On Fri, May 29, 2020 at 8:52 PM Erick Ochoa
<erick.oc...@theobroma-systems.com> wrote:
>
>
>
> This pass is a variant of constant propagation where global
> primitive constants with a single write are propagated to multiple
> read statements.
Just a few small comments while skimming through the code
> ChangeLog:
>
> 2020-05-20 Erick Ochoa <erick.oc...@theobroma.systems.com>
>
> gcc/Makefile.in: Adds new pass
> gcc/passes.def: Same
> gcc/tree-pass.h: Same
> gcc/common.opt: Same
> gcc/cgraph.h: Adds new field to cgraph_node
> gcc/cgraph.c: Same
> gcc/ipa-cp.c: May skip ipa-cp for a function
> if initcall-cp has constants to propagate
> in that same function
> gcc/ipa-initcall-cp.c: New file
> ---
> gcc/Makefile.in | 1 +
> gcc/cgraph.h | 5 +-
> gcc/common.opt | 8 +
> gcc/ipa-cp.c | 2 +-
> gcc/ipa-initcall-cp.c | 1199 +++++++++++++++++++++++++++++++++++++++++
> gcc/passes.def | 1 +
> gcc/tree-pass.h | 1 +
> 7 files changed, 1215 insertions(+), 2 deletions(-)
> create mode 100644 gcc/ipa-initcall-cp.c
>
> diff --git a/gcc/Makefile.in b/gcc/Makefile.in
> index 543b477ff18..b94fb633d78 100644
> --- a/gcc/Makefile.in
> +++ b/gcc/Makefile.in
> @@ -1401,6 +1401,7 @@ OBJS = \
> internal-fn.o \
> ipa-cp.o \
> ipa-sra.o \
> + ipa-initcall-cp.o \
> ipa-devirt.o \
> ipa-fnsummary.o \
> ipa-polymorphic-call.o \
> diff --git a/gcc/cgraph.h b/gcc/cgraph.h
> index 5ddeb65269b..532b4671609 100644
> --- a/gcc/cgraph.h
> +++ b/gcc/cgraph.h
> @@ -937,7 +937,7 @@ struct GTY((tag ("SYMTAB_FUNCTION"))) cgraph_node :
> public symtab_node
> split_part (false), indirect_call_target (false), local (false),
> versionable (false), can_change_signature (false),
> redefined_extern_inline (false), tm_may_enter_irr (false),
> - ipcp_clone (false), m_uid (uid), m_summary_id (-1)
> + ipcp_clone (false), skip_ipa_cp (false), m_uid (uid),
> m_summary_id (-1)
> {}
>
> /* Remove the node from cgraph and all inline clones inlined into it.
> @@ -1539,6 +1539,9 @@ struct GTY((tag ("SYMTAB_FUNCTION"))) cgraph_node
> : public symtab_node
> unsigned tm_may_enter_irr : 1;
> /* True if this was a clone created by ipa-cp. */
> unsigned ipcp_clone : 1;
> + /* True if ipa-initcall-cp and therefore we need to skip ipa-cp for
> cgraph
> + * node. */
> + unsigned skip_ipa_cp : 1;
>
> private:
> /* Unique id of the node. */
> diff --git a/gcc/common.opt b/gcc/common.opt
> index d33383b523c..b2d8d1b0958 100644
> --- a/gcc/common.opt
> +++ b/gcc/common.opt
> @@ -3408,4 +3408,12 @@ fipa-ra
> Common Report Var(flag_ipa_ra) Optimization
> Use caller save register across calls if possible.
>
> +fipa-initcall-cp-dryrun
> +Common Report Var(flag_initcall_cp_dryrun)
> +Run analysis for propagating constants across initialization functions.
> +
> +fipa-initcall-cp
> +Common Report Var(flag_initcall_cp) Optimization
> +Run transformation for propagation constants across initialization
> functions.
> +
> ; This comment is to ensure we retain the blank line above.
> diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
> index c64e9104a94..1036cb0684e 100644
> --- a/gcc/ipa-cp.c
> +++ b/gcc/ipa-cp.c
> @@ -1188,7 +1188,7 @@ initialize_node_lattices (struct cgraph_node *node)
>
> gcc_checking_assert (node->has_gimple_body_p ());
>
> - if (!ipa_get_param_count (info))
> + if (!ipa_get_param_count (info) || node->skip_ipa_cp)
> disable = true;
> else if (node->local)
> {
> diff --git a/gcc/ipa-initcall-cp.c b/gcc/ipa-initcall-cp.c
> new file mode 100644
> index 00000000000..02f70b7a908
> --- /dev/null
> +++ b/gcc/ipa-initcall-cp.c
> @@ -0,0 +1,1199 @@
> +/* Initcall constant propagation pass.
> + Copyright (C) 2019-2020 Free Software Foundation, Inc.
> +
> + Contributed by Erick Ochoa <erick.oc...@theobroma-systems.com>
> +
> +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
> +<http://www.gnu.org/licenses/>. */
> +
> +#include "config.h"
> +#include "system.h"
> +#include "coretypes.h"
> +#include "backend.h"
> +#include "tree.h"
> +#include "tree-eh.h"
> +#include "gimple.h"
> +#include "gimple-expr.h"
> +#include "gimple-iterator.h"
> +#include "predict.h"
> +#include "alloc-pool.h"
> +#include "tree-pass.h"
> +#include "cgraph.h"
> +#include "diagnostic.h"
> +#include "fold-const.h"
> +#include "gimple-fold.h"
> +#include "symbol-summary.h"
> +#include "tree-vrp.h"
> +#include "ipa-prop.h"
> +#include "tree-pretty-print.h"
> +#include "tree-inline.h"
> +#include "ipa-fnsummary.h"
> +#include "ipa-utils.h"
> +#include "tree-ssa-ccp.h"
> +#include "stringpool.h"
> +#include "attribs.h"
> +#include "gimple-pretty-print.h"
> +#include "ssa.h"
> +
> +#define INITCALL_CP_SUFFIX "initcall.cp"
> +
> +typedef vec<struct ipa_ref *> ipa_ref_vec;
> +
> +/* log to dump file */
> +static inline void
> +log (const char *const fmt, ...)
