This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Kill RTL level profiling
- From: Jan Hubicka <jh at suse dot cz>
- To: gcc-patches at gcc dot gnu dot org
- Date: Thu, 28 Jul 2005 14:41:29 +0200
- Subject: Kill RTL level profiling
Hi,
this is the promised patch to remove RTL based profiling. There are two
leftovers:
1) speculative prefetching is not updated to trees so this patch removes it.
The pass was disabled for while as it is usually not win on CPUs with
hardware prefetching and I believe Zdenek has experimental implementation
of it as well as Steven. If one of two turns out to be useful, we cal
always add it back.
2) gcov-4.c testsuite is failing. It demonstrates deeper problem in how
we handle gotos. Since cfgcleanup hapilly forward edges across user
goto statement they often become unreachable in profile even tought
they was reachable in source program. I don't think it is stopper for the
patch because on RTL we got it wrong in many cases too (and some testcases
in testsuite actually contain wrong master profiles).
We might want to get serious about it and do one of:
1) Avoid gimplifier from putting user gotos to CFG edges; avoid cfg_cleanup
being called before profiling. This is implementation I did on
tree-profiling. It works well, but it is bit ugly on restricting
tree_profiling pass placement.
2) Teach cfg_cleanup to not forward across edges with locator attached
and teach tree profiling to split those edges before profiling. We
will probably need to add new cfun variable signaizling that
tree_profiling is expected to diable the forwarding (and
theoretically we might want to make this default for -O0 if we ever
make RTL land safe in this respect so users can breakpoint on their
gotos)
3) Make gcov-4 passing with the invalid count on goto statement and ignore
the problem.
or combination of 1-2 ;)
I believe these two issues are not big ones, so I will push forward the patch
if no one objects tomorrow to unbreak profiling bootstrap. I will also hapilly
work on solving those two issues if these turns out to be important.
Also I am aware of fact that I left the hooks infrasructure in the
sources. I am still running some profile updating sanity check using
the RTL implementation but I will kill them once I am done.
Bootstrapped/regtested i686-pc-gnu-linux with and without profiledbootstrap.
Honza
2005-07-27 Jan Hubicka <jh@suse.cz>
* Makefile.in (rtl-profile.o): Kill all traces of it.
* common.opt (fspeculative-prefetching, ftree-based-profiling): Kill.
* coverage.h (rtl_coverage_counter_ref): Kill.
* opts.c (flag_speculative_prefetching_set): Kill.
(flag_loop_optimize_set): New.
(common_handle_option): Disable loop optimizer when profiling;
do not handle speculative prefetching.
* passes.c (init_optimization_passes): Replace pass_profiling combo
by branch_prob pass.
* profile.c (compute_value_histograms): Update for simplified value
profiles.
(rtl_register_profile_hooks): Kill.
(pass_profiling): Kill.
(rest_of_handle_branch_prob): Do not profile.
* toplev.c (process_options): Remove speculative prefetching.
* toplev.h (flag_tree_based_profiling): Kill.
* tree-profile.c (prepare_instrumented_value,
tree_gen_interval_profiler, tree_gen_pow2_profiler,
tree_gen_one_value_profiler, do_tree_profiling): Update for
simplified datastructures.
* value-prof.c: Add comment that speculative prefetching was dropped;
update rest of file for simplified datastructures.
(NOPREFETCH_RANGE_MIN, NOPREFETCH_RANGE_MAX,
rtl_divmod_values_to_profile, insn_prefetch_values_to_profile,
find_mem_reference_1, find_mem_reference_2, find_mem_reference,
rtl_values_to_profile, rtl_divmod_fixed_value, rtl_mod_pow2,
rtl_mod_subtract, gen_speculative_prefetch,
rtl_divmod_fixed_value_transform, rtl_mod_pow2_value_transform,
rtl_mod_subtract_transform, speculative_prefetching_transform): Kill.
(gate_handle_value_profile_transformations,
rest_of_handle_value_profile_transformations,
pass_value_profile_transformations): Kill.
* value-prof.h (histogram_value_t): Remove IL based unions.
(rtl_register_value_prof_hooks, rtl_register_profile_hooks,
rtl_profile_hooks): Remove hooks.
* invoke.texi (-ftree-based-profiling, -fspeculative-prefetching): Kill.
* inliner-1.c: Do not dump everything.
* tree-prof.exp: Do not pass -ftree-based-profiling.
* bprob.exp: Likewise.
Index: Makefile.in
===================================================================
RCS file: /cvs/gcc/gcc/gcc/Makefile.in,v
retrieving revision 1.1530
diff -c -3 -p -r1.1530 Makefile.in
*** Makefile.in 25 Jul 2005 23:04:33 -0000 1.1530
--- Makefile.in 28 Jul 2005 12:30:04 -0000
*************** OBJS-common = \
*** 966,972 ****
targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o \
varasm.o varray.o vec.o version.o vmsdbgout.o xcoffout.o alloc-pool.o \
et-forest.o cfghooks.o bt-load.o pretty-print.o $(GGC) web.o passes.o \
! rtl-profile.o tree-profile.o rtlhooks.o cfgexpand.o lambda-mat.o \
lambda-trans.o lambda-code.o tree-loop-linear.o tree-ssa-sink.o \
tree-vrp.o tree-stdarg.o tree-cfgcleanup.o tree-ssa-reassoc.o \
tree-ssa-structalias.o tree-object-size.o
--- 966,972 ----
targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o \
varasm.o varray.o vec.o version.o vmsdbgout.o xcoffout.o alloc-pool.o \
et-forest.o cfghooks.o bt-load.o pretty-print.o $(GGC) web.o passes.o \
! tree-profile.o rtlhooks.o cfgexpand.o lambda-mat.o \
lambda-trans.o lambda-code.o tree-loop-linear.o tree-ssa-sink.o \
tree-vrp.o tree-stdarg.o tree-cfgcleanup.o tree-ssa-reassoc.o \
tree-ssa-structalias.o tree-object-size.o
*************** tree-profile.o : tree-profile.c $(CONFIG
*** 2242,2250 ****
$(TM_H) $(RTL_H) $(TREE_H) $(FLAGS_H) output.h $(REGS_H) $(EXPR_H) \
function.h toplev.h $(COVERAGE_H) $(TREE_H) value-prof.h $(TREE_DUMP_H) \
tree-pass.h $(TREE_FLOW_H) $(TIMEVAR_H) $(GGC_H) gt-tree-profile.h
- rtl-profile.o : tree-profile.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
- $(TM_H) $(RTL_H) $(TREE_H) $(FLAGS_H) output.h $(REGS_H) $(EXPR_H) \
- function.h toplev.h $(COVERAGE_H) value-prof.h $(GGC_H)
value-prof.o : value-prof.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) \
$(BASIC_BLOCK_H) hard-reg-set.h value-prof.h $(EXPR_H) output.h $(FLAGS_H) \
$(RECOG_H) insn-config.h $(OPTABS_H) $(REGS_H) $(GGC_H) $(DIAGNOSTIC_H) \
--- 2242,2247 ----
*************** GTFILES = $(srcdir)/input.h $(srcdir)/co
*** 2732,2738 ****
$(srcdir)/tree-iterator.c $(srcdir)/gimplify.c \
$(srcdir)/tree-chrec.h $(srcdir)/tree-vect-generic.c \
$(srcdir)/tree-ssa-operands.h $(srcdir)/tree-ssa-operands.c \
! $(srcdir)/tree-profile.c $(srcdir)/rtl-profile.c $(srcdir)/tree-nested.c \
$(srcdir)/ipa-reference.c \
$(srcdir)/targhooks.c $(out_file) \
@all_gtfiles@
--- 2729,2735 ----
$(srcdir)/tree-iterator.c $(srcdir)/gimplify.c \
$(srcdir)/tree-chrec.h $(srcdir)/tree-vect-generic.c \
$(srcdir)/tree-ssa-operands.h $(srcdir)/tree-ssa-operands.c \
! $(srcdir)/tree-profile.c $(srcdir)/tree-nested.c \
$(srcdir)/ipa-reference.c \
$(srcdir)/targhooks.c $(out_file) \
@all_gtfiles@
Index: common.opt
