/* Transformations based on profile information for values.
- Copyright (C) 2003, 2004 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006, 2007 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 2, or (at your option) any later
+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
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "optabs.h"
#include "regs.h"
#include "ggc.h"
+#include "tree-flow.h"
+#include "tree-flow-inline.h"
+#include "diagnostic.h"
+#include "coverage.h"
+#include "tree.h"
+#include "gcov-io.h"
+#include "cgraph.h"
+#include "timevar.h"
+#include "tree-pass.h"
+#include "toplev.h"
+#include "pointer-set.h"
static struct value_prof_hooks *value_prof_hooks;
following optimizations are implemented (for more detailed descriptions
see comments at value_profile_transformations):
- 1) Division/modulo specialisation. Provided that we can determine that the
+ 1) Division/modulo specialization. Provided that we can determine that the
operands of the division have some special properties, we may use it to
produce more effective code.
2) Speculative prefetching. If we are able to determine that the difference
between addresses accessed by a memory reference is usually constant, we
may add the prefetch instructions.
+ FIXME: This transformation was removed together with RTL based value
+ profiling.
+
+ 3) Indirect/virtual call specialization. If we can determine most
+ common function callee in indirect/virtual call. We can use this
+ information to improve code effectiveness (especially info for
+ inliner).
Every such optimization should add its requirements for profiled values to
insn_values_to_profile function. This function is called from branch_prob
in profile.c and the requested values are instrumented by it in the first
compilation with -fprofile-arcs. The optimization may then read the
gathered data in the second compilation with -fbranch-probabilities.
- The measured data is appended as REG_VALUE_PROFILE note to the instrumented
- insn. The argument to the note consists of an EXPR_LIST where its
- members have the following meaning (from the first to the last):
-
- -- type of information gathered (HIST_TYPE*)
- -- the expression that is profiled
- -- list of counters starting from the first one. */
-
-/* For speculative prefetching, the range in that we do not prefetch (because
- we assume that it will be in cache anyway). The assymetry between min and
- max range is trying to reflect the fact that the sequential prefetching
- of the data is commonly done directly by hardware. Nevertheless, these
- values are just a guess and should of course be target-specific. */
-
-#ifndef NOPREFETCH_RANGE_MIN
-#define NOPREFETCH_RANGE_MIN (-16)
-#endif
-#ifndef NOPREFETCH_RANGE_MAX
-#define NOPREFETCH_RANGE_MAX 32
-#endif
-static void insn_divmod_values_to_profile (rtx, histogram_values *);
-#ifdef HAVE_prefetch
-static bool insn_prefetch_values_to_profile (rtx, histogram_values *);
-static int find_mem_reference_1 (rtx *, void *);
-static void find_mem_reference_2 (rtx, rtx, void *);
-static bool find_mem_reference (rtx, rtx *, int *);
-#endif
-
-static void insn_values_to_profile (rtx, histogram_values *);
-static rtx gen_divmod_fixed_value (enum machine_mode, enum rtx_code, rtx, rtx,
- rtx, gcov_type, int);
-static rtx gen_mod_pow2 (enum machine_mode, enum rtx_code, rtx, rtx, rtx, int);
-static rtx gen_mod_subtract (enum machine_mode, enum rtx_code, rtx, rtx, rtx,
- int, int, int);
-#ifdef HAVE_prefetch
-static rtx gen_speculative_prefetch (rtx, gcov_type, int);
-#endif
-static bool divmod_fixed_value_transform (rtx insn);
-static bool mod_pow2_value_transform (rtx);
-static bool mod_subtract_transform (rtx);
-#ifdef HAVE_prefetch
-static bool speculative_prefetching_transform (rtx);
-#endif
-\f
-/* Find values inside INSN for that we want to measure histograms for
- division/modulo optimization and stores them to VALUES. */
-static void
-insn_divmod_values_to_profile (rtx insn, histogram_values *values)
-{
- rtx set, set_src, op1, op2;
- enum machine_mode mode;
- histogram_value hist;
+ The measured data is pointed to from the histograms
+ field of the statement annotation of the instrumented insns. It is
+ kept as a linked list of struct histogram_value_t's, which contain the
+ same information as above. */
- if (!INSN_P (insn))
- return;
- set = single_set (insn);
- if (!set)
- return;
+static tree tree_divmod_fixed_value (tree, tree, tree, tree,
+ tree, int, gcov_type, gcov_type);
+static tree tree_mod_pow2 (tree, tree, tree, tree, int, gcov_type, gcov_type);
+static tree tree_mod_subtract (tree, tree, tree, tree, int, int, int,
+ gcov_type, gcov_type, gcov_type);
+static bool tree_divmod_fixed_value_transform (tree);
+static bool tree_mod_pow2_value_transform (tree);
+static bool tree_mod_subtract_transform (tree);
+static bool tree_stringops_transform (block_stmt_iterator *);
+static bool tree_ic_transform (tree);
- mode = GET_MODE (SET_DEST (set));
- if (!INTEGRAL_MODE_P (mode))
- return;
+/* Allocate histogram value. */
- set_src = SET_SRC (set);
- switch (GET_CODE (set_src))
- {
- case DIV:
- case MOD:
- case UDIV:
- case UMOD:
- op1 = XEXP (set_src, 0);
- op2 = XEXP (set_src, 1);
- if (side_effects_p (op2))
- return;
-
- /* Check for a special case where the divisor is power of 2. */
- if ((GET_CODE (set_src) == UMOD) && !CONSTANT_P (op2))
- {
- hist = ggc_alloc (sizeof (*hist));
- hist->value = op2;
- hist->seq = NULL_RTX;
- hist->mode = mode;
- hist->insn = insn;
- hist->type = HIST_TYPE_POW2;
- hist->hdata.pow2.may_be_other = 1;
- VEC_safe_push (histogram_value, *values, hist);
- }
-
- /* Check whether the divisor is not in fact a constant. */
- if (!CONSTANT_P (op2))
- {
- hist = ggc_alloc (sizeof (*hist));
- hist->value = op2;
- hist->mode = mode;
- hist->seq = NULL_RTX;
- hist->insn = insn;
- hist->type = HIST_TYPE_SINGLE_VALUE;
- VEC_safe_push (histogram_value, *values, hist);
- }
+static histogram_value
+gimple_alloc_histogram_value (struct function *fun ATTRIBUTE_UNUSED,
+ enum hist_type type, tree stmt, tree value)
+{
+ histogram_value hist = (histogram_value) xcalloc (1, sizeof (*hist));
+ hist->hvalue.value = value;
+ hist->hvalue.stmt = stmt;
+ hist->type = type;
+ return hist;
+}
- /* For mod, check whether it is not often a noop (or replaceable by
- a few subtractions). */
- if (GET_CODE (set_src) == UMOD && !side_effects_p (op1))
- {
- rtx tmp;
-
- hist = ggc_alloc (sizeof (*hist));
- start_sequence ();
- tmp = simplify_gen_binary (DIV, mode, copy_rtx (op1), copy_rtx (op2));
- hist->value = force_operand (tmp, NULL_RTX);
- hist->seq = get_insns ();
- end_sequence ();
- hist->mode = mode;
- hist->insn = insn;
- hist->type = HIST_TYPE_INTERVAL;
- hist->hdata.intvl.int_start = 0;
- hist->hdata.intvl.steps = 2;
- hist->hdata.intvl.may_be_less = 1;
- hist->hdata.intvl.may_be_more = 1;
- VEC_safe_push (histogram_value, *values, hist);
- }
- return;
+/* Hash value for histogram. */
- default:
- return;
- }
+static hashval_t
+histogram_hash (const void *x)
+{
+ return htab_hash_pointer (((const_histogram_value)x)->hvalue.stmt);
}
-#ifdef HAVE_prefetch
-
-/* Called from find_mem_reference through for_each_rtx, finds a memory
- reference. I.e. if *EXPR is a MEM, the reference to this MEM is stored
- to *RET and the traversing of the expression is interrupted by returning 1.
- Otherwise 0 is returned. */
+/* Return nonzero if decl_id of die_struct X is the same as UID of decl *Y. */
static int
-find_mem_reference_1 (rtx *expr, void *ret)
+histogram_eq (const void *x, const void *y)
{
- rtx *mem = ret;
-
- if (GET_CODE (*expr) == MEM)
- {
- *mem = *expr;
- return 1;
- }
- return 0;
+ return ((const_histogram_value) x)->hvalue.stmt == (const_tree)y;
}
-/* Called form find_mem_reference through note_stores to find out whether
- the memory reference MEM is a store. I.e. if EXPR == MEM, the variable
- FMR2_WRITE is set to true. */
+/* Set histogram for STMT. */
-static int fmr2_write;
static void
-find_mem_reference_2 (rtx expr, rtx pat ATTRIBUTE_UNUSED, void *mem)
+set_histogram_value (struct function *fun, tree stmt, histogram_value hist)
{
- if (expr == mem)
- fmr2_write = true;
+ void **loc;
+ if (!hist && !VALUE_HISTOGRAMS (fun))
+ return;
+ if (!VALUE_HISTOGRAMS (fun))
+ VALUE_HISTOGRAMS (fun) = htab_create (1, histogram_hash,
+ histogram_eq, NULL);
+ loc = htab_find_slot_with_hash (VALUE_HISTOGRAMS (fun), stmt,
+ htab_hash_pointer (stmt),
+ hist ? INSERT : NO_INSERT);
+ if (!hist)
+ {
+ if (loc)
+ htab_clear_slot (VALUE_HISTOGRAMS (fun), loc);
+ return;
+ }
+ *loc = hist;
}
-/* Find a memory reference inside INSN, return it in MEM. Set WRITE to true
- if it is a write of the mem. Return false if no memory reference is found,
- true otherwise. */
+/* Get histogram list for STMT. */
-static bool
-find_mem_reference (rtx insn, rtx *mem, int *write)
+histogram_value
+gimple_histogram_value (struct function *fun, tree stmt)
{
- *mem = NULL_RTX;
- for_each_rtx (&PATTERN (insn), find_mem_reference_1, mem);
-
- if (!*mem)
- return false;
-
- fmr2_write = false;
- note_stores (PATTERN (insn), find_mem_reference_2, *mem);
- *write = fmr2_write;
- return true;
+ if (!VALUE_HISTOGRAMS (fun))
+ return NULL;
+ return htab_find_with_hash (VALUE_HISTOGRAMS (fun), stmt,
+ htab_hash_pointer (stmt));
}
-/* Find values inside INSN for that we want to measure histograms for
- a speculative prefetching. Add them to the list VALUES.