> +{
> + if (!dump_file)
> + return;
> +
> + va_list args;
> + va_start (args, fmt);
> + vfprintf (dump_file, fmt, args);
> + va_end (args);
> +}
> +
> +bool
> +reach_nodes_via_bb1 (cgraph_node *n2);
> +static bool
> +bb1_dominates_bb2 (basic_block, basic_block, cgraph_node *);
> +static bool
> +write_before_call (struct ipa_ref *, struct ipa_ref *);
> +static bool
> +call_before_read (struct ipa_ref *, struct ipa_ref *);
> +static hash_map<const char *, unsigned> *clone_num_suffixes1;
> +static hash_map<cgraph_node *, cgraph_node *> *func_to_clone;
> +static vec<struct cgraph_node *>
> +get_nondominated_callees (cgraph_node *caller, cgraph_node *callee,
> + bool *exitEarly = NULL);
> +
> +static void
> +load_function_body_of_ipa_ref (cgraph_node *node)
> +{
> + gcc_assert (in_lto_p);
> + cgraph_node *f_cnode2 = node->ultimate_alias_target ();
> + const char *name = f_cnode2->dump_asm_name ();
> + log ("%s: for function '%s'\n", __func__, name);
> + if (dump_file)
> + {
> + dump_node (f_cnode2->decl, TDF_DETAILS, dump_file);
> + }
> + f_cnode2->get_untransformed_body ();
> +}
> +
> +static void
> +load_function_body_of_ipa_ref (struct ipa_ref *ref)
> +{
> + /* IPA passes don't get the function bodies during LTO. */
> + gcc_assert (in_lto_p);
> +
> + symtab_node *f_node = ref->referring;
> + cgraph_node *f_cnode = dyn_cast<cgraph_node *> (f_node);
> + load_function_body_of_ipa_ref (f_cnode);
> +}
> +
> +static void
> +dump_vnode_ipa_ref (ipa_ref *ref)
> +{
> + if (!dump_file)
> + return;
> + log ("Reference type: %s\n", stringify_ipa_ref_use (ref->use));
> +
> + symtab_node *f_node = ref->referring;
> + log ("Reference function: %s\n", f_node->dump_asm_name ());
> +
> + log ("Gimple stmt (@0x%lx, lto_stmt_uid: %u):\n", (long) ref->stmt,
> + ref->lto_stmt_uid);
> + load_function_body_of_ipa_ref (ref);
> + print_gimple_stmt (dump_file, ref->stmt, 0, TDF_NONE);
> +}
> +
> +/* Return true of all ops of assignment are constants. */
> +static bool
> +gimple_assign_is_single_const (gimple *stmt)
> +{
> + tree op;
> +
> + gcc_assert (gimple_code (stmt) == GIMPLE_ASSIGN);
> +
> + if (gimple_has_volatile_ops (stmt))
> + {
> + log ("has volatile ops!\n");
> + return false;
> + }
> +
> + if (gimple_num_ops (stmt) != 2)
> + {
> + log ("wrong num of ops: %u!\n", gimple_num_ops (stmt));
> + return false;
> + }
So you probably want to check gimple_assign_single_p () which
verifies you are dealing with a plain assignment.
> + op = gimple_op (stmt, 1);
gimple_assign_rhs1 is the type-safe way to access the RHS.
> + if (!tree_code_is_cst (op))
> + {
> + log ("op is not cst!\n");
> + return false;
> + }
> +
> + return true;
> +}
> +
> +/* Collects calls which are not dominated by callee */
> +// assumptions:
> +// * there is a callsite from caller to callee
> +static vec<struct cgraph_node *>
> +get_nondominated_callees (cgraph_node *caller, cgraph_node *callee,
> + bool *exitEarly)
> +{
> + gcc_assert (in_lto_p);
> + caller->get_untransformed_body ();
> +
> + function *func = DECL_STRUCT_FUNCTION (caller->decl);
> + basic_block bb;
> + bool found = false;
> + FOR_EACH_BB_FN (bb, func)
> + {
> + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> + if (gimple_code (stmt) != GIMPLE_CALL)
> + continue;
> +
> + if (dump_file)
> + print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
> + tree rhs = gimple_op (stmt, 1);
gimple_op (stmt, 1) is the function called (gimple_call_fn () btw)
> + if (TREE_CODE (rhs) == RECORD_TYPE)
So this will never be true.
> + {
> + gcc_assert (exitEarly);
> + *exitEarly = true;
> + return vNULL;
> + }
> + tree callee_decl = gimple_call_fndecl (stmt);
> + if (!callee_decl)
> + {
> + gcc_assert (exitEarly);
> + *exitEarly = true;
> + return vNULL;
> + }
> + cgraph_node *callee_node = cgraph_node::get (callee_decl);
> + if (callee_node != callee)
The function is named _nondominated but I see no dominance
checks at all. FOR_EACH_BB_FN does not in any way iterate
over a specific order of basic-blocks.
> + continue;
> +
> + found = true;
> + }
> + if (found)
> + break;
> + }
> +
> + gcc_assert (found);
> +
> + // At this point in the program, we hold a valid bb.
> + // The callee is located in the bb
OK, so with cgraph-edges recorded you could have walked
cgraph edges, finding one with callee and use edge->call_stmt
to find the block. There may be more than one call to callee?
Eventually the caller of this function already knows the callgraph edge...