===================================================================
RCS file: /cvs/gcc/gcc/gcc/common.opt,v
retrieving revision 1.83
diff -c -3 -p -r1.83 common.opt
*** common.opt 21 Jul 2005 17:27:25 -0000 1.83
--- common.opt 28 Jul 2005 12:30:04 -0000
*************** fsingle-precision-constant
*** 779,788 ****
Common Report Var(flag_single_precision_constant)
Convert floating point constants to single precision constants
- fspeculative-prefetching
- Common Report Var(flag_speculative_prefetching)
- Use value profiling for speculative prefetching
-
fsplit-ivs-in-unroller
Common Report Var(flag_split_ivs_in_unroller) Init(1)
Split lifetimes of induction variables when loops are unrolled
--- 779,784 ----
*************** ftrapv
*** 863,872 ****
Common Report Var(flag_trapv)
Trap for signed overflow in addition, subtraction and multiplication
- ftree-based-profiling
- Common Report Var(flag_tree_based_profiling)
- Use tree-ssa based implementation of profiling
-
ftree-ccp
Common Report Var(flag_tree_ccp)
Enable SSA-CCP optimization on trees
--- 859,864 ----
Index: coverage.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/coverage.c,v
retrieving revision 1.54
diff -c -3 -p -r1.54 coverage.c
*** coverage.c 20 Jul 2005 01:18:07 -0000 1.54
--- coverage.c 28 Jul 2005 12:30:04 -0000
*************** coverage_counter_alloc (unsigned counter
*** 410,441 ****
return 1;
}
- /* Generate a MEM rtl to access COUNTER NO. */
-
- rtx
- rtl_coverage_counter_ref (unsigned counter, unsigned no)
- {
- enum machine_mode mode = mode_for_size (GCOV_TYPE_SIZE, MODE_INT, 0);
- rtx ref;
-
- gcc_assert (no < fn_n_ctrs[counter] - fn_b_ctrs[counter]);
- no += prg_n_ctrs[counter] + fn_b_ctrs[counter];
- if (!ctr_labels[counter])
- {
- ctr_labels[counter] = gen_rtx_SYMBOL_REF (Pmode,
- ggc_strdup (IDENTIFIER_POINTER (DECL_NAME
- (tree_ctr_tables[counter]))));
- SYMBOL_REF_FLAGS (ctr_labels[counter]) = SYMBOL_FLAG_LOCAL;
- }
- ref = plus_constant (ctr_labels[counter],
- GCOV_TYPE_SIZE / BITS_PER_UNIT * no);
- ref = gen_rtx_MEM (mode, ref);
- set_mem_alias_set (ref, new_alias_set ());
- MEM_NOTRAP_P (ref) = 1;
-
- return ref;
- }
-
/* Generate a tree to access COUNTER NO. */
tree
--- 410,415 ----
Index: coverage.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/coverage.h,v
retrieving revision 1.11
diff -c -3 -p -r1.11 coverage.h
*** coverage.h 25 Jun 2005 01:59:37 -0000 1.11
--- coverage.h 28 Jul 2005 12:30:04 -0000
*************** extern int coverage_begin_output (void);
*** 38,45 ****
/* Allocate some counters. Repeatable per function. */
extern int coverage_counter_alloc (unsigned /*counter*/, unsigned/*num*/);
/* Use a counter from the most recent allocation. */
- extern rtx rtl_coverage_counter_ref (unsigned /*counter*/, unsigned/*num*/);
- /* Use a counter from the most recent allocation. */
extern tree tree_coverage_counter_ref (unsigned /*counter*/, unsigned/*num*/);
/* Get all the counters for the current function. */
--- 38,43 ----
Index: opts.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/opts.c,v
retrieving revision 1.121
diff -c -3 -p -r1.121 opts.c
*** opts.c 19 Jul 2005 18:45:55 -0000 1.121
--- opts.c 28 Jul 2005 12:30:04 -0000
*************** static const char undocumented_msg[] = N
*** 94,101 ****
static bool profile_arc_flag_set, flag_profile_values_set;
static bool flag_unroll_loops_set, flag_tracer_set;
static bool flag_value_profile_transformations_set;
- bool flag_speculative_prefetching_set;
static bool flag_peel_loops_set, flag_branch_probabilities_set;
/* Input file names. */
const char **in_fnames;
--- 94,101 ----
static bool profile_arc_flag_set, flag_profile_values_set;
static bool flag_unroll_loops_set, flag_tracer_set;
static bool flag_value_profile_transformations_set;
static bool flag_peel_loops_set, flag_branch_probabilities_set;
+ static bool flag_loop_optimize_set;
/* Input file names. */
const char **in_fnames;
*************** common_handle_option (size_t scode, cons
*** 807,812 ****
--- 807,816 ----
flag_branch_probabilities_set = true;
break;
+ case OPT_floop_optimize:
+ flag_loop_optimize_set = true;
+ break;
+
case OPT_fcall_used_:
fix_register (arg, 0, 1);
break;
*************** common_handle_option (size_t scode, cons
*** 883,892 ****
flag_tracer = value;
if (!flag_value_profile_transformations_set)
flag_value_profile_transformations = value;
! #ifdef HAVE_prefetch
! if (0 && !flag_speculative_prefetching_set)
! flag_speculative_prefetching = value;
! #endif
break;
case OPT_fprofile_generate:
--- 887,895 ----
flag_tracer = value;
if (!flag_value_profile_transformations_set)
flag_value_profile_transformations = value;
! /* Old loop optimizer is incompatible with tree profiling. */
! if (!flag_loop_optimize_set)
! flag_loop_optimize = 0;
break;
case OPT_fprofile_generate:
*************** common_handle_option (size_t scode, cons
*** 896,907 ****
flag_profile_values = value;
if (!flag_value_profile_transformations_set)
flag_value_profile_transformations = value;
if (!flag_unroll_loops_set)
flag_unroll_loops = value;
- #ifdef HAVE_prefetch
- if (0 && !flag_speculative_prefetching_set)
- flag_speculative_prefetching = value;
- #endif
break;
case OPT_fprofile_values:
--- 899,904 ----
*************** common_handle_option (size_t scode, cons
*** 927,936 ****
flag_value_profile_transformations_set = true;
break;
- case OPT_fspeculative_prefetching:
- flag_speculative_prefetching_set = true;
- break;
-
case OPT_frandom_seed:
/* The real switch is -fno-random-seed. */
if (value)
--- 924,929 ----
Index: passes.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/passes.c,v
retrieving revision 2.105
diff -c -3 -p -r2.105 passes.c
*** passes.c 19 Jul 2005 18:45:56 -0000 2.105
--- passes.c 28 Jul 2005 12:30:05 -0000
*************** init_optimization_passes (void)
*** 597,603 ****
NEXT_PASS (pass_loop_optimize);
NEXT_PASS (pass_jump_bypass);
NEXT_PASS (pass_cfg);
! NEXT_PASS (pass_profiling);
NEXT_PASS (pass_rtl_ifcvt);
NEXT_PASS (pass_tracer);
/* Perform loop optimizations. It might be better to do them a bit
--- 597,603 ----
NEXT_PASS (pass_loop_optimize);
NEXT_PASS (pass_jump_bypass);
NEXT_PASS (pass_cfg);
! NEXT_PASS (pass_branch_prob);
NEXT_PASS (pass_rtl_ifcvt);
NEXT_PASS (pass_tracer);
/* Perform loop optimizations. It might be better to do them a bit
*************** init_optimization_passes (void)
*** 621,632 ****
NEXT_PASS (pass_postreload);
*p = NULL;
- p = &pass_profiling.sub;
- NEXT_PASS (pass_branch_prob);
- NEXT_PASS (pass_value_profile_transformations);
- NEXT_PASS (pass_remove_death_notes);
- *p = NULL;
-
p = &pass_postreload.sub;
NEXT_PASS (pass_postreload_cse);
NEXT_PASS (pass_gcse2);
--- 621,626 ----
Index: profile.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/profile.c,v
retrieving revision 1.161
diff -c -3 -p -r1.161 profile.c
*** profile.c 15 Jul 2005 23:26:50 -0000 1.161
--- profile.c 28 Jul 2005 12:30:05 -0000
*************** compute_value_histograms (histogram_valu
*** 651,657 ****
gcov_type *histogram_counts[GCOV_N_VALUE_COUNTERS];
gcov_type *act_count[GCOV_N_VALUE_COUNTERS];