- Returns true if such we found any such value, false otherwise. */
+/* Add histogram for STMT. */
-static bool
-insn_prefetch_values_to_profile (rtx insn, histogram_values *values)
+void
+gimple_add_histogram_value (struct function *fun, tree stmt, histogram_value hist)
{
- rtx mem, address;
- int write;
- histogram_value hist;
-
- if (!INSN_P (insn))
- return false;
-
- if (!find_mem_reference (insn, &mem, &write))
- return false;
-
- address = XEXP (mem, 0);
- if (side_effects_p (address))
- return false;
-
- if (CONSTANT_P (address))
- return false;
+ hist->hvalue.next = gimple_histogram_value (fun, stmt);
+ set_histogram_value (fun, stmt, hist);
+}
- hist = ggc_alloc (sizeof (*hist));
- hist->value = address;
- hist->mode = GET_MODE (address);
- hist->seq = NULL_RTX;
- hist->insn = insn;
- hist->type = HIST_TYPE_CONST_DELTA;
- VEC_safe_push (histogram_value, *values, hist);
+/* Remove histogram HIST from STMT's histogram list. */
- return true;
-}
-#endif
-/* Find values inside INSN for that we want to measure histograms and adds
- them to list VALUES (increasing the record of its length in N_VALUES). */
-static void
-insn_values_to_profile (rtx insn, histogram_values *values)
+void
+gimple_remove_histogram_value (struct function *fun, tree stmt, histogram_value hist)
{
- if (flag_value_profile_transformations)
- insn_divmod_values_to_profile (insn, values);
-
-#ifdef HAVE_prefetch
- if (flag_speculative_prefetching)
- insn_prefetch_values_to_profile (insn, values);
+ histogram_value hist2 = gimple_histogram_value (fun, stmt);
+ if (hist == hist2)
+ {
+ set_histogram_value (fun, stmt, hist->hvalue.next);
+ }
+ else
+ {
+ while (hist2->hvalue.next != hist)
+ hist2 = hist2->hvalue.next;
+ hist2->hvalue.next = hist->hvalue.next;
+ }
+ free (hist->hvalue.counters);
+#ifdef ENABLE_CHECKING
+ memset (hist, 0xab, sizeof (*hist));
#endif
+ free (hist);
}
-/* Find list of values for that we want to measure histograms. */
-static void
-rtl_find_values_to_profile (histogram_values *values)
-{
- rtx insn;
- unsigned i;
+/* Lookup histogram of type TYPE in the STMT. */
- life_analysis (NULL, PROP_DEATH_NOTES);
+histogram_value
+gimple_histogram_value_of_type (struct function *fun, tree stmt, enum hist_type type)
+{
+ histogram_value hist;
+ for (hist = gimple_histogram_value (fun, stmt); hist; hist = hist->hvalue.next)
+ if (hist->type == type)
+ return hist;
+ return NULL;
+}
- *values = VEC_alloc (histogram_value, 0);
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- insn_values_to_profile (insn, values);
+/* Dump information about HIST to DUMP_FILE. */
- for (i = 0; i < VEC_length (histogram_value, *values); i++)
+static void
+dump_histogram_value (FILE *dump_file, histogram_value hist)
+{
+ switch (hist->type)
{
- histogram_value hist = VEC_index (histogram_value, *values, i);
+ case HIST_TYPE_INTERVAL:
+ fprintf (dump_file, "Interval counter range %d -- %d",
+ hist->hdata.intvl.int_start,
+ (hist->hdata.intvl.int_start
+ + hist->hdata.intvl.steps - 1));
+ if (hist->hvalue.counters)
+ {
+ unsigned int i;
+ fprintf(dump_file, " [");
+ for (i = 0; i < hist->hdata.intvl.steps; i++)
+ fprintf (dump_file, " %d:"HOST_WIDEST_INT_PRINT_DEC,
+ hist->hdata.intvl.int_start + i,
+ (HOST_WIDEST_INT) hist->hvalue.counters[i]);
+ fprintf (dump_file, " ] outside range:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[i]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
- switch (hist->type)
+ case HIST_TYPE_POW2:
+ fprintf (dump_file, "Pow2 counter ");
+ if (hist->hvalue.counters)
{
- case HIST_TYPE_INTERVAL:
- if (dump_file)
- fprintf (dump_file,
- "Interval counter for insn %d, range %d -- %d.\n",
- INSN_UID ((rtx)hist->insn),
- hist->hdata.intvl.int_start,
- (hist->hdata.intvl.int_start
- + hist->hdata.intvl.steps - 1));
- hist->n_counters = hist->hdata.intvl.steps +
- (hist->hdata.intvl.may_be_less ? 1 : 0) +
- (hist->hdata.intvl.may_be_more ? 1 : 0);
- break;
+ fprintf (dump_file, "pow2:"HOST_WIDEST_INT_PRINT_DEC
+ " nonpow2:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0],
+ (HOST_WIDEST_INT) hist->hvalue.counters[1]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
- case HIST_TYPE_POW2:
- if (dump_file)
- fprintf (dump_file,
- "Pow2 counter for insn %d.\n",
- INSN_UID ((rtx)hist->insn));
- hist->n_counters
- = GET_MODE_BITSIZE (hist->mode)
- + (hist->hdata.pow2.may_be_other ? 1 : 0);
- break;
+ case HIST_TYPE_SINGLE_VALUE:
+ fprintf (dump_file, "Single value ");
+ if (hist->hvalue.counters)
+ {
+ fprintf (dump_file, "value:"HOST_WIDEST_INT_PRINT_DEC
+ " match:"HOST_WIDEST_INT_PRINT_DEC
+ " wrong:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0],
+ (HOST_WIDEST_INT) hist->hvalue.counters[1],
+ (HOST_WIDEST_INT) hist->hvalue.counters[2]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
- case HIST_TYPE_SINGLE_VALUE:
- if (dump_file)
- fprintf (dump_file,
- "Single value counter for insn %d.\n",
- INSN_UID ((rtx)hist->insn));
- hist->n_counters = 3;
- break;
+ case HIST_TYPE_AVERAGE:
+ fprintf (dump_file, "Average value ");
+ if (hist->hvalue.counters)
+ {
+ fprintf (dump_file, "sum:"HOST_WIDEST_INT_PRINT_DEC
+ " times:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0],
+ (HOST_WIDEST_INT) hist->hvalue.counters[1]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
- case HIST_TYPE_CONST_DELTA:
- if (dump_file)
- fprintf (dump_file,
- "Constant delta counter for insn %d.\n",
- INSN_UID ((rtx)hist->insn));
- hist->n_counters = 4;
- break;
+ case HIST_TYPE_IOR:
+ fprintf (dump_file, "IOR value ");
+ if (hist->hvalue.counters)
+ {
+ fprintf (dump_file, "ior:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
- default:
- abort ();
+ case HIST_TYPE_CONST_DELTA:
+ fprintf (dump_file, "Constant delta ");
+ if (hist->hvalue.counters)
+ {
+ fprintf (dump_file, "value:"HOST_WIDEST_INT_PRINT_DEC
+ " match:"HOST_WIDEST_INT_PRINT_DEC
+ " wrong:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0],
+ (HOST_WIDEST_INT) hist->hvalue.counters[1],
+ (HOST_WIDEST_INT) hist->hvalue.counters[2]);
}
- }
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
+ fprintf (dump_file, ".\n");
+ break;
+ case HIST_TYPE_INDIR_CALL:
+ fprintf (dump_file, "Indirect call ");
+ if (hist->hvalue.counters)
+ {
+ fprintf (dump_file, "value:"HOST_WIDEST_INT_PRINT_DEC
+ " match:"HOST_WIDEST_INT_PRINT_DEC
+ " all:"HOST_WIDEST_INT_PRINT_DEC,
+ (HOST_WIDEST_INT) hist->hvalue.counters[0],
+ (HOST_WIDEST_INT) hist->hvalue.counters[1],
+ (HOST_WIDEST_INT) hist->hvalue.counters[2]);
+ }
+ fprintf (dump_file, ".\n");
+ break;
+ }
}
-/* Main entry point. Finds REG_VALUE_PROFILE notes from profiler and uses
- them to identify and exploit properties of values that are hard to analyze
- statically.
-
- We do following transformations:
-
- 1)
-
- x = a / b;
-
- where b is almost always a constant N is transformed to
-
- if (b == N)
- x = a / N;
- else
- x = a / b;
-
- Analogically with %
-
- 2)
+/* Dump all histograms attached to STMT to DUMP_FILE. */
- x = a % b
-
- where b is almost always a power of 2 and the division is unsigned
- TODO -- handle signed case as well
-
- if ((b & (b - 1)) == 0)
- x = a & (b - 1);
- else
- x = x % b;
-
- Note that when b = 0, no error will occur and x = a; this is correct,
- as result of such operation is undefined.
-
- 3)
-
- x = a % b
+void
+dump_histograms_for_stmt (struct function *fun, FILE *dump_file, tree stmt)
+{
+ histogram_value hist;
+ for (hist = gimple_histogram_value (fun, stmt); hist; hist = hist->hvalue.next)
+ dump_histogram_value (dump_file, hist);
+}
- where a is almost always less then b and the division is unsigned
- TODO -- handle signed case as well
+/* Remove all histograms associated with STMT. */
- x = a;
- if (x >= b)
- x %= b;
+void
+gimple_remove_stmt_histograms (struct function *fun, tree stmt)
+{
+ histogram_value val;
+ while ((val = gimple_histogram_value (fun, stmt)) != NULL)
+ gimple_remove_histogram_value (fun, stmt, val);
+}
- 4)
+/* Duplicate all histograms associates with OSTMT to STMT. */
- x = a % b
+void
+gimple_duplicate_stmt_histograms (struct function *fun, tree stmt,
+ struct function *ofun, tree ostmt)
+{
+ histogram_value val;
+ for (val = gimple_histogram_value (ofun, ostmt); val != NULL; val = val->hvalue.next)
+ {
+ histogram_value new = gimple_alloc_histogram_value (fun, val->type, NULL, NULL);
+ memcpy (new, val, sizeof (*val));
+ new->hvalue.stmt = stmt;
+ new->hvalue.counters = xmalloc (sizeof (*new->hvalue.counters) * new->n_counters);
+ memcpy (new->hvalue.counters, val->hvalue.counters, sizeof (*new->hvalue.counters) * new->n_counters);
+ gimple_add_histogram_value (fun, stmt, new);
+ }
+}
- where a is almost always less then 2 * b and the division is unsigned
- TODO -- handle signed case as well
+static bool error_found = false;
- x = a;
- if (x >= b)
- x -= b;
- if (x >= b)
- x %= b;
+/* Helper function for verify_histograms. For each histogram reachable via htab
+ walk verify that it was reached via statement walk. */
- It would be possible to continue analogically for K * b for other small
- K's, but it is probably not useful.
