/* Predictive commoning. Copyright (C) 2005-2014 Free Software Foundation, Inc. This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see . */ /* This file implements the predictive commoning optimization. Predictive commoning can be viewed as CSE around a loop, and with some improvements, as generalized strength reduction-- i.e., reusing values computed in earlier iterations of a loop in the later ones. So far, the pass only handles the most useful case, that is, reusing values of memory references. If you think this is all just a special case of PRE, you are sort of right; however, concentrating on loops is simpler, and makes it possible to incorporate data dependence analysis to detect the opportunities, perform loop unrolling to avoid copies together with renaming immediately, and if needed, we could also take register pressure into account. Let us demonstrate what is done on an example: for (i = 0; i < 100; i++) { a[i+2] = a[i] + a[i+1]; b[10] = b[10] + i; c[i] = c[99 - i]; d[i] = d[i + 1]; } 1) We find data references in the loop, and split them to mutually independent groups (i.e., we find components of a data dependence graph). We ignore read-read dependences whose distance is not constant. (TODO -- we could also ignore antidependences). In this example, we find the following groups: a[i]{read}, a[i+1]{read}, a[i+2]{write} b[10]{read}, b[10]{write} c[99 - i]{read}, c[i]{write} d[i + 1]{read}, d[i]{write} 2) Inside each of the group, we verify several conditions: a) all the references must differ in indices only, and the indices must all have the same step b) the references must dominate loop latch (and thus, they must be ordered by dominance relation). c) the distance of the indices must be a small multiple of the step We are then able to compute the difference of the references (# of iterations before they point to the same place as the first of them). Also, in case there are writes in the loop, we split the groups into chains whose head is the write whose values are used by the reads in the same chain. The chains are then processed independently, making the further transformations simpler. Also, the shorter chains need the same number of registers, but may require lower unrolling factor in order to get rid of the copies on the loop latch. In our example, we get the following chains (the chain for c is invalid). a[i]{read,+0}, a[i+1]{read,-1}, a[i+2]{write,-2} b[10]{read,+0}, b[10]{write,+0} d[i + 1]{read,+0}, d[i]{write,+1} 3) For each read, we determine the read or write whose value it reuses, together with the distance of this reuse. I.e. we take the last reference before it with distance 0, or the last of the references with the smallest positive distance to the read. Then, we remove the references that are not used in any of these chains, discard the empty groups, and propagate all the links so that they point to the single root reference of the chain (adjusting their distance appropriately). Some extra care needs to be taken for references with step 0. In our example (the numbers indicate the distance of the reuse), a[i] --> (*) 2, a[i+1] --> (*) 1, a[i+2] (*) b[10] --> (*) 1, b[10] (*) 4) The chains are combined together if possible. If the corresponding elements of two chains are always combined together with the same operator, we remember just the result of this combination, instead of remembering the values separately. We may need to perform reassociation to enable combining, for example e[i] + f[i+1] + e[i+1] + f[i] can be reassociated as (e[i] + f[i]) + (e[i+1] + f[i+1]) and we can combine the chains for e and f into one chain. 5) For each root reference (end of the chain) R, let N be maximum distance of a reference reusing its value. Variables R0 up to RN are created, together with phi nodes that transfer values from R1 .. RN to R0 .. R(N-1). Initial values are loaded to R0..R(N-1) (in case not all references must necessarily be accessed and they may trap, we may fail here; TODO sometimes, the loads could be guarded by a check for the number of iterations). Values loaded/stored in roots are also copied to RN. Other reads are replaced with the appropriate variable Ri. Everything is put to SSA form. As a small improvement, if R0 is dead after the root (i.e., all uses of the value with the maximum distance dominate the root), we can avoid creating RN and use R0 instead of it. In our example, we get (only the parts concerning a and b are shown): for (i = 0; i < 100; i++) { f = phi (a[0], s); s = phi (a[1], f); x = phi (b[10], x); f = f + s; a[i+2] = f; x = x + i; b[10] = x; } 6) Factor F for unrolling is determined as the smallest common multiple of (N + 1) for each root reference (N for references for that we avoided creating RN). If F and the loop is small enough, loop is unrolled F times. The stores to RN (R0) in the copies of the loop body are periodically replaced with R0, R1, ... (R1, R2, ...), so that they can be coalesced and the copies can be eliminated. TODO -- copy propagation and other optimizations may change the live ranges of the temporary registers and prevent them from being coalesced; this may increase the register pressure. In our case, F = 2 and the (main loop of the) result is for (i = 0; i < ...; i += 2) { f = phi (a[0], f); s = phi (a[1], s); x = phi (b[10], x); f = f + s; a[i+2] = f; x = x + i; b[10] = x; s = s + f; a[i+3] = s; x = x + i; b[10] = x; } TODO -- stores killing other stores can be taken into account, e.g., for (i = 0; i < n; i++) { a[i] = 1; a[i+2] = 2; } can be replaced with t0 = a[0]; t1 = a[1]; for (i = 0; i < n; i++) { a[i] = 1; t2 = 2; t0 = t1; t1 = t2; } a[n] = t0; a[n+1] = t1; The interesting part is that this would generalize store motion; still, since sm is performed elsewhere, it does not seem that important. Predictive commoning can be generalized for arbitrary computations (not just memory loads), and also nontrivial transfer functions (e.g., replacing i * i with ii_last + 2 * i + 1), to generalize strength reduction. */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" #include "tree.h" #include "tm_p.h" #include "cfgloop.h" #include "basic-block.h" #include "tree-ssa-alias.h" #include "internal-fn.h" #include "tree-eh.h" #include "gimple-expr.h" #include "is-a.h" #include "gimple.h" #include "gimplify.h" #include "gimple-iterator.h" #include "gimplify-me.h" #include "gimple-ssa.h" #include "tree-phinodes.h" #include "ssa-iterators.h" #include "stringpool.h" #include "tree-ssanames.h" #include "tree-ssa-loop-ivopts.h" #include "tree-ssa-loop-manip.h" #include "tree-ssa-loop-niter.h" #include "tree-ssa-loop.h" #include "tree-into-ssa.h" #include "expr.h" #include "tree-dfa.h" #include "tree-ssa.h" #include "tree-data-ref.h" #include "tree-scalar-evolution.h" #include "tree-chrec.h" #include "params.h" #include "gimple-pretty-print.h" #include "tree-pass.h" #include "tree-affine.h" #include "tree-inline.h" /* The maximum number of iterations between the considered memory references. */ #define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8) /* Data references (or phi nodes that carry data reference values across loop iterations). */ typedef struct dref_d { /* The reference itself. */ struct data_reference *ref; /* The statement in that the reference appears. */ gimple stmt; /* In case that STMT is a phi node, this field is set to the SSA name defined by it in replace_phis_by_defined_names (in order to avoid pointing to phi node that got reallocated in the meantime). */ tree name_defined_by_phi; /* Distance of the reference from the root of the chain (in number of iterations of the loop). */ unsigned distance; /* Number of iterations