2006-02-14 14:51:51 +01:00
|
|
|
/* Array prefetching.
|
diagnostic.c: Don't include tm.h, tree.h, tm_p.h, langhooks.h or langhooks-def.h.
* diagnostic.c: Don't include tm.h, tree.h, tm_p.h, langhooks.h or
langhooks-def.h.
(diagnostic_initialize): Initialize x_data not last_function.
(diagnostic_report_current_function): Move to tree-diagnostic.c.
(default_diagnostic_starter): Call
diagnostic_report_current_module not
diagnostic_report_current_function.
(diagnostic_report_diagnostic): Initialize x_data not
abstract_origin.
(verbatim): Likewise.
* diagnostic.h (struct diagnostic_info): Change abstract_origin to
x_data.
(struct diagnostic_context): Change last_function to x_data.
(diagnostic_auxiliary_data): Replace with
diagnostic_context_auxiliary_data and
diagnostic_info_auxiliary_data.
(diagnostic_last_function_changed, diagnostic_set_last_function,
diagnostic_report_current_function): Move to tree-diagnostic.h.
(print_declaration, dump_generic_node, print_generic_stmt,
print_generic_stmt_indented, print_generic_expr,
print_generic_decl, debug_c_tree, dump_omp_clauses,
print_call_name, debug_generic_expr, debug_generic_stmt,
debug_tree_chain, default_tree_printer): Move to
tree-pretty-print.h.
(debug_gimple_stmt, debug_gimple_seq, print_gimple_seq,
print_gimple_stmt, print_gimple_expr, dump_gimple_stmt): Move to
gimple-pretty-print.h.
* pretty-print.c: Don't include tree.h
(pp_base_format): Don't handle %K here.
(pp_base_tree_identifier): Move to tree-pretty-print.c.
* pretty-print.h (text_info): Change abstract_origin to x_data.
(pp_tree_identifier, pp_unsupported_tree,
pp_base_tree_identifier): Move to tree-pretty-print.h.
* gimple-pretty-print.h, tree-diagnostic.c, tree-diagnostic.h,
tree-pretty-print.h: New files.
* tree-pretty-print.c: Include tree-pretty-print.h.
(percent_K_format): New. Moved from pretty-print.c.
(pp_base_tree_identifier): Move from pretty-print.c.
* c-objc-common.c: Include tree-pretty-print.h.
(c_tree_printer): Handle %K here.
* langhooks.c: Include tree-diagnostic.h.
(lhd_print_error_function): Use diagnostic_abstract_origin macro.
* toplev.c: Include tree-diagnostic.h and tree-pretty-print.h.
(default_tree_printer): Handle %K using percent_K_format.
(general_init): Use default_tree_diagnostic_starter.
* tree.c: Include tree-diagnostic.h and tree-pretty-print.h.
(free_lang_data): Use default_tree_diagnostic_starter.
* c-pretty-print.c: Include tree-pretty-print.h.
* cfgexpand.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* cgraphunit.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* dwarf2out.c: Include tree-pretty-print.h.
* except.c: Include tree-pretty-print.h.
* gimple-pretty-print.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* gimplify.c: Include tree-pretty-print.h.
* graphite-poly.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-cp.c: Include tree-pretty-print.h.
* ipa-inline.c: Include gimple-pretty-print.h.
* ipa-prop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-pure-const.c: Include gimple-pretty-print.h.
* ipa-struct-reorg.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-type-escape.c: Include tree-pretty-print.h.
* print-rtl.c: Include tree-pretty-print.h.
* print-tree.c: Include gimple-pretty-print.h.
* sese.c: Include tree-pretty-print.h.
* tree-affine.c: Include tree-pretty-print.h.
* tree-browser.c: Include tree-pretty-print.h.
* tree-call-cdce.c: Include gimple-pretty-print.h.
* tree-cfg.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-chrec.c: Include tree-pretty-print.h.
* tree-data-ref.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-dfa.c: Include tree-pretty-print.h.
* tree-if-conv.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-inline.c: Include tree-pretty-print.h.
* tree-into-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-nrv.c: Include tree-pretty-print.h.
* tree-object-size.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-outof-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-parloops.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-predcom.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-scalar-evolution.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-sra.c: Include tree-pretty-print.h.
* tree-ssa-address.c: Include tree-pretty-print.h.
* tree-ssa-alias.c: Include tree-pretty-print.h.
* tree-ssa-ccp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-coalesce.c: Include tree-pretty-print.h.
* tree-ssa-copy.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-copyrename.c: Include tree-pretty-print.h.
* tree-ssa-dce.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-dom.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-dse.c: Include gimple-pretty-print.h.
* tree-ssa-forwprop.c: Include tree-pretty-print.h.
* tree-ssa-ifcombine.c: Include tree-pretty-print.h.
* tree-ssa-live.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-im.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-ivcanon.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-ivopts.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-niter.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-prefetch.c: Include tree-pretty-print.h.
* tree-ssa-math-opts.c: Include gimple-pretty-print.h.
* tree-ssa-operands.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-phiprop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-pre.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-propagate.c: Include gimple-pretty-print.h.
* tree-ssa-reassoc.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-sccvn.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-sink.c: Include gimple-pretty-print.h.
* tree-ssa-ter.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-uninit.c: Include gimple-pretty-print.h.
* tree-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-stdarg.c: Include gimple-pretty-print.h.
* tree-switch-conversion.c: Include gimple-pretty-print.h.
* tree-tailcall.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-data-refs.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-loop-manip.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-loop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-patterns.c: Include gimple-pretty-print.h.
* tree-vect-slp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-stmts.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vectorizer.c: Include tree-pretty-print.h.
* tree-vrp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* value-prof.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* var-tracking.c: Include tree-pretty-print.h.
* Makefile.in (OBJS-common): Add tree-diagnostic.o.
(tree-diagnostic.o): New dependencies.
(c-objc-common.o, c-pretty-print.o, langhooks.o, tree.o,
tree-inline.o, print-tree.o, stor-layout.o, tree-ssa-uninit.o,
tree-ssa.o, tree-into-ssa.o, tree-ssa-ter.o, tree-ssa-coalesce.o,
tree-outof-ssa.o, tree-ssa-forwprop.o, tree-ssa-phiprop.o,
tree-ssa-ifcombine.o, tree-nrv.o, tree-ssa-copy.o,
tree-ssa-propagate.o, tree-ssa-dom.o, tree-ssa-uncprop.o,
tree-ssa-live.o, tree-ssa-copyrename.o, tree-ssa-pre.o,
tree-ssa-sccvn.o, tree-vrp.o, tree-cfg.o, tree-tailcall.o,
tree-ssa-sink.o, tree-if-conv.o, tree-dfa.o, tree-ssa-operands.o,
tree-ssa-address.o, tree-ssa-loop-niter.o,
tree-ssa-loop-ivcanon.o, tree-ssa-loop-prefetch.o, tree-predcom.o,
tree-ssa-loop-ivopts.o, tree-affine.o, tree-ssa-loop-im.o,
tree-ssa-math-opts.o, tree-ssa-alias.o, tree-ssa-reassoc.o,
gimplify.o, tree-browser.o, tree-chrec.o, tree-scalar-evolution.o,
tree-data-ref.o, sese.o, graphite-poly.o, tree-vect-loop.o,
tree-vect-loop-manip.o, tree-vect-patterns.o, tree-vect-slp.o,
tree-vect-stmts.o, tree-vect-data-refs.o, tree-vectorizer.o,
tree-parloops.o, tree-stdarg.o, tree-object-size.o,
gimple-pretty-print.o, tree-pretty-print.o, diagnostic.o,
toplev.o, print-rtl.o, except.o, dwarf2out.o, cgraphunit.o,
ipa-prop.o, ipa-cp.o, ipa-inline.o, ipa-pure-const.o,
ipa-type-escape.o, ipa-struct-reorg.o, tree-ssa-dce.o,
tree-call-cdce.o, tree-ssa-ccp.o, tree-sra.o,
tree-switch-conversion.o, var-tracking.o, value-prof.o,
cfgexpand.o, pretty-print.o): Update dependencies.
cp:
* error.c: Include tree-diagnostic.h and tree-pretty-print.h.
(cp_print_error_function): Use diagnostic_abstract_origin macro.
(cp_printer): Handle %K here using percent_K_format.
* cxx-pretty-print.c: Include tree-pretty-print.h.
* Make-lang.in (cp/error.o, cp/cxx-pretty-print.o): Update
dependencies.
From-SVN: r159685
2010-05-22 00:34:26 +02:00
|
|
|
Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
This file is part of GCC.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
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
|
2007-07-26 10:37:01 +02:00
|
|
|
Free Software Foundation; either version 3, or (at your option) any
|
2006-02-14 14:51:51 +01:00
|
|
|
later version.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
|
|
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
|
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
|
|
for more details.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
You should have received a copy of the GNU General Public License
|
2007-07-26 10:37:01 +02:00
|
|
|
along with GCC; see the file COPYING3. If not see
|
|
|
|
<http://www.gnu.org/licenses/>. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
#include "config.h"
|
|
|
|
#include "system.h"
|
|
|
|
#include "coretypes.h"
|
|
|
|
#include "tm.h"
|
|
|
|
#include "tree.h"
|
|
|
|
#include "tm_p.h"
|
|
|
|
#include "basic-block.h"
|
|
|
|
#include "output.h"
|
diagnostic.c: Don't include tm.h, tree.h, tm_p.h, langhooks.h or langhooks-def.h.
* diagnostic.c: Don't include tm.h, tree.h, tm_p.h, langhooks.h or
langhooks-def.h.
(diagnostic_initialize): Initialize x_data not last_function.
(diagnostic_report_current_function): Move to tree-diagnostic.c.
(default_diagnostic_starter): Call
diagnostic_report_current_module not
diagnostic_report_current_function.
(diagnostic_report_diagnostic): Initialize x_data not
abstract_origin.
(verbatim): Likewise.
* diagnostic.h (struct diagnostic_info): Change abstract_origin to
x_data.
(struct diagnostic_context): Change last_function to x_data.
(diagnostic_auxiliary_data): Replace with
diagnostic_context_auxiliary_data and
diagnostic_info_auxiliary_data.
(diagnostic_last_function_changed, diagnostic_set_last_function,
diagnostic_report_current_function): Move to tree-diagnostic.h.
(print_declaration, dump_generic_node, print_generic_stmt,
print_generic_stmt_indented, print_generic_expr,
print_generic_decl, debug_c_tree, dump_omp_clauses,
print_call_name, debug_generic_expr, debug_generic_stmt,
debug_tree_chain, default_tree_printer): Move to
tree-pretty-print.h.
(debug_gimple_stmt, debug_gimple_seq, print_gimple_seq,
print_gimple_stmt, print_gimple_expr, dump_gimple_stmt): Move to
gimple-pretty-print.h.
* pretty-print.c: Don't include tree.h
(pp_base_format): Don't handle %K here.
(pp_base_tree_identifier): Move to tree-pretty-print.c.
* pretty-print.h (text_info): Change abstract_origin to x_data.
(pp_tree_identifier, pp_unsupported_tree,
pp_base_tree_identifier): Move to tree-pretty-print.h.
* gimple-pretty-print.h, tree-diagnostic.c, tree-diagnostic.h,
tree-pretty-print.h: New files.
* tree-pretty-print.c: Include tree-pretty-print.h.
(percent_K_format): New. Moved from pretty-print.c.
(pp_base_tree_identifier): Move from pretty-print.c.
* c-objc-common.c: Include tree-pretty-print.h.
(c_tree_printer): Handle %K here.
* langhooks.c: Include tree-diagnostic.h.
(lhd_print_error_function): Use diagnostic_abstract_origin macro.
* toplev.c: Include tree-diagnostic.h and tree-pretty-print.h.
(default_tree_printer): Handle %K using percent_K_format.
(general_init): Use default_tree_diagnostic_starter.
* tree.c: Include tree-diagnostic.h and tree-pretty-print.h.
(free_lang_data): Use default_tree_diagnostic_starter.
* c-pretty-print.c: Include tree-pretty-print.h.
* cfgexpand.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* cgraphunit.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* dwarf2out.c: Include tree-pretty-print.h.
* except.c: Include tree-pretty-print.h.
* gimple-pretty-print.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* gimplify.c: Include tree-pretty-print.h.
* graphite-poly.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-cp.c: Include tree-pretty-print.h.
* ipa-inline.c: Include gimple-pretty-print.h.
* ipa-prop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-pure-const.c: Include gimple-pretty-print.h.
* ipa-struct-reorg.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* ipa-type-escape.c: Include tree-pretty-print.h.
* print-rtl.c: Include tree-pretty-print.h.
* print-tree.c: Include gimple-pretty-print.h.
* sese.c: Include tree-pretty-print.h.
* tree-affine.c: Include tree-pretty-print.h.
* tree-browser.c: Include tree-pretty-print.h.
* tree-call-cdce.c: Include gimple-pretty-print.h.
* tree-cfg.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-chrec.c: Include tree-pretty-print.h.
* tree-data-ref.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-dfa.c: Include tree-pretty-print.h.
* tree-if-conv.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-inline.c: Include tree-pretty-print.h.
* tree-into-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-nrv.c: Include tree-pretty-print.h.
* tree-object-size.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-outof-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-parloops.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-predcom.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-scalar-evolution.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-sra.c: Include tree-pretty-print.h.
* tree-ssa-address.c: Include tree-pretty-print.h.
* tree-ssa-alias.c: Include tree-pretty-print.h.
* tree-ssa-ccp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-coalesce.c: Include tree-pretty-print.h.
* tree-ssa-copy.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-copyrename.c: Include tree-pretty-print.h.
* tree-ssa-dce.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-dom.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-dse.c: Include gimple-pretty-print.h.
* tree-ssa-forwprop.c: Include tree-pretty-print.h.
* tree-ssa-ifcombine.c: Include tree-pretty-print.h.
* tree-ssa-live.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-im.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-ivcanon.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-ivopts.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-niter.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-loop-prefetch.c: Include tree-pretty-print.h.
* tree-ssa-math-opts.c: Include gimple-pretty-print.h.
* tree-ssa-operands.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-phiprop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-pre.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-propagate.c: Include gimple-pretty-print.h.
* tree-ssa-reassoc.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-sccvn.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-sink.c: Include gimple-pretty-print.h.
* tree-ssa-ter.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-ssa-uninit.c: Include gimple-pretty-print.h.
* tree-ssa.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-stdarg.c: Include gimple-pretty-print.h.
* tree-switch-conversion.c: Include gimple-pretty-print.h.
* tree-tailcall.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-data-refs.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-loop-manip.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-loop.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-patterns.c: Include gimple-pretty-print.h.
* tree-vect-slp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vect-stmts.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* tree-vectorizer.c: Include tree-pretty-print.h.
* tree-vrp.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* value-prof.c: Include tree-pretty-print.h and
gimple-pretty-print.h.
* var-tracking.c: Include tree-pretty-print.h.
* Makefile.in (OBJS-common): Add tree-diagnostic.o.
(tree-diagnostic.o): New dependencies.
(c-objc-common.o, c-pretty-print.o, langhooks.o, tree.o,
tree-inline.o, print-tree.o, stor-layout.o, tree-ssa-uninit.o,
tree-ssa.o, tree-into-ssa.o, tree-ssa-ter.o, tree-ssa-coalesce.o,
tree-outof-ssa.o, tree-ssa-forwprop.o, tree-ssa-phiprop.o,
tree-ssa-ifcombine.o, tree-nrv.o, tree-ssa-copy.o,
tree-ssa-propagate.o, tree-ssa-dom.o, tree-ssa-uncprop.o,
tree-ssa-live.o, tree-ssa-copyrename.o, tree-ssa-pre.o,
tree-ssa-sccvn.o, tree-vrp.o, tree-cfg.o, tree-tailcall.o,
tree-ssa-sink.o, tree-if-conv.o, tree-dfa.o, tree-ssa-operands.o,
tree-ssa-address.o, tree-ssa-loop-niter.o,
tree-ssa-loop-ivcanon.o, tree-ssa-loop-prefetch.o, tree-predcom.o,
tree-ssa-loop-ivopts.o, tree-affine.o, tree-ssa-loop-im.o,
tree-ssa-math-opts.o, tree-ssa-alias.o, tree-ssa-reassoc.o,
gimplify.o, tree-browser.o, tree-chrec.o, tree-scalar-evolution.o,
tree-data-ref.o, sese.o, graphite-poly.o, tree-vect-loop.o,
tree-vect-loop-manip.o, tree-vect-patterns.o, tree-vect-slp.o,
tree-vect-stmts.o, tree-vect-data-refs.o, tree-vectorizer.o,
tree-parloops.o, tree-stdarg.o, tree-object-size.o,
gimple-pretty-print.o, tree-pretty-print.o, diagnostic.o,
toplev.o, print-rtl.o, except.o, dwarf2out.o, cgraphunit.o,
ipa-prop.o, ipa-cp.o, ipa-inline.o, ipa-pure-const.o,
ipa-type-escape.o, ipa-struct-reorg.o, tree-ssa-dce.o,
tree-call-cdce.o, tree-ssa-ccp.o, tree-sra.o,
tree-switch-conversion.o, var-tracking.o, value-prof.o,
cfgexpand.o, pretty-print.o): Update dependencies.
cp:
* error.c: Include tree-diagnostic.h and tree-pretty-print.h.