> + vec<struct cgraph_node *> callees = vNULL;
> + basic_block bb_c;
> + FOR_EACH_BB_FN (bb_c, func)
> + {
> + bool self_dominates = bb == bb_c;
> + bool bb_dominates_bbc = bb1_dominates_bb2 (bb, bb_c, caller);
> + if (bb_dominates_bbc && !self_dominates)
> + continue;
> +
> + for (gimple_stmt_iterator gsi = gsi_start_bb (bb_c); !gsi_end_p
> (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> + if (gimple_code (stmt) != GIMPLE_CALL)
> + continue;
> +
> + tree callee_decl = gimple_call_fndecl (stmt);
> + cgraph_node *callee_node = cgraph_node::get (callee_decl);
> +
> + if (fndecl_built_in_p (callee_node->decl))
> + continue;
> +
> + if (self_dominates && callee_node == callee)
> + {
> + break;
> + }
> +
> + callees.safe_push (callee_node);
> + }
> + }
> + return callees;
> +}
> +
> +/* Returns true if bb1 dominates bb2
> + * Includes self-dominance */
> +static bool
> +bb1_dominates_bb2 (basic_block bb1, basic_block bb2, cgraph_node *cnode)
> +{
> + // self dominance
> + if (bb1 == bb2)
> + return true;
> +
> + push_cfun (DECL_STRUCT_FUNCTION (cnode->decl));
> +
> + /* If dominator info is not available, we need to calculate it. */
> + if (!dom_info_available_p (CDI_DOMINATORS))
> + calculate_dominance_info (CDI_DOMINATORS);
> +
> + /* Check if the read is dominated by the write. */
> + bool ret = dominated_by_p (CDI_DOMINATORS, bb2, bb1);
> +
> + /* Restore cfun. */
> + pop_cfun ();
push/pop_cfun is _very_ expensive. Dominance query does not
need it set but computation looks at cfun IIRC. Adjust those to
explicitely passed struct function instead.
> +
> + return ret;
> +}
> +
> +/* Return true if write occurs before read.
> + * write and read must be located in the same function
> + * */
> +static bool
> +write_before_read_in_function (struct ipa_ref *write, struct ipa_ref *read)
> +{
> + basic_block w_bb = gimple_bb (write->stmt);
> + basic_block r_bb = gimple_bb (read->stmt);
> +
> + if (w_bb != r_bb)
> + {
> + /*
> + * The dominator framework operates on cfun.
> + * Therefore we need to set cfun accordingly.
> + */
> + symtab_node *w_node = write->referring;
> + cgraph_node *w_cnode = dyn_cast<cgraph_node *> (w_node);
> + return bb1_dominates_bb2 (w_bb, r_bb, w_cnode);
> + }
> +
> + gimple_stmt_iterator gsi;
> + for (gsi = gsi_start_bb (w_bb); !gsi_end_p (gsi); gsi_next (&gsi))
> + {
> + if (gsi_stmt (gsi) == write->stmt)
> + return true;
> + if (gsi_stmt (gsi) == read->stmt)
> + return false;
> + }
> +
> + /* Neither write nor read found it BB. */
> + gcc_unreachable ();
> + return false;
> +}
> +
> +/*
> + * DFS over callees and return if callee is a or b.
> + */
> +static cgraph_node *
> +find_cgraph_in_callee_set (cgraph_node *n, hash_set<cgraph_node *> set,
> + cgraph_node *a, cgraph_node *b)
> +{
> + cgraph_edge *cs;
> + for (cs = n->callees; cs; cs = cs->next_callee)
> + {
> + cgraph_node *callee = cs->callee->function_symbol ();
> + const char *const name = n->dump_asm_name ();
> + const char *const cname = callee->dump_asm_name ();
> + log ("Child of %s: %s\n", name, cname);
> + if (callee == a)
> + return a;
> + if (callee == b)
> + return b;
> + if (!set.contains (callee))
> + continue;
> + return find_cgraph_in_callee_set (callee, set, a, b);
> + }
> + return NULL;
> +}
> +
> +/* Walks back along caller relations until main is found. */
> +static cgraph_node *
> +get_ancestor_main_dfs (hash_set<cgraph_node *> *visited,
> + vec<cgraph_node *> *path, cgraph_node *node)
> +{
> + if (MAIN_NAME_P (DECL_NAME (node->decl)))
> + {
> + path->safe_push (node);
> + return node;
> + }
> +
> + visited->add (node);
> +
> + cgraph_edge *cs;
> + for (cs = node->callers; cs; cs = cs->next_caller)
> + {
> + cgraph_node *caller = cs->caller->function_symbol ();
> + if (visited->contains (caller))
> + continue;
> +
> + cgraph_node *main = get_ancestor_main_dfs (visited, path, caller);
> +
> + if (!main)
> + continue;
> +
> + path->safe_push (node);
> + return main;
> + }
> + return NULL;
> +}
> +
> +/* Returns path from main to cgraph_node in the vector */
> +static cgraph_node *
> +get_path_from_main_to (cgraph_node *node, vec<cgraph_node *> *path)
> +{
> + hash_set<cgraph_node *> visited;
> + cgraph_node *main = get_ancestor_main_dfs (&visited, path, node);
> + visited.empty ();
> + return main;
> +}
> +
> +/*
> + * Verifying that a stmt s1 is dominated by a stmt s2
> + * across function borders is not trivial with the available
> + * infrastructure (dominator algorithm on bb level plus cgraph).
> + * If we rule out external calls/callbacks, then we still need
> + * to guarantee a write on each possible path to the read.
> + *
> + * The implemented solution to this problem, which is of course
> + * too strict,
> + * but also not too compute/memory intensive is as follows:
> + *
> + * - Check if write is reachable by main () by only looking into
> + * the first bb in each function on the path.
> + * - All call stmts between main () and write must not possibly
> + * reach a read. We consider indirect call statements as
> + * possible reaching read (unless we can prove opposite).
> + *
> + * The external calls/callbacks are ruled out as follows:
> + *
> + * - all possible ancestors of read must not be external visible
> + * - all possible ancestors of read must not be function pointers
> + * - Something else?