offset from the first reference in the component. */ double_int offset; /* Number of the reference in a component, in dominance ordering. */ unsigned pos; /* True if the memory reference is always accessed when the loop is entered. */ unsigned always_accessed : 1; } *dref; /* Type of the chain of the references. */ enum chain_type { /* The addresses of the references in the chain are constant. */ CT_INVARIANT, /* There are only loads in the chain. */ CT_LOAD, /* Root of the chain is store, the rest are loads. */ CT_STORE_LOAD, /* A combination of two chains. */ CT_COMBINATION }; /* Chains of data references. */ typedef struct chain { /* Type of the chain. */ enum chain_type type; /* For combination chains, the operator and the two chains that are combined, and the type of the result. */ enum tree_code op; tree rslt_type; struct chain *ch1, *ch2; /* The references in the chain. */ vec refs; /* The maximum distance of the reference in the chain from the root. */ unsigned length; /* The variables used to copy the value throughout iterations. */ vec vars; /* Initializers for the variables. */ vec inits; /* True if there is a use of a variable with the maximal distance that comes after the root in the loop. */ unsigned has_max_use_after : 1; /* True if all the memory references in the chain are always accessed. */ unsigned all_always_accessed : 1; /* True if this chain was combined together with some other chain. */ unsigned combined : 1; } *chain_p; /* Describes the knowledge about the step of the memory references in the component. */ enum ref_step_type { /* The step is zero. */ RS_INVARIANT, /* The step is nonzero. */ RS_NONZERO, /* The step may or may not be nonzero. */ RS_ANY }; /* Components of the data dependence graph. */ struct component { /* The references in the component. */ vec refs; /* What we know about the step of the references in the component. */ enum ref_step_type comp_step; /* Next component in the list. */ struct component *next; }; /* Bitmap of ssa names defined by looparound phi nodes covered by chains. */ static bitmap looparound_phis; /* Cache used by tree_to_aff_combination_expand. */ static struct pointer_map_t *name_expansions; /* Dumps data reference REF to FILE. */ extern void dump_dref (FILE *, dref); void dump_dref (FILE *file, dref ref) { if (ref->ref) { fprintf (file, " "); print_generic_expr (file, DR_REF (ref->ref), TDF_SLIM); fprintf (file, " (id %u%s)\n", ref->pos, DR_IS_READ (ref->ref) ? "" : ", write"); fprintf (file, " offset "); dump_double_int (file, ref->offset, false); fprintf (file, "\n"); fprintf (file, " distance %u\n", ref->distance); } else { if (gimple_code (ref->stmt) == GIMPLE_PHI) fprintf (file, " looparound ref\n"); else fprintf (file, " combination ref\n"); fprintf (file, " in statement "); print_gimple_stmt (file, ref->stmt, 0, TDF_SLIM); fprintf (file, "\n"); fprintf (file, " distance %u\n", ref->distance); } } /* Dumps CHAIN to FILE. */ extern void dump_chain (FILE *, chain_p); void dump_chain (FILE *file, chain_p chain) { dref a; const char *chain_type; unsigned i; tree var; switch (chain->type) { case CT_INVARIANT: chain_type = "Load motion"; break; case CT_LOAD: chain_type = "Loads-only"; break; case CT_STORE_LOAD: chain_type = "Store-loads"; break; case CT_COMBINATION: chain_type = "Combination"; break; default: gcc_unreachable (); } fprintf (file, "%s chain %p%s\n", chain_type, (void *) chain, chain->combined ? " (combined)" : ""); if (chain->type != CT_INVARIANT) fprintf (file, " max distance %u%s\n", chain->length, chain->has_max_use_after ? "" : ", may reuse first"); if (chain->type == CT_COMBINATION) { fprintf (file, " equal to %p %s %p in type ", (void *) chain->ch1, op_symbol_code (chain->op), (void *) chain->ch2); print_generic_expr (file, chain->rslt_type, TDF_SLIM); fprintf (file, "\n"); } if (chain->vars.exists ()) { fprintf (file, " vars"); FOR_EACH_VEC_ELT (chain->vars, i, var) { fprintf (file, " "); print_generic_expr (file, var, TDF_SLIM); } fprintf (file, "\n"); } if (chain->inits.exists ()) { fprintf (file, " inits"); FOR_EACH_VEC_ELT (chain->inits, i, var) { fprintf (file, " "); print_generic_expr (file, var, TDF_SLIM); } fprintf (file, "\n"); } fprintf (file, " references:\n"); FOR_EACH_VEC_ELT (chain->refs, i, a) dump_dref (file, a); fprintf (file, "\n"); } /* Dumps CHAINS to FILE. */ extern void dump_chains (FILE *, vec ); void dump_chains (FILE *file, vec chains) { chain_p chain; unsigned i; FOR_EACH_VEC_ELT (chains, i, chain) dump_chain (file, chain); } /* Dumps COMP to FILE. */ extern void dump_component (FILE *, struct component *); void dump_component (FILE *file, struct component *comp) { dref a; unsigned i; fprintf (file, "Component%s:\n", comp->comp_step == RS_INVARIANT ? " (invariant)" : ""); FOR_EACH_VEC_ELT (comp->refs, i, a) dump_dref (file, a); fprintf (file, "\n"); } /* Dumps COMPS to FILE. */ extern void dump_components (FILE *, struct component *); void dump_components (FILE *file, struct component *comps) { struct component *comp; for (comp = comps; comp; comp = comp->next) dump_component (file, comp); } /* Frees a chain CHAIN. */ static void release_chain (chain_p chain) { dref ref; unsigned i; if (chain == NULL) return; FOR_EACH_VEC_ELT (chain->refs, i, ref) free (ref); chain->refs.release (); chain->vars.release (); chain->inits.release (); free (chain); } /* Frees CHAINS. */ static void release_chains (vec chains) { unsigned i; chain_p chain; FOR_EACH_VEC_ELT (chains, i, chain) release_chain (chain); chains.release (); } /* Frees a component COMP. */ static void release_component (struct component *comp) { comp->refs.release (); free (comp); } /* Frees list of components COMPS. */ static void release_components (struct component *comps) { struct component *act, *next; for (act = comps; act; act = next) { next = act->next; release_component (act); } } /* Finds a root of tree given by FATHERS containing A, and performs path shortening. */ static unsigned component_of (unsigned fathers[], unsigned a) { unsigned root, n; for (root = a; root != fathers[root]; root = fathers[root]) continue; for (; a != root; a = n) { n = fathers[a]; fathers[a] = root; } return root; } /* Join operation for DFU. FATHERS gives the tree, SIZES are sizes of the components, A and B are components to merge. */ static void merge_comps (unsigned fathers[], unsigned sizes[], unsigned a, unsigned b) { unsigned ca = component_of (fathers, a); unsigned cb = component_of (fathers, b); if (ca == cb) return; if (sizes[ca] < sizes[cb]) { sizes[cb] += sizes[ca]; fathers[ca] = cb; } else { sizes[ca] += sizes[cb]; fathers[cb] = ca; } } /* Returns true if A is a reference that is suitable for predictive commoning in the innermost loop that contains it. REF_STEP is set according to the step of the reference A. */ static bool suitable_reference_p (struct data_reference *a, enum ref_step_type *ref_step) { tree ref = DR_REF (a), step = DR_STEP (a); if (!step || TREE_THIS_VOLATILE (ref) || !is_gimple_reg_type (TREE_TYPE (ref)) || tree_could_throw_p (ref)) return false; if (integer_zerop (step)) *ref_step = RS_INVARIANT; else if (integer_nonzerop (step)) *ref_step = RS_NONZERO; else *ref_step = RS_ANY; return true; } /* Stores DR_OFFSET (DR) + DR_INIT (DR) to OFFSET. */ static void aff_combination_dr_offset (struct data_reference *dr, aff_tree *offset) { tree type = TREE_TYPE (DR_OFFSET (dr)); aff_tree delta; tree_to_aff_combination_expand (DR_OFFSET (dr), type, offset, &name_expansions); aff_combination_const (&delta, type, tree_to_double_int (DR_INIT (dr))); aff_combination_add (offset, &delta); } /* Determines number of iterations of the innermost