gcov_type *aact_count;
! histogram_value hist;
for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
n_histogram_counters[t] = 0;
--- 651,657 ----
gcov_type *histogram_counts[GCOV_N_VALUE_COUNTERS];
gcov_type *act_count[GCOV_N_VALUE_COUNTERS];
gcov_type *aact_count;
! histogram_value hist = 0;
for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
n_histogram_counters[t] = 0;
*************** compute_value_histograms (histogram_valu
*** 683,689 ****
for (i = 0; i < VEC_length (histogram_value, values); i++)
{
! rtx hist_list = NULL_RTX;
hist = VEC_index (histogram_value, values, i);
t = (int) hist->type;
--- 683,690 ----
for (i = 0; i < VEC_length (histogram_value, values); i++)
{
! tree stmt = hist->hvalue.stmt;
! stmt_ann_t ann = get_stmt_ann (stmt);
hist = VEC_index (histogram_value, values, i);
t = (int) hist->type;
*************** compute_value_histograms (histogram_valu
*** 691,719 ****
aact_count = act_count[t];
act_count[t] += hist->n_counters;
! if (!ir_type ())
! {
! for (j = hist->n_counters; j > 0; j--)
! hist_list = alloc_EXPR_LIST (0, GEN_INT (aact_count[j - 1]),
! hist_list);
! hist_list = alloc_EXPR_LIST (0,
! copy_rtx (hist->hvalue.rtl.value), hist_list);
! hist_list = alloc_EXPR_LIST (0, GEN_INT (hist->type), hist_list);
! REG_NOTES (hist->hvalue.rtl.insn) =
! alloc_EXPR_LIST (REG_VALUE_PROFILE, hist_list,
! REG_NOTES (hist->hvalue.rtl.insn));
! }
! else
! {
! tree stmt = hist->hvalue.tree.stmt;
! stmt_ann_t ann = get_stmt_ann (stmt);
! hist->hvalue.tree.next = ann->histograms;
! ann->histograms = hist;
! hist->hvalue.tree.counters =
! xmalloc (sizeof (gcov_type) * hist->n_counters);
! for (j = 0; j < hist->n_counters; j++)
! hist->hvalue.tree.counters[j] = aact_count[j];
! }
}
for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
--- 692,703 ----
aact_count = act_count[t];
act_count[t] += hist->n_counters;
! hist->hvalue.next = ann->histograms;
! ann->histograms = hist;
! hist->hvalue.counters =
! xmalloc (sizeof (gcov_type) * hist->n_counters);
! for (j = 0; j < hist->n_counters; j++)
! hist->hvalue.counters[j] = aact_count[j];
}
for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
*************** tree_register_profile_hooks (void)
*** 1324,1378 ****
profile_hooks = &tree_profile_hooks;
}
- /* Set up hooks to enable RTL-based profiling. */
-
- void
- rtl_register_profile_hooks (void)
- {
- gcc_assert (!ir_type ());
- profile_hooks = &rtl_profile_hooks;
- }
- static bool
- gate_handle_profiling (void)
- {
- return optimize > 0
- || (!flag_tree_based_profiling
- && (profile_arc_flag || flag_test_coverage
- || flag_branch_probabilities));
- }
-
- struct tree_opt_pass pass_profiling =
- {
- NULL, /* name */
- gate_handle_profiling, /* gate */
- NULL, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- 0, /* tv_id */
- 0, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- 0, /* todo_flags_start */
- 0, /* todo_flags_finish */
- 0 /* letter */
- };
-
-
/* Do branch profiling and static profile estimation passes. */
static void
rest_of_handle_branch_prob (void)
{
struct loops loops;
- rtl_register_profile_hooks ();
- rtl_register_value_prof_hooks ();
-
- if ((profile_arc_flag || flag_test_coverage || flag_branch_probabilities)
- && !flag_tree_based_profiling)
- branch_prob ();
-
/* Discover and record the loop depth at the head of each basic
block. The loop infrastructure does the real job for us. */
flow_loops_find (&loops);
--- 1308,1320 ----
*************** rest_of_handle_branch_prob (void)
*** 1382,1389 ****
/* Estimate using heuristics if no profiling info is available. */
if (flag_guess_branch_prob
! && (profile_status == PROFILE_ABSENT
! || (profile_status == PROFILE_READ && !flag_tree_based_profiling)))
estimate_probability (&loops);
flow_loops_free (&loops);
--- 1324,1330 ----
/* Estimate using heuristics if no profiling info is available. */
if (flag_guess_branch_prob
! && profile_status == PROFILE_ABSENT)
estimate_probability (&loops);
flow_loops_free (&loops);
Index: toplev.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/toplev.c,v
retrieving revision 1.973
diff -c -3 -p -r1.973 toplev.c
*** toplev.c 22 Jul 2005 17:40:29 -0000 1.973
--- toplev.c 28 Jul 2005 12:30:05 -0000
*************** process_options (void)
*** 1562,1578 ****
if (flag_value_profile_transformations)
flag_profile_values = 1;
- /* Speculative prefetching implies the value profiling. We also switch off
- the prefetching in the loop optimizer, so that we do not emit double
- prefetches. TODO -- we should teach these two to cooperate; the loop
- based prefetching may sometimes do a better job, especially in connection
- with reuse analysis. */
- if (flag_speculative_prefetching)
- {
- flag_profile_values = 1;
- flag_prefetch_loop_arrays = 0;
- }
-
/* Warn about options that are not supported on this machine. */
#ifndef INSN_SCHEDULING
if (flag_schedule_insns || flag_schedule_insns_after_reload)
--- 1562,1567 ----
*************** process_options (void)
*** 1732,1755 ****
warning (0, "-fprefetch-loop-arrays not supported for this target");
flag_prefetch_loop_arrays = 0;
}
- if (flag_speculative_prefetching)
- {
- if (flag_speculative_prefetching_set)
- warning (0, "-fspeculative-prefetching not supported for this target");
- flag_speculative_prefetching = 0;
- }
#else
if (flag_prefetch_loop_arrays && !HAVE_prefetch)
{
warning (0, "-fprefetch-loop-arrays not supported for this target (try -march switches)");
flag_prefetch_loop_arrays = 0;
}
- if (flag_speculative_prefetching && !HAVE_prefetch)
- {
- if (flag_speculative_prefetching_set)
- warning (0, "-fspeculative-prefetching not supported for this target (try -march switches)");
- flag_speculative_prefetching = 0;
- }
#endif
/* This combination of options isn't handled for i386 targets and doesn't
--- 1721,1732 ----
Index: toplev.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/toplev.h,v
retrieving revision 1.140
diff -c -3 -p -r1.140 toplev.h
*** toplev.h 19 Jul 2005 20:19:13 -0000 1.140
--- toplev.h 28 Jul 2005 12:30:05 -0000
*************** extern int flag_unroll_all_loops;
*** 128,134 ****
extern int flag_unswitch_loops;
extern int flag_cprop_registers;
extern int time_report;
- extern int flag_tree_based_profiling;
/* Things to do with target switches. */
extern void print_version (FILE *, const char *);
--- 128,133 ----
Index: tree-profile.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/tree-profile.c,v
retrieving revision 2.14
diff -c -3 -p -r2.14 tree-profile.c
*** tree-profile.c 28 Jun 2005 02:20:29 -0000 2.14
--- tree-profile.c 28 Jul 2005 12:30:05 -0000
*************** static tree
*** 123,129 ****
prepare_instrumented_value (block_stmt_iterator *bsi,
histogram_value value)
{
! tree val = value->hvalue.tree.value;
return force_gimple_operand_bsi (bsi, fold_convert (gcov_type_node, val),
true, NULL_TREE);
}
--- 123,129 ----
prepare_instrumented_value (block_stmt_iterator *bsi,
histogram_value value)
{
! tree val = value->hvalue.value;
return force_gimple_operand_bsi (bsi, fold_convert (gcov_type_node, val),
true, NULL_TREE);
}