> + */
> +static bool
> +write_before_read_across_function_borders (struct ipa_ref *write,
> + struct ipa_ref *read)
> +{
> + symtab_node *w_node = write->referring;
> + cgraph_node *w_cnode = dyn_cast<cgraph_node *> (w_node);
> + symtab_node *r_node = read->referring;
> + cgraph_node *r_cnode = dyn_cast<cgraph_node *> (r_node);
> + cgraph_node *main;
> +
> + /* Get main () function. */
> + vec<cgraph_node *> pw = vNULL;
> + main = get_path_from_main_to (w_cnode, &pw);
> + if (!main)
> + {
> + log ("main () is not ancestor of write -> skipping.\n");
> + return false;
> + }
> +
> + vec<cgraph_node *> pr = vNULL;
> + cgraph_node *main_to_read = get_path_from_main_to (r_cnode, &pr);
> + if (!main_to_read)
> + {
> + log ("main () is not ancestor of read -> skipping.\n");
> + return false;
> + }
> +
> + // Future work will involve looking at unconditional paths.
> + if (!reach_nodes_via_bb1 (w_cnode))
> + return false;
> +
> + int i;
> + cgraph_node *node_in_pr;
> + FOR_EACH_VEC_ELT (pr, i, node_in_pr)
> + {
> + // I think main has to be externally visible.
> + if (node_in_pr == main)
> + continue;
> +
> + /* Assure that all paths to read are not externally visible. */
> + if (cgraph_externally_visible_p (node_in_pr, flag_whole_program))
> + {
> + const char *const name = node_in_pr->dump_asm_name ();
> + log ("%s is externally visible... skipping\n", name);
> + return false;
> + }
> +
> + /* Assure that all paths to read are not
> + * used as function pointers. */
> + if (node_in_pr->address_taken)
> + {
> + const char *const name = node_in_pr->dump_asm_name ();
> + log ("%s might be function pointer... skipping\n", name);
> + return false;
> + }
> + }
> +
> + cgraph_node *caller = main;
> + cgraph_node *node_in_pw;
> + FOR_EACH_VEC_ELT (pw, i, node_in_pw)
> + {
> + gcc_assert (w_cnode != r_cnode);
> + if (node_in_pw == r_cnode && path_exists (r_cnode, w_cnode))
> + {
> + return call_before_read (write, read);
> + }
> +
> + if (node_in_pw == w_cnode && path_exists (w_cnode, r_cnode))
> + {
> + return write_before_call (write, read);
> + }
> +
> + if (node_in_pw == main)
> + {
> + continue;
> + }
> +
> + const char *const cname = caller->dump_asm_name ();
> + const char *const nname = node_in_pw->dump_asm_name ();
> + log ("get_nondominated_callees from %s to %s\n", cname, nname);
> +
> + bool exitEarly = false;
> + vec<cgraph_node *> non_dominated_callees
> + = get_nondominated_callees (caller, node_in_pw, &exitEarly);
> + if (exitEarly)
> + return false;
> + cgraph_node *non_dominated_callee;
> + int j;
> + FOR_EACH_VEC_ELT (non_dominated_callees, j, non_dominated_callee)
> + {
> + const char *const ndname = non_dominated_callee->dump_asm_name ();
> + log ("callee %d %s\n", j, ndname);
> + if (!path_exists (non_dominated_callee, r_cnode))
> + continue;
> +
> + return false;
> + }
> +
> + caller = node_in_pw;
> + }
> + return true;
> +}
> +
> +/* Return callees accessible through bb1 */
> +static vec<struct cgraph_node *>
> +get_bb1_callees (cgraph_node *c, cgraph_node *w_cnode)
> +{
> + vec<struct cgraph_node *> callees = vNULL;
> + if (fndecl_built_in_p (c->decl))
> + return vNULL;
> +
> + const char *cname = c->dump_asm_name ();
> + log ("before get_untransformed_body %s\n", cname);
> +
> + if (!c->definition)
> + return vNULL;
> +
> + push_cfun (DECL_STRUCT_FUNCTION (c->decl));
> +
> + gcc_assert (in_lto_p);
> + c->get_untransformed_body ();
> +
> + pop_cfun ();
> +
> + function *func = DECL_STRUCT_FUNCTION (c->decl);
> + /* Get first BB (after the fake entry BB). */
> + basic_block bb = ENTRY_BLOCK_PTR_FOR_FN (func)->next_bb;
> +
> + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> +
> + if (gimple_code (stmt) != GIMPLE_CALL)
> + continue;
> +
> + tree callee_decl = gimple_call_fndecl (stmt);
> + cgraph_node *callee_node = cgraph_node::get (callee_decl);
> + if (!path_exists (callee_node, w_cnode))
> + continue;
> +
> + callees.safe_push (callee_node);
> + }
> +
> + return callees;
> +}
> +
> +/* Return true if n2 is reachable from n1 by visiting only first basic
> blocks */
> +static bool
> +reach_nodes_via_bb1_dfs (hash_set<cgraph_node *> *visited, cgraph_node *n1,
> + cgraph_node *n2)
> +{
> + if (n1 == n2)
> + return true;
> +
> + visited->add (n1);
> +
> + vec<struct cgraph_node *> callees = get_bb1_callees (n1, n2);
> +
> + int i;
> + cgraph_node *callee;
> + FOR_EACH_VEC_ELT (callees, i, callee)
> + {
> + if (!visited->contains (callee))
> + {
> + bool found = reach_nodes_via_bb1_dfs (visited, callee, n2);
> + if (found)
> + {
> + callees.release ();
> + return true;
> + }
> + }
> + }
> +
> + return false;
> +}
> +
> +/* Returns true if n2 is reachable from main function via following
> call graph
> + * nodes reachable within first basic blocks */
> +bool
> +reach_nodes_via_bb1 (cgraph_node *n2)
> +{
> + vec<cgraph_node *> pr = vNULL;
> + cgraph_node *main = get_path_from_main_to (n2, &pr);
> + hash_set<cgraph_node *> visited;
> + bool path_exists = reach_nodes_via_bb1_dfs (&visited, main, n2);
> + visited.empty ();
> + pr.release ();
> + return path_exists;
> +}
> +
> +/* Determine if write occurs before a function call which contains read*/
> +static bool
> +write_before_call (struct ipa_ref *write, struct ipa_ref *read)
> +{
> + symtab_node *w_node = write->referring;
> + cgraph_node *w_cnode = dyn_cast<cgraph_node *> (w_node);