(cp_print_error_function): Use diagnostic_abstract_origin macro.
(cp_printer): Handle %K here using percent_K_format.
* cxx-pretty-print.c: Include tree-pretty-print.h.
* Make-lang.in (cp/error.o, cp/cxx-pretty-print.o): Update
dependencies.
From-SVN: r159685
2010-05-22 00:34:26 +02:00
|
|
|
#include "tree-pretty-print.h"
|
2006-02-14 14:51:51 +01:00
|
|
|
#include "tree-flow.h"
|
|
|
|
#include "tree-dump.h"
|
|
|
|
#include "timevar.h"
|
|
|
|
#include "cfgloop.h"
|
|
|
|
#include "tree-pass.h"
|
|
|
|
#include "insn-config.h"
|
|
|
|
#include "recog.h"
|
|
|
|
#include "hashtab.h"
|
|
|
|
#include "tree-chrec.h"
|
|
|
|
#include "tree-scalar-evolution.h"
|
2010-07-08 06:22:54 +02:00
|
|
|
#include "diagnostic-core.h"
|
2006-02-14 14:51:51 +01:00
|
|
|
#include "toplev.h"
|
|
|
|
#include "params.h"
|
|
|
|
#include "langhooks.h"
|
2007-01-28 18:40:38 +01:00
|
|
|
#include "tree-inline.h"
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
#include "tree-data-ref.h"
|
gimplify.c: Do not include except.h and optabs.h.
* gimplify.c: Do not include except.h and optabs.h.
(gimplify_body): Do not initialize RTL profiling.
* gimple-low.c: Do not include rtl.h, diagnostic.h, langhooks.h,
langhooks-def.h, timevar.h, except.h, hashtab.h, and expr.h.
* gimple-fold.c: Do not include rtl.h, tm_p.h, ggc.h, basic-block.h,
output.h, expr.h, diagnostic.h, timevar.h, value-prof.h, and
langhooks.h.
* tree-pretty-print.h: Include pretty-print.h.
* gimple-pretty-print.h: Include pretty-print.h.
* tree-pretty-print.c: Do not include diagnostic.h.
* tree-vrp.c: Likewise.
* tree-tailcall.c: Likewise
* tree-scalar-evolution.c: Likewise
* tree-ssa-dse.c: Likewise
* tree-chrec.c: Likewise
* tree-ssa-sccvn.c: Likewise
* tree-ssa-copyrename.c: Likewise
* tree-nomudflap.c: Likewise
* tree-call-cdce.c: Likewise
* tree-stdarg.c: Likewise
* tree-ssa-math-opts.c: Likewise
* tree-nrv.c: Likewise
* tree-ssa-sink.c: Likewise
* tree-browser.c: Likewise
* tree-ssa-loop-ivcanon.c: Likewise
* tree-ssa-loop.c: Likewise
* tree-parloops.c: Likewise
* tree-ssa-address.c: Likewise
* tree-ssa-ifcombine.c: Likewise
* tree-if-conv.c: Likewise
* tree-data-ref.c: Likewise
* tree-affine.c: Likewise
* tree-ssa-phiopt.c: Likewise
* tree-ssa-coalesce.c: Likewise
* tree-ssa-pre.c: Likewise
* tree-ssa-live.c: Likewise
* tree-predcom.c: Likewise
* tree-ssa-forwprop.c: Likewise
* tree-ssa-dce.c: Likewise
* tree-ssa-ter.c: Likewise
* tree-ssa-loop-prefetch.c: Likewise
* tree-optimize.c: Likewise
* tree-ssa-phiprop.c: Likewise
* tree-object-size.c: Likewise
* tree-outof-ssa.c: Likewise
* tree-ssa-structalias.c: Likewise
* tree-switch-conversion.c: Likewise
* tree-ssa-reassoc.c: Likewise
* tree-ssa-operands.c: Likewise
* tree-vectorizer.c: Likewise
* tree-vect-data-refs.c: Likewise
* tree-vect-generic.c: Likewise
* tree-vect-stmts.c: Likewise
* tree-vect-patterns.c: Likewise
* tree-vect-slp.c: Likewise
* tree-vect-loop.c: Likewise
* tree-ssa-loop-ivopts.c: Likewise
* tree-ssa-loop-im.c: Likewise
* tree-ssa-loop-niter.c: Likewise
* tree-ssa-loop-unswitch.c: Likewise
* tree-ssa-loop-manip.c: Likewise
* tree-ssa-loop-ch.c: Likewise
* tree-dump.c: Likewise
* tree-complex.c: Likewise
* tree-into-ssa.c: Do not include diagnostic.h and expr.h.
* tree-ssa-uninit.c: Likewise
* tree-ssa-threadupdate.c: Likewise
* tree-ssa-uncprop.c: Likewise
* tree-ssa-ccp.c: Likewise
* tree-ssa-dom.c: Likewise
* tree-ssa-propagate.c: Likewise
* tree-ssa-alias.c: Likewise
* tree-dfa.c: Likewise
* tree-cfgcleanup.c: Likewise
* tree-sra.c: Likewise
* tree-ssa-copy.c: Likewise
* tree-ssa.c: Likewise
* tree-profile.c: Likewise
* tree-cfg.c: Likewise
* tree-ssa-threadedge.c: Likewise
* tree-vect-loop-manip.c: Likewise
* tree-inline.c: Do not include diagnostic.h and expr.h.
Include rtl.h.
(copy_decl_for_dup_finish): Do not use NULL_RTX.
* tree-loop-linear.c: Do not include diagnostic.h, expr.h,
and optabs.h.
* tree-loop-distribution.c: Likewise.
From-SVN: r160125
2010-06-02 00:00:56 +02:00
|
|
|
|
|
|
|
|
|
|
|
/* FIXME: Needed for optabs, but this should all be moved to a TBD interface
|
|
|
|
between the GIMPLE and RTL worlds. */
|
|
|
|
#include "expr.h"
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
#include "optabs.h"
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
/* This pass inserts prefetch instructions to optimize cache usage during
|
|
|
|
accesses to arrays in loops. It processes loops sequentially and:
|
|
|
|
|
|
|
|
1) Gathers all memory references in the single loop.
|
|
|
|
2) For each of the references it decides when it is profitable to prefetch
|
|
|
|
it. To do it, we evaluate the reuse among the accesses, and determines
|
|
|
|
two values: PREFETCH_BEFORE (meaning that it only makes sense to do
|
|
|
|
prefetching in the first PREFETCH_BEFORE iterations of the loop) and
|
|
|
|
PREFETCH_MOD (meaning that it only makes sense to prefetch in the
|
|
|
|
iterations of the loop that are zero modulo PREFETCH_MOD). For example
|
|
|
|
(assuming cache line size is 64 bytes, char has size 1 byte and there
|
|
|
|
is no hardware sequential prefetch):
|
|
|
|
|
|
|
|
char *a;
|
|
|
|
for (i = 0; i < max; i++)
|
|
|
|
{
|
|
|
|
a[255] = ...; (0)
|
|
|
|
a[i] = ...; (1)
|
|
|
|
a[i + 64] = ...; (2)
|
|
|
|
a[16*i] = ...; (3)
|
|
|
|
a[187*i] = ...; (4)
|
|
|
|
a[187*i + 50] = ...; (5)
|
|
|
|
}
|
|
|
|
|
|
|
|
(0) obviously has PREFETCH_BEFORE 1
|
|
|
|
(1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
|
|
|
|
location 64 iterations before it, and PREFETCH_MOD 64 (since
|
|
|
|
it hits the same cache line otherwise).
|
|
|
|
(2) has PREFETCH_MOD 64
|
|
|
|
(3) has PREFETCH_MOD 4
|
|
|
|
(4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
|
|
|
|
the cache line accessed by (4) is the same with probability only
|
|
|
|
7/32.
|
|
|
|
(5) has PREFETCH_MOD 1 as well.
|
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
Additionally, we use data dependence analysis to determine for each
|
|
|
|
reference the distance till the first reuse; this information is used
|
|
|
|
to determine the temporality of the issued prefetch instruction.
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
3) We determine how much ahead we need to prefetch. The number of
|
|
|
|
iterations needed is time to fetch / time spent in one iteration of
|
|
|
|
the loop. The problem is that we do not know either of these values,
|
|
|
|
so we just make a heuristic guess based on a magic (possibly)
|
|
|
|
target-specific constant and size of the loop.
|
|
|
|
|
|
|
|
4) Determine which of the references we prefetch. We take into account
|
|
|
|
that there is a maximum number of simultaneous prefetches (provided
|
|
|
|
by machine description). We prefetch as many prefetches as possible
|
|
|
|
while still within this bound (starting with those with lowest
|
|
|
|
prefetch_mod, since they are responsible for most of the cache
|
|
|
|
misses).
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
|
|
|
|
and PREFETCH_BEFORE requirements (within some bounds), and to avoid
|
|
|
|
prefetching nonaccessed memory.
|
|
|
|
TODO -- actually implement peeling.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
6) We actually emit the prefetch instructions. ??? Perhaps emit the
|
|
|
|
prefetch instructions with guards in cases where 5) was not sufficient
|
|
|
|
to satisfy the constraints?
|
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
A cost model is implemented to determine whether or not prefetching is
|
|
|
|
profitable for a given loop. The cost model has three heuristics:
|
|
|
|
|
|
|
|
1. Function trip_count_to_ahead_ratio_too_small_p implements a
|
|
|
|
heuristic that determines whether or not the loop has too few
|
|
|
|
iterations (compared to ahead). Prefetching is not likely to be
|
|
|
|
beneficial if the trip count to ahead ratio is below a certain
|
|
|
|
minimum.
|
|
|
|
|
|
|
|
2. Function mem_ref_count_reasonable_p implements a heuristic that
|
|
|
|
determines whether the given loop has enough CPU ops that can be
|
|
|
|
overlapped with cache missing memory ops. If not, the loop
|
|
|
|
won't benefit from prefetching. In the implementation,
|
|
|
|
prefetching is not considered beneficial if the ratio between
|
|
|
|
the instruction count and the mem ref count is below a certain
|
|
|
|
minimum.
|
|
|
|
|
|
|
|
3. Function insn_to_prefetch_ratio_too_small_p implements a
|
|
|
|
heuristic that disables prefetching in a loop if the prefetching
|
|
|
|
cost is above a certain limit. The relative prefetching cost is
|
|
|
|
estimated by taking the ratio between the prefetch count and the
|
|
|
|
total intruction count (this models the I-cache cost).
|
|
|
|
|
2009-06-08 18:00:13 +02:00
|
|
|
The limits used in these heuristics are defined as parameters with
|
2009-11-25 11:55:54 +01:00
|
|
|
reasonable default values. Machine-specific default values will be
|
2009-06-08 18:00:13 +02:00
|
|
|
added later.
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
Some other TODO:
|
|
|
|
-- write and use more general reuse analysis (that could be also used
|
|
|
|
in other cache aimed loop optimizations)
|
|
|
|
-- make it behave sanely together with the prefetches given by user
|
|
|
|
(now we just ignore them; at the very least we should avoid
|
|
|
|
optimizing loops in that user put his own prefetches)
|
|
|
|
-- we assume cache line size alignment of arrays; this could be
|
|
|
|
improved. */
|
|
|
|
|
|
|
|
/* Magic constants follow. These should be replaced by machine specific
|
|
|
|
numbers. */
|
|
|
|
|
|
|
|
/* True if write can be prefetched by a read prefetch. */
|
|
|
|
|
|
|
|
#ifndef WRITE_CAN_USE_READ_PREFETCH
|
|
|
|
#define WRITE_CAN_USE_READ_PREFETCH 1
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* True if read can be prefetched by a write prefetch. */
|
|
|
|
|
|
|
|
#ifndef READ_CAN_USE_WRITE_PREFETCH
|
|
|
|
#define READ_CAN_USE_WRITE_PREFETCH 0
|
|
|
|
#endif
|
|
|
|
|
params.c (set_param_value): Initialize the "set" field.
* params.c (set_param_value): Initialize the "set" field.
* params.h (struct param_info): Add "set" field.
(PARAM_SET_P): New macro.
(PREFETCH_LATENCY, SIMULTANEOUS_PREFETCHES, L1_CACHE_SIZE,
L1_CACHE_LINE_SIZE): New macros.
* toplev.c (DEFPARAM): Initialize the "set" field.
* tree-ssa-loop-prefetch.c (PREFETCH_LATENCY,
SIMULTANEOUS_PREFETCHES): Removed.
(PREFETCH_BLOCK): Use L1_CACHE_LINE_SIZE.
(tree_ssa_prefetch_arrays): Dump the values of the parameters.
* config/sparc/sparc.c: Include params.h.
(sparc_override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sparc/sparc.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.c: Include params.h.
(k8_cost): Change default value for SIMULTANEOUS_PREFETCHES.
(override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sh/sh.h (SIMULTANEOUS_PREFETCHES): Removed.
(OPTIMIZATION_OPTIONS): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.c (ia64_optimization_options): Set
SIMULTANEOUS_PREFETCHES and L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.h (SIMULTANEOUS_PREFETCHES, PREFETCH_BLOCK):
Removed.
* params.def (PARAM_PREFETCH_LATENCY, PARAM_SIMULTANEOUS_PREFETCHES,
PARAM_L1_CACHE_SIZE, PARAM_L1_CACHE_LINE_SIZE): New params.
* doc/invoke.texi: Document new params.
From-SVN: r118728
2006-11-12 20:17:02 +01:00
|
|
|
/* The size of the block loaded by a single prefetch. Usually, this is
|
|
|
|
the same as cache line size (at the moment, we only consider one level
|
|
|
|
of cache hierarchy). */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
#ifndef PREFETCH_BLOCK
|
params.c (set_param_value): Initialize the "set" field.
* params.c (set_param_value): Initialize the "set" field.
* params.h (struct param_info): Add "set" field.
(PARAM_SET_P): New macro.
(PREFETCH_LATENCY, SIMULTANEOUS_PREFETCHES, L1_CACHE_SIZE,
L1_CACHE_LINE_SIZE): New macros.
* toplev.c (DEFPARAM): Initialize the "set" field.
* tree-ssa-loop-prefetch.c (PREFETCH_LATENCY,
SIMULTANEOUS_PREFETCHES): Removed.
(PREFETCH_BLOCK): Use L1_CACHE_LINE_SIZE.
(tree_ssa_prefetch_arrays): Dump the values of the parameters.
* config/sparc/sparc.c: Include params.h.
(sparc_override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sparc/sparc.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.c: Include params.h.
(k8_cost): Change default value for SIMULTANEOUS_PREFETCHES.
(override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sh/sh.h (SIMULTANEOUS_PREFETCHES): Removed.
(OPTIMIZATION_OPTIONS): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.c (ia64_optimization_options): Set
SIMULTANEOUS_PREFETCHES and L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.h (SIMULTANEOUS_PREFETCHES, PREFETCH_BLOCK):
Removed.
* params.def (PARAM_PREFETCH_LATENCY, PARAM_SIMULTANEOUS_PREFETCHES,
PARAM_L1_CACHE_SIZE, PARAM_L1_CACHE_LINE_SIZE): New params.
* doc/invoke.texi: Document new params.
From-SVN: r118728
2006-11-12 20:17:02 +01:00
|
|
|
#define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
|
2006-02-14 14:51:51 +01:00
|
|
|
#endif
|
|
|
|
|
|
|
|
/* Do we have a forward hardware sequential prefetching? */
|
|
|
|
|
|
|
|
#ifndef HAVE_FORWARD_PREFETCH
|
|
|
|
#define HAVE_FORWARD_PREFETCH 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* Do we have a backward hardware sequential prefetching? */
|
|
|
|
|
|
|
|
#ifndef HAVE_BACKWARD_PREFETCH
|
|
|
|
#define HAVE_BACKWARD_PREFETCH 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* In some cases we are only able to determine that there is a certain
|
|
|
|
probability that the two accesses hit the same cache line. In this
|
|
|
|
case, we issue the prefetches for both of them if this probability
|
2008-06-06 07:42:00 +02:00
|
|
|
is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
#ifndef ACCEPTABLE_MISS_RATE
|
|
|
|
#define ACCEPTABLE_MISS_RATE 50
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifndef HAVE_prefetch
|
|
|
|
#define HAVE_prefetch 0
|
|
|
|
#endif
|
|
|
|
|
invoke.texi (l1-cache-size): Update documentation.