*************** prepare_instrumented_value (block_stmt_i
*** 135,141 ****
static void
tree_gen_interval_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.tree.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
--- 135,141 ----
static void
tree_gen_interval_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
*************** tree_gen_interval_profiler (histogram_va
*** 162,168 ****
static void
tree_gen_pow2_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.tree.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
--- 162,168 ----
static void
tree_gen_pow2_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
*************** tree_gen_pow2_profiler (histogram_value
*** 185,191 ****
static void
tree_gen_one_value_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.tree.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
--- 185,191 ----
static void
tree_gen_one_value_profiler (histogram_value value, unsigned tag, unsigned base)
{
! tree stmt = value->hvalue.stmt;
block_stmt_iterator bsi = bsi_for_stmt (stmt);
tree ref = tree_coverage_counter_ref (tag, base), ref_ptr;
tree args, call, val;
*************** tree_gen_const_delta_profiler (histogram
*** 225,232 ****
static bool
do_tree_profiling (void)
{
! if (flag_tree_based_profiling
! && (profile_arc_flag || flag_test_coverage || flag_branch_probabilities))
{
tree_register_profile_hooks ();
tree_register_value_prof_hooks ();
--- 225,231 ----
static bool
do_tree_profiling (void)
{
! if (profile_arc_flag || flag_test_coverage || flag_branch_probabilities)
{
tree_register_profile_hooks ();
tree_register_value_prof_hooks ();
Index: value-prof.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/value-prof.c,v
retrieving revision 1.31
diff -c -3 -p -r1.31 value-prof.c
*** value-prof.c 5 Jul 2005 16:20:30 -0000 1.31
--- value-prof.c 28 Jul 2005 12:30:05 -0000
*************** static histogram_values static_values =
*** 61,66 ****
--- 61,68 ----
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.
Every such optimization should add its requirements for profiled values to
insn_values_to_profile function. This function is called from branch_prob
*************** static histogram_values static_values =
*** 68,135 ****
compilation with -fprofile-arcs. The optimization may then read the
gathered data in the second compilation with -fbranch-probabilities.
! There are currently two versions, RTL-based and tree-based. Over time
! the RTL-based version may go away.
!
! In the RTL-based version, 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.
!
! In the tree-based version, 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. */
! /* For speculative prefetching, the range in that we do not prefetch (because
! we assume that it will be in cache anyway). The asymmetry 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.
!
! FIXME: There is no tree form of speculative prefetching as yet.
!
! FIXME: A better approach to instrumentation in the profile-generation
! pass is to generate calls to magic library functions (to be added to
! libgcc) rather than inline code. This approach will probably be
! necessary to get tree-based speculative prefetching working in a useful
! fashion, as inline code bloats things so much the rest of the compiler has
! serious problems dealing with it (judging from the rtl behavior). */
!
! #ifndef NOPREFETCH_RANGE_MIN
! #define NOPREFETCH_RANGE_MIN (-16)
! #endif
! #ifndef NOPREFETCH_RANGE_MAX
! #define NOPREFETCH_RANGE_MAX 32
! #endif
!
! static void rtl_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 rtl_values_to_profile (rtx, histogram_values *);
! static rtx rtl_divmod_fixed_value (enum machine_mode, enum rtx_code, rtx, rtx,
! rtx, gcov_type, int);
! static rtx rtl_mod_pow2 (enum machine_mode, enum rtx_code, rtx, rtx, rtx, int);
! static rtx rtl_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 rtl_divmod_fixed_value_transform (rtx);
! static bool rtl_mod_pow2_value_transform (rtx);
! static bool rtl_mod_subtract_transform (rtx);
! #ifdef HAVE_prefetch
! static bool speculative_prefetching_transform (rtx);
! #endif
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);
--- 70,81 ----
compilation with -fprofile-arcs. The optimization may then read the
gathered data in the second compilation with -fbranch-probabilities.
! 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. */
!
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 bool tree_divmod_fixed_value_tran
*** 139,1027 ****
static bool tree_mod_pow2_value_transform (tree);
static bool tree_mod_subtract_transform (tree);
-
- /* Find values inside INSN for that we want to measure histograms for
- division/modulo optimization and stores them to VALUES. */
- static void
- rtl_divmod_values_to_profile (rtx insn, histogram_values *values)
- {
- rtx set, set_src, op1, op2;
- enum machine_mode mode;
- histogram_value hist;
-
- if (!INSN_P (insn))
- return;
-
- set = single_set (insn);
- if (!set)
- return;
-
- mode = GET_MODE (SET_DEST (set));
- if (!INTEGRAL_MODE_P (mode))
- return;
-
- 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->hvalue.rtl.value = op2;
- hist->hvalue.rtl.seq = NULL_RTX;
- hist->hvalue.rtl.mode = mode;
- hist->hvalue.rtl.insn = insn;
- hist->type = HIST_TYPE_POW2;
- VEC_safe_push (histogram_value, heap, *values, hist);
- }
-
- /* Check whether the divisor is not in fact a constant. */
- if (!CONSTANT_P (op2))
- {
- hist = ggc_alloc (sizeof (*hist));
- hist->hvalue.rtl.value = op2;
- hist->hvalue.rtl.mode = mode;
- hist->hvalue.rtl.seq = NULL_RTX;
- hist->hvalue.rtl.insn = insn;
- hist->type = HIST_TYPE_SINGLE_VALUE;
- VEC_safe_push (histogram_value, heap, *values, 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->hvalue.rtl.value = force_operand (tmp, NULL_RTX);
- hist->hvalue.rtl.seq = get_insns ();
- end_sequence ();
- hist->hvalue.rtl.mode = mode;
- hist->hvalue.rtl.insn = insn;
- hist->type = HIST_TYPE_INTERVAL;
- hist->hdata.intvl.int_start = 0;
- hist->hdata.intvl.steps = 2;
- VEC_safe_push (histogram_value, heap, *values, hist);
- }
- return;
-
- default:
- return;
- }
- }
-
- #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. */
-
- static int
- find_mem_reference_1 (rtx *expr, void *ret)
- {
- rtx *mem = ret;
-
- if (MEM_P (*expr))
- {
- *mem = *expr;
- return 1;
- }
- return 0;
- }
-
- /* 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. */
-
- static int fmr2_write;
- static void
- find_mem_reference_2 (rtx expr, rtx pat ATTRIBUTE_UNUSED, void *mem)
- {
- if (expr == mem)
- fmr2_write = true;
- }
-
- /* 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. */
-
- static bool
- find_mem_reference (rtx insn, rtx *mem, int *write)
- {
- *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;
- }
-
- /* 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. */
-
- static bool
- insn_prefetch_values_to_profile (rtx insn, histogram_values* values)
- {
- rtx mem, address;
- int write;
- histogram_value hist;
-
- /* It only makes sense to look for memory references in ordinary insns. */