enclosing loop before B refers to exactly the same location as A and stores it to OFF. If A and B do not have the same step, they never meet, or anything else fails, returns false, otherwise returns true. Both A and B are assumed to satisfy suitable_reference_p. */ static bool determine_offset (struct data_reference *a, struct data_reference *b, double_int *off) { aff_tree diff, baseb, step; tree typea, typeb; /* Check that both the references access the location in the same type. */ typea = TREE_TYPE (DR_REF (a)); typeb = TREE_TYPE (DR_REF (b)); if (!useless_type_conversion_p (typeb, typea)) return false; /* Check whether the base address and the step of both references is the same. */ if (!operand_equal_p (DR_STEP (a), DR_STEP (b), 0) || !operand_equal_p (DR_BASE_ADDRESS (a), DR_BASE_ADDRESS (b), 0)) return false; if (integer_zerop (DR_STEP (a))) { /* If the references have loop invariant address, check that they access exactly the same location. */ *off = double_int_zero; return (operand_equal_p (DR_OFFSET (a), DR_OFFSET (b), 0) && operand_equal_p (DR_INIT (a), DR_INIT (b), 0)); } /* Compare the offsets of the addresses, and check whether the difference is a multiple of step. */ aff_combination_dr_offset (a, &diff); aff_combination_dr_offset (b, &baseb); aff_combination_scale (&baseb, double_int_minus_one); aff_combination_add (&diff, &baseb); tree_to_aff_combination_expand (DR_STEP (a), TREE_TYPE (DR_STEP (a)), &step, &name_expansions); return aff_combination_constant_multiple_p (&diff, &step, off); } /* Returns the last basic block in LOOP for that we are sure that it is executed whenever the loop is entered. */ static basic_block last_always_executed_block (struct loop *loop) { unsigned i; vec exits = get_loop_exit_edges (loop); edge ex; basic_block last = loop->latch; FOR_EACH_VEC_ELT (exits, i, ex) last = nearest_common_dominator (CDI_DOMINATORS, last, ex->src); exits.release (); return last; } /* Splits dependence graph on DATAREFS described by DEPENDS to components. */ static struct component * split_data_refs_to_components (struct loop *loop, vec datarefs, vec depends) { unsigned i, n = datarefs.length (); unsigned ca, ia, ib, bad; unsigned *comp_father = XNEWVEC (unsigned, n + 1); unsigned *comp_size = XNEWVEC (unsigned, n + 1); struct component **comps; struct data_reference *dr, *dra, *drb; struct data_dependence_relation *ddr; struct component *comp_list = NULL, *comp; dref dataref; basic_block last_always_executed = last_always_executed_block (loop); FOR_EACH_VEC_ELT (datarefs, i, dr) { if (!DR_REF (dr)) { /* A fake reference for call or asm_expr that may clobber memory; just fail. */ goto end; } /* predcom pass isn't prepared to handle calls with data references. */ if (is_gimple_call (DR_STMT (dr))) goto end; dr->aux = (void *) (size_t) i; comp_father[i] = i; comp_size[i] = 1; } /* A component reserved for the "bad" data references. */ comp_father[n] = n; comp_size[n] = 1; FOR_EACH_VEC_ELT (datarefs, i, dr) { enum ref_step_type dummy; if (!suitable_reference_p (dr, &dummy)) { ia = (unsigned) (size_t) dr->aux; merge_comps (comp_father, comp_size, n, ia); } } FOR_EACH_VEC_ELT (depends, i, ddr) { double_int dummy_off; if (DDR_ARE_DEPENDENT (ddr) == chrec_known) continue; dra = DDR_A (ddr); drb = DDR_B (ddr); ia = component_of (comp_father, (unsigned) (size_t) dra->aux); ib = component_of (comp_father, (unsigned) (size_t) drb->aux); if (ia == ib) continue; bad = component_of (comp_father, n); /* If both A and B are reads, we may ignore unsuitable dependences. */ if (DR_IS_READ (dra) && DR_IS_READ (drb)) { if (ia == bad || ib == bad || !determine_offset (dra, drb, &dummy_off)) continue; } /* If A is read and B write or vice versa and there is unsuitable dependence, instead of merging both components into a component that will certainly not pass suitable_component_p, just put the read into bad component, perhaps at least the write together with all the other data refs in it's component will be optimizable. */ else if (DR_IS_READ (dra) && ib != bad) { if (ia == bad) continue; else if (!determine_offset (dra, drb, &dummy_off)) { merge_comps (comp_father, comp_size, bad, ia); continue; } } else if (DR_IS_READ (drb) && ia != bad) { if (ib == bad) continue; else if (!determine_offset (dra, drb, &dummy_off)) { merge_comps (comp_father, comp_size, bad, ib); continue; } } merge_comps (comp_father, comp_size, ia, ib); } comps = XCNEWVEC (struct component *, n); bad = component_of (comp_father, n); FOR_EACH_VEC_ELT (datarefs, i, dr) { ia = (unsigned) (size_t) dr->aux; ca = component_of (comp_father, ia); if (ca == bad) continue; comp = comps[ca]; if (!comp) { comp = XCNEW (struct component); comp->refs.create (comp_size[ca]); comps[ca] = comp; } dataref = XCNEW (struct dref_d); dataref->ref = dr; dataref->stmt = DR_STMT (dr); dataref->offset = double_int_zero; dataref->distance = 0; dataref->always_accessed = dominated_by_p (CDI_DOMINATORS, last_always_executed, gimple_bb (dataref->stmt)); dataref->pos = comp->refs.length (); comp->refs.quick_push (dataref); } for (i = 0; i < n; i++) { comp = comps[i]; if (comp) { comp->next = comp_list; comp_list = comp; } } free (comps); end: free (comp_father); free (comp_size); return comp_list; } /* Returns true if the component COMP satisfies the conditions described in 2) at the beginning of this file. LOOP is the current loop. */ static bool suitable_component_p (struct loop *loop, struct component *comp) { unsigned i; dref a, first; basic_block ba, bp = loop->header; bool ok, has_write = false; FOR_EACH_VEC_ELT (comp->refs, i, a) { ba = gimple_bb (a->stmt); if (!just_once_each_iteration_p (loop, ba)) return false; gcc_assert (dominated_by_p (CDI_DOMINATORS, ba, bp)); bp = ba; if (DR_IS_WRITE (a->ref)) has_write = true; } first = comp->refs[0]; ok = suitable_reference_p (first->ref, &comp->comp_step); gcc_assert (ok); first->offset = double_int_zero; for (i = 1; comp->refs.iterate (i, &a); i++) { if (!determine_offset (first->ref, a->ref, &a->offset)) return false; #ifdef ENABLE_CHECKING { enum ref_step_type a_step; ok = suitable_reference_p (a->ref, &a_step); gcc_assert (ok && a_step == comp->comp_step); } #endif } /* If there is a write inside the component, we must know whether the step is nonzero or not -- we would not otherwise be able to recognize whether the value accessed by reads comes from the OFFSET-th iteration or the previous one. */ if (has_write && comp->comp_step == RS_ANY) return false; return true; } /* Check the conditions on references inside each of components COMPS, and remove the unsuitable components from the list. The new list of components is returned. The conditions are described in 2) at the beginning of this file. LOOP is the current loop. */ static struct component * filter_suitable_components (struct loop *loop, struct component *comps) { struct component **comp, *act; for (comp = &comps; *comp; ) { act = *comp; if (suitable_component_p (loop, act)) comp = &act->next; else { dref ref; unsigned i; *comp = act->next; FOR_EACH_VEC_ELT (act->refs, i, ref) free (ref); release_component (act); } } return comps; } /* Compares two drefs A and B by their offset and position. Callback for qsort. */ static int order_drefs (const void *a, const void *b) { const dref *const da = (const dref *) a; const dref *const db = (const dref *) b; int offcmp = (*da)->offset.scmp ((*db)->offset); if (offcmp != 0) return offcmp; return (*da)->pos - (*db)->pos; } /* Returns root of the CHAIN. */ static inline dref get_chain_root (chain_p chain) { return chain->refs[0]; } /* Adds REF to the chain CHAIN. */ static void add_ref_to_chain (chain_p chain, dref ref) { dref