* doc/invoke.texi (l1-cache-size): Update documentation.
(l2-cache-size): Document.
* params.h (L2_CACHE_SIZE): New macro.
* tree-ssa-loop-prefetch.c (L1_CACHE_SIZE_BYTES): Reflect
that L1_CACHE_SIZE is in kB now.
(L2_CACHE_SIZE_BYTES): New macro.
(tree_ssa_prefetch_arrays): Show size in kB.
* config/i386/i386.h (struct processor_costs): Add l1_cache_size
and l2_cache_size fields.
* config/i386/driver-i386.c (describe_cache): Detect cache size
in kB.
* config/i386/i386.c (size_cost, i386_cost, i486_cost,pentium_cost,
pentiumpro_cost, geode_cost, k6_cost, athlon_cost, k8_cost,
amdfam10_cost, pentium4_cost, nocona_cost, core2_cost,
generic64_cost, generic32_cost): Add l1_cache_size and l2_cache_size.
(override_options): Set l1-cache-size and l2-cache-size to default
values if not specified otherwise.
* params.def (PARAM_L1_CACHE_SIZE): Change to set in kB.
(PARAM_L2_CACHE_SIZE): New.
From-SVN: r127117
2007-08-01 12:57:03 +02:00
|
|
|
#define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
|
|
|
|
#define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
|
|
|
|
/* We consider a memory access nontemporal if it is not reused sooner than
|
|
|
|
after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
|
|
|
|
accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
|
|
|
so that we use nontemporal prefetches e.g. if single memory location
|
|
|
|
is accessed several times in a single iteration of the loop. */
|
|
|
|
#define NONTEMPORAL_FRACTION 16
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
/* In case we have to emit a memory fence instruction after the loop that
|
|
|
|
uses nontemporal stores, this defines the builtin to use. */
|
|
|
|
|
|
|
|
#ifndef FENCE_FOLLOWING_MOVNT
|
|
|
|
#define FENCE_FOLLOWING_MOVNT NULL_TREE
|
|
|
|
#endif
|
|
|
|
|
2010-05-07 19:26:02 +02:00
|
|
|
/* It is not profitable to prefetch when the trip count is not at
|
|
|
|
least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance.
|
|
|
|
For example, in a loop with a prefetch ahead distance of 10,
|
|
|
|
supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is
|
|
|
|
profitable to prefetch when the trip count is greater or equal to
|
|
|
|
40. In that case, 30 out of the 40 iterations will benefit from
|
|
|
|
prefetching. */
|
|
|
|
|
|
|
|
#ifndef TRIP_COUNT_TO_AHEAD_RATIO
|
|
|
|
#define TRIP_COUNT_TO_AHEAD_RATIO 4
|
|
|
|
#endif
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* The group of references between that reuse may occur. */
|
|
|
|
|
|
|
|
struct mem_ref_group
|
|
|
|
{
|
|
|
|
tree base; /* Base of the reference. */
|
2010-05-19 12:36:40 +02:00
|
|
|
tree step; /* Step of the reference. */
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref *refs; /* References in the group. */
|
|
|
|
struct mem_ref_group *next; /* Next group of references. */
|
|
|
|
};
|
|
|
|
|
|
|
|
/* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
|
|
|
|
|
|
|
|
#define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
|
|
|
|
|
2010-05-18 00:33:53 +02:00
|
|
|
/* Do not generate a prefetch if the unroll factor is significantly less
|
|
|
|
than what is required by the prefetch. This is to avoid redundant
|
2010-06-10 00:51:46 +02:00
|
|
|
prefetches. For example, when prefetch_mod is 16 and unroll_factor is
|
|
|
|
2, prefetching requires unrolling the loop 16 times, but
|
|
|
|
the loop is actually unrolled twice. In this case (ratio = 8),
|
2010-05-18 00:33:53 +02:00
|
|
|
prefetching is not likely to be beneficial. */
|
|
|
|
|
|
|
|
#ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO
|
2010-06-10 00:51:46 +02:00
|
|
|
#define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4
|
2010-05-18 00:33:53 +02:00
|
|
|
#endif
|
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Some of the prefetch computations have quadratic complexity. We want to
|
|
|
|
avoid huge compile times and, therefore, want to limit the amount of
|
|
|
|
memory references per loop where we consider prefetching. */
|
|
|
|
|
|
|
|
#ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP
|
|
|
|
#define PREFETCH_MAX_MEM_REFS_PER_LOOP 200
|
|
|
|
#endif
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* The memory reference. */
|
|
|
|
|
|
|
|
struct mem_ref
|
|
|
|
{
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple stmt; /* Statement in that the reference appears. */
|
2006-02-14 14:51:51 +01:00
|
|
|
tree mem; /* The reference. */
|
|
|
|
HOST_WIDE_INT delta; /* Constant offset of the reference. */
|
|
|
|
struct mem_ref_group *group; /* The group of references it belongs to. */
|
|
|
|
unsigned HOST_WIDE_INT prefetch_mod;
|
|
|
|
/* Prefetch only each PREFETCH_MOD-th
|
|
|
|
iteration. */
|
|
|
|
unsigned HOST_WIDE_INT prefetch_before;
|
|
|
|
/* Prefetch only first PREFETCH_BEFORE
|
|
|
|
iterations. */
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
unsigned reuse_distance; /* The amount of data accessed before the first
|
|
|
|
reuse of this value. */
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref *next; /* The next reference in the group. */
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
unsigned write_p : 1; /* Is it a write? */
|
|
|
|
unsigned independent_p : 1; /* True if the reference is independent on
|
|
|
|
all other references inside the loop. */
|
|
|
|
unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */
|
|
|
|
unsigned storent_p : 1; /* True if we changed the store to a
|
|
|
|
nontemporal one. */
|
2006-02-14 14:51:51 +01:00
|
|
|
};
|
|
|
|
|
cfgcleanup.c, [...]: Fix comment typos.
* cfgcleanup.c, config/dfp-bit.c, expr.c, fold-const.c,
jump.c, mips-tfile.c, omp-low.c, sched-int.h,
tree-ssa-loop-prefetch.c, tree-vrp.c: Fix comment typos.
From-SVN: r112971
2006-04-15 23:45:06 +02:00
|
|
|
/* Dumps information about reference REF to FILE. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static void
|
|
|
|
dump_mem_ref (FILE *file, struct mem_ref *ref)
|
|
|
|
{
|
|
|
|
fprintf (file, "Reference %p:\n", (void *) ref);
|
|
|
|
|
|
|
|
fprintf (file, " group %p (base ", (void *) ref->group);
|
|
|
|
print_generic_expr (file, ref->group->base, TDF_SLIM);
|
|
|
|
fprintf (file, ", step ");
|
2010-05-19 12:36:40 +02:00
|
|
|
if (cst_and_fits_in_hwi (ref->group->step))
|
|
|
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step));
|
|
|
|
else
|
|
|
|
print_generic_expr (file, ref->group->step, TDF_TREE);
|
2006-02-14 14:51:51 +01:00
|
|
|
fprintf (file, ")\n");
|
|
|
|
|
2006-11-11 11:15:18 +01:00
|
|
|
fprintf (file, " delta ");
|
2006-02-14 14:51:51 +01:00
|
|
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
|
|
|
|
fprintf (file, "\n");
|
|
|
|
|
|
|
|
fprintf (file, " %s\n", ref->write_p ? "write" : "read");
|
|
|
|
|
|
|
|
fprintf (file, "\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
|
|
|
|
exist. */
|
|
|
|
|
|
|
|
static struct mem_ref_group *
|
2010-05-19 12:36:40 +02:00
|
|
|
find_or_create_group (struct mem_ref_group **groups, tree base, tree step)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
struct mem_ref_group *group;
|
|
|
|
|
|
|
|
for (; *groups; groups = &(*groups)->next)
|
|
|
|
{
|
2010-05-19 12:36:40 +02:00
|
|
|
if (operand_equal_p ((*groups)->step, step, 0)
|
2006-02-14 14:51:51 +01:00
|
|
|
&& operand_equal_p ((*groups)->base, base, 0))
|
|
|
|
return *groups;
|
|
|
|
|
2010-05-19 12:36:40 +02:00
|
|
|
/* If step is an integer constant, keep the list of groups sorted
|
|
|
|
by decreasing step. */
|
|
|
|
if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step)
|
|
|
|
&& int_cst_value ((*groups)->step) < int_cst_value (step))
|
2006-02-14 14:51:51 +01:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
group = XNEW (struct mem_ref_group);
|
2006-02-14 14:51:51 +01:00
|
|
|
group->base = base;
|
|
|
|
group->step = step;
|
|
|
|
group->refs = NULL;
|
|
|
|
group->next = *groups;
|
|
|
|
*groups = group;
|
|
|
|
|
|
|
|
return group;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Records a memory reference MEM in GROUP with offset DELTA and write status
|
|
|
|
WRITE_P. The reference occurs in statement STMT. */
|
|
|
|
|
|
|
|
static void
|
2008-07-28 16:33:56 +02:00
|
|
|
record_ref (struct mem_ref_group *group, gimple stmt, tree mem,
|
2006-02-14 14:51:51 +01:00
|
|
|
HOST_WIDE_INT delta, bool write_p)
|
|
|
|
{
|
|
|
|
struct mem_ref **aref;
|
|
|
|
|
|
|
|
/* Do not record the same address twice. */
|
|
|
|
for (aref = &group->refs; *aref; aref = &(*aref)->next)
|
|
|
|
{
|
|
|
|
/* It does not have to be possible for write reference to reuse the read
|
|
|
|
prefetch, or vice versa. */
|
|
|
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
|
|
|
&& write_p
|
|
|
|
&& !(*aref)->write_p)
|
|
|
|
continue;
|
|
|
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
|
|
|
&& !write_p
|
|
|
|
&& (*aref)->write_p)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if ((*aref)->delta == delta)
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
(*aref) = XNEW (struct mem_ref);
|
2006-02-14 14:51:51 +01:00
|
|
|
(*aref)->stmt = stmt;
|
|
|
|
(*aref)->mem = mem;
|
|
|
|
(*aref)->delta = delta;
|
|
|
|
(*aref)->write_p = write_p;
|
|
|
|
(*aref)->prefetch_before = PREFETCH_ALL;
|
|
|
|
(*aref)->prefetch_mod = 1;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
(*aref)->reuse_distance = 0;
|
2006-02-14 14:51:51 +01:00
|
|
|
(*aref)->issue_prefetch_p = false;
|
|
|
|
(*aref)->group = group;
|
|
|
|
(*aref)->next = NULL;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
(*aref)->independent_p = false;
|
|
|
|
(*aref)->storent_p = false;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
dump_mem_ref (dump_file, *aref);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Release memory references in GROUPS. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
release_mem_refs (struct mem_ref_group *groups)
|
|
|
|
{
|
|
|
|
struct mem_ref_group *next_g;
|
|
|
|
struct mem_ref *ref, *next_r;
|
|
|
|
|
|
|
|
for (; groups; groups = next_g)
|
|
|
|
{
|
|
|
|
next_g = groups->next;
|
|
|
|
for (ref = groups->refs; ref; ref = next_r)
|
|
|
|
{
|
|
|
|
next_r = ref->next;
|
|
|
|
free (ref);
|
|
|
|
}
|
|
|
|
free (groups);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* A structure used to pass arguments to idx_analyze_ref. */
|
|
|
|
|
|
|
|
struct ar_data
|
|
|
|
{
|
|
|
|
struct loop *loop; /* Loop of the reference. */
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple stmt; /* Statement of the reference. */
|
2010-05-19 12:36:40 +02:00
|
|
|
tree *step; /* Step of the memory reference. */
|
2006-02-14 14:51:51 +01:00
|
|
|
HOST_WIDE_INT *delta; /* Offset of the memory reference. */
|
|
|
|
};
|
|
|
|
|
|
|
|
/* Analyzes a single INDEX of a memory reference to obtain information
|
|
|
|
described at analyze_ref. Callback for for_each_index. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
idx_analyze_ref (tree base, tree *index, void *data)
|
|
|
|
{
|
2007-06-06 22:18:47 +02:00
|
|
|
struct ar_data *ar_data = (struct ar_data *) data;
|
2006-02-14 14:51:51 +01:00
|
|
|
tree ibase, step, stepsize;
|
2010-05-19 12:36:40 +02:00
|
|
|
HOST_WIDE_INT idelta = 0, imult = 1;
|
2006-02-14 14:51:51 +01:00
|
|
|
affine_iv iv;
|
|
|
|
|
2010-07-05 14:25:20 +02:00
|
|
|
if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF)
|
2006-02-14 14:51:51 +01:00
|
|
|
return false;
|
|
|
|
|
2009-03-04 18:50:20 +01:00
|
|
|
if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
|
2010-05-19 12:36:40 +02:00
|
|
|
*index, &iv, true))
|
2006-02-14 14:51:51 +01:00
|
|
|
return false;
|
|
|
|
ibase = iv.base;
|
|
|
|
step = iv.step;
|
|
|
|
|
[multiple changes]
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
Zdenek Dvorak <dvorakz@suse.cz>
Richard Guenther <rguenther@suse.de>
Kaz Kojima <kkojima@gcc.gnu.org>
* tree-vrp.c (compare_values_warnv): Convert val2 to
the type of val1.
(extract_range_from_assert): Create
POINTER_PLUS_EXPR for pointer types.
(extract_range_from_binary_expr): Handle
only POINTER_PLUS_EXPR, MIN_EXPR, and MAX_EXPR
for pointer types.
* doc/c-tree.texi (POINTER_PLUS_EXPR): Document.
* tree-ssa-loop-niter.c (split_to_var_and_offset): Handle
POINTER_PLUS_EXPR as PLUS_EXPR.
(number_of_iterations_lt_to_ne):
For pointer types, use sizetype when
creating MINUS_EXPR/PLUS_EXPRs.
(assert_loop_rolls_lt): For pointer types, use sizetype when
creating MINUS_EXPR/PLUS_EXPRs.
(number_of_iterations_le): Likewise.
(expand_simple_operations): POINTER_PLUS_EXPR are simple also.
(derive_constant_upper_bound): Handle POINTER_PLUS_EXPR just
like PLUS_EXPR and MINUS_EXPR.
* tree-pretty-print.c (dump_generic_node): Handle
POINTER_PLUS_EXPR.
(op_prio): Likewise.
(op_symbol_1): Likewise.
* optabs.c (optab_for_tree_code): Likewise.
* tree-ssa-loop-manip.c (create_iv): Handle pointer base
specially.
* tree-tailcall.c (process_assignment): Mention
POINTER_PLUS_EXPR in a TODO comment.
* tree.c (build2_stat): Assert when trying to use PLUS_EXPR or
MINUS_EXPR with a pointer. Also assert for POINTER_PLUS_EXPR
not used with a pointer and an integer type.
* tree-scalar-evolution.c (add_to_evolution_1): Convert the
increment using chrec_convert_rhs instead of chrec_convert.
(follow_ssa_edge_in_rhs): Handle POINTER_PLUS_EXPR like
PLUS_EXPR except for the right hand side's type will be
sizetype.
(interpret_rhs_modify_stmt): Handle POINTER_PLUS_EXPR.
(fold_used_pointer_cast): Kill.
(pointer_offset_p): Kill.
(fold_used_pointer): Kill.
(pointer_used_p): Kill.
(analyze_scalar_evolution_1 <case GIMPLE_MODIFY_STMT>): Don't
call fold_used_pointer.
(instantiate_parameters_1): Convert the increment
using chrec_convert_rhs instead of chrec_convert.
Handle POINTER_PLUS_EXPR as PLUS_EXPR.
* builtins.c (get_pointer_alignment): Handle POINTER_PLUS_EXPR
instead of PLUS_EXPR.
(expand_builtin_strcat): Create a POINTER_PLUS_EXPR instead of
PLUS_EXPR for pointers.
(std_gimplify_va_arg_expr): Likewise.
(fold_builtin_memory_op): Likewise.
(fold_builtin_strstr): Likewise.
(fold_builtin_strchr): Likewise.
(fold_builtin_strrchr): Likewise.
(fold_builtin_strpbrk): Likewise.