- if (!NONJUMP_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 = ggc_alloc (sizeof (*hist));
- hist->hvalue.rtl.value = address;
- hist->hvalue.rtl.mode = GET_MODE (address);
- hist->hvalue.rtl.seq = NULL_RTX;
- hist->hvalue.rtl.insn = insn;
- hist->type = HIST_TYPE_CONST_DELTA;
- VEC_safe_push (histogram_value, heap, *values, hist);
-
- 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
- rtl_values_to_profile (rtx insn, histogram_values *values)
- {
- if (flag_value_profile_transformations)
- rtl_divmod_values_to_profile (insn, values);
-
- #ifdef HAVE_prefetch
- if (flag_speculative_prefetching)
- insn_prefetch_values_to_profile (insn, values);
- #endif
- }
-
- /* 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, libcall_level;
- histogram_value hist;
-
- life_analysis (NULL, PROP_DEATH_NOTES);
-
- *values = NULL;
- libcall_level = 0;
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- rtl_values_to_profile (insn, values);
- static_values = *values;
-
- for (i = 0; VEC_iterate (histogram_value, *values, i, hist); i++)
- {
- switch (hist->type)
- {
- case HIST_TYPE_INTERVAL:
- if (dump_file)
- fprintf (dump_file,
- "Interval counter for insn %d, range %d -- %d.\n",
- INSN_UID ((rtx)hist->hvalue.rtl.insn),
- hist->hdata.intvl.int_start,
- (hist->hdata.intvl.int_start
- + hist->hdata.intvl.steps - 1));
- hist->n_counters = hist->hdata.intvl.steps + 2;
- break;
-
- case HIST_TYPE_POW2:
- if (dump_file)
- fprintf (dump_file,
- "Pow2 counter for insn %d.\n",
- INSN_UID ((rtx)hist->hvalue.rtl.insn));
- hist->n_counters = 2;
- break;
-
- case HIST_TYPE_SINGLE_VALUE:
- if (dump_file)
- fprintf (dump_file,
- "Single value counter for insn %d.\n",
- INSN_UID ((rtx)hist->hvalue.rtl.insn));
- hist->n_counters = 3;
- break;
-
- case HIST_TYPE_CONST_DELTA:
- if (dump_file)
- fprintf (dump_file,
- "Constant delta counter for insn %d.\n",
- INSN_UID ((rtx)hist->hvalue.rtl.insn));
- hist->n_counters = 4;
- break;
-
- default:
- gcc_unreachable ();
- }
- }
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
- }
-
- /* 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)
-
- 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
-
- where a is almost always less then b and the division is unsigned
- TODO -- handle signed case as well
-
- x = a;
- if (x >= b)
- x %= b;
-
- 4)
-
- x = a % b
-
- where a is almost always less then 2 * b and the division is unsigned
- TODO -- handle signed case as well
-
- x = a;
- if (x >= b)
- x -= b;
- if (x >= b)
- x %= b;
-
- It would be possible to continue analogically for K * b for other small
- K's, but it is probably not useful.
-
- 5)
-
- 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).
-
- 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
-
- TODO:
-
- 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. */
-
- static bool
- rtl_value_profile_transformations (void)
- {
- rtx insn, next;
- int changed = false;
-
- for (insn = get_insns (); insn; insn = next)
- {
- 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;
-
- /* Ignore cold areas -- we are growing a code. */
- if (!maybe_hot_bb_p (BLOCK_FOR_INSN (insn)))
- continue;
-
- if (dump_file)
- {
- fprintf (dump_file, "Trying transformations on insn %d\n",
- INSN_UID (insn));
- print_rtl_single (dump_file, insn);
- }
-
- /* Transformations: */
- if (flag_value_profile_transformations
- && (rtl_mod_subtract_transform (insn)
- || rtl_divmod_fixed_value_transform (insn)
- || rtl_mod_pow2_value_transform (insn)))
- changed = true;
- #ifdef HAVE_prefetch
- if (flag_speculative_prefetching
- && speculative_prefetching_transform (insn))
- changed = true;
- #endif
- }
-
- if (changed)
- {
- commit_edge_insertions ();
- allocate_reg_info (max_reg_num (), FALSE, FALSE);
- }
-
- 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
- rtl_divmod_fixed_value (enum machine_mode mode, enum rtx_code operation,
- rtx target, rtx op1, rtx op2, gcov_type value,
- int prob)
- {
- rtx tmp, tmp1, 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;
-
- 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);
-
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
- }
-
- /* Do transform 1) on INSN if applicable. */
- static bool
- rtl_divmod_fixed_value_transform (rtx insn)
- {
- rtx set, set_src, set_dest, op1, op2, value, histogram;
- enum rtx_code code;
- enum machine_mode mode;
- gcov_type val, count, all;
- edge e;
- int prob;
-
- set = single_set (insn);
- if (!set)
- return false;
-
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
-
- if (code != DIV && code != MOD && code != UDIV && code != UMOD)
- 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;
-
- 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));
-
- /* We require that count be 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)
- return false;
-
- if (dump_file)
- fprintf (dump_file, "Div/mod by constant transformation on insn %d\n",
- INSN_UID (insn));
-
- /* 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 (
- rtl_divmod_fixed_value (mode, code, set_dest,
- op1, op2, val, prob), e);
-
- 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
- rtl_mod_pow2 (enum machine_mode mode, enum rtx_code operation, rtx target,
- rtx op1, rtx op2, int prob)
- {
- 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);
-
- emit_label (end_label);
-
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
- }
-
- /* Do transform 2) on INSN if applicable. */
- static bool
- rtl_mod_pow2_value_transform (rtx insn)
- {
- 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 all, prob;
-
- set = single_set (insn);
- if (!set)
- return false;
-
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
-
- if (code != UMOD)
- 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;
-
- 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);
- count = INTVAL (XEXP (histogram, 0));
-
- 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)
- return false;
-
- if (dump_file)
- fprintf (dump_file, "Mod power of 2 transformation on insn %d\n",
- INSN_UID (insn));
-
- /* Compute probability of taking the optimal path. */
- all = count + wrong_values;
- 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 (
- rtl_mod_pow2 (mode, code, set_dest, op1, op2, prob), e);
-
- 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
- rtl_mod_subtract (enum machine_mode mode, enum rtx_code operation,
- rtx target, rtx op1, rtx op2, int sub, int prob1, int prob2)
- {
- rtx tmp, tmp1, jump;
- rtx end_label = gen_label_rtx ();
- rtx sequence;
- int i;
-
- start_sequence ();
-
- if (!REG_P (op2))
- {
- tmp = gen_reg_rtx (mode);
- emit_move_insn (tmp, copy_rtx (op2));
- }
- 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);
-
- /* 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));
-
- for (i = 0; i < sub; i++)
- {
- 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));
- }
-
- 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);
-
- sequence = get_insns ();
- end_sequence ();
- rebuild_jump_labels (sequence);
- return sequence;
- }
-
- /* Do transforms 3) and 4) on INSN if applicable. */
- static bool