+static int
+visit_hist (void **slot, void *data)
+{
+ struct pointer_set_t *visited = (struct pointer_set_t *) data;
+ histogram_value hist = *(histogram_value *) slot;
+ if (!pointer_set_contains (visited, hist))
+ {
+ error ("Dead histogram");
+ dump_histogram_value (stderr, hist);
+ debug_generic_stmt (hist->hvalue.stmt);
+ error_found = true;
+ }
+ return 1;
+}
- 5)
+/* Verify sanity of the histograms. */
- Read or write of mem[address], where the value of address changes usually
- by a constant C != 0 between the following accesses to the computation; with
- -fspeculative-prefetching we then add a prefetch of address + C before
- the insn. This handles prefetching of several interesting cases in addition
- to a simple prefetching for addresses that are induction variables, e. g.
- linked lists allocated sequentially (even in case they are processed
- recursively).
+void
+verify_histograms (void)
+{
+ basic_block bb;
+ block_stmt_iterator bsi;
+ histogram_value hist;
+ struct pointer_set_t *visited_hists;
+
+ error_found = false;
+ visited_hists = pointer_set_create ();
+ FOR_EACH_BB (bb)
+ for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ {
+ tree stmt = bsi_stmt (bsi);
+
+ for (hist = gimple_histogram_value (cfun, stmt); hist; hist = hist->hvalue.next)
+ {
+ if (hist->hvalue.stmt != stmt)
+ {
+ error ("Histogram value statement does not correspond to statement"
+ " it is associated with");
+ debug_generic_stmt (stmt);
+ dump_histogram_value (stderr, hist);
+ error_found = true;
+ }
+ pointer_set_insert (visited_hists, hist);
+ }
+ }
+ if (VALUE_HISTOGRAMS (cfun))
+ htab_traverse (VALUE_HISTOGRAMS (cfun), visit_hist, visited_hists);
+ pointer_set_destroy (visited_hists);
+ if (error_found)
+ internal_error ("verify_histograms failed");
+}
- TODO -- we should also check whether there is not (usually) a small
- difference with the adjacent memory references, so that we do
- not issue overlapping prefetches. Also we should employ some
- heuristics to eliminate cases where prefetching evidently spoils
- the code.
- -- it should somehow cooperate with the loop optimizer prefetching
+/* Helper function for verify_histograms. For each histogram reachable via htab
+ walk verify that it was reached via statement walk. */
- TODO:
+static int
+free_hist (void **slot, void *data ATTRIBUTE_UNUSED)
+{
+ histogram_value hist = *(histogram_value *) slot;
+ free (hist->hvalue.counters);
+#ifdef ENABLE_CHECKING
+ memset (hist, 0xab, sizeof (*hist));
+#endif
+ free (hist);
+ return 1;
+}
- There are other useful cases that could be handled by a similar mechanism,
- for example:
-
- for (i = 0; i < n; i++)
- ...
-
- transform to (for constant N):
-
- if (n == N)
- for (i = 0; i < N; i++)
- ...
- else
- for (i = 0; i < n; i++)
- ...
- making unroller happy. Since this may grow the code significantly,
- we would have to be very careful here. */
+void
+free_histograms (void)
+{
+ if (VALUE_HISTOGRAMS (cfun))
+ {
+ htab_traverse (VALUE_HISTOGRAMS (cfun), free_hist, NULL);
+ htab_delete (VALUE_HISTOGRAMS (cfun));
+ VALUE_HISTOGRAMS (cfun) = NULL;
+ }
+}
+/* The overall number of invocations of the counter should match execution count
+ of basic block. Report it as error rather than internal error as it might
+ mean that user has misused the profile somehow. */
static bool
-rtl_value_profile_transformations (void)
+check_counter (tree stmt, const char * name, gcov_type all, gcov_type bb_count)
{
- rtx insn, next;
- int changed = false;
-
- for (insn = get_insns (); insn; insn = next)
+ if (all != bb_count)
{
- next = NEXT_INSN (insn);
-
- if (!INSN_P (insn))
- continue;
-
- /* Scan for insn carrying a histogram. */
- if (!find_reg_note (insn, REG_VALUE_PROFILE, 0))
- continue;
+ location_t * locus;
+ locus = (stmt != NULL && EXPR_HAS_LOCATION (stmt)
+ ? EXPR_LOCUS (stmt)
+ : &DECL_SOURCE_LOCATION (current_function_decl));
+ error ("%HCorrupted value profile: %s profiler overall count (%d) does not match BB count (%d)",
+ locus, name, (int)all, (int)bb_count);
+ return true;
+ }
+ return false;
+}
- /* Ignore cold areas -- we are growing a code. */
- if (!maybe_hot_bb_p (BLOCK_FOR_INSN (insn)))
- continue;
+/* Tree based transformations. */
+static bool
+tree_value_profile_transformations (void)
+{
+ basic_block bb;
+ block_stmt_iterator bsi;
+ bool changed = false;
- if (dump_file)
+ FOR_EACH_BB (bb)
+ {
+ for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
{
- fprintf (dump_file, "Trying transformations on insn %d\n",
- INSN_UID (insn));
- print_rtl_single (dump_file, insn);
- }
+ tree stmt = bsi_stmt (bsi);
+ histogram_value th = gimple_histogram_value (cfun, stmt);
+ if (!th)
+ continue;
- /* Transformations: */
- if (flag_value_profile_transformations
- && (mod_subtract_transform (insn)
- || divmod_fixed_value_transform (insn)
- || mod_pow2_value_transform (insn)))
- changed = true;
-#ifdef HAVE_prefetch
- if (flag_speculative_prefetching
- && speculative_prefetching_transform (insn))
- changed = true;
-#endif
+ if (dump_file)
+ {
+ fprintf (dump_file, "Trying transformations on stmt ");
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ dump_histograms_for_stmt (cfun, dump_file, stmt);
+ }
+
+ /* Transformations: */
+ /* The order of things in this conditional controls which
+ transformation is used when more than one is applicable. */
+ /* It is expected that any code added by the transformations
+ will be added before the current statement, and that the
+ current statement remain valid (although possibly
+ modified) upon return. */
+ if (flag_value_profile_transformations
+ && (tree_mod_subtract_transform (stmt)
+ || tree_divmod_fixed_value_transform (stmt)
+ || tree_mod_pow2_value_transform (stmt)
+ || tree_stringops_transform (&bsi)
+ || tree_ic_transform (stmt)))
+ {
+ stmt = bsi_stmt (bsi);
+ changed = true;
+ /* Original statement may no longer be in the same block. */
+ if (bb != bb_for_stmt (stmt))
+ {
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+ }
+ }
+ }
}
if (changed)
{
- commit_edge_insertions ();
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
+ counts_to_freqs ();
}
return changed;
}
-/* Generate code for transformation 1 (with MODE and OPERATION, operands OP1
- and OP2, whose value is expected to be VALUE, result TARGET and
- probability of taking the optimal path PROB). */
-static rtx
-gen_divmod_fixed_value (enum machine_mode mode, enum rtx_code operation,
- rtx target, rtx op1, rtx op2, gcov_type value,
- int prob)
+/* Generate code for transformation 1 (with OPERATION, operands OP1
+ and OP2, whose value is expected to be VALUE, parent modify-expr STMT and
+ probability of taking the optimal path PROB, which is equivalent to COUNT/ALL
+ within roundoff error). This generates the result into a temp and returns
+ the temp; it does not replace or alter the original STMT. */
+static tree
+tree_divmod_fixed_value (tree stmt, tree operation,
+ tree op1, tree op2, tree value, int prob, gcov_type count,
+ gcov_type all)
{
- rtx tmp, tmp1, jump;
- rtx neq_label = gen_label_rtx ();
- rtx end_label = gen_label_rtx ();
- rtx sequence;
-
- start_sequence ();
+ tree stmt1, stmt2, stmt3;
+ tree tmp1, tmp2, tmpv;
+ tree label_decl1 = create_artificial_label ();
+ tree label_decl2 = create_artificial_label ();
+ tree label1, label2;
+ tree bb1end, bb2end, bb3end;
+ basic_block bb, bb2, bb3, bb4;
+ tree optype = TREE_TYPE (operation);
+ edge e12, e13, e23, e24, e34;
+ block_stmt_iterator bsi;
+
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+
+ tmpv = create_tmp_var (optype, "PROF");
+ tmp1 = create_tmp_var (optype, "PROF");
+ stmt1 = build_gimple_modify_stmt (tmpv, fold_convert (optype, value));
+ stmt2 = build_gimple_modify_stmt (tmp1, op2);
+ stmt3 = build3 (COND_EXPR, void_type_node,
+ build2 (NE_EXPR, boolean_type_node, tmp1, tmpv),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt3, BSI_SAME_STMT);
+ bb1end = stmt3;
+
+ tmp2 = create_tmp_var (optype, "PROF");
+ label1 = build1 (LABEL_EXPR, void_type_node, label_decl1);
+ stmt1 = build_gimple_modify_stmt (tmp2,
+ build2 (TREE_CODE (operation), optype,
+ op1, tmpv));
+ bsi_insert_before (&bsi, label1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb2end = stmt1;
+
+ label2 = build1 (LABEL_EXPR, void_type_node, label_decl2);
+ stmt1 = build_gimple_modify_stmt (tmp2,
+ build2 (TREE_CODE (operation), optype,
+ op1, op2));
+ bsi_insert_before (&bsi, label2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb3end = stmt1;