(expand_builtin_memory_chk): Likewise.
(fold_builtin_memory_chk): Likewise.
(std_expand_builtin_va_start): Use
sizetype for the call to make_tree and then convert
to the pointer type.
(fold_builtin_memchr): Use POINTER_PLUS_EXPR
instead of PLUS_EXPR for adding to a pointer.
(std_gimplify_va_arg_expr): Use fold_build2 for
the creating of POINTER_PLUS_EXPR. For the BIT_AND_EXPR, cast
the operands to sizetype first and then cast the BIT_AND_EXPR
back to the pointer type.
* fold-const.c (build_range_check): Handle pointer types
specially.
(extract_array_ref): Look for POINTER_PLUS_EXPR instead
of PLUS_EXPR's. Make sure the offset is converted to
sizetype.
(try_move_mult_to_index): Strip the NOPs from the offset.
Remove code argument and replace all uses with PLUS_EXPR.
(fold_to_nonsharp_ineq_using_bound): Handle pointer types
specially. Don't use a pointer type for MINUS_EXPR.
(fold_unary): Handle for (T1)(X op Y),
only p+ as that is the only as that can be handled for
binary operators now.
(fold_binary <case POINTER_PLUS_EXPR>): Add folding of
POINTER_PLUS_EXPR.
<case PLUS_EXPR>: Add folding of PTR+INT into
PTR p+ INT.
Don't call try_move_mult_to_index.
<case MINUS_EXPR>: Fold (PTR0 p+ A) - (PTR1 p+ B)
into (PTR0 - PTR1) + (A - B). Fold (PTR0 p+ A) - PTR1 into
(PTR0 - PTR1) + A iff (PTR0 - PTR1) simplifies.
Don't call try_move_mult_to_index.
(tree_expr_nonnegative_warnv_p): Handle POINTER_PLUS_EXPR.
(tree_expr_nonzero_p): Likewise.
(fold_indirect_ref_1): Look at POINTER_PLUS_EXPR instead
of PLUS_EXPR for the complex expression folding.
* tree-chrec.c (chrec_fold_plus_poly_poly): If the
first chrec is a pointer type, then the second should
be sizetype and not the first's type.
For POINTER_PLUS_EXPR, use a different right hand side type.
Handle POINTER_PLUS_EXPR like PLUS_EXPR.
(chrec_fold_plus_1): For POINTER_PLUS_EXPR, use a
different right hand side type.
Handle POINTER_PLUS_EXPR like PLUS_EXPR.
(chrec_fold_plus): For pointer types, use POINTER_PLUS_EXPR
instead of PLUS_EXPR.
When either operand is zero, convert the other operand.
(chrec_apply): Use chrec_convert_rhs
on the argument x instead of chrec_convert.
(reset_evolution_in_loop): For pointer types, the new_evol
should be sizetype.
(convert_affine_scev): For POINTER_PLUS_EXPR, use a
different right hand side type.
Handle POINTER_PLUS_EXPR like PLUS_EXPR.
(chrec_convert_rhs): New function.
(chrec_convert_aggressive): For POINTER_PLUS_EXPR, use a
different right hand side type.
Handle POINTER_PLUS_EXPR like PLUS_EXPR.
* tree-chrec.h (chrec_convert_rhs): New prototype.
(build_polynomial_chrec): For pointer types, the right hand
* tree-ssa-ccp.c (maybe_fold_stmt_indirect): Look for
POINTER_PLUS_EXPR instead of PLUS_EXPR's.
Remove subtraction case as it is always addition now.
Make sure the offset is converted to sizetype.
(fold_stmt_r): Don't handle PLUS_EXPR/MINUS_EXPR specially.
Handle POINTER_PLUS_EXPR like PLUS_EXPR was handled before.
* tree-ssa-loop-ivopts.c (determine_base_object): Abort for
PLUS_EXPR in pointer type.
Handle POINTER_PLUS_EXPR.
(tree_to_aff_combination): Likewise.
(force_expr_to_var_cost): Likewise.
(force_expr_to_var_cost): Likewise. Create a POINTER_PLUS_EXPR
instead of PLUS_EXPR for pointers.
* c-format.c (check_format_arg): Handle POINTER_PLUS_EXPR
instead of PLUS_EXPR of pointer types.
* tree-stdarg.c (va_list_counter_bump): Handle POINTER_PLUS_EXPR
as PLUS_EXPR.
(check_va_list_escapes): Likewise.
(check_all_va_list_escapes): Likewise.
* dwarf2out.c (loc_descriptor_from_tree_1):
Handle POINT_PLUS_EXPR as a PLUS_EXPR.
* expr.c (expand_expr_real_1): Handle POINTER_PLUS_EXPR.
(string_constant): Likewise.
* tree-ssa-address.c (tree_mem_ref_addr): When adding
the offset to the base, use POINTER_PLUS_EXPR.
(add_to_parts): Convert the index to sizetype.
(create_mem_ref): Create A POINTER_PLUS_EXPR for the one case.
* matrix-reorg.c (collect_data_for_malloc_call): Stmt
will now only be either INDIRECT_REF and POINTER_PLUS_EXPR.
Offset only holds something for PLUS_EXPR.
(ssa_accessed_in_tree): Handle POINTER_PLUS_EXPR just as
a PLUS_EXPR.
(analyze_transpose): POINTER_PLUS_EXPR will only show up now
and not PLUS_EXPR.
(analyze_accesses_for_modify_stmt): Likewise.
Remove comment about the type being integral type as it is
wrong now.
(can_calculate_expr_before_stmt): Handle POINTER_PLUS_EXPR as
PLUS_EXPR.
(transform_access_sites): POINTER_PLUS_EXPR will only show up now
and not PLUS_EXPR.
Correct the type which the artimentic is done in (is now
sizetype).
Reindent one loop.
* tree-data-ref.c (split_constant_offset): Handle
POINTER_PLUS_EXPR
* tree-affine.c (tree_to_aff_combination): Likewise.
* c-typeck.c (build_unary_op): For pointers create the increment
as a sizetype. Create a POINTER_PLUS_EXPR instead of PLUS_EXPR
for pointers.
* gimplify.c (gimplify_self_mod_expr): Create a
POINTER_PLUS_EXPR instead of PLUS_EXPR for pointers.
(gimplify_omp_atomic_fetch_op): Handle POINTER_PLUS_EXPR.
* tree.def (POINTER_PLUS_EXPR): New tree code.
* tree-predcom.c (ref_at_iteration): If we have a pointer
type do the multiplication in sizetype.
* tree-mudflap.c (mf_xform_derefs_1): Create a
POINTER_PLUS_EXPR instead of PLUS_EXPR for pointers.
* tree-ssa-forwprop.c
(forward_propagate_addr_into_variable_array_index):
Don't expect there to be a cast for the index as that
does not exist anymore.
(forward_propagate_addr_expr_1): Check for POINTER_PLUS_EXPR
instead of PLUS_EXPR.
Don't check for the first operand of the POINTER_PLUS_EXPR
was the index as it cannot be.
Call forward_propagate_addr_into_variable_array_index with
the SSA_NAME instead of the statement.
* varasm.c (const_hash_1): Handle POINTER_PLUS_EXPR.
(compare_constant): Likewise.
(copy_constant): Likewise.
(compute_reloc_for_constant): Likewise.
(output_addressed_constants): Likewise.
(initializer_constant_valid_p): Likewise.
* tree-ssa.c (tree_ssa_useless_type_conversion_1):
Convert the MIN/MAX of the inner type to the outer
type before comparing them.
* tree-ssa-loop-prefetch.c (idx_analyze_ref): Handle
POINTER_PLUS_EXPR instead of PLUS_EXPR.
(issue_prefetch_ref): Create a POINTER_PLUS_EXPR instead
of PLUS_EXPR for pointers.
* tree-inline.c (estimate_num_insns_1): Handle
POINTER_PLUS_EXPR.
* tree-vect-transform.c (vect_create_addr_base_for_vector_ref):
Create a POINTER_PLUS_EXPR instead of PLUS_EXPR for pointers.
(bump_vector_ptr): Create a POINTER_PLUS_EXPR
instead of PLUS_EXPR for the pointer increment statement.
(vect_update_ivs_after_vectorizer): For pointer types, create
POINTER_PLUS_EXPR instead of PLUS_EXPR and also create
MULT_EXPR in sizetype.
(vect_gen_niters_for_prolog_loop): Add a cast when creating
byte_misalign.
* tree-object-size.c (plus_expr_object_size): Handle
POINTER_PLUS_EXPR instead of PLUS_EXPR. Removing all the extra
code which is trying to figure out which side is a pointer and
is the index.
(check_for_plus_in_loops_1): Likewise.
(check_for_plus_in_loops): Likewise.
* c-common.c (pointer_int_sum): Create a
POINTER_PLUS_EXPR instead of PLUS_EXPR for pointers.
* tree-ssa-structalias.c (handle_ptr_arith): Handle
only POINTER_PLUS_EXPR. Removing all the extra
code which is trying to figure out which side is a pointer and
is the index.
* tree-cfg.c (verify_expr): Add extra checking for pointers and
PLUS_EXPR and MINUS_EXPR.
Also add checking to make sure the operands of POINTER_PLUS_EXPR
are correct.
* config/frv/frv.c (frv_expand_builtin_va_start): Use sizetype
with make_tree, instead of a pointer type.
* config/s390/s390.c (s390_va_start): Use POINTER_PLUS_EXPR
for pointers instead of PLUS_EXPR.
(s390_gimplify_va_arg): Likewise.
* config/spu/spu.c (spu_va_start): Create POINTER_PLUS_EXPR
instead of PLUS_EXPR when doing addition on pointer
types. Use sizetype for the second operand.
(spu_gimplify_va_arg_expr): Likewise.
* config/sparc/sparc.c (sparc_gimplify_va_arg): Use
POINTER_PLUS_EXPR instead of PLUS_EXPR when the operand was
a pointer. Don't create a BIT_AND_EXPR for pointer types.
* config/i386/i386.c (ix86_va_start): Use POINTER_PLUS_EXPR
for the pointer addition and also use size_int/sizetype
for the offset.
(ix86_gimplify_va_arg): Likewise.
Perform BIT_AND_EXPR on sizetype arguments.
* config/sh/sh.c (sh_va_start): Call make_tree with sizetype
and convert its result to a pointer type. Use POINTER_PLUS_EXPR
for the pointer additions and also use size_int for the offsets.
(sh_gimplify_va_arg_expr): Use POINTER_PLUS_EXPR for the pointer
additions and also use size_int for the offsets. Perform
BIT_AND_EXPR on sizetype arguments.
* config/ia64/ia64.c (ia64_gimplify_va_arg): Use
POINTER_PLUS_EXPR for pointers and create the
BIT_AND_EXPR in sizetype.
* config/rs6000/rs6000.c (rs6000_va_start): Use POINTER_PLUS_EXPR
instead of PLUS_EXPR for pointer addition.
(rs6000_va_start): Likewise.
Also use sizetype for the offset.
* config/pa/pa.c (reloc_needed): Handle POINTER_PLUS_EXPR
as PLUS_EXPR/MINUS_EXPR.
(hppa_gimplify_va_arg_expr): Don't create MINUS_EXPR or
PLUS_EXPR for pointers, instead use POINTER_PLUS_EXPR.
Don't use BIT_AND_EXPR on a pointer type, convert the
expression to sizetype first.
* config/mips/mips.c (mips_va_start): Use POINTER_PLUS_EXPR
for pointers.
(mips_gimplify_va_arg_expr): Likewise.
Don't create BIT_AND_EXPR in a pointer type.
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
* trans-intrinsic.c (gfc_conv_intrinsic_repeat): Use
POINTER_PLUS_EXPR instead of PLUS_EXPR for pointer addition.
* trans-expr.c (gfc_trans_string_copy): Create
POINTER_PLUS_EXPR instead of a PLUS_EXPR
for pointer types.
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
* typeck.c (build_binary_op): For templates build the
expression in pieces to avoid the assert in build2_stat.
(get_member_function_from_ptrfunc):
Change over to using POINTER_PLUS_EXPR and convert
the second operand to sizetype.
* typeck2.c (build_m_component_ref): Likewise.
* init.c (expand_virtual_init): Create a POINTER_PLUS_EXPR
instead of PLUS_EXPR for pointers.
(build_new_1): Likewise.
(build_vec_delete_1): Likewise.
(build_vec_delete): Likewise.
* class.c (build_base_path): Likewise.
(build_base_path): Likewise.
(convert_to_base_statically): Likewise.
(fixed_type_or_null): Handle POINTER_PLUS_EXPR.
(get_vtbl_decl_for_binfo): Handle POINTER_PLUS_EXPR
instead of PLUS_EXPR.
(dfs_accumulate_vtbl_inits): Create a POINTER_PLUS_EXPR
instead of PLUS_EXPR for pointers.
* call.c (build_special_member_call): Likewise.
* rtti.c (build_headof): Likewise.
Use sizetype instead of ptrdiff_type_node.
(tinfo_base_init): Create a POINTER_PLUS_EXPR
instead of PLUS_EXPR for pointers.
* except.c (expand_start_catch_block): Do a
NEGATIVE and then a POINTER_PLUS_EXPR instead
of a MINUS_EXPR.
* cp-gimplify.c (cxx_omp_clause_apply_fn): Convert
PLUS_EXPR on pointer types over to use
POINTER_PLUS_EXPR and remove the conversion
to the pointer types.
* method.c (thunk_adjust): Use POINTER_PLUS_EXPR for
adding to a pointer type. Use size_int instead of
ssize_int. Convert the index to sizetype before
adding it to the pointer.
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
* trans.c (Attribute_to_gnu): When subtracting an
offset from a pointer, use POINTER_PLUS_EXPR with
NEGATE_EXPR instead of MINUS_EXPR.
(gnat_to_gnu): Likewise.
* utils.c (convert): When converting between
thin pointers, use POINTER_PLUS_EXPR and sizetype
for the offset.
* utils2.c (known_alignment): POINTER_PLUS_EXPR
have the same semantics as PLUS_EXPR for alignment.
(build_binary_op): Add support for the semantics of
POINTER_PLUS_EXPR's operands.
When adding an offset to a pointer, use POINTER_PLUS_EXPR.
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
* class.c (make_class_data): Build the index in sizetype.
Use POINTER_PLUS_EXPR instead of PLUS_EXPR when
adding to a pointer type.
(build_symbol_entry): Likewise.
* expr.c (build_java_arrayaccess): Likewise.
(build_field_ref): Likewise.
(build_known_method_ref): Likewise.
(build_invokevirtual): Likewise.
* except.c (build_exception_object_ref): Do a
NEGATIVE and then a POINTER_PLUS_EXPR instead
of a MINUS_EXPR.
2007-06-15 Andrew Pinski <andrew_pinski@playstation.sony.com>
PR tree-opt/32225
* gcc.c-torture/compile/20070605-1.c: New test.
* gcc.c-torture/compile/20070603-1.c: New testcase.
* gcc.c-torture/compile/20070603-2.c: New testcase.
* gcc.c-torture/compile/20070531-1.c: New test.
PR tree-opt/32167
* gcc.c-torture/compile/20070531-2.c: New test.
PR tree-opt/32144
* gcc.c-torture/compile/20070529-1.c: New test.
PR tree-opt/32145
* gcc.c-torture/compile/20070529-2.c: New test.
PR tree-opt/32015
* gcc.c-torture/compile/20070520-1.c: New test.
* g++.dg/ext/java-1.C: New test.
* gcc.dg/vect/vect-106.c: We are now able to vectorize two
loops instead of one. Remove the "can't determine dependence"
check.
* gcc.dg/tree-ssa/20030815-1.c: Remove testcase which is no longer
needed as the cast is gone in the first place.
* gcc.dg/max-1.c: Change local variable a to be a global one.
* gcc.dg/tree-ssa/ssa-pre-8.c: Update testcase since we don't
have a cast which is PREd.
From-SVN: r125755
2007-06-16 07:42:36 +02:00
|
|
|
if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
|
2006-02-14 14:51:51 +01:00
|
|
|
&& cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
|
|
|
|
{
|
|
|
|
idelta = int_cst_value (TREE_OPERAND (ibase, 1));
|
|
|
|
ibase = TREE_OPERAND (ibase, 0);
|
|
|
|
}
|
|
|
|
if (cst_and_fits_in_hwi (ibase))
|
|
|
|
{
|
|
|
|
idelta += int_cst_value (ibase);
|
tree-ssa-loop-niter.c (inverse, [...]): Use build_int_cst instead of build_int_cst_type.