> + symtab_node *r_node = read->referring;
> + cgraph_node *r_cnode = dyn_cast<cgraph_node *> (r_node);
> +
> + gcc_assert (path_exists (w_cnode, r_cnode));
> + gcc_assert (w_cnode != r_cnode);
> +
> + basic_block w_bb = gimple_bb (write->stmt);
> + basic_block r_bb = gimple_bb (read->stmt);
> +
> + gcc_assert (in_lto_p);
> + w_cnode->get_untransformed_body ();
> +
> + function *func = DECL_STRUCT_FUNCTION (w_cnode->decl);
> + basic_block c_bb;
> + vec<struct cgraph_node *> callees = vNULL;
> + FOR_EACH_BB_FN (c_bb, func)
> + {
> + bool self_dominates = w_bb == c_bb;
> + bool w_bb_dominates_c_bb = bb1_dominates_bb2 (w_bb, c_bb, w_cnode);
> + if (w_bb_dominates_c_bb && !self_dominates)
> + continue;
> +
> + for (gimple_stmt_iterator gsi = gsi_start_bb (c_bb); !gsi_end_p
> (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> +
> + if (stmt == write->stmt)
> + {
> + break;
> + }
> +
> + if (gimple_code (stmt) != GIMPLE_CALL)
> + {
> + continue;
> + }
> + if (dump_file)
> + print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
> + tree rhs = gimple_op (stmt, 1);
> +
> + if (rhs && TREE_CODE (rhs) == POINTER_TYPE)
> + return false;
> +
> + tree callee_decl = gimple_call_fndecl (stmt);
> + if (!callee_decl)
> + return false;
> +
> + cgraph_node *callee_node = cgraph_node::get (callee_decl);
> + const char *identifier
> + = IDENTIFIER_POINTER (DECL_NAME (callee_node->decl));
> + const char *sigsetjmp = "sigsetjmp";
> + if (strstr (identifier, sigsetjmp) != NULL)
> + return false;
> +
> + if (fndecl_built_in_p (callee_node->decl))
> + continue;
> +
> + log ("found callee %s\n", callee_node->dump_asm_name ());
> + callees.safe_push (callee_node);
> + }
> + }
> + cgraph_node *callee;
> + int i;
> + FOR_EACH_VEC_ELT (callees, i, callee)
> + {
> + if (path_exists (callee, r_cnode))
> + {
> + return false;
> + }
> + }
> +
> + return true;
> +}
> +
> +/* Determine if call which contains write occurs before a read*/
> +static bool
> +call_before_read (struct ipa_ref *write, struct ipa_ref *read)
> +{
> + symtab_node *w_node = write->referring;
> + cgraph_node *w_cnode = dyn_cast<cgraph_node *> (w_node);
> + symtab_node *r_node = read->referring;
> + cgraph_node *r_cnode = dyn_cast<cgraph_node *> (r_node);
> +
> + gcc_assert (path_exists (r_cnode, w_cnode));
> + gcc_assert (w_cnode != r_cnode);
> +
> + basic_block w_bb = gimple_bb (write->stmt);
> + basic_block r_bb = gimple_bb (read->stmt);
> +
> + gcc_assert (in_lto_p);
> + r_cnode->get_untransformed_body ();
> +
> + function *func = DECL_STRUCT_FUNCTION (r_cnode->decl);
> + basic_block c_bb;
> + FOR_EACH_BB_FN (c_bb, func)
> + {
> + bool self_dominates = c_bb == r_bb;
> + bool call_dominates_read = bb1_dominates_bb2 (c_bb, r_bb, r_cnode);
> + if (!call_dominates_read && !self_dominates)
> + continue;
> +
> + for (gimple_stmt_iterator gsi = gsi_start_bb (c_bb); !gsi_end_p
> (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> +
> + // self dominance
> + if (stmt == read->stmt)
> + break;
> +
> + if (gimple_code (stmt) != GIMPLE_CALL)
> + continue;
> +
> + if (dump_file)
> + print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
> + tree rhs = gimple_op (stmt, 1);
> + if (TREE_CODE (rhs) == POINTER_TYPE)
> + return false;
> + tree callee_decl = gimple_call_fndecl (stmt);
> +
> + cgraph_node *callee_node = cgraph_node::get (callee_decl);
> + const char *identifier
> + = IDENTIFIER_POINTER (DECL_NAME (callee_node->decl));
> + const char *sigsetjmp = "sigsetjmp";
> + if (strstr (identifier, sigsetjmp) != NULL)
> + return false;
> +
> + if (path_exists (callee_node, w_cnode))
> + return true;
> + }
> + }
> + return false;
> +}
> +
> +/* Determine if write occurs before a read*/
> +static bool
> +write_before_read (struct ipa_ref *write, struct ipa_ref *read)
> +{
> + symtab_node *w_node = write->referring;
> + cgraph_node *w_cnode = dyn_cast<cgraph_node *> (w_node);
> + symtab_node *r_node = read->referring;
> + cgraph_node *r_cnode = dyn_cast<cgraph_node *> (r_node);
> +
> + if (w_cnode == r_cnode)
> + /* Within the same function. */
> + return write_before_read_in_function (write, read);
> +
> + /* Not within the same function. */
> + return write_before_read_across_function_borders (write, read);
> +}
> +
> +/* Determine if write occurs before all reads*/
> +static bool
> +write_before_all_reads (struct ipa_ref *write, const ipa_ref_vec &reads)
> +{
> + int i;
> + struct ipa_ref *read;
> +
> + FOR_EACH_VEC_ELT (reads, i, read)
> + {
> + load_function_body_of_ipa_ref (read);
> + if (!write_before_read (write, read))
> + return false;
> + }
> + return true;
> +}
> +
> +static void
> +propagate_constant_to_read (tree write_val, struct ipa_ref *ref,
> + hash_set<cgraph_node *> *func)
> +{
> + gcc_assert (ref->use == IPA_REF_LOAD);
> + load_function_body_of_ipa_ref (ref);
> +
> + gimple *read_stmt = ref->stmt;
> +
> + gcc_assert (gimple_code (read_stmt) == GIMPLE_ASSIGN);
> + gcc_assert (gimple_num_ops (read_stmt) == 2);
> +
> + tree old_lhs = gimple_op (read_stmt, 0);
> + symtab_node *r_node = ref->referring;
> + cgraph_node *r_cnode = dyn_cast<cgraph_node *> (r_node);
> +
> + cgraph_node **possible_clone = func_to_clone->get (r_cnode);
> + if (!possible_clone)
> + {
> + static const char *const suffix = INITCALL_CP_SUFFIX;
> + // callers has to be vNULL, otherwise, we will be
> + // analyzing clones...