- rtl_mod_subtract_transform (rtx insn)
- {
- rtx set, set_src, set_dest, op1, op2, 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)
- return false;
-
- set_src = SET_SRC (set);
- set_dest = SET_DEST (set);
- code = GET_CODE (set_src);
- mode = GET_MODE (set_dest);
-
- if (code != UMOD)
- 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;
-
- if (!histogram)
- return false;
-
- histogram = XEXP (XEXP (histogram, 0), 1);
- histogram = XEXP (histogram, 1);
-
- all = 0;
- for (i = 0; i < 2; i++)
- {
- counts[i] = INTVAL (XEXP (histogram, 0));
- all += counts[i];
- histogram = XEXP (histogram, 1);
- }
- 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++)
- {
- count += counts[i];
- if (count * 2 >= all)
- break;
- }
-
- if (i == 2)
- return false;
-
- if (dump_file)
- fprintf (dump_file, "Mod subtract transformation on insn %d\n",
- INSN_UID (insn));
-
- /* 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;
-
- e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
- delete_insn (insn);
-
- insert_insn_on_edge (
- rtl_mod_subtract (mode, code, set_dest,
- op1, op2, i, prob1, prob2), e);
-
- 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 rtx
- gen_speculative_prefetch (rtx address, gcov_type delta, int write)
- {
- 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);
- else
- {
- tmp = simplify_gen_binary (PLUS, Pmode,
- copy_rtx (address), GEN_INT (delta));
- tmp = force_operand (tmp, NULL);
- }
- 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;
- }
-
- /* Do transform 5) on INSN if applicable. */
-
- static bool
- speculative_prefetching_transform (rtx insn)
- {
- rtx histogram, value;
- gcov_type val, count, all;
- edge e;
- rtx mem, address;
- int write;
-
- if (!maybe_hot_bb_p (BLOCK_FOR_INSN (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;
-
- 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;
-
- if (!histogram)
- 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)
- return false;
-
- /* We require that count be 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 (address, value) || 2 * count < all)
- 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)
- return false;
-
- 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));
-
- insert_insn_on_edge (gen_speculative_prefetch (address, val, write), e);
-
- return true;
- }
- #endif /* HAVE_prefetch */
-
/* Tree based transformations. */
static bool
tree_value_profile_transformations (void)
--- 85,90 ----
*************** tree_value_profile_transformations (void
*** 1070,1077 ****
/* Free extra storage from compute_value_histograms. */
while (th)
{
! free (th->hvalue.tree.counters);
! th = th->hvalue.tree.next;
}
ann->histograms = 0;
}
--- 133,140 ----
/* Free extra storage from compute_value_histograms. */
while (th)
{
! free (th->hvalue.counters);
! th = th->hvalue.next;
}
ann->histograms = 0;
}
*************** tree_divmod_fixed_value_transform (tree
*** 1205,1221 ****
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.tree.next)
if (histogram->type == HIST_TYPE_SINGLE_VALUE)
break;
if (!histogram)
return false;
! value = histogram->hvalue.tree.value;
! val = histogram->hvalue.tree.counters[0];
! count = histogram->hvalue.tree.counters[1];
! all = histogram->hvalue.tree.counters[2];
/* We require that count is at least half of all; this means
that for the transformation to fire the value must be constant
--- 268,284 ----
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.next)
if (histogram->type == HIST_TYPE_SINGLE_VALUE)
break;
if (!histogram)
return false;
! value = histogram->hvalue.value;
! val = histogram->hvalue.counters[0];
! count = histogram->hvalue.counters[1];
! all = histogram->hvalue.counters[2];
/* We require that count is at least half of all; this means
that for the transformation to fire the value must be constant
*************** tree_mod_pow2_value_transform (tree stmt
*** 1371,1386 ****
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.tree.next)
if (histogram->type == HIST_TYPE_POW2)
break;
if (!histogram)
return false;
! value = histogram->hvalue.tree.value;
! wrong_values = histogram->hvalue.tree.counters[0];
! count = histogram->hvalue.tree.counters[1];
/* 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)
--- 434,449 ----
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.next)
if (histogram->type == HIST_TYPE_POW2)
break;
if (!histogram)
return false;
! value = histogram->hvalue.value;
! wrong_values = histogram->hvalue.counters[0];
! count = histogram->hvalue.counters[1];
/* 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)
*************** tree_mod_subtract_transform (tree stmt)
*** 1551,1571 ****
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.tree.next)
if (histogram->type == HIST_TYPE_INTERVAL)
break;
if (!histogram)
return false;
! value = histogram->hvalue.tree.value;
all = 0;
wrong_values = 0;
for (i = 0; i < histogram->hdata.intvl.steps; i++)
! all += histogram->hvalue.tree.counters[i];
! wrong_values += histogram->hvalue.tree.counters[i];
! wrong_values += histogram->hvalue.tree.counters[i+1];
all += wrong_values;
/* We require that we use just subtractions in at least 50% of all
--- 614,634 ----
if (!ann->histograms)
return false;
! for (histogram = ann->histograms; histogram; histogram = histogram->hvalue.next)
if (histogram->type == HIST_TYPE_INTERVAL)
break;
if (!histogram)
return false;
! value = histogram->hvalue.value;
all = 0;
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];
all += wrong_values;
/* We require that we use just subtractions in at least 50% of all
*************** tree_mod_subtract_transform (tree stmt)
*** 1573,1579 ****
count = 0;
for (i = 0; i < histogram->hdata.intvl.steps; i++)
{
! count += histogram->hvalue.tree.counters[i];
if (count * 2 >= all)
break;
}
--- 636,642 ----
count = 0;
for (i = 0; i < histogram->hdata.intvl.steps; i++)
{
! count += histogram->hvalue.counters[i];
if (count * 2 >= all)
break;
}
*************** tree_mod_subtract_transform (tree stmt)
*** 1587,1608 ****
}
/* Compute probability of taking the optimal path(s). */
! prob1 = (histogram->hvalue.tree.counters[0] * REG_BR_PROB_BASE + all / 2) / all;
! prob2 = (histogram->hvalue.tree.counters[1] * REG_BR_PROB_BASE + all / 2) / all;
/* 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,
! histogram->hvalue.tree.counters[0],
! histogram->hvalue.tree.counters[1], all);
TREE_OPERAND (modify, 1) = result;
return true;
}
!
! /* Connection to the outside world. */
! /* Struct for IR-dependent hooks. */
struct value_prof_hooks {
/* Find list of values for which we want to measure histograms. */
void (*find_values_to_profile) (histogram_values *);
--- 650,669 ----
}
/* Compute probability of taking the optimal path(s). */
! prob1 = (histogram->hvalue.counters[0] * REG_BR_PROB_BASE + all / 2) / all;
! prob2 = (histogram->hvalue.counters[1] * REG_BR_PROB_BASE + all / 2) / all;
/* 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,
! histogram->hvalue.counters[0],
! histogram->hvalue.counters[1], all);
TREE_OPERAND (modify, 1) = result;
return true;
}
!