* tree-ssa-loop-niter.c (inverse, number_of_iterations_ne,
assert_no_overflow_lt, assert_loop_rolls_lt, number_of_iterations_lt,
number_of_iterations_le, number_of_iterations_cond, find_loop_niter,
estimate_numbers_of_iterations_loop): Use build_int_cst instead of
build_int_cst_type.
* tree-chrec.c (chrec_fold_multiply_poly_poly,
chrec_fold_multiply): Same.
* tree-ssa-loop-ivopts.c (strip_offset_1, force_expr_to_var_cost,
force_expr_to_var_cost): Same.
* tree-mudflap.c (mf_xform_derefs_1): Same.
* tree-ssa-loop-prefetch.c (idx_analyze_ref): Same.
From-SVN: r112755
2006-04-07 17:44:59 +02:00
|
|
|
ibase = build_int_cst (TREE_TYPE (ibase), 0);
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (TREE_CODE (base) == ARRAY_REF)
|
|
|
|
{
|
|
|
|
stepsize = array_ref_element_size (base);
|
|
|
|
if (!cst_and_fits_in_hwi (stepsize))
|
|
|
|
return false;
|
|
|
|
imult = int_cst_value (stepsize);
|
2010-05-25 13:18:07 +02:00
|
|
|
step = fold_build2 (MULT_EXPR, sizetype,
|
|
|
|
fold_convert (sizetype, step),
|
|
|
|
fold_convert (sizetype, stepsize));
|
2006-02-14 14:51:51 +01:00
|
|
|
idelta *= imult;
|
|
|
|
}
|
|
|
|
|
2010-05-25 13:18:07 +02:00
|
|
|
if (*ar_data->step == NULL_TREE)
|
|
|
|
*ar_data->step = step;
|
|
|
|
else
|
|
|
|
*ar_data->step = fold_build2 (PLUS_EXPR, sizetype,
|
|
|
|
fold_convert (sizetype, *ar_data->step),
|
|
|
|
fold_convert (sizetype, step));
|
2006-02-14 14:51:51 +01:00
|
|
|
*ar_data->delta += idelta;
|
|
|
|
*index = ibase;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2006-09-18 03:54:33 +02:00
|
|
|
/* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
|
2006-02-14 14:51:51 +01:00
|
|
|
STEP are integer constants and iter is number of iterations of LOOP. The
|
2006-09-18 03:54:33 +02:00
|
|
|
reference occurs in statement STMT. Strips nonaddressable component
|
|
|
|
references from REF_P. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static bool
|
2006-09-18 03:54:33 +02:00
|
|
|
analyze_ref (struct loop *loop, tree *ref_p, tree *base,
|
2010-05-19 12:36:40 +02:00
|
|
|
tree *step, HOST_WIDE_INT *delta,
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple stmt)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
struct ar_data ar_data;
|
|
|
|
tree off;
|
|
|
|
HOST_WIDE_INT bit_offset;
|
2006-09-18 03:54:33 +02:00
|
|
|
tree ref = *ref_p;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2010-05-19 12:36:40 +02:00
|
|
|
*step = NULL_TREE;
|
2006-02-14 14:51:51 +01:00
|
|
|
*delta = 0;
|
|
|
|
|
2010-07-21 17:44:24 +02:00
|
|
|
/* First strip off the component references. Ignore bitfields.
|
|
|
|
Also strip off the real and imagine parts of a complex, so that
|
|
|
|
they can have the same base. */
|
|
|
|
if (TREE_CODE (ref) == REALPART_EXPR
|
|
|
|
|| TREE_CODE (ref) == IMAGPART_EXPR
|
|
|
|
|| (TREE_CODE (ref) == COMPONENT_REF
|
|
|
|
&& DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1))))
|
|
|
|
{
|
|
|
|
if (TREE_CODE (ref) == IMAGPART_EXPR)
|
|
|
|
*delta += int_size_in_bytes (TREE_TYPE (ref));
|
|
|
|
ref = TREE_OPERAND (ref, 0);
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2006-09-18 03:54:33 +02:00
|
|
|
*ref_p = ref;
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
|
|
|
|
{
|
|
|
|
off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
|
|
|
|
bit_offset = TREE_INT_CST_LOW (off);
|
|
|
|
gcc_assert (bit_offset % BITS_PER_UNIT == 0);
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
*delta += bit_offset / BITS_PER_UNIT;
|
|
|
|
}
|
|
|
|
|
|
|
|
*base = unshare_expr (ref);
|
|
|
|
ar_data.loop = loop;
|
|
|
|
ar_data.stmt = stmt;
|
|
|
|
ar_data.step = step;
|
|
|
|
ar_data.delta = delta;
|
|
|
|
return for_each_index (base, idx_analyze_ref, &ar_data);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Record a memory reference REF to the list REFS. The reference occurs in
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
LOOP in statement STMT and it is write if WRITE_P. Returns true if the
|
|
|
|
reference was recorded, false otherwise. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
static bool
|
2006-02-14 14:51:51 +01:00
|
|
|
gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
|
2008-07-28 16:33:56 +02:00
|
|
|
tree ref, bool write_p, gimple stmt)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
2010-05-19 12:36:40 +02:00
|
|
|
tree base, step;
|
|
|
|
HOST_WIDE_INT delta;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref_group *agrp;
|
|
|
|
|
2008-06-24 12:06:17 +02:00
|
|
|
if (get_base_address (ref) == NULL)
|
|
|
|
return false;
|
|
|
|
|
2006-09-18 03:54:33 +02:00
|
|
|
if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
return false;
|
2010-05-19 12:36:40 +02:00
|
|
|
/* If analyze_ref fails the default is a NULL_TREE. We can stop here. */
|
|
|
|
if (step == NULL_TREE)
|
|
|
|
return false;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2010-06-10 00:56:08 +02:00
|
|
|
/* Limit non-constant step prefetching only to the innermost loops. */
|
|
|
|
if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
|
|
|
|
return false;
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
|
|
|
|
are integer constants. */
|
|
|
|
agrp = find_or_create_group (refs, base, step);
|
|
|
|
record_ref (agrp, stmt, ref, delta, write_p);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
|
|
|
return true;
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
/* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
|
|
|
|
true if there are no other memory references inside the loop. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static struct mem_ref_group *
|
2009-06-08 18:00:13 +02:00
|
|
|
gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
basic_block *body = get_loop_body_in_dom_order (loop);
|
|
|
|
basic_block bb;
|
|
|
|
unsigned i;
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple_stmt_iterator bsi;
|
|
|
|
gimple stmt;
|
|
|
|
tree lhs, rhs;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref_group *refs = NULL;
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
*no_other_refs = true;
|
2009-06-08 18:00:13 +02:00
|
|
|
*ref_count = 0;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Scan the loop body in order, so that the former references precede the
|
|
|
|
later ones. */
|
|
|
|
for (i = 0; i < loop->num_nodes; i++)
|
|
|
|
{
|
|
|
|
bb = body[i];
|
|
|
|
if (bb->loop_father != loop)
|
|
|
|
continue;
|
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
2008-07-28 16:33:56 +02:00
|
|
|
stmt = gsi_stmt (bsi);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
{
|
re PR middle-end/13146 (inheritance for nonoverlapping_component_refs_p)
2009-04-03 Richard Guenther <rguenther@suse.de>
PR middle-end/13146
PR tree-optimization/23940
PR tree-optimization/33237
PR middle-end/33974
PR middle-end/34093
PR tree-optimization/36201
PR tree-optimization/36230
PR tree-optimization/38049
PR tree-optimization/38207
PR tree-optimization/38230
PR tree-optimization/38301
PR tree-optimization/38585
PR middle-end/38895
PR tree-optimization/38985
PR tree-optimization/39299
* tree-ssa-structalias.h: Remove.
* tree-ssa-operands.h (NULL_USE_OPERAND_P): Make of type use_operand_p.
(NULL_DEF_OPERAND_P): Make of type def_operand_p.
(struct vuse_element_d): Remove.
(struct vuse_vec_d): Likewise.
(VUSE_VECT_NUM_ELEM, VUSE_VECT_ELEMENT_NC, VUSE_ELEMENT_PTR_NC,
VUSE_ELEMENT_VAR_NC, VUSE_VECT_ELEMENT, VUSE_ELEMENT_PTR,
SET_VUSE_VECT_ELEMENT, SET_VUSE_ELEMENT_VAR, SET_VUSE_ELEMENT_PTR,
VUSE_ELEMENT_VAR): Likewise.
(struct voptype_d): Likewise.
(NUM_VOP_FREE_BUCKETS): Likewise.
(struct ssa_operands): Remove vop_free_buckets and mpt_table fields.
(struct stmt_operands_d): Remove.
(VUSE_OP_PTR, VUSE_OP, SET_VUSE_OP, VUSE_NUM, VUSE_VECT,
VDEF_RESULT_PTR, VDEF_RESULT, VDEF_OP_PTR, VDEF_OP, SET_VDEF_OP,
VDEF_NUM, VDEF_VECT): Likewise.
(copy_virtual_operands): Remove.
(operand_build_cmp): Likewise.
(create_ssa_artificial_load_stmt): Likewise.
(enum ssa_op_iter_type): Remove ssa_op_iter_vdef.
(struct ssa_operand_iterator_d): Remove vuses, vdefs, mayusesm
vuse_index and mayuse_index members. Pack and move done and iter_type
members to the front.
(SSA_OP_VMAYUSE): Remove.
(SSA_OP_VIRTUAL_USES): Adjust.
(FOR_EACH_SSA_VDEF_OPERAND): Remove.
(unlink_stmt_vdef): Declare.
(add_to_addressable_set): Remove.
* tree-vrp.c (stmt_interesting_for_vrp): Adjust.
(vrp_visit_stmt): Likewise.
* doc/tree-ssa.texi (Alias analysis): Update.
* doc/invoke.texi (max-aliased-vops): Remove docs.
(avg-aliased-vops): Likewise.
* tree-into-ssa.c (syms_to_rename): Remove.
(need_to_update_vops_p): Likewise.
(need_to_initialize_update_ssa_p): Rename to ...
(update_ssa_initialized_fn): ... this. Track function we are
initialized for.
(symbol_marked_for_renaming): Simplify.
(add_new_name_mapping): Do not set need_to_update_vops_p.
(dump_currdefs): Use SYMS_TO_RENAME.
(rewrite_update_stmt): Always walk all uses/defs.
(dump_update_ssa): Adjust.
(init_update_ssa): Take function argument. Track what we are
initialized for.
(delete_update_ssa): Reset SYMS_TO_RENAME and update_ssa_initialized_fn.
(create_new_def_for): Initialize for cfun, assert we are initialized
for cfun.
(mark_sym_for_renaming): Simplify.
(mark_set_for_renaming): Do not initialize update-ssa.
(need_ssa_update_p): Simplify. Take function argument.
(name_mappings_registered_p): Assert we ask for the correct function.
(name_registered_for_update_p): Likewise.
(ssa_names_to_replace): Likewise.
(release_ssa_name_after_update_ssa): Likewise.
(update_ssa): Likewise. Use SYMS_TO_RENAME.
(dump_decl_set): Do not print a newline.
(debug_decl_set): Do it here.
(dump_update_ssa): And here.
* tree-ssa-loop-im.c (move_computations): Adjust.
(movement_possibility): Likewise.
(determine_max_movement): Likewise.
(gather_mem_refs_stmt): Likewise.
* tree-dump.c (dequeue_and_dump): Do not handle SYMBOL_MEMORY_TAG
or NAME_MEMORY_TAG.
* tree-complex.c (update_all_vops): Remove.
(expand_complex_move): Adjust.
* tree-ssa-loop-niter.c (chain_of_csts_start): Use NULL_TREE.
Simplify test for memory referencing statement. Exclude
non-invariant ADDR_EXPRs.
* tree-pretty-print.c (dump_generic_node): Do not handle memory tags.
* tree-loop-distribution.c (generate_memset_zero): Adjust.
(rdg_flag_uses): Likewise.
* tree-tailcall.c (suitable_for_tail_opt_p): Remove memory-tag
related code.
(tree_optimize_tail_calls_1): Also split the
edge from the entry block if we have degenerate PHI nodes in
the first basic block.
* tree.c (init_ttree): Remove memory-tag related code.
(tree_code_size): Likewise.
(tree_node_structure): Likewise.
(build7_stat): Re-write to be build6_stat.
* tree.h (MTAG_P, TREE_MEMORY_TAG_CHECK, TMR_TAG): Remove.
(SSA_VAR_P): Adjust.
(struct tree_memory_tag): Remove.
(struct tree_memory_partition_tag): Likewise.
(union tree_node): Adjust.
(build7): Re-write to be build6.
* tree-pass.h (pass_reset_cc_flags): Remove.
(TODO_update_address_taken): New flag.
(pass_simple_dse): Remove.
* ipa-cp.c (ipcp_update_callgraph): Update SSA form.
* params.h (MAX_ALIASED_VOPS): Remove.
(AVG_ALIASED_VOPS): Likewise.
* omp-low.c (expand_omp_taskreg): Update SSA form.
* tree-ssa-dse.c (dse_optimize_stmt): Properly query if the rhs
aliases the lhs in a copy stmt.
* tree-ssa-dse.c (struct address_walk_data): Remove.
(memory_ssa_name_same): Likewise.
(memory_address_same): Likewise.
(get_kill_of_stmt_lhs): Likewise.
(dse_possible_dead_store_p): Simplify, use the oracle. Handle
unused stores. Look through PHI nodes into post-dominated regions.
(dse_optimize_stmt): Simplify. Properly remove stores.
(tree_ssa_dse): Compute dominators.
(execute_simple_dse): Remove.
(pass_simple_dse): Likewise.
* ipa-reference.c (scan_stmt_for_static_refs): Open-code
gimple_loaded_syms and gimple_stored_syms computation.
* toplev.c (dump_memory_report): Dump alias and pta stats.
* tree-ssa-sccvn.c (vn_reference_compute_hash): Simplify.
(vn_reference_eq): Likewise.
(vuses_to_vec, copy_vuses_from_stmt, vdefs_to_vec,
copy_vdefs_from_stmt, shared_lookup_vops, shared_vuses_from_stmt,
valueize_vuses): Remove.
(get_def_ref_stmt_vuses): Simplify. Rename to ...
(get_def_ref_stmt_vuse): ... this.
(vn_reference_lookup_2): New function.
(vn_reference_lookup_pieces): Use walk_non_aliased_vuses for
walking equivalent vuses. Simplify.
(vn_reference_lookup): Likewise.
(vn_reference_insert): Likewise.
(vn_reference_insert_pieces): Likewise.
(visit_reference_op_call): Simplify.
(visit_reference_op_load): Likewise.
(visit_reference_op_store): Likewise.
(init_scc_vn): Remove shared_lookup_vuses initialization.
(free_scc_vn): Remove shared_lookup_vuses freeing.
(sort_vuses, sort_vuses_heap): Remove.
(get_ref_from_reference_ops): Export.
* tree-ssa-sccvn.h (struct vn_reference_s): Replace vuses
vector with single vuse pointer.
(vn_reference_lookup_pieces, vn_reference_lookup,
vn_reference_insert, vn_reference_insert_pieces): Adjust prototypes.
(shared_vuses_from_stmt): Remove.
(get_ref_from_reference_ops): Declare.
* tree-ssa-loop-manip.c (slpeel_can_duplicate_loop_p): Adjust.
* tree-ssa-copyrename.c (copy_rename_partition_coalesce): Remove
memory-tag related code.
* tree-ssa-ccp.c (get_symbol_constant_value): Remove memory-tag code.
(likely_value): Add comment, skip static-chain of call statements.
(surely_varying_stmt_p): Adjust.
(gimplify_and_update_call_from_tree): Likewise.
(execute_fold_all_builtins): Do not rebuild alias info.
(gimplify_and_update_call_from_tree): Properly update VOPs.
* tree-ssa-loop-ivopts.c (get_ref_tag): Remove.
(copy_ref_info): Remove memory-tag related code.
* tree-call-cdce.c (tree_call_cdce): Rename the VOP.
* ipa-pure-const.c (check_decl): Remove memory-tag related code.
(check_stmt): Open-code gimple_loaded_syms and gimple_stored_syms
computation.
* tree-ssa-dom.c (gimple_p): Remove typedef.
(eliminate_redundant_computations): Adjust.
(record_equivalences_from_stmt): Likewise.
(avail_expr_hash): Likewise.
(avail_expr_eq): Likewise.
* tree-ssa-propagate.c (update_call_from_tree): Properly
update VOPs.
(stmt_makes_single_load): Likewise.
(stmt_makes_single_store): Likewise.
* tree-ssa-alias.c: Rewrite completely.