> + // and we don't want that...
> + // but that means that we will need to update the callers
> + // later... in update_callgraph
> + cgraph_node *clone
> + = r_cnode->create_version_clone_with_body (vNULL, NULL, NULL, NULL,
> + NULL, suffix, NULL);
> + clone->get_untransformed_body ();
> + func_to_clone->put (r_cnode, clone);
> + }
> +
> + possible_clone = func_to_clone->get (r_cnode);
> + cgraph_node *clone = *possible_clone;
> +
> + // Build new stmt and replace old.
> + gimple_stmt_iterator gsi;
> + basic_block bb;
> + // Let's create a clone with body here...
> + // The clone should not have callers as
> + // to not interfere with the current
> + // analysis.
> + // The callers will be set at the end...
> +
> + push_cfun (DECL_STRUCT_FUNCTION (clone->decl));
> + function *clone_func = DECL_STRUCT_FUNCTION (clone->decl);
> + bool found = false;
> + FOR_EACH_BB_FN (bb, clone_func)
> + {
> + for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> + if (gimple_code (stmt) != GIMPLE_ASSIGN)
> + continue;
> + tree old_val_clone = gimple_op (stmt, 1);
> + tree lhs = gimple_op (stmt, 0);
> +
> + if (TREE_CODE (old_val_clone) != VAR_DECL)
> + continue;
> +
> + tree old_val = gimple_op (read_stmt, 1);
> + if (IDENTIFIER_POINTER (DECL_NAME (old_val_clone))
> + != IDENTIFIER_POINTER (DECL_NAME (old_val)))
> + continue;
> +
> + found = true;
> +
> + gimple_stmt_iterator gsi2 = gsi_for_stmt (stmt);
> + tree new_lhs = make_ssa_name (TREE_TYPE (lhs));
> + gimple *new_read_stmt = gimple_build_assign (new_lhs, write_val);
> + gsi_insert_before (&gsi2, new_read_stmt, GSI_SAME_STMT);
> +
> + gimple *use_stmt;
> + imm_use_iterator use_iter;
> + FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
> + {
> + use_operand_p use_p;
> + FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter)
> + SET_USE (use_p, new_lhs);
> + update_stmt (use_stmt);
> + }
> + }
> +
> + if (found)
> + break;
> + }
> + gcc_assert (found);
> + pop_cfun ();
> +}
> +
> +static bool
> +ipa_initcall_get_writes_and_reads (varpool_node *vnode, ipa_ref_vec
> *writes,
> + ipa_ref_vec *reads)
> +{
> + int i;
> + struct ipa_ref *ref;
> +
> + log ("%s for variable '%s'.\n", __func__, vnode->name ());
> +
> + /* Only IPA_REF_STORE and IPA_REF_LOAD left. */
> + for (i = 0; vnode->iterate_referring (i, ref); i++)
> + {
> + symtab_node *f_node = ref->referring;
> + cgraph_node *f_cnode = dyn_cast<cgraph_node *> (f_node);
> +
> + if (!f_cnode)
> + {
> + log ("skipping variable %s due to static initialization\n",
> + vnode->name ());
> + return false;
> + }
> + // it is possible that f_cnode is NULL if the dyn_cast fails.
> + // If the dyn_cast fails, this is an example of static
> initialization.
> + const char *identifier = IDENTIFIER_POINTER (DECL_NAME
> (f_cnode->decl));
> + static const char *const suffix = INITCALL_CP_SUFFIX;
> + if (strstr (identifier, suffix) != NULL)
> + continue;
> +
> + if (ref->use == IPA_REF_STORE)
> + writes->safe_push (ref);
> +
> + if (ref->use == IPA_REF_LOAD)
> + reads->safe_push (ref);
> + }
> + return true;
> +}
> +
> +static void
> +propagate_constant_to_reads (tree write_val, const ipa_ref_vec
> &reads_original,
> + hash_set<cgraph_node *> *funcs)
> +{
> + /* Iterate over reads and replace them by constant. */
> + struct ipa_ref *ref;
> + int i;
> + FOR_EACH_VEC_ELT (reads_original, i, ref)
> + {
> + propagate_constant_to_read (write_val, ref, funcs);
> + }
> +}
> +
> +/*
> + * Extracts the assigned constant, iff the given statement
> + * is a constant assignment. Returns NULL_TREE otherwise.