+
+ /* Fix CFG. */
+ /* Edge e23 connects bb2 to bb3, etc. */
+ e12 = split_block (bb, bb1end);
+ bb2 = e12->dest;
+ bb2->count = count;
+ e23 = split_block (bb2, bb2end);
+ bb3 = e23->dest;
+ bb3->count = all - count;
+ e34 = split_block (bb3, bb3end);
+ bb4 = e34->dest;
+ bb4->count = all;
+
+ e12->flags &= ~EDGE_FALLTHRU;
+ e12->flags |= EDGE_FALSE_VALUE;
+ e12->probability = prob;
+ e12->count = count;
+
+ e13 = make_edge (bb, bb3, EDGE_TRUE_VALUE);
+ e13->probability = REG_BR_PROB_BASE - prob;
+ e13->count = all - count;
+
+ remove_edge (e23);
- if (!REG_P (op2))
- {
- tmp = gen_reg_rtx (mode);
- emit_move_insn (tmp, copy_rtx (op2));
- }
- else
- tmp = op2;
-
- do_compare_rtx_and_jump (tmp, GEN_INT (value), NE, 0, mode, NULL_RTX,
- NULL_RTX, neq_label);
-
- /* Add branch probability to jump we just created. */
- jump = get_last_insn ();
- REG_NOTES (jump) = gen_rtx_EXPR_LIST (REG_BR_PROB,
- GEN_INT (REG_BR_PROB_BASE - prob),
- REG_NOTES (jump));
-
- tmp1 = simplify_gen_binary (operation, mode,
- copy_rtx (op1), GEN_INT (value));
- tmp1 = force_operand (tmp1, target);
- if (tmp1 != target)
- emit_move_insn (copy_rtx (target), copy_rtx (tmp1));
-
- emit_jump_insn (gen_jump (end_label));
- emit_barrier ();
-
- emit_label (neq_label);
- tmp1 = simplify_gen_binary (operation, mode,
- copy_rtx (op1), copy_rtx (tmp));
- tmp1 = force_operand (tmp1, target);
- if (tmp1 != target)
- emit_move_insn (copy_rtx (target), copy_rtx (tmp1));
-
- emit_label (end_label);
+ e24 = make_edge (bb2, bb4, EDGE_FALLTHRU);
+ e24->probability = REG_BR_PROB_BASE;
+ e24->count = count;
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
+ e34->probability = REG_BR_PROB_BASE;
+ e34->count = all - count;
+
+ return tmp2;
}
/* Do transform 1) on INSN if applicable. */
static bool
-divmod_fixed_value_transform (rtx insn)
+tree_divmod_fixed_value_transform (tree stmt)
{
- rtx set, set_src, set_dest, op1, op2, value, histogram;
- enum rtx_code code;
- enum machine_mode mode;
+ histogram_value histogram;
+ enum tree_code code;
gcov_type val, count, all;
- edge e;
+ tree modify, op, op1, op2, result, value, tree_val;
int prob;
- set = single_set (insn);
- if (!set)
+ modify = stmt;
+ if (TREE_CODE (stmt) == RETURN_EXPR
+ && TREE_OPERAND (stmt, 0)
+ && TREE_CODE (TREE_OPERAND (stmt, 0)) == GIMPLE_MODIFY_STMT)
+ modify = TREE_OPERAND (stmt, 0);
+ if (TREE_CODE (modify) != GIMPLE_MODIFY_STMT)
return false;
-
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
+ op = GIMPLE_STMT_OPERAND (modify, 1);
+ if (!INTEGRAL_TYPE_P (TREE_TYPE (op)))
+ return false;
+ code = TREE_CODE (op);
- if (code != DIV && code != MOD && code != UDIV && code != UMOD)
+ if (code != TRUNC_DIV_EXPR && code != TRUNC_MOD_EXPR)
return false;
- op1 = XEXP (set_src, false);
- op2 = XEXP (set_src, 1);
-
- for (histogram = REG_NOTES (insn);
- histogram;
- histogram = XEXP (histogram, 1))
- if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
- && XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_SINGLE_VALUE))
- break;
+ op1 = TREE_OPERAND (op, 0);
+ op2 = TREE_OPERAND (op, 1);
+
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_SINGLE_VALUE);
if (!histogram)
return false;
- histogram = XEXP (XEXP (histogram, 0), 1);
- value = XEXP (histogram, 0);
- histogram = XEXP (histogram, 1);
- val = INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- count = INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- all = INTVAL (XEXP (histogram, 0));
+ value = histogram->hvalue.value;
+ val = histogram->hvalue.counters[0];
+ count = histogram->hvalue.counters[1];
+ all = histogram->hvalue.counters[2];
+ gimple_remove_histogram_value (cfun, stmt, histogram);
/* We require that count is at least half of all; this means
that for the transformation to fire the value must be constant
at least 50% of time (and 75% gives the guarantee of usage). */
- if (!rtx_equal_p (op2, value) || 2 * count < all)
+ if (simple_cst_equal (op2, value) != 1 || 2 * count < all
+ || !maybe_hot_bb_p (bb_for_stmt (stmt)))
return false;
- if (dump_file)
- fprintf (dump_file, "Div/mod by constant transformation on insn %d\n",
- INSN_UID (insn));
+ if (check_counter (stmt, "value", all, bb_for_stmt (stmt)->count))
+ return false;
/* Compute probability of taking the optimal path. */
prob = (count * REG_BR_PROB_BASE + all / 2) / all;
- e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
- delete_insn (insn);
-
- insert_insn_on_edge (
- gen_divmod_fixed_value (mode, code, set_dest,
- op1, op2, val, prob), e);
+ tree_val = build_int_cst_wide (get_gcov_type (),
+ (unsigned HOST_WIDE_INT) val,
+ val >> (HOST_BITS_PER_WIDE_INT - 1) >> 1);
+ result = tree_divmod_fixed_value (stmt, op, op1, op2, tree_val, prob, count, all);
+
+ if (dump_file)
+ {
+ fprintf (dump_file, "Div/mod by constant ");
+ print_generic_expr (dump_file, value, TDF_SLIM);
+ fprintf (dump_file, "=");
+ print_generic_expr (dump_file, tree_val, TDF_SLIM);
+ fprintf (dump_file, " transformation on insn ");
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ }
+
+ GIMPLE_STMT_OPERAND (modify, 1) = result;
return true;
}
-/* Generate code for transformation 2 (with MODE and OPERATION, operands OP1
- and OP2, result TARGET and probability of taking the optimal path PROB). */
-static rtx
-gen_mod_pow2 (enum machine_mode mode, enum rtx_code operation, rtx target,
- rtx op1, rtx op2, int prob)
+/* Generate code for transformation 2 (with OPERATION, operands OP1
+ and OP2, parent modify-expr STMT and probability of taking the optimal
+ path PROB, which is equivalent to COUNT/ALL within roundoff error).
+ This generates the result into a temp and returns
+ the temp; it does not replace or alter the original STMT. */
+static tree
+tree_mod_pow2 (tree stmt, tree operation, tree op1, tree op2, int prob,
+ gcov_type count, gcov_type all)
{
- rtx tmp, tmp1, tmp2, tmp3, jump;
- rtx neq_label = gen_label_rtx ();
- rtx end_label = gen_label_rtx ();
- rtx sequence;
-
- start_sequence ();
-
- if (!REG_P (op2))
- {
- tmp = gen_reg_rtx (mode);
- emit_move_insn (tmp, copy_rtx (op2));
- }
- else
- tmp = op2;
-
- tmp1 = expand_simple_binop (mode, PLUS, tmp, constm1_rtx, NULL_RTX,
- 0, OPTAB_WIDEN);
- tmp2 = expand_simple_binop (mode, AND, tmp, tmp1, NULL_RTX,
- 0, OPTAB_WIDEN);
- do_compare_rtx_and_jump (tmp2, const0_rtx, NE, 0, mode, NULL_RTX,
- NULL_RTX, neq_label);
-
- /* Add branch probability to jump we just created. */
- jump = get_last_insn ();
- REG_NOTES (jump) = gen_rtx_EXPR_LIST (REG_BR_PROB,
- GEN_INT (REG_BR_PROB_BASE - prob),
- REG_NOTES (jump));
-
- tmp3 = expand_simple_binop (mode, AND, op1, tmp1, target,
- 0, OPTAB_WIDEN);
- if (tmp3 != target)
- emit_move_insn (copy_rtx (target), tmp3);
- emit_jump_insn (gen_jump (end_label));
- emit_barrier ();
-
- emit_label (neq_label);
- tmp1 = simplify_gen_binary (operation, mode, copy_rtx (op1), copy_rtx (tmp));
- tmp1 = force_operand (tmp1, target);
- if (tmp1 != target)
- emit_move_insn (target, tmp1);
+ tree stmt1, stmt2, stmt3, stmt4;
+ tree tmp2, tmp3;
+ tree label_decl1 = create_artificial_label ();
+ tree label_decl2 = create_artificial_label ();
+ tree label1, label2;
+ tree bb1end, bb2end, bb3end;
+ basic_block bb, bb2, bb3, bb4;
+ tree optype = TREE_TYPE (operation);
+ edge e12, e13, e23, e24, e34;
+ block_stmt_iterator bsi;
+ tree result = create_tmp_var (optype, "PROF");
+
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+
+ tmp2 = create_tmp_var (optype, "PROF");
+ tmp3 = create_tmp_var (optype, "PROF");
+ stmt2 = build_gimple_modify_stmt (tmp2,
+ build2 (PLUS_EXPR, optype, op2,
+ build_int_cst (optype, -1)));
+ stmt3 = build_gimple_modify_stmt (tmp3,
+ build2 (BIT_AND_EXPR, optype, tmp2, op2));
+ stmt4 = build3 (COND_EXPR, void_type_node,
+ build2 (NE_EXPR, boolean_type_node,
+ tmp3, build_int_cst (optype, 0)),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt3, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt4, BSI_SAME_STMT);
+ bb1end = stmt4;
+
+ /* tmp2 == op2-1 inherited from previous block */
+ label1 = build1 (LABEL_EXPR, void_type_node, label_decl1);
+ stmt1 = build_gimple_modify_stmt (result,
+ build2 (BIT_AND_EXPR, optype, op1, tmp2));
+ bsi_insert_before (&bsi, label1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb2end = stmt1;
+
+ label2 = build1 (LABEL_EXPR, void_type_node, label_decl2);
+ stmt1 = build_gimple_modify_stmt (result,
+ build2 (TREE_CODE (operation), optype,
+ op1, op2));