(debug_memory_partitions, dump_mem_ref_stats, debug_mem_ref_stats,
debug_mem_sym_stats, dump_mem_sym_stats_for_var,
debug_all_mem_sym_stats, debug_mp_info, update_mem_sym_stats_from_stmt,
delete_mem_ref_stats, create_tag_raw, dump_points_to_info,
dump_may_aliases_for, debug_may_aliases_for, new_type_alias):
Remove public functions.
(pass_reset_cc_flags): Remove.
(pass_build_alias): Move ...
* tree-ssa-structalias.c (pass_build_alias): ... here.
* tree-ssa-alias.c (may_be_aliased): Move ...
* tree-flow-inline.h (may_be_aliased): ... here.
tree-ssa-alias.c (struct count_ptr_d, count_ptr_derefs,
count_uses_and_derefs): Move ...
* gimple.c: ... here.
* gimple.h (count_uses_and_derefs): Declare.
* tree-ssa-alias.c (dump_alias_stats, ptr_deref_may_alias_global_p,
ptr_deref_may_alias_decl_p, ptr_derefs_may_alias_p,
same_type_for_tbaa, nonaliasing_component_refs_p, decl_refs_may_alias_p,
indirect_ref_may_alias_decl_p, indirect_refs_may_alias_p,
ref_maybe_used_by_call_p, ref_maybe_used_by_stmt_p,
call_may_clobber_ref_p, stmt_may_clobber_ref_p, maybe_skip_until,
get_continuation_for_phi, walk_non_aliased_vuses, walk_aliased_vdefs):
New functions.
* tree-dfa.c (refs_may_alias_p): Move ...
* tree-ssa-alias.c (refs_may_alias_p): ... here. Extend.
* tree-ssa-alias.h: New file.
* tree-ssa-sink.c (is_hidden_global_store): Adjust.
(statement_sink_location): Likewise.
* opts.c (decode_options): Do not adjust max-aliased-vops or
avg-aliased-vops values.
* timevar.def (TV_TREE_MAY_ALIAS): Remove.
(TV_CALL_CLOBBER): Likewise.
(TV_FLOW_SENSITIVE): Likewise.
(TV_FLOW_INSENSITIVE): Likewise.
(TV_MEMORY_PARTITIONING): Likewise.
(TV_ALIAS_STMT_WALK): New timevar.
* tree-ssa-loop-ivcanon.c (empty_loop_p): Adjust.
* tree-ssa-address.c (create_mem_ref_raw): Use build6.
(get_address_description): Remove memory-tag related code.
* tree-ssa-ifcombine.c (bb_no_side_effects_p): Adjust.
* treestruct.def (TS_MEMORY_TAG, TS_MEMORY_PARTITION_TAG): Remove.
* tree-eh.c (cleanup_empty_eh): Do not leave stale SSA_NAMEs
and immediate uses in statements. Document.
* gimple-pretty-print.c (dump_gimple_mem_ops): Adjust.
(dump_symbols): Remove.
(dump_gimple_mem_ops): Do not dump loaded or stored syms.
* alias.c (get_deref_alias_set): New function split out from ...
(get_alias_set): ... here.
* alias.h (get_deref_alias_set): Declare.
* tree-vect-data-refs.c (vect_create_data_ref_ptr): Remove unused
type parameter. Remove restrict pointer handling. Create a
ref-all pointer in case type-based alias sets do not conflict.
(vect_analyze_data_refs): Remove SMT related code.
* tree-vect-stmts.c (vectorizable_store): Re-instantiate TBAA assert.
(vectorizable_load): Likewise.
* tree-data-ref.h (struct dr_alias): Remove symbol_tag field.
(DR_SYMBOL_TAG, DR_VOPS): Remove.
* tree-data-ref.c (dr_may_alias_p): Use the alias-oracle.
Ignore vops and SMTs.
(dr_analyze_alias): Likewise..
(free_data_ref): Likewise.
(create_data_ref): Likewise.
(analyze_all_data_dependences): Likewise.
(get_references_in_stmt): Adjust.
* tree-flow-inline.h (gimple_aliases_computed_p,
gimple_addressable_vars, gimple_call_clobbered_vars,
gimple_call_used_vars, gimple_global_var, may_aliases, memory_partition,
factoring_name_p, mark_call_clobbered, clear_call_clobbered,
compare_ssa_operands_equal, symbol_mem_tag, set_symbol_mem_tag,
gimple_mem_ref_stats): Remove.
(gimple_vop): New function.
(op_iter_next_use): Remove vuses and mayuses cases.
(op_iter_next_def): Remove vdefs case.
(op_iter_next_tree): Remove vuses, mayuses and vdefs cases.
(clear_and_done_ssa_iter): Do not set removed fields.
(op_iter_init): Likewise. Skip vuse and/or vdef if requested.
Assert we are not iterating over vuses or vdefs if not also
iterating over uses or defs.
(op_iter_init_use): Likewise.
(op_iter_init_def): Likewise.
(op_iter_next_vdef): Remove.
(op_iter_next_mustdef): Likewise.
(op_iter_init_vdef): Likewise.
(compare_ssa_operands_equal): Likewise.
(link_use_stmts_after): Handle vuse operand.
(is_call_used): Use is_call_clobbered.
(is_call_clobbered): Global variables are always call clobbered,
query the call-clobbers bitmap.
(mark_call_clobbered): Ignore global variables.
(clear_call_clobbered): Likewise.
* tree-ssa-coalesce.c (create_outofssa_var_map): Adjust
virtual operands sanity check.
* tree.def (NAME_MEMORY_TAG, SYMBOL_MEMORY_TAG, MEMORY_PARTITION_TAG):
Remove.
(TARGET_MEM_REF): Remove TMR_TAG operand.
* tree-dfa.c (add_referenced_var): Initialize call-clobber state.
Remove call-clobber related code.
(remove_referenced_var): Likewise. Do not clear mpt or symbol_mem_tag.
(dump_variable): Do not dump SMTs, memory stats, may-aliases or
partitions or escape reason.
(get_single_def_stmt, get_single_def_stmt_from_phi,
get_single_def_stmt_with_phi): Remove.
(dump_referenced_vars): Tidy.
(get_ref_base_and_extent): Allow bare decls.
(collect_dfa_stats): Adjust.
* graphite.c (rename_variables_in_stmt): Adjust.
(graphite_copy_stmts_from_block): Likewise.
(translate_clast): Likewise.
* tree-ssa-pre.c (struct bb_bitmap_sets): Add expr_dies bitmap.
(EXPR_DIES): New.
(translate_vuse_through_block): Use the oracle.
(phi_translate_1): Adjust.
(value_dies_in_block_x): Use the oracle. Cache the outcome
in EXPR_DIES.
(valid_in_sets): Check if the VUSE for
a REFERENCE is available.
(eliminate): Do not remove stmts during elimination,
instead queue and remove them afterwards.
(do_pre): Do not rebuild alias info.
(pass_pre): Run TODO_rebuild_alias before PRE.
* tree-ssa-live.c (remove_unused_locals): Remove memory-tag code.
* tree-sra.c (sra_walk_function): Use gimple_references_memory_p.
(mark_all_v_defs_stmt): Remove.
(mark_all_v_defs_seq): Adjust.
(sra_replace): Likewise.
(scalarize_use): Likewise.
(scalarize_copy): Likewise.
(scalarize_init): Likewise.
(scalarize_ldst): Likewise.
(todoflags): Remove.
(tree_sra): Do not rebuild alias info.
(tree_sra_early): Adjust.
(pass_sra): Run TODO_update_address_taken before SRA.
* tree-predcom.c (set_alias_info): Remove.
(prepare_initializers_chain): Do not call it.
(mark_virtual_ops_for_renaming): Adjust.
(mark_virtual_ops_for_renaming_list): Remove.
(initialize_root_vars): Adjust.
(initialize_root_vars_lm): Likewise.
(prepare_initializers_chain): Likewise.
* tree-ssa-copy.c (may_propagate_copy): Remove memory-tag related code.
(may_propagate_copy_into_stmt): Likewise.
(merge_alias_info): Do nothing for now.
(propagate_tree_value_into_stmt): Adjust.
(stmt_may_generate_copy): Likewise.
* tree-ssa-forwprop.c (tidy_after_forward_propagate_addr): Do
not mark symbols for renaming.
(forward_propagate_addr_expr): Match up push/pop_stmt_changes
with the same statement, make sure to update the new pointed-to one.
* tree-ssa-dce.c (eliminate_unnecessary_stmts): Do not copy
call statements, do not mark symbols for renaming.
(mark_operand_necessary): Dump something.
(ref_may_be_aliased): New function.
(mark_aliased_reaching_defs_necessary_1): New helper function.
(mark_aliased_reaching_defs_necessary): Likewise.
(mark_all_reaching_defs_necessary_1): Likewise.
(mark_all_reaching_defs_necessary): Likewise.
(propagate_necessity): Do not process virtual PHIs. For
non-aliased loads mark all reaching definitions as necessary.
For aliased loads and stores mark the immediate dominating
aliased clobbers as necessary.
(visited): New global static.
(perform_tree_ssa_dce): Free visited bitmap after propagating
necessity.
(remove_dead_phis): Perform simple dead virtual PHI removal.
(remove_dead_stmt): Properly unlink virtual operands when
removing stores.
(eliminate_unnecessary_stmts): Schedule PHI removal after
stmt removal.
* tree-ssa-ter.c (is_replaceable_p): Adjust.
(process_replaceable): Likewise.
(find_replaceable_in_bb): Likewise.
* tree-ssa.c (verify_ssa_name): Verify all VOPs are
based on the single gimple vop.
(verify_flow_insensitive_alias_info): Remove.
(verify_flow_sensitive_alias_info): Likewise.
(verify_call_clobbering): Likewise.
(verify_memory_partitions): Likewise.
(verify_alias_info): Likewise.
(verify_ssa): Adjust..
(execute_update_addresses_taken): Export. Update SSA
manually. Optimize only when optimizing. Use a local bitmap.
(pass_update_address_taken): Remove TODO_update_ssa, add
TODO_dump_func.
(pass_update_address_taken): Just use TODO_update_address_taken.
(init_tree_ssa): Do not initialize addressable_vars.
(verify_ssa): Verify new VUSE / VDEF properties.
Verify that all stmts definitions have the stmt as SSA_NAME_DEF_STMT.
Do not call verify_alias_info.
(delete_tree_ssa): Clear the VUSE, VDEF operands.
Do not free the loaded and stored syms bitmaps. Reset the escaped
and callused solutions. Do not free addressable_vars.
Remove memory-tag related code.
(warn_uninitialized_var): Aliases are always available.
* tree-ssa-loop-prefetch.c (gather_memory_references): Adjust.
* lambda-code.c (can_put_in_inner_loop): Adjust.
(can_put_after_inner_loop): Likewise.
(perfect_nestify): Likewise.
* tree-vect-stmts.c (vect_stmt_relevant_p): Adjust.
(vect_gen_widened_results_half): Remove CALL_EXPR handling.
(vectorizable_conversion): Do not mark symbols for renaming.
* tree-inline.c (remap_gimple_stmt): Clear VUSE/VDEF.
(expand_call_inline): Unlink the calls virtual operands before
replacing it.
(tree_function_versioning): Do not call update_ssa if we are not
updating clones. Simplify.
* tree-ssa-phiprop.c (phivn_valid_p): Adjust.
(propagate_with_phi): Likewise..
* tree-outof-ssa.c (create_temp): Remove memory tag and call
clobber code. Assert we are not aliased or global.
* tree-flow.h: Include tree-ssa-alias.h
(enum escape_type): Remove.
(struct mem_sym_stats_d): Likewise.
(struct mem_ref_stats_d): Likewise.
(struct gimple_df): Add vop member. Remove global_var,
call_clobbered_vars, call_used_vars, addressable_vars,
aliases_compted_p and mem_ref_stats members. Add syms_to_rename,
escaped and callused members.
(struct ptr_info_def): Remove all members, add points-to solution
member pt.
(struct var_ann_d): Remove in_vuse_list, in_vdef_list,
call_clobbered, escape_mask, mpt and symbol_mem_tag members.
* Makefile.in (TREE_FLOW_H): Add tree-ssa-alias.h.
(tree-ssa-structalias.o): Remove tree-ssa-structalias.h.
(tree-ssa-alias.o): Likewise.
(toplev.o): Add tree-ssa-alias.h
(GTFILES): Remove tree-ssa-structalias.h, add tree-ssa-alias.h.
* gimple.c (gimple_set_bb): Fix off-by-one error.
(is_gimple_reg): Do not handle memory tags.
(gimple_copy): Also copy virtual operands.
Delay updating the statement. Do not reset loaded and stored syms.
(gimple_set_stored_syms): Remove.
(gimple_set_loaded_syms): Likewise.
(gimple_call_copy_skip_args): Copy the virtual operands
and mark the new statement modified.
* tree-ssa-structalias.c (may_alias_p): Remove.
(set_uids_in_ptset): Take the alias set to prune with as
parameter. Fold in the alias test of may_alias_p.
(compute_points_to_sets): Compute whether a ptr is dereferenced
in a local sbitmap.
(process_constraint): Deal with &ANYTHING on the lhs, reject all
other ADDRESSOF constraints on the lhs.
(get_constraint_for_component_ref): Assert that we don't get
ADDRESSOF constraints from the base of the reference.
Properly generate UNKNOWN_OFFSET for DEREF if needed.
(struct variable_info): Remove collapsed_to member.
(get_varinfo_fc): Remove.
(new_var_info): Do not set collapsed_to.
(dump_constraint): Do not follow cycles.
(dump_constraint_graph): Likewise.
(build_pred_graph): Likewise.
(build_succ_graph): Likewise.
(rewrite_constraints): Likewise.
(do_simple_structure_copy): Remove.
(do_rhs_deref_structure_copy): Remove.
(do_lhs_deref_structure_copy): Remove.
(collapse_rest_of_var): Remove.
(do_structure_copy): Re-implement.
(pta_stats): New global variable.
(dump_pta_stats): New function.
(struct constraint_expr): Make offset signed.
(UNKNOWN_OFFSET): Define special value.
(dump_constraint): Dump UNKNOWN_OFFSET as UNKNOWN.
(solution_set_expand): New helper function split out from ...
(do_sd_constraint): ... here.
(solution_set_add): Handle UNKNOWN_OFFSET. Handle negative offsets.
(do_ds_constraint): Likewise.
(do_sd_constraint): Likewise. Do not special-case ESCAPED = *ESCAPED
and CALLUSED = *CALLUSED.
(set_union_with_increment): Make inc argument signed.
(type_safe): Remove.
(get_constraint_for_ptr_offset): Handle unknown and negative
constant offsets.
(first_vi_for_offset): Handle offsets before start. Bail
out early for offsets beyond the variable extent.
(first_or_preceding_vi_for_offset): New function.
(init_base_vars): Add ESCAPED = ESCAPED + UNKNOWN_OFFSET constraint.
Together with ESCAPED = *ESCAPED this properly computes reachability.
(find_what_var_points_to): New function.
(find_what_p_points_to): Implement in terms of find_what_var_points_to.
(pt_solution_reset, pt_solution_empty_p, pt_solution_includes_global,
pt_solution_includes_1, pt_solution_includes, pt_solutions_intersect_1,
pt_solutions_intersect): New functions.
(compute_call_used_vars): Remove.
(compute_may_aliases): New main entry into PTA computation.
* gimple.h (gimple_p): New typedef.
(struct gimple_statement_base): Remove references_memory_p.
(struct gimple_statement_with_memory_ops_base): Remove
vdef_ops, vuse_ops, stores and loads members. Add vdef and vuse
members.
(gimple_vuse_ops, gimple_set_vuse_ops, gimple_vdef_ops,
gimple_set_vdef_ops, gimple_loaded_syms, gimple_stored_syms,
gimple_set_references_memory): Remove.
(gimple_vuse_op, gimple_vdef_op, gimple_vuse, gimple_vdef,
gimple_vuse_ptr, gimple_vdef_ptri, gimple_set_vuse, gimple_set_vdef):
New functions.
* tree-cfg.c (move_block_to_fn): Fix off-by-one error.
(verify_expr): Allow RESULT_DECL.
(gimple_duplicate_bb): Do not copy virtual operands.
(gimple_duplicate_sese_region): Adjust.
(gimple_duplicate_sese_tail): Likewise.
(mark_virtual_ops_in_region): Remove.
(move_sese_region_to_fn): Do not call it.
* passes.c (init_optimization_passes): Remove pass_reset_cc_flags
and pass_simple_dse.
(execute_function_todo): Handle TODO_update_address_taken,
call execute_update_addresses_taken for TODO_rebuild_alias.