struct value_prof_hooks {
/* Find list of values for which we want to measure histograms. */
void (*find_values_to_profile) (histogram_values *);
*************** struct value_prof_hooks {
*** 1611,1630 ****
statically. See value-prof.c for more detail. */
bool (*value_profile_transformations) (void);
};
-
- /* Hooks for RTL-based versions (the only ones that currently work). */
- static struct value_prof_hooks rtl_value_prof_hooks =
- {
- rtl_find_values_to_profile,
- rtl_value_profile_transformations
- };
-
- void
- rtl_register_value_prof_hooks (void)
- {
- value_prof_hooks = &rtl_value_prof_hooks;
- gcc_assert (!ir_type ());
- }
/* Find values inside STMT for that we want to measure histograms for
division/modulo optimization. */
--- 672,677 ----
*************** tree_divmod_values_to_profile (tree stmt
*** 1662,1669 ****
/* Check for the case where the divisor is the same value most
of the time. */
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.tree.value = divisor;
! hist->hvalue.tree.stmt = stmt;
hist->type = HIST_TYPE_SINGLE_VALUE;
VEC_quick_push (histogram_value, *values, hist);
}
--- 709,716 ----
/* Check for the case where the divisor is the same value most
of the time. */
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.value = divisor;
! hist->hvalue.stmt = stmt;
hist->type = HIST_TYPE_SINGLE_VALUE;
VEC_quick_push (histogram_value, *values, hist);
}
*************** tree_divmod_values_to_profile (tree stmt
*** 1675,1688 ****
{
/* Check for a special case where the divisor is power of 2. */
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.tree.value = divisor;
! hist->hvalue.tree.stmt = stmt;
hist->type = HIST_TYPE_POW2;
VEC_quick_push (histogram_value, *values, hist);
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.tree.stmt = stmt;
! hist->hvalue.tree.value
= build2 (TRUNC_DIV_EXPR, type, op0, divisor);
hist->type = HIST_TYPE_INTERVAL;
hist->hdata.intvl.int_start = 0;
--- 722,735 ----
{
/* Check for a special case where the divisor is power of 2. */
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.value = divisor;
! hist->hvalue.stmt = stmt;
hist->type = HIST_TYPE_POW2;
VEC_quick_push (histogram_value, *values, hist);
hist = ggc_alloc (sizeof (*hist));
! hist->hvalue.stmt = stmt;
! hist->hvalue.value
= build2 (TRUNC_DIV_EXPR, type, op0, divisor);
hist->type = HIST_TYPE_INTERVAL;
hist->hdata.intvl.int_start = 0;
*************** tree_find_values_to_profile (histogram_v
*** 1728,1734 ****
if (dump_file)
{
fprintf (dump_file, "Interval counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.tree.stmt,
TDF_SLIM);
fprintf (dump_file, ", range %d -- %d.\n",
hist->hdata.intvl.int_start,
--- 775,781 ----
if (dump_file)
{
fprintf (dump_file, "Interval counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.stmt,
TDF_SLIM);
fprintf (dump_file, ", range %d -- %d.\n",
hist->hdata.intvl.int_start,
*************** tree_find_values_to_profile (histogram_v
*** 1742,1748 ****
if (dump_file)
{
fprintf (dump_file, "Pow2 counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.tree.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 2;
--- 789,795 ----
if (dump_file)
{
fprintf (dump_file, "Pow2 counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 2;
*************** tree_find_values_to_profile (histogram_v
*** 1752,1758 ****
if (dump_file)
{
fprintf (dump_file, "Single value counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.tree.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 3;
--- 799,805 ----
if (dump_file)
{
fprintf (dump_file, "Single value counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 3;
*************** tree_find_values_to_profile (histogram_v
*** 1762,1768 ****
if (dump_file)
{
fprintf (dump_file, "Constant delta counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.tree.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 4;
--- 809,815 ----
if (dump_file)
{
fprintf (dump_file, "Constant delta counter for tree ");
! print_generic_expr (dump_file, hist->hvalue.stmt, TDF_SLIM);
fprintf (dump_file, ".\n");
}
hist->n_counters = 4;
*************** value_profile_transformations (void)
*** 1801,1839 ****
return retval;
}
- static bool
- gate_handle_value_profile_transformations (void)
- {
- return flag_branch_probabilities
- && flag_profile_values
- && !flag_tree_based_profiling
- && (flag_value_profile_transformations
- || flag_speculative_prefetching);
- }
-
-
- /* Do optimizations based on expression value profiles. */
- static void
- rest_of_handle_value_profile_transformations (void)
- {
- if (value_profile_transformations ())
- cleanup_cfg (CLEANUP_EXPENSIVE);
- }
-
- struct tree_opt_pass pass_value_profile_transformations =
- {
- "vpt", /* name */
- gate_handle_value_profile_transformations, /* gate */
- rest_of_handle_value_profile_transformations, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_VPT, /* tv_id */
- 0, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- 0, /* todo_flags_start */
- TODO_dump_func, /* todo_flags_finish */
- 'V' /* letter */
- };
--- 848,851 ----
Index: value-prof.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/value-prof.h,v
retrieving revision 1.15
diff -c -3 -p -r1.15 value-prof.h
*** value-prof.h 25 Jun 2005 02:01:59 -0000 1.15
--- value-prof.h 28 Jul 2005 12:30:05 -0000
*************** enum hist_type
*** 40,61 ****
/* The value to measure. */
struct histogram_value_t
{
! union
{
! struct
! {
! rtx value; /* The value to profile. */
! rtx seq; /* Insns required to count the profiled value. */
! rtx insn; /* Insn before that to measure. */
! enum machine_mode mode; /* Mode of value to profile. */
! } rtl;
! struct
! {
! tree value; /* The value to profile. */
! tree stmt; /* Insn containing the value. */
! gcov_type *counters; /* Pointer to first counter. */
! struct histogram_value_t *next; /* Linked list pointer. */
! } tree;
} hvalue;
enum hist_type type; /* Type of information to measure. */
unsigned n_counters; /* Number of required counters. */
--- 40,51 ----
/* The value to measure. */
struct histogram_value_t
{
! struct
{
! tree value; /* The value to profile. */
! tree stmt; /* Insn containing the value. */
! gcov_type *counters; /* Pointer to first counter. */
! struct histogram_value_t *next; /* Linked list pointer. */
} hvalue;
enum hist_type type; /* Type of information to measure. */
unsigned n_counters; /* Number of required counters. */
*************** DEF_VEC_ALLOC_P(histogram_value,heap);
*** 77,83 ****
typedef VEC(histogram_value,heap) *histogram_values;
/* Hooks registration. */
- extern void rtl_register_value_prof_hooks (void);
extern void tree_register_value_prof_hooks (void);
/* IR-independent entry points. */
--- 67,72 ----
*************** extern void init_branch_prob (void);
*** 113,125 ****
extern void branch_prob (void);
extern void end_branch_prob (void);
extern void tree_register_profile_hooks (void);
- extern void rtl_register_profile_hooks (void);
/* In tree-profile.c. */
extern struct profile_hooks tree_profile_hooks;
- /* In rtl-profile.c. */
- extern struct profile_hooks rtl_profile_hooks;
-
#endif /* GCC_VALUE_PROF_H */
--- 102,110 ----
Index: doc/invoke.texi
===================================================================
RCS file: /cvs/gcc/gcc/gcc/doc/invoke.texi,v
retrieving revision 1.655
diff -c -3 -p -r1.655 invoke.texi
*** doc/invoke.texi 25 Jul 2005 19:42:02 -0000 1.655
--- doc/invoke.texi 28 Jul 2005 12:30:07 -0000
*************** Objective-C and Objective-C++ Dialects}.
*** 281,287 ****
-ftree-vectorizer-verbose=@var{n} @gol
-fdump-tree-storeccp@r{[}-@var{n}@r{]} @gol
-feliminate-dwarf2-dups -feliminate-unused-debug-types @gol
! -feliminate-unused-debug-symbols -fmem-report -fprofile-arcs -ftree-based-profiling @gol