+ bsi_insert_before (&bsi, label2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb3end = stmt1;
+
+ /* Fix CFG. */
+ /* Edge e23 connects bb2 to bb3, etc. */
+ e12 = split_block (bb, bb1end);
+ bb2 = e12->dest;
+ bb2->count = count;
+ e23 = split_block (bb2, bb2end);
+ bb3 = e23->dest;
+ bb3->count = all - count;
+ e34 = split_block (bb3, bb3end);
+ bb4 = e34->dest;
+ bb4->count = all;
+
+ e12->flags &= ~EDGE_FALLTHRU;
+ e12->flags |= EDGE_FALSE_VALUE;
+ e12->probability = prob;
+ e12->count = count;
+
+ e13 = make_edge (bb, bb3, EDGE_TRUE_VALUE);
+ e13->probability = REG_BR_PROB_BASE - prob;
+ e13->count = all - count;
+
+ remove_edge (e23);
- emit_label (end_label);
+ e24 = make_edge (bb2, bb4, EDGE_FALLTHRU);
+ e24->probability = REG_BR_PROB_BASE;
+ e24->count = count;
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
+ e34->probability = REG_BR_PROB_BASE;
+ e34->count = all - count;
+
+ return result;
}
/* Do transform 2) on INSN if applicable. */
static bool
-mod_pow2_value_transform (rtx insn)
+tree_mod_pow2_value_transform (tree stmt)
{
- rtx set, set_src, set_dest, op1, op2, value, histogram;
- enum rtx_code code;
- enum machine_mode mode;
- gcov_type wrong_values, count;
- edge e;
- int i, all, prob;
-
- set = single_set (insn);
- if (!set)
- return false;
+ histogram_value histogram;
+ enum tree_code code;
+ gcov_type count, wrong_values, all;
+ tree modify, op, op1, op2, result, value;
+ int prob;
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
+ modify = stmt;
+ if (TREE_CODE (stmt) == RETURN_EXPR
+ && TREE_OPERAND (stmt, 0)
+ && TREE_CODE (TREE_OPERAND (stmt, 0)) == GIMPLE_MODIFY_STMT)
+ modify = TREE_OPERAND (stmt, 0);
+ if (TREE_CODE (modify) != GIMPLE_MODIFY_STMT)
+ return false;
+ op = GIMPLE_STMT_OPERAND (modify, 1);
+ if (!INTEGRAL_TYPE_P (TREE_TYPE (op)))
+ return false;
+ code = TREE_CODE (op);
- if (code != UMOD)
+ if (code != TRUNC_MOD_EXPR || !TYPE_UNSIGNED (TREE_TYPE (op)))
return false;
- op1 = XEXP (set_src, 0);
- op2 = XEXP (set_src, 1);
-
- for (histogram = REG_NOTES (insn);
- histogram;
- histogram = XEXP (histogram, 1))
- if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
- && XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_POW2))
- break;
+ op1 = TREE_OPERAND (op, 0);
+ op2 = TREE_OPERAND (op, 1);
+
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_POW2);
if (!histogram)
return false;
- histogram = XEXP (XEXP (histogram, 0), 1);
- value = XEXP (histogram, 0);
- histogram = XEXP (histogram, 1);
- wrong_values =INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
+ value = histogram->hvalue.value;
+ wrong_values = histogram->hvalue.counters[0];
+ count = histogram->hvalue.counters[1];
- count = 0;
- for (i = 0; i < GET_MODE_BITSIZE (mode); i++)
- {
- count += INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- }
+ gimple_remove_histogram_value (cfun, stmt, histogram);
- if (!rtx_equal_p (op2, value))
- return false;
-
- /* We require that we hit a power of two at least half of all evaluations. */
- if (count < wrong_values)
+ /* We require that we hit a power of 2 at least half of all evaluations. */
+ if (simple_cst_equal (op2, value) != 1 || count < wrong_values
+ || !maybe_hot_bb_p (bb_for_stmt (stmt)))
return false;
if (dump_file)
- fprintf (dump_file, "Mod power of 2 transformation on insn %d\n",
- INSN_UID (insn));
+ {
+ fprintf (dump_file, "Mod power of 2 transformation on insn ");
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ }
/* Compute probability of taking the optimal path. */
all = count + wrong_values;
+
+ if (check_counter (stmt, "pow2", all, bb_for_stmt (stmt)->count))
+ return false;
+
prob = (count * REG_BR_PROB_BASE + all / 2) / all;
- e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
- delete_insn (insn);
-
- insert_insn_on_edge (
- gen_mod_pow2 (mode, code, set_dest, op1, op2, prob), e);
+ result = tree_mod_pow2 (stmt, op, op1, op2, prob, count, all);
+
+ GIMPLE_STMT_OPERAND (modify, 1) = result;
return true;
}
-/* Generate code for transformations 3 and 4 (with MODE and OPERATION,
- operands OP1 and OP2, result TARGET, at most SUB subtractions, and
- probability of taking the optimal path(s) PROB1 and PROB2). */
-static rtx
-gen_mod_subtract (enum machine_mode mode, enum rtx_code operation,
- rtx target, rtx op1, rtx op2, int sub, int prob1, int prob2)
+/* Generate code for transformations 3 and 4 (with OPERATION, operands OP1
+ and OP2, parent modify-expr STMT, and NCOUNTS the number of cases to
+ support. Currently only NCOUNTS==0 or 1 is supported and this is
+ built into this interface. The probabilities of taking the optimal
+ paths are PROB1 and PROB2, which are equivalent to COUNT1/ALL and
+ COUNT2/ALL respectively within roundoff error). This generates the
+ result into a temp and returns the temp; it does not replace or alter
+ the original STMT. */
+/* FIXME: Generalize the interface to handle NCOUNTS > 1. */
+
+static tree
+tree_mod_subtract (tree stmt, tree operation, tree op1, tree op2,
+ int prob1, int prob2, int ncounts,
+ gcov_type count1, gcov_type count2, gcov_type all)
{
- rtx tmp, tmp1, jump;
- rtx end_label = gen_label_rtx ();
- rtx sequence;
- int i;
+ tree stmt1, stmt2, stmt3;
+ tree tmp1;
+ tree label_decl1 = create_artificial_label ();
+ tree label_decl2 = create_artificial_label ();
+ tree label_decl3 = create_artificial_label ();
+ tree label1, label2, label3;
+ tree bb1end, bb2end = NULL_TREE, bb3end;
+ basic_block bb, bb2, bb3, bb4;
+ tree optype = TREE_TYPE (operation);
+ edge e12, e23 = 0, e24, e34, e14;
+ block_stmt_iterator bsi;
+ tree result = create_tmp_var (optype, "PROF");
+
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+
+ tmp1 = create_tmp_var (optype, "PROF");
+ stmt1 = build_gimple_modify_stmt (result, op1);
+ stmt2 = build_gimple_modify_stmt (tmp1, op2);
+ stmt3 = build3 (COND_EXPR, void_type_node,
+ build2 (LT_EXPR, boolean_type_node, result, tmp1),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt3, BSI_SAME_STMT);
+ bb1end = stmt3;
+
+ if (ncounts) /* Assumed to be 0 or 1 */
+ {
+ label1 = build1 (LABEL_EXPR, void_type_node, label_decl1);
+ stmt1 = build_gimple_modify_stmt (result,
+ build2 (MINUS_EXPR, optype,
+ result, tmp1));
+ stmt2 = build3 (COND_EXPR, void_type_node,
+ build2 (LT_EXPR, boolean_type_node, result, tmp1),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, label1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bb2end = stmt2;
+ }
- start_sequence ();
-
- if (!REG_P (op2))
+ /* Fallback case. */
+ label2 = build1 (LABEL_EXPR, void_type_node, label_decl2);
+ stmt1 = build_gimple_modify_stmt (result,
+ build2 (TREE_CODE (operation), optype,
+ result, tmp1));
+ bsi_insert_before (&bsi, label2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb3end = stmt1;
+
+ label3 = build1 (LABEL_EXPR, void_type_node, label_decl3);
+ bsi_insert_before (&bsi, label3, BSI_SAME_STMT);
+
+ /* Fix CFG. */
+ /* Edge e23 connects bb2 to bb3, etc. */
+ /* However block 3 is optional; if it is not there, references
+ to 3 really refer to block 2. */
+ e12 = split_block (bb, bb1end);
+ bb2 = e12->dest;
+ bb2->count = all - count1;
+
+ if (ncounts) /* Assumed to be 0 or 1. */
{
- tmp = gen_reg_rtx (mode);
- emit_move_insn (tmp, copy_rtx (op2));
+ e23 = split_block (bb2, bb2end);
+ bb3 = e23->dest;
+ bb3->count = all - count1 - count2;
}
- else
- tmp = op2;
- emit_move_insn (target, copy_rtx (op1));
- do_compare_rtx_and_jump (target, tmp, LTU, 0, mode, NULL_RTX,
- NULL_RTX, end_label);
+ e34 = split_block (ncounts ? bb3 : bb2, bb3end);
+ bb4 = e34->dest;
+ bb4->count = all;
- /* Add branch probability to jump we just created. */
- jump = get_last_insn ();
- REG_NOTES (jump) = gen_rtx_EXPR_LIST (REG_BR_PROB,
- GEN_INT (prob1), REG_NOTES (jump));
+ e12->flags &= ~EDGE_FALLTHRU;
+ e12->flags |= EDGE_FALSE_VALUE;
+ e12->probability = REG_BR_PROB_BASE - prob1;
+ e12->count = all - count1;
- for (i = 0; i < sub; i++)
+ e14 = make_edge (bb, bb4, EDGE_TRUE_VALUE);
+ e14->probability = prob1;
+ e14->count = count1;
+
+ if (ncounts) /* Assumed to be 0 or 1. */
{
- tmp1 = expand_simple_binop (mode, MINUS, target, tmp, target,
- 0, OPTAB_WIDEN);
- if (tmp1 != target)
- emit_move_insn (target, tmp1);
- do_compare_rtx_and_jump (target, tmp, LTU, 0, mode, NULL_RTX,
- NULL_RTX, end_label);
-
- /* Add branch probability to jump we just created. */
- jump = get_last_insn ();
- REG_NOTES (jump) = gen_rtx_EXPR_LIST (REG_BR_PROB,
- GEN_INT (prob2), REG_NOTES (jump));