> + */
> +static tree
> +extract_constant_from_initcall_write (struct ipa_ref *write)
> +{
> + symtab_node *f_node = write->referring;
> + cgraph_node *f_cnode = dyn_cast<cgraph_node *> (f_node);
> +
> + tree decl = f_cnode->decl;
> + if (TREE_CODE (decl) != FUNCTION_DECL)
> + {
> + log ("Not function decl -> skipping.\n");
> + return NULL_TREE;
> + }
> +
> + if (!f_cnode->has_gimple_body_p ())
> + log ("Does NOT have gimple body!\n");
> + if (f_cnode->inlined_to)
> + log ("Inlined To\n");
> +
> + log ("%s: for writer gimple:\n", __func__);
> + dump_vnode_ipa_ref (write);
> +
> + /* Get the write function's body. */
> + load_function_body_of_ipa_ref (write);
> +
> + gimple *stmt = write->stmt;
> +
> + /* Verify that we have an assignment. */
> + if (gimple_code (stmt) != GIMPLE_ASSIGN)
> + {
> + log ("Write stmt not assignment -> skipping.\n");
> + return NULL_TREE;
> + }
> +
> + /* Check if write's LHS (vnode) is not volatile. */
> + tree lhs = gimple_assign_lhs (stmt);
> + if (TREE_THIS_VOLATILE (TREE_TYPE (lhs)))
> + {
> + log ("Variable volatile -> skipping.\n");
> + return NULL_TREE;
> + }
> +
> + /* Check if RHS of write stmt is constant. */
> + if (!gimple_assign_is_single_const (stmt))
> + {
> + log ("Found non-const operands.\n");
> + return NULL_TREE;
> + }
> +
> + /* Extract the constant. */
> + tree write_val = gimple_op (stmt, 1);
> +
> + if (dump_file)
> + {
> + log ("Const op:\n");
> + dump_node (write_val, TDF_DETAILS, dump_file);
> + }
> +
> + return write_val;
> +}
> +
> +static void
> +ipa_initcall_cp_execute_for_var (varpool_node *vnode,
> + hash_set<cgraph_node *> *update_functions)
> +{
> + ipa_ref_vec writes = vNULL;
> + ipa_ref_vec reads = vNULL;
> + struct ipa_ref *write;
> + tree write_val;
> + gimple_stmt_iterator gsi;
> + bool remove_permanently;
> +
> + log ("%s for variable '%s'.\n", __func__, vnode->name ());
> +
> + if (dump_file)
> + {
> + dump_node (vnode->decl, TDF_DETAILS, dump_file);
> + }
> +
> + /* Variable must not be externally visible. */
> + if (vnode->externally_visible_p ())
> + {
> + log ("\texternally visible -> skipping\n");
> + return;
> + }
> +
> + /* All refs must be explicit. */
> + if (!vnode->all_refs_explicit_p ())
> + {
> + log ("Not explicit variable refs -> skipping.\n");
> + return;
> + }
> +
> + /* Check if any ref->use is IPA_REF_ALIAS. */
> + if (vnode->has_aliases_p ())
> + {
> + log ("Found IPA_REF_ALIAS -> skipping.\n");
> + return;
> + }
> +
> + /* Check if any ref->use is IPA_REF_ADDR. */
> + if (vnode->alias || vnode->address_matters_p ())
> + {
> + log ("Found IPA_REF_ADDR -> skipping.\n");
> + return;
> + }
> +
> + /* We don't touch arrays. */
> + if (TREE_CODE (TREE_TYPE (vnode->decl)) == ARRAY_TYPE)
> + {
> + log ("Variable is array -> skipping.\n");
> + return;
> + }
> +
> + /* We don't touch structs. */
> + if (TREE_CODE (TREE_TYPE (vnode->decl)) == RECORD_TYPE)
> + {
> + log ("Variable is struct -> skipping.\n");
> + return;
> + }
> +
> + /* We don't touch unions. */
> + if (TREE_CODE (TREE_TYPE (vnode->decl)) == UNION_TYPE)
> + {
> + log ("Variable is union -> skipping.\n");
> + return;
> + }
> +
> + /* Get all refs (must be REF_STORE or REF_LOAD). */
> + bool continue_work
> + = ipa_initcall_get_writes_and_reads (vnode, &writes, &reads);
> + if (!continue_work)
> + {
> + log ("Static initialization -> skipping.\n");
> + writes.release ();
> + reads.release ();
> + return;
> + }
> +
> + if (writes.length () > 1)
> + {
> + /* More than one writer. */
> + log ("More than one writer -> skipping.\n");
IPA analysis already computes whether all accesses to a variable
happen in one function, I suppose computing whether all
writes happen in one function would be easy as well.
> + writes.release ();
> + reads.release ();
> + return;
> + }
> + else if (writes.length () < 1)
> + {
> + /* No writer. */
> + log ("No writer -> skipping.\n");
> + writes.release ();
> + reads.release ();
> + return;
> + }
> +
> + /*
> + * Note:
> + * Limiting ourselves to only one write is not necessary in general,
> + * but good enough to address the typical init () case.
> + * Big advantage is, that it makes the following code much easier.
> + */
> +
> + /* Get (only) write ref. */
> + write = writes.pop ();
> +
> + /* Check if write's RHS is a constant and get it. */
> + write_val = extract_constant_from_initcall_write (write);
> + if (write_val == NULL_TREE)
> + {
> + log ("Write's RHS is not constant -> skipping.\n");
> + writes.release ();
> + reads.release ();
> + return;
> + }
> +
> + /* Assure all reads are after the write. */
> + if (!write_before_all_reads (write, reads))
> + {
> + log ("Write not guaranteed to be before read -> skipping.\n");
> + writes.release ();
> + reads.release ();
> + return;
> + }
> +
> + /* Propagate constant to reads. */
> + if (!flag_initcall_cp_dryrun)
> + propagate_constant_to_reads (write_val, reads, update_functions);
> +
> + /* Finally remove the write. */
> + gsi = gsi_for_stmt (write->stmt);
> + remove_permanently = false; // XXX: fails with true?