(execute_todo): Adjust.
(execute_one_pass): Init dump files early.
* ipa-struct-reorg.c (finalize_var_creation): Do not mark vars
call-clobbered.
(create_general_new_stmt): Clear vops.
* tree-ssa-reassoc.c (get_rank): Adjust.
* tree-vect-slp.c (vect_create_mask_and_perm): Do not mark
symbols for renaming.
* params.def (PARAM_MAX_ALIASED_VOPS): Remove.
(PARAM_AVG_ALIASED_VOPS): Likewise.
* tree-ssanames.c (init_ssanames): Allocate SYMS_TO_RENAME.
(duplicate_ssa_name_ptr_info): No need to copy the shared bitmaps.
* tree-ssa-operands.c: Simplify for new virtual operand
representation.
(operand_build_cmp, copy_virtual_operands,
create_ssa_artificial_load_stmt, add_to_addressable_set,
gimple_add_to_addresses_taken): Remove public functions.
(unlink_stmt_vdef): New function.
* gcc.dg/pr19633-1.c: Adjust.
* gcc.dg/torture/pta-callused-1.c: Likewise.
* gcc.dg/torture/pr39074-2.c: Likewise.
* gcc.dg/torture/pr39074.c: Likewise.
* gcc.dg/torture/pta-ptrarith-3.c: New testcase.
* gcc.dg/torture/pr30375.c: Adjust.
* gcc.dg/torture/pr33563.c: Likewise.
* gcc.dg/torture/pr33870.c: Likewise.
* gcc.dg/torture/pr33560.c: Likewise.
* gcc.dg/torture/pta-structcopy-1.c: New testcase.
* gcc.dg/torture/ssa-pta-fn-1.c: Likewise.
* gcc.dg/tree-ssa/alias-15.c: Remove.
* gcc.dg/tree-ssa/ssa-dce-4.c: New testcase.
* gcc.dg/tree-ssa/pr26421.c: Adjust.
* gcc.dg/tree-ssa/ssa-fre-10.c: XFAIL.
* gcc.dg/tree-ssa/ssa-dce-5.c: New testcase.
* gcc.dg/tree-ssa/pr23382.c: Adjust.
* gcc.dg/tree-ssa/ssa-fre-20.c: New testcase.
* gcc.dg/tree-ssa/alias-16.c: Adjust.
* gcc.dg/tree-ssa/ssa-fre-13.c: Likewise.
* gcc.dg/tree-ssa/ssa-fre-14.c: Likewise.
* gcc.dg/tree-ssa/alias-18.c: Likewise.
* gcc.dg/tree-ssa/ssa-fre-15.c: Likewise.
* gcc.dg/tree-ssa/ssa-lim-3.c: Likewise.
* gcc.dg/tree-ssa/alias-19.c: Likewise.
* gcc.dg/tree-ssa/pta-ptrarith-1.c: New testcase.
* gcc.dg/tree-ssa/pr13146.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-23.c: Likewise.
* gcc.dg/tree-ssa/pta-ptrarith-2.c: Likewise.
* gcc.dg/tree-ssa/ssa-fre-18.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-24.c: New XFAILed testcase.
* gcc.dg/tree-ssa/ssa-fre-19.c: New testcase.
* gcc.dg/tree-ssa/alias-20.c: Likewise.
* gcc.dg/tree-ssa/ssa-dse-12.c: Likewise.
* gcc.dg/tree-ssa/pr38895.c: Likewise.
* gcc.dg/uninit-B.c: XFAIL.
* gcc.dg/vect/no-vfa-vect-43.c: Adjust.
* gcc.dg/uninit-pr19430.c: XFAIL.
* g++.dg/tree-ssa/pr13146.C: New testcase.
* g++.dg/opt/pr36187.C: Adjust.
* g++.dg/torture/20090329-1.C: New testcase.
From-SVN: r145494
2009-04-03 12:24:28 +02:00
|
|
|
if (gimple_vuse (stmt)
|
2008-07-28 16:33:56 +02:00
|
|
|
|| (is_gimple_call (stmt)
|
|
|
|
&& !(gimple_call_flags (stmt) & ECF_CONST)))
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
*no_other_refs = false;
|
|
|
|
continue;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
lhs = gimple_assign_lhs (stmt);
|
|
|
|
rhs = gimple_assign_rhs1 (stmt);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (REFERENCE_CLASS_P (rhs))
|
2009-06-08 18:00:13 +02:00
|
|
|
{
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
|
|
|
rhs, false, stmt);
|
2009-06-08 18:00:13 +02:00
|
|
|
*ref_count += 1;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
if (REFERENCE_CLASS_P (lhs))
|
2009-06-08 18:00:13 +02:00
|
|
|
{
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
|
|
|
lhs, true, stmt);
|
2009-06-08 18:00:13 +02:00
|
|
|
*ref_count += 1;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
free (body);
|
|
|
|
|
|
|
|
return refs;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Prune the prefetch candidate REF using the self-reuse. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
prune_ref_by_self_reuse (struct mem_ref *ref)
|
|
|
|
{
|
2010-05-19 12:36:40 +02:00
|
|
|
HOST_WIDE_INT step;
|
|
|
|
bool backward;
|
|
|
|
|
|
|
|
/* If the step size is non constant, we cannot calculate prefetch_mod. */
|
|
|
|
if (!cst_and_fits_in_hwi (ref->group->step))
|
|
|
|
return;
|
|
|
|
|
|
|
|
step = int_cst_value (ref->group->step);
|
|
|
|
|
|
|
|
backward = step < 0;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (step == 0)
|
|
|
|
{
|
|
|
|
/* Prefetch references to invariant address just once. */
|
|
|
|
ref->prefetch_before = 1;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (backward)
|
|
|
|
step = -step;
|
|
|
|
|
|
|
|
if (step > PREFETCH_BLOCK)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if ((backward && HAVE_BACKWARD_PREFETCH)
|
|
|
|
|| (!backward && HAVE_FORWARD_PREFETCH))
|
|
|
|
{
|
|
|
|
ref->prefetch_before = 1;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ref->prefetch_mod = PREFETCH_BLOCK / step;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Divides X by BY, rounding down. */
|
|
|
|
|
|
|
|
static HOST_WIDE_INT
|
|
|
|
ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
|
|
|
|
{
|
|
|
|
gcc_assert (by > 0);
|
|
|
|
|
|
|
|
if (x >= 0)
|
|
|
|
return x / by;
|
|
|
|
else
|
|
|
|
return (x + by - 1) / by;
|
|
|
|
}
|
|
|
|
|
2009-11-25 11:55:54 +01:00
|
|
|
/* Given a CACHE_LINE_SIZE and two inductive memory references
|
|
|
|
with a common STEP greater than CACHE_LINE_SIZE and an address
|
|
|
|
difference DELTA, compute the probability that they will fall
|
2010-07-02 18:34:38 +02:00
|
|
|
in different cache lines. Return true if the computed miss rate
|
|
|
|
is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the
|
|
|
|
number of distinct iterations after which the pattern repeats itself.
|
2009-08-13 23:37:24 +02:00
|
|
|
ALIGN_UNIT is the unit of alignment in bytes. */
|
|
|
|
|
2010-07-02 18:34:38 +02:00
|
|
|
static bool
|
|
|
|
is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size,
|
2009-08-13 23:37:24 +02:00
|
|
|
HOST_WIDE_INT step, HOST_WIDE_INT delta,
|
|
|
|
unsigned HOST_WIDE_INT distinct_iters,
|
|
|
|
int align_unit)
|
|
|
|
{
|
|
|
|
unsigned align, iter;
|
2010-07-02 18:34:38 +02:00
|
|
|
int total_positions, miss_positions, max_allowed_miss_positions;
|
2009-08-13 23:37:24 +02:00
|
|
|
int address1, address2, cache_line1, cache_line2;
|
|
|
|
|
2010-07-02 18:34:29 +02:00
|
|
|
/* It always misses if delta is greater than or equal to the cache
|
|
|
|
line size. */
|
2010-07-02 18:34:38 +02:00
|
|
|
if (delta >= (HOST_WIDE_INT) cache_line_size)
|
|
|
|
return false;
|
2010-07-02 18:34:29 +02:00
|
|
|
|
2009-08-13 23:37:24 +02:00
|
|
|
miss_positions = 0;
|
2010-07-02 18:34:38 +02:00
|
|
|
total_positions = (cache_line_size / align_unit) * distinct_iters;
|
|
|
|
max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000;
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2009-08-13 23:37:24 +02:00
|
|
|
/* Iterate through all possible alignments of the first
|
|
|
|
memory reference within its cache line. */
|
|
|
|
for (align = 0; align < cache_line_size; align += align_unit)
|
|
|
|
|
|
|
|
/* Iterate through all distinct iterations. */
|
|
|
|
for (iter = 0; iter < distinct_iters; iter++)
|
|
|
|
{
|
|
|
|
address1 = align + step * iter;
|
|
|
|
address2 = address1 + delta;
|
|
|
|
cache_line1 = address1 / cache_line_size;
|
|
|
|
cache_line2 = address2 / cache_line_size;
|
|
|
|
if (cache_line1 != cache_line2)
|
2010-07-02 18:34:38 +02:00
|
|
|
{
|
|
|
|
miss_positions += 1;
|
|
|
|
if (miss_positions > max_allowed_miss_positions)
|
|
|
|
return false;
|
|
|
|
}
|
2009-08-13 23:37:24 +02:00
|
|
|
}
|
2010-07-02 18:34:38 +02:00
|
|
|
return true;
|
2009-08-13 23:37:24 +02:00
|
|
|
}
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Prune the prefetch candidate REF using the reuse with BY.
|
|
|
|
If BY_IS_BEFORE is true, BY is before REF in the loop. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
|
|
|
|
bool by_is_before)
|
|
|
|
{
|
2010-05-19 12:36:40 +02:00
|
|
|
HOST_WIDE_INT step;
|
|
|
|
bool backward;
|
2006-02-14 14:51:51 +01:00
|
|
|
HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
|
|
|
|
HOST_WIDE_INT delta = delta_b - delta_r;
|
|
|
|
HOST_WIDE_INT hit_from;
|
|
|
|
unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
|
2009-08-13 23:37:24 +02:00
|
|
|
HOST_WIDE_INT reduced_step;
|
|
|
|
unsigned HOST_WIDE_INT reduced_prefetch_block;
|
|
|
|
tree ref_type;
|
|
|
|
int align_unit;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2010-05-19 12:36:40 +02:00
|
|
|
/* If the step is non constant we cannot calculate prefetch_before. */
|
|
|
|
if (!cst_and_fits_in_hwi (ref->group->step)) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
step = int_cst_value (ref->group->step);
|
|
|
|
|
|
|
|
backward = step < 0;
|
|
|
|
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
if (delta == 0)
|
|
|
|
{
|
|
|
|
/* If the references has the same address, only prefetch the
|
|
|
|
former. */
|
|
|
|
if (by_is_before)
|
|
|
|
ref->prefetch_before = 0;
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!step)
|
|
|
|
{
|
|
|
|
/* If the reference addresses are invariant and fall into the
|
|
|
|
same cache line, prefetch just the first one. */
|
|
|
|
if (!by_is_before)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (ddown (ref->delta, PREFETCH_BLOCK)
|
|
|
|
!= ddown (by->delta, PREFETCH_BLOCK))
|
|
|
|
return;
|
|
|
|
|
|
|
|
ref->prefetch_before = 0;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Only prune the reference that is behind in the array. */
|
|
|
|
if (backward)
|
|
|
|
{
|
|
|
|
if (delta > 0)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* Transform the data so that we may assume that the accesses
|
|
|
|
are forward. */
|
|
|
|
delta = - delta;
|
|
|
|
step = -step;
|
|
|
|
delta_r = PREFETCH_BLOCK - 1 - delta_r;
|
|
|
|
delta_b = PREFETCH_BLOCK - 1 - delta_b;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
if (delta < 0)
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Check whether the two references are likely to hit the same cache
|
|
|
|
line, and how distant the iterations in that it occurs are from
|
|
|
|
each other. */
|
|
|
|
|
|
|
|
if (step <= PREFETCH_BLOCK)
|
|
|
|
{
|
|
|
|
/* The accesses are sure to meet. Let us check when. */
|
|
|
|
hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
|
|
|
|
prefetch_before = (hit_from - delta_r + step - 1) / step;
|
|
|
|
|
2010-05-11 09:29:59 +02:00
|
|
|
/* Do not reduce prefetch_before if we meet beyond cache size. */
|
2010-05-14 17:48:51 +02:00
|
|
|
if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step))
|
2010-05-11 09:29:59 +02:00
|
|
|
prefetch_before = PREFETCH_ALL;
|
2006-02-14 14:51:51 +01:00
|
|
|
if (prefetch_before < ref->prefetch_before)
|
|
|
|
ref->prefetch_before = prefetch_before;
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2009-11-25 11:55:54 +01:00
|
|
|
/* A more complicated case with step > prefetch_block. First reduce
|
2009-08-13 23:37:24 +02:00
|
|
|
the ratio between the step and the cache line size to its simplest
|
2009-11-25 11:55:54 +01:00
|
|
|
terms. The resulting denominator will then represent the number of
|
|
|
|
distinct iterations after which each address will go back to its
|
|
|
|
initial location within the cache line. This computation assumes
|
2009-08-13 23:37:24 +02:00
|
|
|
that PREFETCH_BLOCK is a power of two. */
|
2006-02-14 14:51:51 +01:00
|
|
|
prefetch_block = PREFETCH_BLOCK;
|
2009-08-13 23:37:24 +02:00
|
|
|
reduced_prefetch_block = prefetch_block;
|
|
|
|
reduced_step = step;
|
|
|
|
while ((reduced_step & 1) == 0
|
|
|
|
&& reduced_prefetch_block > 1)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
2009-08-13 23:37:24 +02:00
|
|
|
reduced_step >>= 1;
|
|
|
|
reduced_prefetch_block >>= 1;
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
prefetch_before = delta / step;
|
|
|
|
delta %= step;
|
2009-08-13 23:37:24 +02:00
|
|
|
ref_type = TREE_TYPE (ref->mem);
|
|
|
|
align_unit = TYPE_ALIGN (ref_type) / 8;
|
2010-07-02 18:34:38 +02:00
|
|
|
if (is_miss_rate_acceptable (prefetch_block, step, delta,
|
|
|
|
reduced_prefetch_block, align_unit))
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
2010-05-11 09:29:59 +02:00
|
|
|
/* Do not reduce prefetch_before if we meet beyond cache size. */
|
|
|
|
if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK)
|
|
|
|
prefetch_before = PREFETCH_ALL;
|
2006-02-14 14:51:51 +01:00
|
|
|
if (prefetch_before < ref->prefetch_before)
|
|
|
|
ref->prefetch_before = prefetch_before;
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Try also the following iteration. */
|
|
|
|
prefetch_before++;
|
|
|
|
delta = step - delta;
|
2010-07-02 18:34:38 +02:00
|
|
|
if (is_miss_rate_acceptable (prefetch_block, step, delta,
|
|
|
|
reduced_prefetch_block, align_unit))
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
if (prefetch_before < ref->prefetch_before)
|
|
|
|
ref->prefetch_before = prefetch_before;
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* The ref probably does not reuse by. */
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Prune the prefetch candidate REF using the reuses with other references
|
|
|
|
in REFS. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
|
|
|
|
{
|
|
|
|
struct mem_ref *prune_by;
|
|
|
|
bool before = true;
|
|
|
|
|
|
|
|
prune_ref_by_self_reuse (ref);
|
|
|
|
|
|
|
|
for (prune_by = refs; prune_by; prune_by = prune_by->next)
|
|
|
|
{
|
|
|
|
if (prune_by == ref)
|
|
|
|
{
|
|
|
|
before = false;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
|
|
|
&& ref->write_p
|
|
|
|
&& !prune_by->write_p)
|
|
|
|
continue;
|
|
|
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
|
|
|
&& !ref->write_p
|
|
|
|
&& prune_by->write_p)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
prune_ref_by_group_reuse (ref, prune_by, before);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Prune the prefetch candidates in GROUP using the reuse analysis. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
prune_group_by_reuse (struct mem_ref_group *group)
|
|
|
|
{
|
|
|
|
struct mem_ref *ref_pruned;
|
|
|
|
|
|
|
|
for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
|
|
|
|
{
|
|
|
|
prune_ref_by_reuse (ref_pruned, group->refs);
|
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
{
|
|
|
|
fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
|
|
|
|
|
|
|
|
if (ref_pruned->prefetch_before == PREFETCH_ALL
|
|
|
|
&& ref_pruned->prefetch_mod == 1)
|
|
|
|
fprintf (dump_file, " no restrictions");
|
|
|
|
else if (ref_pruned->prefetch_before == 0)
|
|
|
|
fprintf (dump_file, " do not prefetch");
|
|
|
|
else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
|
|
|
|
fprintf (dump_file, " prefetch once");
|
|
|
|
else
|
|
|
|
{
|
|
|
|
if (ref_pruned->prefetch_before != PREFETCH_ALL)
|
|
|
|
{
|
|
|
|
fprintf (dump_file, " prefetch before ");
|
|
|
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
|
|
|
ref_pruned->prefetch_before);
|
|
|
|
}
|
|
|
|
if (ref_pruned->prefetch_mod != 1)
|
|
|
|
{
|
|
|
|
fprintf (dump_file, " prefetch mod ");
|
|
|
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
|
|
|
ref_pruned->prefetch_mod);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf (dump_file, "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
prune_by_reuse (struct mem_ref_group *groups)
|
|
|
|
{
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
prune_group_by_reuse (groups);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Returns true if we should issue prefetch for REF. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
should_issue_prefetch_p (struct mem_ref *ref)
|
|
|
|
{
|
|
|
|
/* For now do not issue prefetches for only first few of the
|
|
|
|
iterations. */
|
|
|
|
if (ref->prefetch_before != PREFETCH_ALL)
|
2010-05-11 09:28:42 +02:00
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, "Ignoring %p due to prefetch_before\n",
|
|
|
|
(void *) ref);
|
|
|
|
return false;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
/* Do not prefetch nontemporal stores. */
|
|
|
|
if (ref->storent_p)
|
2010-05-11 09:28:42 +02:00
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref);
|
|
|
|
return false;
|
|
|
|
}
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Decide which of the prefetch candidates in GROUPS to prefetch.