+ e23->flags &= ~EDGE_FALLTHRU;
+ e23->flags |= EDGE_FALSE_VALUE;
+ e23->count = all - count1 - count2;
+ e23->probability = REG_BR_PROB_BASE - prob2;
+
+ e24 = make_edge (bb2, bb4, EDGE_TRUE_VALUE);
+ e24->probability = prob2;
+ e24->count = count2;
}
- tmp1 = simplify_gen_binary (operation, mode, copy_rtx (target), copy_rtx (tmp));
- tmp1 = force_operand (tmp1, target);
- if (tmp1 != target)
- emit_move_insn (target, tmp1);
-
- emit_label (end_label);
+ e34->probability = REG_BR_PROB_BASE;
+ e34->count = all - count1 - count2;
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
+ return result;
}
/* Do transforms 3) and 4) on INSN if applicable. */
static bool
-mod_subtract_transform (rtx insn)
+tree_mod_subtract_transform (tree stmt)
{
- rtx set, set_src, set_dest, op1, op2, value, histogram;
- enum rtx_code code;
- enum machine_mode mode;
- gcov_type wrong_values, counts[2], count, all;
- edge e;
- int i, prob1, prob2;
-
- set = single_set (insn);
- if (!set)
+ histogram_value histogram;
+ enum tree_code code;
+ gcov_type count, wrong_values, all;
+ tree modify, op, op1, op2, result, value;
+ int prob1, prob2;
+ unsigned int i, steps;
+ gcov_type count1, count2;
+
+ modify = stmt;
+ if (TREE_CODE (stmt) == RETURN_EXPR
+ && TREE_OPERAND (stmt, 0)
+ && TREE_CODE (TREE_OPERAND (stmt, 0)) == GIMPLE_MODIFY_STMT)
+ modify = TREE_OPERAND (stmt, 0);
+ if (TREE_CODE (modify) != GIMPLE_MODIFY_STMT)
return false;
-
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
+ op = GIMPLE_STMT_OPERAND (modify, 1);
+ if (!INTEGRAL_TYPE_P (TREE_TYPE (op)))
+ return false;
+ code = TREE_CODE (op);
- if (code != UMOD)
+ if (code != TRUNC_MOD_EXPR || !TYPE_UNSIGNED (TREE_TYPE (op)))
return false;
- op1 = XEXP (set_src, 0);
- op2 = XEXP (set_src, 1);
-
- for (histogram = REG_NOTES (insn);
- histogram;
- histogram = XEXP (histogram, 1))
- if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
- && XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_INTERVAL))
- break;
+ op1 = TREE_OPERAND (op, 0);
+ op2 = TREE_OPERAND (op, 1);
+
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_INTERVAL);
if (!histogram)
return false;
- histogram = XEXP (XEXP (histogram, 0), 1);
- value = XEXP (histogram, 0);
- histogram = XEXP (histogram, 1);
-
+ value = histogram->hvalue.value;
all = 0;
- for (i = 0; i < 2; i++)
+ wrong_values = 0;
+ for (i = 0; i < histogram->hdata.intvl.steps; i++)
+ all += histogram->hvalue.counters[i];
+
+ wrong_values += histogram->hvalue.counters[i];
+ wrong_values += histogram->hvalue.counters[i+1];
+ steps = histogram->hdata.intvl.steps;
+ all += wrong_values;
+ count1 = histogram->hvalue.counters[0];
+ count2 = histogram->hvalue.counters[1];
+
+ /* Compute probability of taking the optimal path. */
+ if (check_counter (stmt, "interval", all, bb_for_stmt (stmt)->count))
{
- counts[i] = INTVAL (XEXP (histogram, 0));
- all += counts[i];
- histogram = XEXP (histogram, 1);
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ return false;
}
- wrong_values = INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- wrong_values += INTVAL (XEXP (histogram, 0));
- all += wrong_values;
/* We require that we use just subtractions in at least 50% of all
evaluations. */
count = 0;
- for (i = 0; i < 2; i++)
+ for (i = 0; i < histogram->hdata.intvl.steps; i++)
{
- count += counts[i];
+ count += histogram->hvalue.counters[i];
if (count * 2 >= all)
break;
}
-
- if (i == 2)
+ if (i == steps
+ || !maybe_hot_bb_p (bb_for_stmt (stmt)))
return false;
+ gimple_remove_histogram_value (cfun, stmt, histogram);
if (dump_file)
- fprintf (dump_file, "Mod subtract transformation on insn %d\n",
- INSN_UID (insn));
+ {
+ fprintf (dump_file, "Mod subtract transformation on insn ");
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ }
/* Compute probability of taking the optimal path(s). */
- prob1 = (counts[0] * REG_BR_PROB_BASE + all / 2) / all;
- prob2 = (counts[1] * REG_BR_PROB_BASE + all / 2) / all;
+ prob1 = (count1 * REG_BR_PROB_BASE + all / 2) / all;
+ prob2 = (count2 * REG_BR_PROB_BASE + all / 2) / all;
- e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
- delete_insn (insn);
-
- insert_insn_on_edge (
- gen_mod_subtract (mode, code, set_dest,
- op1, op2, i, prob1, prob2), e);
+ /* In practice, "steps" is always 2. This interface reflects this,
+ and will need to be changed if "steps" can change. */
+ result = tree_mod_subtract (stmt, op, op1, op2, prob1, prob2, i,
+ count1, count2, all);
+
+ GIMPLE_STMT_OPERAND (modify, 1) = result;
return true;
}
-#ifdef HAVE_prefetch
-/* Generate code for transformation 5 for mem with ADDRESS and a constant
- step DELTA. WRITE is true if the reference is a store to mem. */
+static struct cgraph_node** pid_map = NULL;
+
+/* Initialize map of pids (pid -> cgraph node) */
+
+static void
+init_pid_map (void)
+{
+ struct cgraph_node *n;
+
+ if (pid_map != NULL)
+ return;
+
+ pid_map
+ = (struct cgraph_node**) xmalloc (sizeof (struct cgraph_node*) * cgraph_max_pid);
+
+ for (n = cgraph_nodes; n; n = n->next)
+ {
+ if (n->pid != -1)
+ pid_map [n->pid] = n;
+ }
+}
+
+/* Return cgraph node for function with pid */
-static rtx
-gen_speculative_prefetch (rtx address, gcov_type delta, int write)
+static inline struct cgraph_node*
+find_func_by_pid (int pid)
{
- rtx tmp;
- rtx sequence;
-
- /* TODO: we do the prefetching for just one iteration ahead, which
- often is not enough. */
- start_sequence ();
- if (offsettable_address_p (0, VOIDmode, address))
- tmp = plus_constant (copy_rtx (address), delta);
+ init_pid_map ();
+
+ return pid_map [pid];
+}
+
+/* Do transformation
+
+ if (actual_callee_addres == addres_of_most_common_function/method)
+ do direct call
else
+ old call
+ */
+
+static tree
+tree_ic (tree stmt, tree call, struct cgraph_node* direct_call,
+ int prob, gcov_type count, gcov_type all)
+{
+ tree stmt1, stmt2, stmt3;
+ tree tmp1, tmpv, tmp;
+ tree label_decl1 = create_artificial_label ();
+ tree label_decl2 = create_artificial_label ();
+ tree label1, label2;
+ tree bb1end, bb2end, bb3end;
+ tree new_call;
+ basic_block bb, bb2, bb3, bb4;
+ tree optype = build_pointer_type (void_type_node);
+ edge e12, e13, e23, e24, e34;
+ block_stmt_iterator bsi;
+ int region;
+
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+
+ tmpv = create_tmp_var (optype, "PROF");
+ tmp1 = create_tmp_var (optype, "PROF");
+ stmt1 = build_gimple_modify_stmt (tmpv,
+ unshare_expr (CALL_EXPR_FN (call)));
+ tmp = fold_convert (optype, build_addr (direct_call->decl,
+ current_function_decl));
+ stmt2 = build_gimple_modify_stmt (tmp1, tmp);
+ stmt3 = build3 (COND_EXPR, void_type_node,
+ build2 (NE_EXPR, boolean_type_node, tmp1, tmpv),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt3, BSI_SAME_STMT);
+ bb1end = stmt3;
+
+ label1 = build1 (LABEL_EXPR, void_type_node, label_decl1);
+ stmt1 = unshare_expr (stmt);
+ new_call = get_call_expr_in (stmt1);
+ CALL_EXPR_FN (new_call) = build_addr (direct_call->decl,
+ current_function_decl);
+ bsi_insert_before (&bsi, label1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bb2end = stmt1;
+
+ label2 = build1 (LABEL_EXPR, void_type_node, label_decl2);
+ bsi_insert_before (&bsi, label2, BSI_SAME_STMT);
+ bb3end = stmt;
+
+ /* Fix CFG. */
+ /* Edge e23 connects bb2 to bb3, etc. */
+ e12 = split_block (bb, bb1end);
+ bb2 = e12->dest;
+ bb2->count = count;
+ e23 = split_block (bb2, bb2end);
+ bb3 = e23->dest;
+ bb3->count = all - count;
+ e34 = split_block (bb3, bb3end);
+ bb4 = e34->dest;
+ bb4->count = all;
+
+ e12->flags &= ~EDGE_FALLTHRU;
+ e12->flags |= EDGE_FALSE_VALUE;
+ e12->probability = prob;
+ e12->count = count;
+
+ e13 = make_edge (bb, bb3, EDGE_TRUE_VALUE);
+ e13->probability = REG_BR_PROB_BASE - prob;
+ e13->count = all - count;
+
+ remove_edge (e23);
+
+ e24 = make_edge (bb2, bb4, EDGE_FALLTHRU);
+ e24->probability = REG_BR_PROB_BASE;
+ e24->count = count;
+ e34->probability = REG_BR_PROB_BASE;
+ e34->count = all - count;
+
+ /* Fix eh edges */
+ region = lookup_stmt_eh_region (stmt);
+ if (region >=0 && tree_could_throw_p (stmt1))
{
- tmp = simplify_gen_binary (PLUS, Pmode,
- copy_rtx (address), GEN_INT (delta));
- tmp = force_operand (tmp, NULL);
+ add_stmt_to_eh_region (stmt1, region);
+ make_eh_edges (stmt1);
}
- if (! (*insn_data[(int)CODE_FOR_prefetch].operand[0].predicate)
- (tmp, insn_data[(int)CODE_FOR_prefetch].operand[0].mode))
- tmp = force_reg (Pmode, tmp);
- emit_insn (gen_prefetch (tmp, GEN_INT (write), GEN_INT (3)));
