + vals[j] = n; + } + assert(j == new_arity); + new_psi = new_r_Psi( + current_ir_graph, get_nodes_block(psi), + new_arity, conds, vals, get_irn_mode(psi) + ); + ir_fprintf(stderr, "Folded %+F into new %+F\n", psi, new_psi); + exchange(psi, new_psi); + free(vals); + free(conds); + return new_psi; +} + + +/* + * Merge consecutive psi inputs if the data inputs are the same + */ +static void meld_psi(ir_node* psi) +{ + int arity = get_Psi_n_conds(psi); + int new_arity; + ir_node** conds; + ir_node** vals; + ir_node* cond; + ir_node* val; + int i; + int j; + ir_node* new_psi; + + new_arity = 1; + val = get_Psi_val(psi, 0); + ir_fprintf(stderr, "Pred 0 of %+F is %+F\n", psi, val); + for (i = 1; i < arity; ++i) { + ir_node* v = get_Psi_val(psi, i); + ir_fprintf(stderr, "Pred %2d of %+F is %+F\n", i, psi, v); + if (val != v) { + val = v; + ++new_arity; + } + } + ir_fprintf(stderr, "Default of %+F is %+F\n", psi, get_Psi_default(psi)); + if (val == get_Psi_default(psi)) --new_arity; + + ir_fprintf(stderr, "Melding Psi %+F from %d conds to %d\n", + psi, arity, new_arity + ); + + if (new_arity == arity) return; + + // If all data inputs of the Psi are equal, exchange the Psi with that value + if (new_arity == 0) { + exchange(psi, val); + return; + } + + conds = malloc(sizeof(*conds) * new_arity); + vals = malloc(sizeof(*vals) * (new_arity + 1)); + cond = get_Psi_cond(psi, 0); + val = get_Psi_val(psi, 0); + j = 0; + for (i = 1; i < arity; ++i) { + ir_node* v = get_Psi_val(psi, i); + + if (v == val) { + cond = new_r_Or( + current_ir_graph, get_nodes_block(psi), + cond, get_Psi_cond(psi, i), mode_b + ); + } else { + conds[j] = cond; + vals[j] = val; + ++j; + val = v; + } + } + if (val != get_Psi_default(psi)) { + conds[j] = cond; + vals[j] = val; + ++j; + } + vals[j] = get_Psi_default(psi); + assert(j == new_arity); + new_psi = new_r_Psi( + current_ir_graph, get_nodes_block(psi), + new_arity, conds, vals, get_irn_mode(psi) + ); + ir_fprintf(stderr, "Molded %+F into %+F\n", psi, new_psi); + exchange(psi, new_psi); +} + + +static void optimise_psis(ir_node* node, void* env) +{ + if (get_irn_op(node) != op_Psi) return; +#if 1 + node = fold_psi(node); +#endif +#if 1 + meld_psi(node); +#endif +} + + +void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params) +{ + ir_fprintf(stderr, "Running if-conversion on %+F\n", irg); + + dump_ir_block_graph(irg, "_00_pre"); + + normalize_one_return(irg); + remove_critical_cf_edges(irg); + + dump_ir_block_graph(irg, "_01_normal"); + + compute_cdep(irg); + compute_doms(irg); + + irg_block_walk_graph(irg, init_block_link, NULL, NULL); + irg_walk_graph(irg, collect_phis, NULL, NULL); + irg_block_walk_graph(irg, NULL, if_conv_walker, NULL); + + dump_ir_block_graph(irg, "_02_ifconv"); + + local_optimize_graph(irg); + irg_walk_graph(irg, NULL, optimise_psis, NULL); + + dump_ir_block_graph(irg, "_03_postifconv"); + + free_dom(irg); + free_cdep(irg); +} + +#else + /** * @file ifconv.c * If conversion. @@ -995,4 +1527,8 @@ void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params) DBG((dbg, LEVEL_1, "muxes made: %d\n", muxes_made)); obstack_free(&obst, NULL); + + dump_ir_block_graph(irg, "_ifconv_hack"); } + +#endif -- 2.20.1