root = get_chain_root (chain); double_int dist; gcc_assert (root->offset.sle (ref->offset)); dist = ref->offset - root->offset; if (double_int::from_uhwi (MAX_DISTANCE).ule (dist)) { free (ref); return; } gcc_assert (dist.fits_uhwi ()); chain->refs.safe_push (ref); ref->distance = dist.to_uhwi (); if (ref->distance >= chain->length) { chain->length = ref->distance; chain->has_max_use_after = false; } if (ref->distance == chain->length && ref->pos > root->pos) chain->has_max_use_after = true; chain->all_always_accessed &= ref->always_accessed; } /* Returns the chain for invariant component COMP. */ static chain_p make_invariant_chain (struct component *comp) { chain_p chain = XCNEW (struct chain); unsigned i; dref ref; chain->type = CT_INVARIANT; chain->all_always_accessed = true; FOR_EACH_VEC_ELT (comp->refs, i, ref) { chain->refs.safe_push (ref); chain->all_always_accessed &= ref->always_accessed; } return chain; } /* Make a new chain rooted at REF. */ static chain_p make_rooted_chain (dref ref) { chain_p chain = XCNEW (struct chain); chain->type = DR_IS_READ (ref->ref) ? CT_LOAD : CT_STORE_LOAD; chain->refs.safe_push (ref); chain->all_always_accessed = ref->always_accessed; ref->distance = 0; return chain; } /* Returns true if CHAIN is not trivial. */ static bool nontrivial_chain_p (chain_p chain) { return chain != NULL && chain->refs.length () > 1; } /* Returns the ssa name that contains the value of REF, or NULL_TREE if there is no such name. */ static tree name_for_ref (dref ref) { tree name; if (is_gimple_assign (ref->stmt)) { if (!ref->ref || DR_IS_READ (ref->ref)) name = gimple_assign_lhs (ref->stmt); else name = gimple_assign_rhs1 (ref->stmt); } else name = PHI_RESULT (ref->stmt); return (TREE_CODE (name) == SSA_NAME ? name : NULL_TREE); } /* Returns true if REF is a valid initializer for ROOT with given DISTANCE (in iterations of the innermost enclosing loop). */ static bool valid_initializer_p (struct data_reference *ref, unsigned distance, struct data_reference *root) { aff_tree diff, base, step; double_int off; /* Both REF and ROOT must be accessing the same object. */ if (!operand_equal_p (DR_BASE_ADDRESS (ref), DR_BASE_ADDRESS (root), 0)) return false; /* The initializer is defined outside of loop, hence its address must be invariant inside the loop. */ gcc_assert (integer_zerop (DR_STEP (ref))); /* If the address of the reference is invariant, initializer must access exactly the same location. */ if (integer_zerop (DR_STEP (root))) return (operand_equal_p (DR_OFFSET (ref), DR_OFFSET (root), 0) && operand_equal_p (DR_INIT (ref), DR_INIT (root), 0)); /* Verify that this index of REF is equal to the root's index at -DISTANCE-th iteration. */ aff_combination_dr_offset (root, &diff); aff_combination_dr_offset (ref, &base); aff_combination_scale (&base, double_int_minus_one); aff_combination_add (&diff, &base); tree_to_aff_combination_expand (DR_STEP (root), TREE_TYPE (DR_STEP (root)), &step, &name_expansions); if (!aff_combination_constant_multiple_p (&diff, &step, &off)) return false; if (off != double_int::from_uhwi (distance)) return false; return true; } /* Finds looparound phi node of LOOP that copies the value of REF, and if its initial value is correct (equal to initial value of REF shifted by one iteration), returns the phi node. Otherwise, NULL_TREE is returned. ROOT is the root of the current chain. */ static gimple find_looparound_phi (struct loop *loop, dref ref, dref root) { tree name, init, init_ref; gimple phi = NULL, init_stmt; edge latch = loop_latch_edge (loop); struct data_reference init_dr; gimple_stmt_iterator psi; if (is_gimple_assign (ref->stmt)) { if (DR_IS_READ (ref->ref)) name = gimple_assign_lhs (ref->stmt); else name = gimple_assign_rhs1 (ref->stmt); } else name = PHI_RESULT (ref->stmt); if (!name) return NULL; for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) { phi = gsi_stmt (psi); if (PHI_ARG_DEF_FROM_EDGE (phi, latch) == name) break; } if (gsi_end_p (psi)) return NULL; init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); if (TREE_CODE (init) != SSA_NAME) return NULL; init_stmt = SSA_NAME_DEF_STMT (init); if (gimple_code (init_stmt) != GIMPLE_ASSIGN) return NULL; gcc_assert (gimple_assign_lhs (init_stmt) == init); init_ref = gimple_assign_rhs1 (init_stmt); if (!REFERENCE_CLASS_P (init_ref) && !DECL_P (init_ref)) return NULL; /* Analyze the behavior of INIT_REF with respect to LOOP (innermost loop enclosing PHI). */ memset (&init_dr, 0, sizeof (struct data_reference)); DR_REF (&init_dr) = init_ref; DR_STMT (&init_dr) = phi; if (!dr_analyze_innermost (&init_dr, loop)) return NULL; if (!valid_initializer_p (&init_dr, ref->distance + 1, root->ref)) return NULL; return phi; } /* Adds a reference for the looparound copy of REF in PHI to CHAIN. */ static void insert_looparound_copy (chain_p chain, dref ref, gimple phi) { dref nw = XCNEW (struct dref_d), aref; unsigned i; nw->stmt = phi; nw->distance = ref->distance + 1; nw->always_accessed = 1; FOR_EACH_VEC_ELT (chain->refs, i, aref) if (aref->distance >= nw->distance) break; chain->refs.safe_insert (i, nw); if (nw->distance > chain->length) { chain->length = nw->distance; chain->has_max_use_after = false; } } /* For references in CHAIN that are copied around the LOOP (created previously by PRE, or by user), add the results of such copies to the chain. This enables us to remove the copies by unrolling, and may need less registers (also, it may allow us to combine chains together). */ static void add_looparound_copies (struct loop *loop, chain_p chain) { unsigned i; dref ref, root = get_chain_root (chain); gimple phi; FOR_EACH_VEC_ELT (chain->refs, i, ref) { phi = find_looparound_phi (loop, ref, root); if (!phi) continue; bitmap_set_bit (looparound_phis, SSA_NAME_VERSION (PHI_RESULT (phi))); insert_looparound_copy (chain, ref, phi); } } /* Find roots of the values and determine distances in the component COMP. The references are redistributed into CHAINS. LOOP is the current loop. */ static void determine_roots_comp (struct loop *loop, struct component *comp, vec *chains) { unsigned i; dref a; chain_p chain = NULL; double_int last_ofs = double_int_zero; /* Invariants are handled specially. */ if (comp->comp_step == RS_INVARIANT) { chain = make_invariant_chain (comp); chains->safe_push (chain); return; } comp->refs.qsort (order_drefs); FOR_EACH_VEC_ELT (comp->refs, i, a) { if (!chain || DR_IS_WRITE (a->ref) || double_int::from_uhwi (MAX_DISTANCE).ule (a->offset - last_ofs)) { if (nontrivial_chain_p (chain)) { add_looparound_copies (loop, chain); chains->safe_push (chain); } else release_chain (chain); chain = make_rooted_chain (a); last_ofs = a->offset; continue; } add_ref_to_chain (chain, a); } if (nontrivial_chain_p (chain)) { add_looparound_copies (loop, chain); chains->safe_push (chain); } else release_chain (chain); } /* Find roots of the values and determine distances in components COMPS, and separates the references to CHAINS. LOOP is the current loop. */ static void determine_roots (struct loop *loop, struct component *comps, vec *chains) { struct component *comp; for (comp = comps; comp; comp = comp->next) determine_roots_comp (loop, comp, chains); } /* Replace the reference in statement STMT with temporary variable NEW_TREE. If SET is true, NEW_TREE is instead initialized to the value of the reference in the statement. IN_LHS is true if the reference is in the lhs of STMT, false if it is in rhs. */ static void replace_ref_with (gimple stmt, tree new_tree, bool set, bool in_lhs) { tree val; gimple new_stmt; gimple_stmt_iterator bsi, psi; if (gimple_code (stmt) == GIMPLE_PHI) { gcc_assert (!in_lhs && !set); val = PHI_RESULT (stmt); bsi = gsi_after_labels (gimple_bb (stmt)); psi = gsi_for_stmt (stmt); remove_phi_node (&psi, false); /* Turn the phi node into GIMPLE_ASSIGN. */ new_stmt = gimple_build_assign (val, new_tree); gsi_insert_before (&bsi, new_stmt, GSI_NEW_STMT); return; } /* Since the reference is of gimple_reg type, it should only appear as lhs or rhs of modify statement. */ gcc_assert (is_gimple_assign (stmt)); bsi = gsi_for_stmt (stmt); /* If we do not need to initialize NEW_TREE, just replace the use of OLD. */ if (!set) { gcc_assert (!in_lhs); gimple_assign_set_rhs_from_tree (&bsi, new_tree); stmt = gsi_stmt (bsi); update_stmt (stmt); return; } if (in_lhs) { /* We have statement OLD = VAL If OLD is a memory reference, then VAL is gimple_val, and we transform this to OLD = VAL NEW = VAL Otherwise, we are replacing a combination chain, VAL is the expression that performs the combination, and OLD is an SSA name. In this case, we transform the assignment to OLD = VAL NEW = OLD */ val = gimple_assign_lhs (stmt); if (TREE_CODE (val) != SSA_NAME) { val = gimple_assign_rhs1 (stmt); gcc_assert (gimple_assign_single_p (stmt)); if (TREE_CLOBBER_P (val)) val = get_or_create_ssa_default_def (cfun, SSA_NAME_VAR (new_tree)); else gcc_assert (gimple_assign_copy_p (stmt)); } } else { /* VAL = OLD is transformed to VAL = OLD NEW = VAL */ val = gimple_assign_lhs (stmt); } new_stmt = gimple_build_assign (new_tree, unshare_expr (val)); gsi_insert_after (&bsi, new_stmt, GSI_NEW_STMT); } /* Returns a memory reference to DR in the ITER-th iteration of the loop it was analyzed in. Append init stmts to STMTS. */ static tree ref_at_iteration (data_reference_p dr, int iter, gimple_seq *stmts) { tree off = DR_OFFSET (dr); tree coff = DR_INIT (dr); if (iter == 0) ; else if (TREE_CODE (DR_STEP (dr)) == INTEGER_CST) coff = size_binop (PLUS_EXPR, coff, size_binop (MULT_EXPR, DR_STEP (dr), ssize_int (iter))); else off = size_binop (PLUS_EXPR, off, size_binop (MULT_EXPR, DR_STEP (dr), ssize_int (iter))); tree addr = fold_build_pointer_plus (DR_BASE_ADDRESS (dr), off); addr = force_gimple_operand_1 (unshare_expr (addr), stmts, is_gimple_mem_ref_addr, NULL_TREE); tree alias_ptr = fold_convert (reference_alias_ptr_type (DR_REF (dr)), coff); /* While data-ref analysis punts on bit offsets it still handles bitfield accesses at byte boundaries. Cope with that. Note that we cannot simply re-apply the outer COMPONENT_REF because the byte-granular portion of it is already applied via DR_INIT and DR_OFFSET, so simply build a BIT_FIELD_REF knowing that the bits start at offset zero. */ if (TREE_CODE (DR_REF (dr)) == COMPONENT_REF && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (dr), 1))) { tree field = TREE_OPERAND (DR_REF (dr), 1); return build3 (BIT_FIELD_REF, TREE_TYPE (DR_REF (dr)), build2 (MEM_REF, DECL_BIT_FIELD_TYPE (field), addr, alias_ptr), DECL_SIZE (field), bitsize_zero_node); } else return fold_build2 (MEM_REF, TREE_TYPE (DR_REF (dr)), addr, alias_ptr); } /* Get the initialization expression for the INDEX-th temporary variable of CHAIN. */ static tree get_init_expr (chain_p chain, unsigned index) { if (chain->type == CT_COMBINATION) { tree e1 = get_init_expr (chain->ch1, index); tree e2 = get_init_expr (chain->ch2, index); return fold_build2 (chain->op, chain->rslt_type, e1, e2); } else return chain->inits[index]; } /* Returns a new temporary variable used for the I-th variable carrying value of REF. The variable's uid is marked in TMP_VARS. */ static tree predcom_tmp_var (tree ref, unsigned i, bitmap tmp_vars) { tree type = TREE_TYPE (ref); /* We never access the components of the temporary variable in predictive commoning. */ tree var = create_tmp_reg (type, get_lsm_tmp_name (ref, i)); bitmap_set_bit (tmp_vars, DECL_UID (var)); return var; } /* Creates the variables for CHAIN, as well as phi nodes for them and initialization on entry to LOOP. Uids of the newly created temporary variables are marked in TMP_VARS. */ static void initialize_root_vars (struct loop *loop, chain_p chain, bitmap tmp_vars) { unsigned i; unsigned n = chain->length; dref root = get_chain_root (chain); bool reuse_first = !chain->has_max_use_after; tree ref, init, var, next; gimple phi; gimple_seq stmts; edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop); /* If N == 0, then all the references are within the single iteration. And since this is an nonempty chain, reuse_first cannot be true. */ gcc_assert (n > 0 || !reuse_first); chain->vars.create (n + 1); if (chain->type == CT_COMBINATION) ref = gimple_assign_lhs (root->stmt); else ref = DR_REF (root->ref); for (i = 0; i < n + (reuse_first ? 0 : 1); i++) { var = predcom_tmp_var (ref, i, tmp_vars); chain->vars.quick_push (var); } if (reuse_first) chain->vars.quick_push (chain->vars[0]); FOR_EACH_VEC_ELT (chain->vars, i, var) chain->vars[i] = make_ssa_name (var, NULL); for (i = 0; i < n; i++) { var = chain->vars[i]; next = chain->vars[i + 1]; init = get_init_expr (chain, i); init = force_gimple_operand (init, &stmts, true, NULL_TREE); if (stmts) gsi_insert_seq_on_edge_immediate (entry, stmts); phi = create_phi_node (var, loop->header); add_phi_arg (phi, init, entry, UNKNOWN_LOCATION); add_phi_arg (phi, next, latch, UNKNOWN_LOCATION); } } /* Create the variables and initialization statement for root of chain CHAIN. Uids of the newly created temporary variables are marked in TMP_VARS. */ static void initialize_root (struct loop *loop, chain_p chain, bitmap tmp_vars) { dref root = get_chain_root (chain); bool in_lhs = (chain->type == CT_STORE_LOAD || chain->type == CT_COMBINATION); initialize_root_vars (loop, chain, tmp_vars); replace_ref_with (root->stmt, chain->vars[chain->length], true, in_lhs); } /* Initializes a variable for load motion for ROOT and prepares phi nodes and initialization on entry to LOOP if necessary. The ssa name for the variable is stored in VARS. If WRITTEN is true, also a phi node to copy its value around the loop is created. Uid of the newly created temporary variable is marked in TMP_VARS. INITS is the list containing the (single) initializer. */ static void initialize_root_vars_lm (struct loop *loop, dref root, bool written, vec *vars, vec inits, bitmap tmp_vars) { unsigned i; tree ref = DR_REF (root->ref), init, var, next; gimple_seq stmts; gimple phi; edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop); /* Find the initializer for the variable, and check that it cannot trap. */ init = inits[0]; vars->create (written ? 2 : 1); var = predcom_tmp_var (ref, 0, tmp_vars); vars->quick_push (var); if (written) vars->quick_push ((*vars)[0]); FOR_EACH_VEC_ELT (*vars, i, var) (*vars)[i] = make_ssa_name (var, NULL); var = (*vars)[0]; init = force_gimple_operand (init, &stmts, written, NULL_TREE); if (stmts) gsi_insert_seq_on_edge_immediate (entry, stmts); if (written) { next = (*vars)[1]; phi = create_phi_node (var, loop->header); add_phi_arg (phi, init, entry, UNKNOWN_LOCATION); add_phi_arg (phi, next, latch, UNKNOWN_LOCATION); } else { gimple init_stmt = gimple_build_assign (var, init); gsi_insert_on_edge_immediate (entry, init_stmt); } } /* Execute load motion for references in chain CHAIN. Uids of the newly created temporary variables are marked in TMP_VARS. */ static void execute_load_motion (struct loop *loop, chain_p chain, bitmap tmp_vars) { auto_vec vars; dref a; unsigned n_writes = 0, ridx, i; tree var; gcc_assert (chain->type == CT_INVARIANT); gcc_assert (!chain->combined); FOR_EACH_VEC_ELT (chain->refs, i, a) if (DR_IS_WRITE (a->ref)) n_writes++; /* If there are no reads in the loop, there is nothing to do. */ if (n_writes == chain->refs.length ()) return; initialize_root_vars_lm (loop, get_chain_root (chain), n_writes > 0, &vars, chain->inits, tmp_vars); ridx = 0; FOR_EACH_VEC_ELT (chain->refs, i, a) { bool is_read = DR_IS_READ (a->ref); if (DR_IS_WRITE (a->ref)) { n_writes--; if (n_writes) { var = vars[0]; var = make_ssa_name (SSA_NAME_VAR (var), NULL); vars[0] = var; } else ridx = 1; } replace_ref_with (a->stmt, vars[ridx], !is_read, !is_read); } } /* Returns the single statement in that NAME is used, excepting the looparound phi nodes contained in one of the chains. If there is no such statement, or more statements, NULL is returned. */ static gimple single_nonlooparound_use (tree name) { use_operand_p use; imm_use_iterator it; gimple stmt, ret = NULL; FOR_EACH_IMM_USE_FAST (use, it, name) { stmt = USE_STMT (use); if (gimple_code (stmt) == GIMPLE_PHI) { /* Ignore uses in looparound phi nodes. Uses in other phi nodes could not be processed anyway, so just fail for them. */ if (bitmap_bit_p (looparound_phis, SSA_NAME_VERSION (PHI_RESULT (stmt)))) continue; return NULL; } else if (is_gimple_debug (stmt)) continue; else if (ret != NULL) return NULL; else ret = stmt; } return ret; } /* Remove statement STMT, as well as the chain of assignments in that it is used. */ static void remove_stmt (gimple stmt) { tree name; gimple next; gimple_stmt_iterator psi; if (gimple_code (stmt) == GIMPLE_PHI) { name = PHI_RESULT (stmt); next = single_nonlooparound_use (name); reset_debug_uses (stmt); psi = gsi_for_stmt (stmt); remove_phi_node (&psi, true); if (!next || !gimple_assign_ssa_name_copy_p (next) || gimple_assign_rhs1 (next) != name) return; stmt = next; } while (1) { gimple_stmt_iterator bsi; bsi = gsi_for_stmt (stmt); name = gimple_assign_lhs (stmt); gcc_assert (TREE_CODE (name) == SSA_NAME); next = single_nonlooparound_use (name); reset_debug_uses (stmt); unlink_stmt_vdef (stmt); gsi_remove (&bsi, true); release_defs (stmt); if (!next || !gimple_assign_ssa_name_copy_p (next) || gimple_assign_rhs1 (next) != name) return; stmt = next; } } /* Perform the predictive commoning optimization for a chain CHAIN. Uids of the newly created temporary variables are marked in TMP_VARS.*/ static void execute_pred_commoning_chain (struct loop *loop, chain_p chain, bitmap tmp_vars) { unsigned i; dref a; tree var; if (chain->combined) { /* For combined chains, just remove the statements that are used to compute the values of the expression (except for the root one). */ for (i = 1; chain->refs.iterate (i, &a); i++) remove_stmt (a->stmt); } else { /* For non-combined chains, set up the variables that hold its value, and replace the uses of the original references by these variables. */ initialize_root (loop, chain, tmp_vars); for (i = 1; chain->refs.iterate (i, &a); i++) { var = chain->vars[chain->length - a->distance]; replace_ref_with (a->stmt, var, false, false); } } } /* Determines the unroll factor necessary to remove as many temporary variable copies as possible. CHAINS is the list of chains that will be optimized. */ static unsigned determine_unroll_factor (vec chains) { chain_p chain; unsigned factor = 1, af, nfactor, i; unsigned max = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES); FOR_EACH_VEC_ELT (chains, i, chain) { if (chain->type == CT_INVARIANT || chain->combined) continue; /* The best unroll factor for this chain is equal to the number of temporary variables that we create for it. */ af = chain->length; if (chain->has_max_use_after) af++; nfactor = factor * af / gcd (factor, af); if (nfactor <= max) factor = nfactor; } return factor; } /* Perform the predictive commoning optimization for CHAINS. Uids of the newly created temporary variables are marked in TMP_VARS. */ static void execute_pred_commoning (struct loop *loop, vec chains, bitmap tmp_vars) { chain_p chain; unsigned i; FOR_EACH_VEC_ELT (chains, i, chain) { if (chain->type == CT_INVARIANT) execute_load_motion (loop, chain, tmp_vars); else execute_pred_commoning_chain (loop, chain, tmp_vars); } update_ssa (TODO_update_ssa_only_virtuals); } /* For each reference in CHAINS, if its defining statement is phi node, record the ssa name that is defined by it. */ static void replace_phis_by_defined_names (vec chains) { chain_p chain; dref a; unsigned i, j; FOR_EACH_VEC_ELT (chains, i, chain) FOR_EACH_VEC_ELT (chain->refs, j, a) { if (gimple_code (a->stmt) == GIMPLE_PHI) { a->name_defined_by_phi = PHI_RESULT (a->stmt); a->stmt = NULL; } } } /* For each reference in CHAINS, if name_defined_by_phi is not NULL, use it to set the stmt field. */ static void replace_names_by_phis (vec chains) { chain_p chain; dref a; unsigned i, j; FOR_EACH_VEC_ELT (chains, i, chain) FOR_EACH_VEC_ELT (chain->refs, j, a) if (a->stmt == NULL) { a->stmt = SSA_NAME_DEF_STMT (a->name_defined_by_phi); gcc_assert (gimple_code (a->stmt) == GIMPLE_PHI); a->name_defined_by_phi = NULL_TREE; } } /* Wrapper over execute_pred_commoning, to pass it as a callback to tree_transform_and_unroll_loop. */ struct epcc_data { vec chains; bitmap tmp_vars; }; static void execute_pred_commoning_cbck (struct loop *loop, void *data) { struct epcc_data *const dta = (struct epcc_data *) data; /* Restore phi nodes that were replaced by ssa names before tree_transform_and_unroll_loop (see detailed description in tree_predictive_commoning_loop). */ replace_names_by_phis (dta->chains); execute_pred_commoning (loop, dta->chains, dta->tmp_vars); } /* Base NAME and all the names in the chain of phi nodes that use it on variable VAR. The phi nodes are recognized by being in the copies of the header of the LOOP. */ static void base_names_in_chain_on (struct loop *loop, tree name, tree var) { gimple stmt, phi; imm_use_iterator iter; replace_ssa_name_symbol (name, var); while (1) { phi = NULL; FOR_EACH_IMM_USE_STMT (stmt, iter, name) { if (gimple_code (stmt) == GIMPLE_PHI && flow_bb_inside_loop_p (loop, gimple_bb (stmt))) { phi = stmt; BREAK_FROM_IMM_USE_STMT (iter); } } if (!phi) return; name = PHI_RESULT (phi); replace_ssa_name_symbol (name, var); } } /* Given an unrolled LOOP after predictive commoning, remove the register copies arising from phi nodes by changing the base variables of SSA names. TMP_VARS is the set of the temporary variables for those we want to perform this. */ static void eliminate_temp_copies (struct loop *loop, bitmap tmp_vars) { edge e; gimple phi, stmt; tree name, use, var; gimple_stmt_iterator psi; e = loop_latch_edge (loop); for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) { phi = gsi_stmt (psi); name = PHI_RESULT (phi); var = SSA_NAME_VAR (name); if (!var || !bitmap_bit_p (tmp_vars, DECL_UID (var))) continue; use = PHI_ARG_DEF_FROM_EDGE (phi, e); gcc_assert (TREE_CODE (use) == SSA_NAME); /* Base all the ssa names in the ud and du chain of NAME on VAR. */ stmt = SSA_NAME_DEF_STMT (use); while (gimple_code (stmt) == GIMPLE_PHI /* In case we could not unroll the loop enough to eliminate all copies, we may reach the loop header before the defining statement (in that case, some register copies will be present in loop latch in the final code, corresponding to the newly created looparound phi nodes). */ && gimple_bb (stmt) != loop->header) { gcc_assert (single_pred_p (gimple_bb (stmt))); use = PHI_ARG_DEF (stmt, 0); stmt = SSA_NAME_DEF_STMT (use); } base_names_in_chain_on (loop, use, var); } } /* Returns true if CHAIN is suitable to be combined. */ static bool chain_can_be_combined_p (chain_p chain) { return (!chain->combined && (chain->type == CT_LOAD || chain->type == CT_COMBINATION)); } /* Returns the modify statement that uses NAME. Skips over assignment statements, NAME is replaced with the actual name used in the returned statement. */ static gimple find_use_stmt (tree *name) { gimple stmt; tree rhs, lhs; /* Skip over assignments. */ while (1) { stmt = single_nonlooparound_use (*name); if (!stmt) return NULL; if (gimple_code (stmt) != GIMPLE_ASSIGN) return NULL; lhs = gimple_assign_lhs (stmt); if (TREE_CODE (lhs) != SSA_NAME) return NULL; if (gimple_assign_copy_p (stmt)) { rhs = gimple_assign_rhs1 (stmt); if (rhs != *name) return NULL; *name = lhs; } else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)) == GIMPLE_BINARY_RHS) return stmt; else return NULL; } } /* Returns true if we may perform reassociation for operation CODE in TYPE. */ static bool may_reassociate_p (tree type, enum tree_code code) { if (FLOAT_TYPE_P (type) && !flag_unsafe_math_optimizations) return false; return (commutative_tree_code (code) && associative_tree_code (code)); } /* If the operation used in STMT is associative and commutative, go through the tree of the same operations and returns its root. Distance to the root is stored in DISTANCE. */ static gimple find_associative_operation_root (gimple stmt, unsigned *distance) { tree lhs; gimple next; enum tree_code code = gimple_assign_rhs_code (stmt); tree type = TREE_TYPE (gimple_assign_lhs (stmt)); unsigned dist = 0; if (!may_reassociate_p (type, code)) return NULL; while (1) { lhs = gimple_assign_lhs (stmt); gcc_assert (TREE_CODE (lhs) == SSA_NAME); next = find_use_stmt (&lhs); if (!next || gimple_assign_rhs_code (next) != code) break; stmt = next; dist++; } if (distance) *distance = dist; return stmt; } /* Returns the common statement in that NAME1 and NAME2 have a use. If there is no such statement, returns NULL_TREE. In case the operation used on NAME1 and NAME2 is associative and commutative, returns the root of the tree formed by this operation instead of the statement that uses NAME1 or NAME2. */ static gimple find_common_use_stmt (tree *name1, tree *name2) { gimple stmt1, stmt2; stmt1 = find_use_stmt (name1); if (!stmt1) return NULL; stmt2 = find_use_stmt (name2); if (!stmt2) return NULL; if (stmt1 == stmt2) return stmt1; stmt1 = find_associative_operation_root (stmt1, NULL); if (!stmt1) return NULL; stmt2 = find_associative_operation_root (stmt2, NULL); if (!stmt2) return NULL; return (stmt1 == stmt2 ? stmt1 : NULL); } /* Checks whether R1 and R2 are combined together using CODE, with the result in RSLT_TYPE, in order R1 CODE R2 if SWAP is false and in order R2 CODE R1 if it is true. If CODE is ERROR_MARK, set these values instead. */ static bool combinable_refs_p (dref r1, dref r2, enum tree_code *code, bool *swap, tree *rslt_type) { enum tree_code acode; bool aswap; tree atype; tree name1, name2; gimple stmt; name1 = name_for_ref (r1); name2 = name_for_ref (r2); gcc_assert (name1 != NULL_TREE && name2 != NULL_TREE); stmt = find_common_use_stmt (&name1, &name2); if (!stmt /* A simple post-dominance check - make sure the combination is executed under the same condition as the references. */ || (gimple_bb (stmt) != gimple_bb (r1->stmt) && gimple_bb (stmt) != gimple_bb (r2->stmt))) return false; acode = gimple_assign_rhs_code (stmt); aswap = (!commutative_tree_code (acode) && gimple_assign_rhs1 (stmt) != name1); atype = TREE_TYPE (gimple_assign_lhs (stmt)); if (*code == ERROR_MARK) { *code = acode; *swap = aswap; *rslt_type = atype; return true; } return (*code == acode && *swap == aswap && *rslt_type == atype); } /* Remove OP from the operation on rhs of STMT, and replace STMT with an assignment of the remaining operand. */ static void remove_name_from_operation (gimple stmt, tree op) { tree other_op; gimple_stmt_iterator si; gcc_assert (is_gimple_assign (stmt)); if (gimple_assign_rhs1 (stmt) == op) other_op = gimple_assign_rhs2 (stmt); else other_op = gimple_assign_rhs1 (stmt); si = gsi_for_stmt (stmt); gimple_assign_set_rhs_from_tree (&si, other_op); /* We should not have reallocated STMT. */ gcc_assert (gsi_stmt (si) == stmt); update_stmt (stmt); } /* Reassociates the expression in that NAME1 and NAME2 are used so that they are combined in a single statement, and returns this statement. */ static gimple reassociate_to_the_same_stmt (tree name1, tree name2) { gimple stmt1, stmt2, root1, root2, s1, s2; gimple new_stmt, tmp_stmt; tree new_name, tmp_name, var, r1, r2; unsigned dist1, dist2; enum tree_code code; tree type = TREE_TYPE (name1); gimple_stmt_iterator bsi; stmt1 = find_use_stmt (&name1); stmt2 = find_use_stmt (&name2); root1 = find_associative_operation_root (stmt1, &dist1); root2 = find_associative_operation_root (stmt2, &dist2); code = gimple_assign_rhs_code (stmt1); gcc_assert (root1 && root2 && root1 == root2 && code == gimple_assign_rhs_code (stmt2)); /* Find the root of the nearest expression in that both NAME1 and NAME2 are used. */ r1 = name1; s1 = stmt1; r2 = name2; s2 = stmt2; while (dist1 > dist2) { s1 = find_use_stmt (&r1); r1 = gimple_assign_lhs (s1); dist1--; } while (dist2 > dist1) { s2 = find_use_stmt (&r2); r2 = gimple_assign_lhs (s2); dist2--; } while (s1 != s2) { s1 = find_use_stmt (&r1); r1 = gimple_assign_lhs (s1); s2 = find_use_stmt (&r2); r2 = gimple_assign_lhs (s2); } /* Remove NAME1 and NAME2 from the statements in that they are used currently. */ remove_name_from_operation (stmt1, name1); remove_name_from_operation (stmt2, name2); /* Insert the new statement combining NAME1 and NAME2 before S1, and combine it with the rhs of S1. */ var = create_tmp_reg (type, "predreastmp"); new_name = make_ssa_name (var, NULL); new_stmt = gimple_build_assign_with_ops (code, new_name, name1, name2); var = create_tmp_reg (type, "predreastmp"); tmp_name = make_ssa_name (var, NULL); /* Rhs of S1 may now be either a binary expression with operation CODE, or gimple_val (in case that stmt1 == s1 or stmt2 == s1, so that name1 or name2 was removed from it). */ tmp_stmt = gimple_build_assign_with_ops (gimple_assign_rhs_code (s1), tmp_name, gimple_assign_rhs1 (s1), gimple_assign_rhs2 (s1)); bsi = gsi_for_stmt (s1); gimple_assign_set_rhs_with_ops (&bsi, code, new_name, tmp_name); s1 = gsi_stmt (bsi); update_stmt (s1); gsi_insert_before (&bsi, new_stmt, GSI_SAME_STMT); gsi_insert_before (&bsi, tmp_stmt, GSI_SAME_STMT); return new_stmt; } /* Returns the statement that combines references R1 and R2. In case R1 and R2 are not used in the same statement, but they are used with an associative and commutative operation in the same expression, reassociate the expression so that they are used in the same statement. */ static gimple stmt_combining_refs (dref r1, dref r2) { gimple stmt1, stmt2; tree name1 = name_for_ref (r1); tree name2 = name_for_ref (r2); stmt1 = find_use_stmt (&name1); stmt2 = find_use_stmt (&name2); if (stmt1 == stmt2) return stmt1; return reassociate_to_the_same_stmt (name1, name2); } /* Tries to combine chains CH1 and CH2 together. If this succeeds, the description of the new chain is returned, otherwise we return NULL. */ static chain_p combine_chains (chain_p ch1, chain_p ch2) { dref r1, r2, nw; enum tree_code op = ERROR_MARK; bool swap = false; chain_p new_chain; unsigned i; gimple root_stmt; tree rslt_type = NULL_TREE; if (ch1 == ch2) return NULL; if (ch1->length != ch2->length) return NULL; if (ch1->refs.length () != ch2->refs.length ()) return NULL; for (i = 0; (ch1->refs.iterate (i, &r1) && ch2->refs.iterate (i, &r2)); i++) { if (r1->distance != r2->distance) return NULL; if (!combinable_refs_p (r1, r2, &op, &swap, &rslt_type)) return NULL; } if (swap) { chain_p tmp = ch1; ch1 = ch2; ch2 = tmp; } new_chain = XCNEW (struct chain); new_chain->type = CT_COMBINATION; new_chain->op = op; new_chain->ch1 = ch1; new_chain->ch2 = ch2; new_chain->rslt_type = rslt_type; new_chain->length = ch1->length; for (i = 0; (ch1->refs.iterate (i, &r1) && ch2->refs.iterate (i, &r2)); i++) { nw = XCNEW (struct dref_d); nw->stmt = stmt_combining_refs (r1, r2); nw->distance = r1->distance; new_chain->refs.safe_push (nw); } new_chain->has_max_use_after = false; root_stmt = get_chain_root (new_chain)->stmt; for (i = 1; new_chain->refs.iterate (i, &nw); i++) { if (nw->distance == new_chain->length && !stmt_dominates_stmt_p (nw->stmt, root_stmt)) { new_chain->has_max_use_after = true; break; } } ch1->combined = true; ch2->combined = true; return new_chain; } /* Try to combine the CHAINS. */ static void try_combine_chains (vec *chains) { unsigned i, j; chain_p ch1, ch2, cch; auto_vec worklist; FOR_EACH_VEC_ELT (*chains, i, ch1) if (chain_can_be_combined_p (ch1)) worklist.safe_push (ch1); while (!worklist.is_empty ()) { ch1 = worklist.pop (); if (!chain_can_be_combined_p (ch1)) continue; FOR_EACH_VEC_ELT (*chains, j, ch2) { if (!chain_can_be_combined_p (ch2)) continue; cch = combine_chains (ch1, ch2); if (cch) { worklist.safe_push (cch); chains->safe_push (cch); break; } } } } /* Prepare initializers for CHAIN in LOOP. Returns false if this is impossible because one of these initializers may trap, true otherwise. */ static bool prepare_initializers_chain (struct loop *loop, chain_p chain) { unsigned i, n = (chain->type == CT_INVARIANT) ? 1 : chain->length; struct data_reference *dr = get_chain_root (chain)->ref; tree init; gimple_seq stmts; dref laref; edge entry = loop_preheader_edge (loop); /* Find the initializers for the variables, and check that they cannot trap. */ chain->inits.create (n); for (i = 0; i < n; i++) chain->inits.quick_push (NULL_TREE); /* If we have replaced some looparound phi nodes, use their initializers instead of creating our own. */ FOR_EACH_VEC_ELT (chain->refs, i, laref) { if (gimple_code (laref->stmt) != GIMPLE_PHI) continue; gcc_assert (laref->distance > 0); chain->inits[n - laref->distance] = PHI_ARG_DEF_FROM_EDGE (laref->stmt, entry); } for (i = 0; i < n; i++) { if (chain->inits[i] != NULL_TREE) continue; init = ref_at_iteration (dr, (int) i - n, &stmts); if (!chain->all_always_accessed && tree_could_trap_p (init)) return false; if (stmts) gsi_insert_seq_on_edge_immediate (entry, stmts); chain->inits[i] = init; } return true; } /* Prepare initializers for CHAINS in LOOP, and free chains that cannot be used because the initializers might trap. */ static void prepare_initializers (struct loop *loop, vec chains) { chain_p chain; unsigned i; for (i = 0; i < chains.length (); ) { chain = chains[i]; if (prepare_initializers_chain (loop, chain)) i++; else { release_chain (chain); chains.unordered_remove (i); } } } /* Performs predictive commoning for LOOP. Returns true if LOOP was unrolled. */ static bool tree_predictive_commoning_loop (struct loop *loop) { vec datarefs; vec dependences; struct component *components; vec chains = vNULL; unsigned unroll_factor; struct tree_niter_desc desc; bool unroll = false; edge exit; bitmap tmp_vars; if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Processing loop %d\n", loop->num); /* Find the data references and split them into components according to their dependence relations. */ auto_vec loop_nest; dependences.create (10); datarefs.create (10); if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, &dependences)) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Cannot analyze data dependencies\n"); free_data_refs (datarefs); free_dependence_relations (dependences); return false; } if (dump_file && (dump_flags & TDF_DETAILS)) dump_data_dependence_relations (dump_file, dependences); components = split_data_refs_to_components (loop, datarefs, dependences); loop_nest.release (); free_dependence_relations (dependences); if (!components) { free_data_refs (datarefs); free_affine_expand_cache (&name_expansions); return false; } if (dump_file && (dump_flags & TDF_DETAILS)) { fprintf (dump_file, "Initial state:\n\n"); dump_components (dump_file, components); } /* Find the suitable components and split them into chains. */ components = filter_suitable_components (loop, components); tmp_vars = BITMAP_ALLOC (NULL); looparound_phis = BITMAP_ALLOC (NULL); determine_roots (loop, components, &chains); release_components (components); if (!chains.exists ()) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Predictive commoning failed: no suitable chains\n"); goto end; } prepare_initializers (loop, chains); /* Try to combine the chains that are always worked with together. */ try_combine_chains (&chains); if (dump_file && (dump_flags & TDF_DETAILS)) { fprintf (dump_file, "Before commoning:\n\n"); dump_chains (dump_file, chains); } /* Determine the unroll factor, and if the loop should be unrolled, ensure that its number of iterations is divisible by the factor. */ unroll_factor = determine_unroll_factor (chains); scev_reset (); unroll = (unroll_factor > 1 && can_unroll_loop_p (loop, unroll_factor, &desc)); exit = single_dom_exit (loop); /* Execute the predictive commoning transformations, and possibly unroll the loop. */ if (unroll) { struct epcc_data dta; if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unrolling %u times.\n", unroll_factor); dta.chains = chains; dta.tmp_vars = tmp_vars; update_ssa (TODO_update_ssa_only_virtuals); /* Cfg manipulations performed in tree_transform_and_unroll_loop before execute_pred_commoning_cbck is called may cause phi nodes to be reallocated, which is a problem since CHAINS may point to these statements. To fix this, we store the ssa names defined by the phi nodes here instead of the phi nodes themselves, and restore the phi nodes in execute_pred_commoning_cbck. A bit hacky. */ replace_phis_by_defined_names (chains); tree_transform_and_unroll_loop (loop, unroll_factor, exit, &desc, execute_pred_commoning_cbck, &dta); eliminate_temp_copies (loop, tmp_vars); } else { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Executing predictive commoning without unrolling.\n"); execute_pred_commoning (loop, chains, tmp_vars); } end: ; release_chains (chains); free_data_refs (datarefs); BITMAP_FREE (tmp_vars); BITMAP_FREE (looparound_phis); free_affine_expand_cache (&name_expansions); return unroll; } /* Runs predictive commoning. */ unsigned tree_predictive_commoning (void) { bool unrolled = false; struct loop *loop; unsigned ret = 0; initialize_original_copy_tables (); FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) if (optimize_loop_for_speed_p (loop)) { unrolled |= tree_predictive_commoning_loop (loop); } if (unrolled) { scev_reset (); ret = TODO_cleanup_cfg; } free_original_copy_tables (); return ret; } /* Predictive commoning Pass. */ static unsigned run_tree_predictive_commoning (void) { if (!current_loops) return 0; return tree_predictive_commoning (); } static bool gate_tree_predictive_commoning (void) { return flag_predictive_commoning != 0; } namespace { const pass_data pass_data_predcom = { GIMPLE_PASS, /* type */ "pcom", /* name */ OPTGROUP_LOOP, /* optinfo_flags */ true, /* has_gate */ true, /* has_execute */ TV_PREDCOM, /* tv_id */ PROP_cfg, /* properties_required */ 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ TODO_update_ssa_only_virtuals, /* todo_flags_finish */ }; class pass_predcom : public gimple_opt_pass { public: pass_predcom (gcc::context *ctxt) : gimple_opt_pass (pass_data_predcom, ctxt) {} /* opt_pass methods: */ bool gate () { return gate_tree_predictive_commoning (); } unsigned int execute () { return run_tree_predictive_commoning (); } }; // class pass_predcom } // anon namespace gimple_opt_pass * make_pass_predcom (gcc::context *ctxt) { return new pass_predcom (ctxt); }