- sequence = get_insns ();
- end_sequence ();
-
- return sequence;
+
+ if (region >=0 && tree_could_throw_p (stmt))
+ {
+ tree_purge_dead_eh_edges (bb4);
+ make_eh_edges (stmt);
+ }
+
+ return stmt1;
}
-/* Do transform 5) on INSN if applicable. */
+/*
+ For every checked indirect/virtual call determine if most common pid of
+ function/class method has probability more than 50%. If yes modify code of
+ this call to:
+ */
static bool
-speculative_prefetching_transform (rtx insn)
+tree_ic_transform (tree stmt)
{
- rtx histogram, value;
+ histogram_value histogram;
gcov_type val, count, all;
- edge e;
- rtx mem, address;
- int write;
+ int prob;
+ tree call, callee, modify;
+ struct cgraph_node *direct_call;
+
+ call = get_call_expr_in (stmt);
- if (!maybe_hot_bb_p (BLOCK_FOR_INSN (insn)))
+ if (!call || TREE_CODE (call) != CALL_EXPR)
return false;
- if (!find_mem_reference (insn, &mem, &write))
+ callee = CALL_EXPR_FN (call);
+
+ if (TREE_CODE (callee) == ADDR_EXPR)
return false;
- address = XEXP (mem, 0);
- if (side_effects_p (address))
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_INDIR_CALL);
+ if (!histogram)
return false;
-
- if (CONSTANT_P (address))
+
+ val = histogram->hvalue.counters [0];
+ count = histogram->hvalue.counters [1];
+ all = histogram->hvalue.counters [2];
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+
+ if (4 * count <= 3 * all)
return false;
- for (histogram = REG_NOTES (insn);
- histogram;
- histogram = XEXP (histogram, 1))
- if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
- && XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_CONST_DELTA))
- break;
+ prob = (count * REG_BR_PROB_BASE + all / 2) / all;
+ direct_call = find_func_by_pid ((int)val);
- if (!histogram)
+ if (direct_call == NULL)
return false;
- histogram = XEXP (XEXP (histogram, 0), 1);
- value = XEXP (histogram, 0);
- histogram = XEXP (histogram, 1);
- /* Skip last value referenced. */
- histogram = XEXP (histogram, 1);
- val = INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- count = INTVAL (XEXP (histogram, 0));
- histogram = XEXP (histogram, 1);
- all = INTVAL (XEXP (histogram, 0));
-
- /* With that few executions we do not really have a reason to optimize the
- statement, and more importantly, the data about differences of addresses
- are spoiled by the first item that had no previous value to compare
- with. */
- if (all < 4)
+ modify = tree_ic (stmt, call, direct_call, prob, count, all);
+
+ if (dump_file)
+ {
+ fprintf (dump_file, "Indirect call -> direct call ");
+ print_generic_expr (dump_file, call, TDF_SLIM);
+ fprintf (dump_file, "=> ");
+ print_generic_expr (dump_file, direct_call->decl, TDF_SLIM);
+ fprintf (dump_file, " transformation on insn ");
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ fprintf (dump_file, " to ");
+ print_generic_stmt (dump_file, modify, TDF_SLIM);
+ }
+
+ return true;
+}
+
+/* Return true if the stringop CALL with FNDECL shall be profiled. */
+static bool
+interesting_stringop_to_profile_p (tree fndecl, tree call)
+{
+ enum built_in_function fcode = DECL_FUNCTION_CODE (fndecl);
+
+ if (fcode != BUILT_IN_MEMSET && fcode != BUILT_IN_MEMCPY
+ && fcode != BUILT_IN_BZERO)
return false;
- /* We require that count is at least half of all; this means
- that for the transformation to fire the value must be constant
- at least 50% of time (and 75% gives the garantee of usage). */
- if (!rtx_equal_p (address, value) || 2 * count < all)
+ switch (fcode)
+ {
+ case BUILT_IN_MEMCPY:
+ case BUILT_IN_MEMPCPY:
+ return validate_arglist (call,
+ POINTER_TYPE, POINTER_TYPE, INTEGER_TYPE,
+ VOID_TYPE);
+ case BUILT_IN_MEMSET:
+ return validate_arglist (call,
+ POINTER_TYPE, INTEGER_TYPE, INTEGER_TYPE,
+ VOID_TYPE);
+ case BUILT_IN_BZERO:
+ return validate_arglist (call, POINTER_TYPE, INTEGER_TYPE,
+ VOID_TYPE);
+ default:
+ gcc_unreachable ();
+ }
+}
+
+/* Convert stringop (..., size)
+ into
+ if (size == VALUE)
+ stringop (...., VALUE);
+ else
+ stringop (...., size);
+ assuming constant propagation of VALUE will happen later.
+*/
+static void
+tree_stringop_fixed_value (tree stmt, tree value, int prob, gcov_type count,
+ gcov_type all)
+{
+ tree stmt1, stmt2, stmt3;
+ tree tmp1, tmpv;
+ tree label_decl1 = create_artificial_label ();
+ tree label_decl2 = create_artificial_label ();
+ tree label1, label2;
+ tree bb1end, bb2end;
+ basic_block bb, bb2, bb3, bb4;
+ edge e12, e13, e23, e24, e34;
+ block_stmt_iterator bsi;
+ tree call = get_call_expr_in (stmt);
+ tree blck_size = CALL_EXPR_ARG (call, 2);
+ tree optype = TREE_TYPE (blck_size);
+ int region;
+
+ bb = bb_for_stmt (stmt);
+ bsi = bsi_for_stmt (stmt);
+
+ if (bsi_end_p (bsi))
+ {
+ edge_iterator ei;
+ for (ei = ei_start (bb->succs); (e34 = ei_safe_edge (ei)); )
+ if (!e34->flags & EDGE_ABNORMAL)
+ break;
+ }
+ else
+ {
+ e34 = split_block (bb, stmt);
+ bsi = bsi_for_stmt (stmt);
+ }
+ bb4 = e34->dest;
+
+ tmpv = create_tmp_var (optype, "PROF");
+ tmp1 = create_tmp_var (optype, "PROF");
+ stmt1 = build_gimple_modify_stmt (tmpv, fold_convert (optype, value));
+ stmt2 = build_gimple_modify_stmt (tmp1, blck_size);
+ stmt3 = build3 (COND_EXPR, void_type_node,
+ build2 (NE_EXPR, boolean_type_node, tmp1, tmpv),
+ NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt2, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt3, BSI_SAME_STMT);
+ bb1end = stmt3;
+
+ label1 = build1 (LABEL_EXPR, void_type_node, label_decl1);
+ stmt1 = unshare_expr (stmt);
+ call = get_call_expr_in (stmt1);
+ CALL_EXPR_ARG (call, 2) = value;
+ bsi_insert_before (&bsi, label1, BSI_SAME_STMT);
+ bsi_insert_before (&bsi, stmt1, BSI_SAME_STMT);
+ region = lookup_stmt_eh_region (stmt);
+ if (region >= 0)
+ add_stmt_to_eh_region (stmt1, region);
+ bb2end = stmt1;
+ label2 = build1 (LABEL_EXPR, void_type_node, label_decl2);
+ bsi_insert_before (&bsi, label2, BSI_SAME_STMT);
+
+ /* Fix CFG. */
+ /* Edge e23 connects bb2 to bb3, etc. */
+ e12 = split_block (bb, bb1end);
+ bb2 = e12->dest;
+ bb2->count = count;
+ e23 = split_block (bb2, bb2end);
+ bb3 = e23->dest;
+ bb3->count = all - count;
+
+ e12->flags &= ~EDGE_FALLTHRU;
+ e12->flags |= EDGE_FALSE_VALUE;
+ e12->probability = prob;
+ e12->count = count;
+
+ e13 = make_edge (bb, bb3, EDGE_TRUE_VALUE);
+ e13->probability = REG_BR_PROB_BASE - prob;
+ e13->count = all - count;
+
+ remove_edge (e23);
+
+ e24 = make_edge (bb2, bb4, EDGE_FALLTHRU);
+ e24->probability = REG_BR_PROB_BASE;
+ e24->count = count;
+
+ e34->probability = REG_BR_PROB_BASE;
+ e34->count = all - count;
+}
+
+/* Find values inside STMT for that we want to measure histograms for
+ division/modulo optimization. */
+static bool
+tree_stringops_transform (block_stmt_iterator *bsi)
+{
+ tree stmt = bsi_stmt (*bsi);
+ tree call = get_call_expr_in (stmt);
+ tree fndecl;
+ tree blck_size;
+ enum built_in_function fcode;
+ histogram_value histogram;
+ gcov_type count, all, val;
+ tree value;
+ tree dest, src;
+ unsigned int dest_align, src_align;
+ int prob;
+ tree tree_val;
+
+ if (!call)
+ return false;
+ fndecl = get_callee_fndecl (call);
+ if (!fndecl)
+ return false;
+ fcode = DECL_FUNCTION_CODE (fndecl);
+ if (!interesting_stringop_to_profile_p (fndecl, call))
return false;
- /* If the difference is too small, it does not make too much sense to
- prefetch, as the memory is probably already in cache. */
- if (val >= NOPREFETCH_RANGE_MIN && val <= NOPREFETCH_RANGE_MAX)
+ if (fcode == BUILT_IN_BZERO)
+ blck_size = CALL_EXPR_ARG (call, 1);
+ else
+ blck_size = CALL_EXPR_ARG (call, 2);
+ if (TREE_CODE (blck_size) == INTEGER_CST)
return false;
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_SINGLE_VALUE);
+ if (!histogram)
+ return false;
+ value = histogram->hvalue.value;
+ val = histogram->hvalue.counters[0];
+ count = histogram->hvalue.counters[1];
+ all = histogram->hvalue.counters[2];
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ /* We require that count is at least half of all; this means
+ that for the transformation to fire the value must be constant
+ at least 80% of time. */
+ if ((6 * count / 5) < all || !maybe_hot_bb_p (bb_for_stmt (stmt)))
+ return false;
+ if (check_counter (stmt, "value", all, bb_for_stmt (stmt)->count))
+ return false;
+ prob = (count * REG_BR_PROB_BASE + all / 2) / all;
+ dest = CALL_EXPR_ARG (call, 0);
+ dest_align = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+ switch (fcode)
+ {
+ case BUILT_IN_MEMCPY:
+ case BUILT_IN_MEMPCPY:
+ src = CALL_EXPR_ARG (call, 1);