> + // gsi_remove (&gsi, remove_permanently);
> +
> + log ("Eliminated variable '%s'.\n\n", vnode->name ());
> +
> + writes.release ();
> + reads.release ();
> +}
> +
> +/* Replace functions with clones */
> +bool
> +update_callgraph (cgraph_node *const &r_cnode, cgraph_node **clone_ptr,
> void *)
> +{
> + vec<cgraph_edge *> callers = r_cnode->collect_callers ();
> + cgraph_node *clone = *clone_ptr;
> + cgraph_edge *e;
> + int i;
> + profile_count count = r_cnode->count;
> +
> + FOR_EACH_VEC_ELT (callers, i, e)
> + e->redirect_callee (clone);
> +
> + for (e = clone->callers; e; e = e->next_caller)
> + {
> + e->caller->get_untransformed_body ();
> + function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl);
> + gimple_call_set_fndecl (e->call_stmt, clone->decl);
> + maybe_clean_eh_stmt_fn (inner_function, e->call_stmt);
> + }
> +
> + r_cnode->skip_ipa_cp = true;
> + push_cfun (DECL_STRUCT_FUNCTION (r_cnode->decl));
> +
> + log ("dumping function %s\n", r_cnode->dump_asm_name ());
> + function *func = DECL_STRUCT_FUNCTION (r_cnode->decl);
> + basic_block bb;
> + FOR_EACH_BB_FN (bb, func)
> + {
> + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
> + gsi_next (&gsi))
> + {
> + gimple *stmt = gsi_stmt (gsi);
> + if (dump_file)
> + print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
> + }
> + }
> + if (dom_info_available_p (CDI_DOMINATORS))
> + free_dominance_info (CDI_DOMINATORS);
> + pop_cfun ();
> + return true;
> +}
> +
> +static unsigned int
> +ipa_initcall_cp_execute (void)
> +{
> + varpool_node *vnode;
> +
> + clone_num_suffixes1 = new hash_map<const char *, unsigned>;
> + hash_set<cgraph_node *> update_functions;
> + func_to_clone = new hash_map<cgraph_node *, cgraph_node *>;
> + FOR_EACH_VARIABLE (vnode)
> + {
> + ipa_initcall_cp_execute_for_var (vnode, &update_functions);
> + }
> +
> + if (!flag_initcall_cp_dryrun)
> + func_to_clone->traverse<void *, update_callgraph> (NULL);
> +
> + delete clone_num_suffixes1;
> + delete func_to_clone;
> +
> + return 0;
> +}
> +
> +namespace {
> +
> +const pass_data pass_data_ipa_initcall_cp = {
> + SIMPLE_IPA_PASS,
> + "initcall-cp",
> + OPTGROUP_NONE,
> + TV_NONE,
> + (PROP_cfg | PROP_ssa),
> + 0,
> + 0,
> + 0,
> + (TODO_update_ssa | TODO_cleanup_cfg | TODO_dump_symtab
> + | TODO_rebuild_cgraph_edges | TODO_discard_function),
> +};
> +
> +class pass_ipa_initcall_cp : public ipa_opt_pass_d
> +{
> +public:
> + pass_ipa_initcall_cp (gcc::context *ctxt)
> + : ipa_opt_pass_d (pass_data_ipa_initcall_cp, ctxt, NULL, NULL,
> NULL, NULL,
> + NULL, NULL, 0, NULL, NULL)
> + {}
> +
> + /* opt_pass methods: */
> + virtual bool gate (function *)
> + {
> + return in_lto_p && (flag_initcall_cp || flag_initcall_cp_dryrun);
> + }
> +
> + virtual unsigned int execute (function *)
> + {
> + return ipa_initcall_cp_execute ();
> + }
> +
> +}; // class pass_ipa_initcall_cp
> +
> +} // namespace
> +
> +ipa_opt_pass_d *
> +make_pass_ipa_initcall_cp (gcc::context *ctxt)
> +{
> + return new pass_ipa_initcall_cp (ctxt);
> +}
> diff --git a/gcc/passes.def b/gcc/passes.def
> index 2bf2cb78fc5..62609951bac 100644
> --- a/gcc/passes.def
> +++ b/gcc/passes.def
> @@ -149,6 +149,7 @@ along with GCC; see the file COPYING3. If not see
> NEXT_PASS (pass_ipa_profile);
> NEXT_PASS (pass_ipa_icf);
> NEXT_PASS (pass_ipa_devirt);
> + NEXT_PASS (pass_ipa_initcall_cp);
> NEXT_PASS (pass_ipa_cp);
> NEXT_PASS (pass_ipa_sra);
> NEXT_PASS (pass_ipa_cdtor_merge);
> diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
> index a1207a20a3c..65e09c657b7 100644
> --- a/gcc/tree-pass.h
> +++ b/gcc/tree-pass.h
> @@ -501,6 +501,7 @@ extern ipa_opt_pass_d *make_pass_ipa_fn_summary
> (gcc::context *ctxt);
> extern ipa_opt_pass_d *make_pass_ipa_inline (gcc::context *ctxt);
> extern simple_ipa_opt_pass *make_pass_ipa_free_lang_data (gcc::context
> *ctxt);
> extern simple_ipa_opt_pass *make_pass_ipa_free_fn_summary
> (gcc::context *ctxt);
> +extern ipa_opt_pass_d *make_pass_ipa_initcall_cp (gcc::context *ctxt);
> extern ipa_opt_pass_d *make_pass_ipa_cp (gcc::context *ctxt);
> extern ipa_opt_pass_d *make_pass_ipa_sra (gcc::context *ctxt);
> extern ipa_opt_pass_d *make_pass_ipa_icf (gcc::context *ctxt);
> --
> 2.18.1
>