-frandom-seed=@var{string} -fsched-verbose=@var{n} @gol
-ftest-coverage -ftime-report -fvar-tracking @gol
-g -g@var{level} -gcoff -gdwarf-2 @gol
--- 281,287 ----
-ftree-vectorizer-verbose=@var{n} @gol
-fdump-tree-storeccp@r{[}-@var{n}@r{]} @gol
-feliminate-dwarf2-dups -feliminate-unused-debug-types @gol
! -feliminate-unused-debug-symbols -fmem-report -fprofile-arcs @gol
-frandom-seed=@var{string} -fsched-verbose=@var{n} @gol
-ftest-coverage -ftime-report -fvar-tracking @gol
-g -g@var{level} -gcoff -gdwarf-2 @gol
*************** Objective-C and Objective-C++ Dialects}.
*** 325,331 ****
-fsched-stalled-insns=@var{n} -sched-stalled-insns-dep=@var{n} @gol
-fsched2-use-superblocks @gol
-fsched2-use-traces -freschedule-modulo-scheduled-loops @gol
! -fsignaling-nans -fsingle-precision-constant -fspeculative-prefetching @gol
-fstrength-reduce -fstrict-aliasing -ftracer -fthread-jumps @gol
-funroll-all-loops -funroll-loops -fpeel-loops @gol
-fsplit-ivs-in-unroller -funswitch-loops @gol
--- 325,331 ----
-fsched-stalled-insns=@var{n} -sched-stalled-insns-dep=@var{n} @gol
-fsched2-use-superblocks @gol
-fsched2-use-traces -freschedule-modulo-scheduled-loops @gol
! -fsignaling-nans -fsingle-precision-constant @gol
-fstrength-reduce -fstrict-aliasing -ftracer -fthread-jumps @gol
-funroll-all-loops -funroll-loops -fpeel-loops @gol
-fsplit-ivs-in-unroller -funswitch-loops @gol
*************** executed. When an arc is the only exit
*** 3579,3595 ****
instrumentation code can be added to the block; otherwise, a new basic
block must be created to hold the instrumentation code.
- @item -ftree-based-profiling
- @opindex ftree-based-profiling
- This option is used in addition to @option{-fprofile-arcs} or
- @option{-fbranch-probabilities} to control whether those optimizations
- are performed on a tree-based or rtl-based internal representation.
- If you use this option when compiling with @option{-fprofile-arcs},
- you must also use it when compiling later with @option{-fbranch-probabilities}.
- Currently the tree-based optimization is in an early stage of
- development, and this option is recommended only for those people
- working on improving it.
-
@need 2000
@item -ftest-coverage
@opindex ftest-coverage
--- 3579,3584 ----
*************** The following options are enabled: @code
*** 5328,5335 ****
Enable profile feedback directed optimizations, and optimizations
generally profitable only with profile feedback available.
! The following options are enabled: @code{-fbranch-probabilities},
! @code{-fvpt}, @code{-funroll-loops}, @code{-fpeel-loops}, @code{-ftracer}.
@end table
--- 5317,5325 ----
Enable profile feedback directed optimizations, and optimizations
generally profitable only with profile feedback available.
! The following options are enabled: @code{-fbranch-probabilities}, @code{-fvpt},
! @code{-funroll-loops}, @code{-fpeel-loops}, @code{-ftracer},
! @code{-fno-loop-optimize}.
@end table
*************** and actually performs the optimizations
*** 5526,5548 ****
Currently the optimizations include specialization of division operation
using the knowledge about the value of the denominator.
- @item -fspeculative-prefetching
- @opindex fspeculative-prefetching
- If combined with @option{-fprofile-arcs}, it instructs the compiler to add
- a code to gather information about addresses of memory references in the
- program.
-
- With @option{-fbranch-probabilities}, it reads back the data gathered
- and issues prefetch instructions according to them. In addition to the opportunities
- noticed by @option{-fprefetch-loop-arrays}, it also notices more complicated
- memory access patterns---for example accesses to the data stored in linked
- list whose elements are usually allocated sequentially.
-
- In order to prevent issuing double prefetches, usage of
- @option{-fspeculative-prefetching} implies @option{-fno-prefetch-loop-arrays}.
-
- Enabled with @option{-fprofile-generate} and @option{-fprofile-use}.
-
@item -frename-registers
@opindex frename-registers
Attempt to avoid false dependencies in scheduled code by making use
--- 5516,5521 ----
Index: testsuite/g++.dg/bprob/bprob.exp
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/g++.dg/bprob/bprob.exp,v
retrieving revision 1.11
diff -c -3 -p -r1.11 bprob.exp
*** testsuite/g++.dg/bprob/bprob.exp 25 Jun 2005 01:45:09 -0000 1.11
--- testsuite/g++.dg/bprob/bprob.exp 28 Jul 2005 12:30:08 -0000
*************** load_lib profopt.exp
*** 52,61 ****
set profile_options "-fprofile-arcs"
set feedback_options "-fbranch-probabilities"
- if {[check_profiling_available "-ftree-based-profiling"]} {
- lappend profile_options "-ftree-based-profiling -fprofile-arcs"
- lappend feedback_options "-ftree-based-profiling -fbranch-probabilities"
- }
# Main loop.
foreach profile_option $profile_options feedback_option $feedback_options {
--- 52,57 ----
Index: testsuite/gcc.dg/tree-prof/inliner-1.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/gcc.dg/tree-prof/inliner-1.c,v
retrieving revision 1.2
diff -c -3 -p -r1.2 inliner-1.c
*** testsuite/gcc.dg/tree-prof/inliner-1.c 27 Jul 2005 07:40:25 -0000 1.2
--- testsuite/gcc.dg/tree-prof/inliner-1.c 28 Jul 2005 12:30:09 -0000
***************
*** 1,4 ****
! /* { dg-options "-O2 -fdump-tree-optimized -fdump-tree-all" } */
int a;
int b[100];
void abort (void);
--- 1,4 ----
! /* { dg-options "-O2 -fdump-tree-optimized" } */
int a;
int b[100];
void abort (void);
Index: testsuite/gcc.dg/tree-prof/tree-prof.exp
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/gcc.dg/tree-prof/tree-prof.exp,v
retrieving revision 1.3
diff -c -3 -p -r1.3 tree-prof.exp
*** testsuite/gcc.dg/tree-prof/tree-prof.exp 25 Jun 2005 01:45:16 -0000 1.3
--- testsuite/gcc.dg/tree-prof/tree-prof.exp 28 Jul 2005 12:30:09 -0000
***************
*** 20,26 ****
load_lib target-supports.exp
# Some targets don't support tree profiling.
! if { ![check_profiling_available "-ftree-based-profiling"] } {
return
}
--- 20,26 ----
load_lib target-supports.exp
# Some targets don't support tree profiling.
! if { ![check_profiling_available ""] } {
return
}
*************** load_lib profopt.exp
*** 41,48 ****
# These are globals used by profopt-execute. The first is options
# needed to generate profile data, the second is options to use the
# profile data.
! set profile_option "-ftree-based-profiling -fprofile-generate"
! set feedback_option "-ftree-based-profiling -fprofile-use"
foreach src [lsort [glob -nocomplain $srcdir/$subdir/*.c]] {
# If we're only testing specific files and this isn't one of them, skip it.
--- 41,48 ----
# These are globals used by profopt-execute. The first is options
# needed to generate profile data, the second is options to use the
# profile data.
! set profile_option "-fprofile-generate"
! set feedback_option "-fprofile-use"
foreach src [lsort [glob -nocomplain $srcdir/$subdir/*.c]] {
# If we're only testing specific files and this isn't one of them, skip it.
Index: testsuite/gcc.misc-tests/bprob.exp
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/gcc.misc-tests/bprob.exp,v
retrieving revision 1.12
diff -c -3 -p -r1.12 bprob.exp
*** testsuite/gcc.misc-tests/bprob.exp 25 Jun 2005 01:45:17 -0000 1.12
--- testsuite/gcc.misc-tests/bprob.exp 28 Jul 2005 12:30:09 -0000
*************** load_lib profopt.exp
*** 49,58 ****
set profile_options "-fprofile-arcs"
set feedback_options "-fbranch-probabilities"
- if {[check_profiling_available "-ftree-based-profiling"]} {
- lappend profile_options "-ftree-based-profiling -fprofile-arcs"
- lappend feedback_options "-ftree-based-profiling -fbranch-probabilities"
- }
foreach profile_option $profile_options feedback_option $feedback_options {
foreach src [lsort [glob -nocomplain $srcdir/$subdir/bprob-*.c]] {
--- 49,54 ----