+ src_align = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
+ if (!can_move_by_pieces (val, MIN (dest_align, src_align)))
+ return false;
+ break;
+ case BUILT_IN_MEMSET:
+ if (!can_store_by_pieces (val, builtin_memset_read_str,
+ CALL_EXPR_ARG (call, 1),
+ dest_align, true))
+ return false;
+ break;
+ case BUILT_IN_BZERO:
+ if (!can_store_by_pieces (val, builtin_memset_read_str,
+ integer_zero_node,
+ dest_align, true))
+ return false;
+ break;
+ default:
+ gcc_unreachable ();
+ }
+ tree_val = build_int_cst_wide (get_gcov_type (),
+ (unsigned HOST_WIDE_INT) val,
+ val >> (HOST_BITS_PER_WIDE_INT - 1) >> 1);
if (dump_file)
- fprintf (dump_file, "Speculative prefetching for insn %d\n",
- INSN_UID (insn));
-
- e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
+ {
+ fprintf (dump_file, "Single value %i stringop transformation on ",
+ (int)val);
+ print_generic_stmt (dump_file, stmt, TDF_SLIM);
+ }
+ tree_stringop_fixed_value (stmt, tree_val, prob, count, all);
- insert_insn_on_edge (gen_speculative_prefetch (address, val, write), e);
-
return true;
}
-#endif /* HAVE_prefetch */
-\f
-/* Connection to the outside world. */
-/* Struct for IR-dependent hooks. */
+
+void
+stringop_block_profile (tree stmt, unsigned int *expected_align,
+ HOST_WIDE_INT *expected_size)
+{
+ histogram_value histogram;
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_AVERAGE);
+ if (!histogram)
+ *expected_size = -1;
+ else if (!histogram->hvalue.counters[1])
+ {
+ *expected_size = -1;
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ }
+ else
+ {
+ gcov_type size;
+ size = ((histogram->hvalue.counters[0]
+ + histogram->hvalue.counters[1] / 2)
+ / histogram->hvalue.counters[1]);
+ /* Even if we can hold bigger value in SIZE, INT_MAX
+ is safe "infinity" for code generation strategies. */
+ if (size > INT_MAX)
+ size = INT_MAX;
+ *expected_size = size;
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ }
+ histogram = gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE_IOR);
+ if (!histogram)
+ *expected_align = 0;
+ else if (!histogram->hvalue.counters[0])
+ {
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ *expected_align = 0;
+ }
+ else
+ {
+ gcov_type count;
+ int alignment;
+
+ count = histogram->hvalue.counters[0];
+ alignment = 1;
+ while (!(count & alignment)
+ && (alignment * 2 * BITS_PER_UNIT))
+ alignment <<= 1;
+ *expected_align = alignment * BITS_PER_UNIT;
+ gimple_remove_histogram_value (cfun, stmt, histogram);
+ }
+}
+
struct value_prof_hooks {
/* Find list of values for which we want to measure histograms. */
void (*find_values_to_profile) (histogram_values *);
statically. See value-prof.c for more detail. */
bool (*value_profile_transformations) (void);
};
+\f
+/* Find values inside STMT for that we want to measure histograms for
+ division/modulo optimization. */
+static void
+tree_divmod_values_to_profile (tree stmt, histogram_values *values)
+{
+ tree assign, lhs, rhs, divisor, op0, type;
+ histogram_value hist;
-/* Hooks for RTL-based versions (the only ones that currently work). */
-static struct value_prof_hooks rtl_value_prof_hooks =
+ if (TREE_CODE (stmt) == RETURN_EXPR)
+ assign = TREE_OPERAND (stmt, 0);
+ else
+ assign = stmt;
+
+ if (!assign
+ || TREE_CODE (assign) != GIMPLE_MODIFY_STMT)
+ return;
+ lhs = GIMPLE_STMT_OPERAND (assign, 0);
+ type = TREE_TYPE (lhs);
+ if (!INTEGRAL_TYPE_P (type))
+ return;
+
+ rhs = GIMPLE_STMT_OPERAND (assign, 1);
+ switch (TREE_CODE (rhs))
+ {
+ case TRUNC_DIV_EXPR:
+ case TRUNC_MOD_EXPR:
+ divisor = TREE_OPERAND (rhs, 1);
+ op0 = TREE_OPERAND (rhs, 0);
+
+ VEC_reserve (histogram_value, heap, *values, 3);
+
+ if (is_gimple_reg (divisor))
+ /* Check for the case where the divisor is the same value most
+ of the time. */
+ VEC_quick_push (histogram_value, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_SINGLE_VALUE,
+ stmt, divisor));
+
+ /* For mod, check whether it is not often a noop (or replaceable by
+ a few subtractions). */
+ if (TREE_CODE (rhs) == TRUNC_MOD_EXPR
+ && TYPE_UNSIGNED (type))
+ {
+ tree val;
+ /* Check for a special case where the divisor is power of 2. */
+ VEC_quick_push (histogram_value, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_POW2,
+ stmt, divisor));
+
+ val = build2 (TRUNC_DIV_EXPR, type, op0, divisor);
+ hist = gimple_alloc_histogram_value (cfun, HIST_TYPE_INTERVAL,
+ stmt, val);
+ hist->hdata.intvl.int_start = 0;
+ hist->hdata.intvl.steps = 2;
+ VEC_quick_push (histogram_value, *values, hist);
+ }
+ return;
+
+ default:
+ return;
+ }
+}
+
+/* Find calls inside STMT for that we want to measure histograms for
+ indirect/virtual call optimization. */
+
+static void
+tree_indirect_call_to_profile (tree stmt, histogram_values *values)
{
- rtl_find_values_to_profile,
- rtl_value_profile_transformations
-};
+ tree call;
+ tree callee;
+
+ call = get_call_expr_in (stmt);
+
+ if (!call || TREE_CODE (call) != CALL_EXPR)
+ return;
+
+ callee = CALL_EXPR_FN (call);
+
+ if (TREE_CODE (callee) == ADDR_EXPR)
+ return;
+
+ VEC_reserve (histogram_value, heap, *values, 3);
-void
-rtl_register_value_prof_hooks (void)
+ VEC_quick_push (histogram_value, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_INDIR_CALL,
+ stmt, callee));
+
+ return;
+}
+
+/* Find values inside STMT for that we want to measure histograms for
+ string operations. */
+static void
+tree_stringops_values_to_profile (tree stmt, histogram_values *values)
{
- value_prof_hooks = &rtl_value_prof_hooks;
- if (ir_type ())
- abort ();
+ tree call = get_call_expr_in (stmt);
+ tree fndecl;
+ tree blck_size;
+ tree dest;
+ enum built_in_function fcode;
+
+ if (!call)
+ return;
+ fndecl = get_callee_fndecl (call);
+ if (!fndecl)
+ return;
+ fcode = DECL_FUNCTION_CODE (fndecl);
+
+ if (!interesting_stringop_to_profile_p (fndecl, call))
+ return;
+
+ dest = CALL_EXPR_ARG (call, 0);
+ if (fcode == BUILT_IN_BZERO)
+ blck_size = CALL_EXPR_ARG (call, 1);
+ else
+ blck_size = CALL_EXPR_ARG (call, 2);
+
+ if (TREE_CODE (blck_size) != INTEGER_CST)
+ {
+ VEC_safe_push (histogram_value, heap, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_SINGLE_VALUE,
+ stmt, blck_size));
+ VEC_safe_push (histogram_value, heap, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_AVERAGE,
+ stmt, blck_size));
+ }
+ if (TREE_CODE (blck_size) != INTEGER_CST)
+ VEC_safe_push (histogram_value, heap, *values,
+ gimple_alloc_histogram_value (cfun, HIST_TYPE_IOR,
+ stmt, dest));
}
-\f
-/* Tree-based versions are stubs for now. */
+
+/* Find values inside STMT for that we want to measure histograms and adds
+ them to list VALUES. */
+
static void
-tree_find_values_to_profile (histogram_values *values ATTRIBUTE_UNUSED)
+tree_values_to_profile (tree stmt, histogram_values *values)
{
- abort ();
+ if (flag_value_profile_transformations)
+ {
+ tree_divmod_values_to_profile (stmt, values);
+ tree_stringops_values_to_profile (stmt, values);
+ tree_indirect_call_to_profile (stmt, values);
+ }
}
-static bool
-tree_value_profile_transformations (void)
+static void
+tree_find_values_to_profile (histogram_values *values)
{
- abort ();
+ basic_block bb;
+ block_stmt_iterator bsi;
+ unsigned i;
+ histogram_value hist = NULL;
+
+ *values = NULL;
+ FOR_EACH_BB (bb)
+ for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ tree_values_to_profile (bsi_stmt (bsi), values);
+
+ for (i = 0; VEC_iterate (histogram_value, *values, i, hist); i++)
+ {
+ switch (hist->type)
+ {
+ case HIST_TYPE_INTERVAL:
+ hist->n_counters = hist->hdata.intvl.steps + 2;
+ break;
+
+ case HIST_TYPE_POW2:
+ hist->n_counters = 2;
+ break;
+
+ case HIST_TYPE_SINGLE_VALUE:
+ hist->n_counters = 3;
+ break;
+
+ case HIST_TYPE_CONST_DELTA:
+ hist->n_counters = 4;
+ break;
+
+ case HIST_TYPE_INDIR_CALL:
+ hist->n_counters = 3;
+ break;
+
+ case HIST_TYPE_AVERAGE:
+ hist->n_counters = 2;
+ break;
+
+ case HIST_TYPE_IOR:
+ hist->n_counters = 1;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+ if (dump_file)
+ {
+ fprintf (dump_file, "Stmt ");
+ print_generic_expr (dump_file, hist->hvalue.stmt, TDF_SLIM);
+ dump_histogram_value (dump_file, hist);
+ }
+ }
}
static struct value_prof_hooks tree_value_prof_hooks = {
void
tree_register_value_prof_hooks (void)
{
+ gcc_assert (current_ir_type () == IR_GIMPLE);
value_prof_hooks = &tree_value_prof_hooks;
- if (!ir_type ())
- abort ();
}
\f
/* IR-independent entry points. */
{
return (value_prof_hooks->value_profile_transformations) ();
}
+\f
OSDN Git Service