|
|
|
|
AHEAD is the number of iterations to prefetch ahead (which corresponds
|
|
|
|
to the number of simultaneous instances of one prefetch running at a
|
|
|
|
time). UNROLL_FACTOR is the factor by that the loop is going to be
|
|
|
|
unrolled. Returns true if there is anything to prefetch. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
|
|
|
|
unsigned ahead)
|
|
|
|
{
|
2006-11-12 21:05:49 +01:00
|
|
|
unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
|
|
|
|
unsigned slots_per_prefetch;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref *ref;
|
|
|
|
bool any = false;
|
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
/* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
|
|
|
|
remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
/* The prefetch will run for AHEAD iterations of the original loop, i.e.,
|
|
|
|
AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
|
|
|
|
it will need a prefetch slot. */
|
|
|
|
slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
|
2006-02-14 14:51:51 +01:00
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
2006-11-12 21:05:49 +01:00
|
|
|
fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
|
|
|
|
slots_per_prefetch);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
/* For now we just take memory references one by one and issue
|
|
|
|
prefetches for as many as possible. The groups are sorted
|
|
|
|
starting with the largest step, since the references with
|
builtins.c, [...]: Fix comment typos.
* builtins.c, c-pragma.h, c-typeck.c, cgraph.c, cgraphunit.c,
combine.c, common.opt, config/dfp-bit.c, config/i386/i386.c,
config/m68k/m68k.c, config/m68k/m68k.md, config/mt/mt.c,
config/mt/mt.h, config/s390/s390.md, df-core.c, df-problems.c,
df-scan.c, df.h, diagnostic.c, expr.c, function.h, gimplify.c,
loop-invariant.c, omp-low.c, opts.c, passes.c,
rtl-factoring.c, rtlanal.c, struct-equiv.c, tree-cfgcleanup.c,
tree-ssa-loop-niter.c, tree-ssa-loop-prefetch.c,
tree-ssa-structalias.c, tree-ssa-threadedge.c,
tree-ssa-threadupdate.c, tree-vect-patterns.c,
tree-vect-transform.c, tree-vectorizer.h, tree-vrp.c,
unwind-dw2.c: Fix comment typos. Follow spelling conventions.
From-SVN: r111721
2006-03-05 00:05:24 +01:00
|
|
|
large step are more likely to cause many cache misses. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
|
|
|
{
|
|
|
|
if (!should_issue_prefetch_p (ref))
|
|
|
|
continue;
|
|
|
|
|
2010-05-18 00:33:53 +02:00
|
|
|
/* The loop is far from being sufficiently unrolled for this
|
|
|
|
prefetch. Do not generate prefetch to avoid many redudant
|
|
|
|
prefetches. */
|
|
|
|
if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO)
|
|
|
|
continue;
|
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
/* If we need to prefetch the reference each PREFETCH_MOD iterations,
|
|
|
|
and we unroll the loop UNROLL_FACTOR times, we need to insert
|
|
|
|
ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
|
|
|
|
iteration. */
|
2006-02-14 14:51:51 +01:00
|
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
|
|
|
/ ref->prefetch_mod);
|
2006-11-12 21:05:49 +01:00
|
|
|
prefetch_slots = n_prefetches * slots_per_prefetch;
|
|
|
|
|
|
|
|
/* If more than half of the prefetches would be lost anyway, do not
|
|
|
|
issue the prefetch. */
|
|
|
|
if (2 * remaining_prefetch_slots < prefetch_slots)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
ref->issue_prefetch_p = true;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
if (remaining_prefetch_slots <= prefetch_slots)
|
|
|
|
return true;
|
|
|
|
remaining_prefetch_slots -= prefetch_slots;
|
2006-02-14 14:51:51 +01:00
|
|
|
any = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return any;
|
|
|
|
}
|
|
|
|
|
2010-06-14 22:51:26 +02:00
|
|
|
/* Return TRUE if no prefetch is going to be generated in the given
|
|
|
|
GROUPS. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
nothing_to_prefetch_p (struct mem_ref_group *groups)
|
|
|
|
{
|
|
|
|
struct mem_ref *ref;
|
|
|
|
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
|
|
|
if (should_issue_prefetch_p (ref))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Estimate the number of prefetches in the given GROUPS.
|
|
|
|
UNROLL_FACTOR is the factor by which LOOP was unrolled. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2009-06-08 18:00:13 +02:00
|
|
|
static int
|
2010-06-14 22:51:26 +02:00
|
|
|
estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
struct mem_ref *ref;
|
2010-06-14 22:51:26 +02:00
|
|
|
unsigned n_prefetches;
|
2009-06-08 18:00:13 +02:00
|
|
|
int prefetch_count = 0;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
|
|
|
if (should_issue_prefetch_p (ref))
|
2010-06-14 22:51:26 +02:00
|
|
|
{
|
|
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
|
|
|
/ ref->prefetch_mod);
|
|
|
|
prefetch_count += n_prefetches;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2009-06-08 18:00:13 +02:00
|
|
|
return prefetch_count;
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Issue prefetches for the reference REF into loop as decided before.
|
|
|
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
|
builtins.c, [...]: Fix comment typos.
* builtins.c, config/arm/arm.c, config/i386/cygwin.h,
config/i386/i386.c, config/ia64/ia64.c, config/s390/fixdfdi.h,
config/sh/sh.c, config/sh/sh.h, df-scan.c, except.c,
haifa-sched.c, optabs.c, rtl.h, sched-deps.c, sched-int.h,
sched-rgn.c, tree-inline.h, tree-ssa-dom.c,
tree-ssa-loop-prefetch.c, tree-ssa-operands.c,
tree-vect-patterns.c, tree-vrp.c: Fix comment typos. Follow
spelling convensions.
* config/ia64/ia64.opt, doc/contrib.texi, doc/invoke.texi,
doc/passes.texi, doc/tm.texi, doc/tree-ssa.texi: Fix comment
typos. Follow spelling conventions.
From-SVN: r112782
2006-04-08 19:01:17 +02:00
|
|
|
is the factor by which LOOP was unrolled. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static void
|
|
|
|
issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
|
|
|
|
{
|
|
|
|
HOST_WIDE_INT delta;
|
2010-05-19 12:36:40 +02:00
|
|
|
tree addr, addr_base, write_p, local, forward;
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple prefetch;
|
|
|
|
gimple_stmt_iterator bsi;
|
2006-02-14 14:51:51 +01:00
|
|
|
unsigned n_prefetches, ap;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
fprintf (dump_file, "Issued%s prefetch for %p.\n",
|
|
|
|
nontemporal ? " nontemporal" : "",
|
|
|
|
(void *) ref);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
bsi = gsi_for_stmt (ref->stmt);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
|
|
|
/ ref->prefetch_mod);
|
|
|
|
addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
|
2008-07-28 16:33:56 +02:00
|
|
|
addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
|
|
|
|
true, NULL, true, GSI_SAME_STMT);
|
2006-11-12 21:05:49 +01:00
|
|
|
write_p = ref->write_p ? integer_one_node : integer_zero_node;
|
2010-07-15 20:47:23 +02:00
|
|
|
local = nontemporal ? integer_zero_node : integer_three_node;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
for (ap = 0; ap < n_prefetches; ap++)
|
|
|
|
{
|
2010-05-19 12:36:40 +02:00
|
|
|
if (cst_and_fits_in_hwi (ref->group->step))
|
|
|
|
{
|
|
|
|
/* Determine the address to prefetch. */
|
|
|
|
delta = (ahead + ap * ref->prefetch_mod) *
|
|
|
|
int_cst_value (ref->group->step);
|
|
|
|
addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
|
|
|
|
addr_base, size_int (delta));
|
|
|
|
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
|
|
|
|
true, GSI_SAME_STMT);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* The step size is non-constant but loop-invariant. We use the
|
|
|
|
heuristic to simply prefetch ahead iterations ahead. */
|
|
|
|
forward = fold_build2 (MULT_EXPR, sizetype,
|
|
|
|
fold_convert (sizetype, ref->group->step),
|
|
|
|
fold_convert (sizetype, size_int (ahead)));
|
|
|
|
addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base,
|
|
|
|
forward);
|
|
|
|
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true,
|
|
|
|
NULL, true, GSI_SAME_STMT);
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Create the prefetch instruction. */
|
2008-07-28 16:33:56 +02:00
|
|
|
prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
|
|
|
|
3, addr, write_p, local);
|
|
|
|
gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Issue prefetches for the references in GROUPS into loop as decided before.
|
|
|
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
|
|
|
|
factor by that LOOP was unrolled. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
issue_prefetches (struct mem_ref_group *groups,
|
|
|
|
unsigned unroll_factor, unsigned ahead)
|
|
|
|
{
|
|
|
|
struct mem_ref *ref;
|
|
|
|
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
|
|
|
if (ref->issue_prefetch_p)
|
|
|
|
issue_prefetch_ref (ref, unroll_factor, ahead);
|
|
|
|
}
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
/* Returns true if REF is a memory write for that a nontemporal store insn
|
|
|
|
can be used. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
nontemporal_store_p (struct mem_ref *ref)
|
|
|
|
{
|
|
|
|
enum machine_mode mode;
|
|
|
|
enum insn_code code;
|
|
|
|
|
|
|
|
/* REF must be a write that is not reused. We require it to be independent
|
|
|
|
on all other memory references in the loop, as the nontemporal stores may
|
|
|
|
be reordered with respect to other memory references. */
|
|
|
|
if (!ref->write_p
|
|
|
|
|| !ref->independent_p
|
|
|
|
|| ref->reuse_distance < L2_CACHE_SIZE_BYTES)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/* Check that we have the storent instruction for the mode. */
|
|
|
|
mode = TYPE_MODE (TREE_TYPE (ref->mem));
|
|
|
|
if (mode == BLKmode)
|
|
|
|
return false;
|
|
|
|
|
optabs.h (optab_handler, [...]): Turn into inline functions that return an insn code.
gcc/
* optabs.h (optab_handler, convert_optab_handler): Turn into
inline functions that return an insn code.
(set_optab_handler, set_convert_optab_handler): New functions.
* builtins.c: Replace optab_handler(X)->insn_code with
optab_handler or set_optab_handler thoughout. Likewise
convert_optab_handler(X)->insn_code with convert_optab_handler
and set_convert_optab_handler.
* expmed.c, expr.c, genopinit.c, ifcvt.c, optabs.c, reload.c,
reload1.c, stmt.c, targhooks.c, tree-ssa-loop-prefetch.c,
tree-ssa-math-opts.c, tree-vect-data-refs.c, tree-vect-generic.c,
tree-vect-loop.c, tree-vect-patterns.c, tree-vect-slp.c,
tree-vect-stmts.c, config/m32c/m32c.c, config/rs6000/rs6000.c,
config/spu/spu.c: Likewise.
From-SVN: r161808
2010-07-05 00:13:09 +02:00
|
|
|
code = optab_handler (storent_optab, mode);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
return code != CODE_FOR_nothing;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* If REF is a nontemporal store, we mark the corresponding modify statement
|
|
|
|
and return true. Otherwise, we return false. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
mark_nontemporal_store (struct mem_ref *ref)
|
|
|
|
{
|
|
|
|
if (!nontemporal_store_p (ref))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
|
|
|
|
(void *) ref);
|
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple_assign_set_nontemporal_move (ref->stmt, true);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
ref->storent_p = true;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Issue a memory fence instruction after LOOP. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
emit_mfence_after_loop (struct loop *loop)
|
|
|
|
{
|
|
|
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
|
|
|
edge exit;
|
2008-07-28 16:33:56 +02:00
|
|
|
gimple call;
|
|
|
|
gimple_stmt_iterator bsi;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
|
|
|
{
|
2008-07-28 16:33:56 +02:00
|
|
|
call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
|
|
|
if (!single_pred_p (exit->dest)
|
|
|
|
/* If possible, we prefer not to insert the fence on other paths
|
|
|
|
in cfg. */
|
|
|
|
&& !(exit->flags & EDGE_ABNORMAL))
|
|
|
|
split_loop_exit_edge (exit);
|
2008-07-28 16:33:56 +02:00
|
|
|
bsi = gsi_after_labels (exit->dest);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
2008-07-28 16:33:56 +02:00
|
|
|
gsi_insert_before (&bsi, call, GSI_NEW_STMT);
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
mark_virtual_ops_for_renaming (call);
|
|
|
|
}
|
|
|
|
|
|
|
|
VEC_free (edge, heap, exits);
|
|
|
|
update_ssa (TODO_update_ssa_only_virtuals);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Returns true if we can use storent in loop, false otherwise. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
may_use_storent_in_loop_p (struct loop *loop)
|
|
|
|
{
|
|
|
|
bool ret = true;
|
|
|
|
|
|
|
|
if (loop->inner != NULL)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/* If we must issue a mfence insn after using storent, check that there
|
|
|
|
is a suitable place for it at each of the loop exits. */
|
|
|
|
if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
|
|
|
{
|
|
|
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
|
|
|
unsigned i;
|
|
|
|
edge exit;
|
|
|
|
|
|
|
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
|
|
|
if ((exit->flags & EDGE_ABNORMAL)
|
|
|
|
&& exit->dest == EXIT_BLOCK_PTR)
|
|
|
|
ret = false;
|
|
|
|
|
|
|
|
VEC_free (edge, heap, exits);
|
|
|
|
}
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
|
|
|
|
references in the loop. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
|
|
|
|
{
|
|
|
|
struct mem_ref *ref;
|
|
|
|
bool any = false;
|
|
|
|
|
|
|
|
if (!may_use_storent_in_loop_p (loop))
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (; groups; groups = groups->next)
|
|
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
|
|
|
any |= mark_nontemporal_store (ref);
|
|
|
|
|
|
|
|
if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
|
|
|
emit_mfence_after_loop (loop);
|
|
|
|
}
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Determines whether we can profitably unroll LOOP FACTOR times, and if
|
|
|
|
this is the case, fill in DESC by the description of number of
|
|
|
|
iterations. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
|
|
|
|
unsigned factor)
|
|
|
|
{
|
|
|
|
if (!can_unroll_loop_p (loop, factor, desc))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/* We only consider loops without control flow for unrolling. This is not
|
|
|
|
a hard restriction -- tree_unroll_loop works with arbitrary loops
|
|
|
|
as well; but the unrolling/prefetching is usually more profitable for
|
|
|
|
loops consisting of a single basic block, and we want to limit the
|
|
|
|
code growth. */
|
|
|
|
if (loop->num_nodes > 2)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Determine the coefficient by that unroll LOOP, from the information
|
|
|
|
contained in the list of memory references REFS. Description of
|
2007-03-01 23:14:23 +01:00
|
|
|
umber of iterations of LOOP is stored to DESC. NINSNS is the number of
|
|
|
|
insns of the LOOP. EST_NITER is the estimated number of iterations of
|
|
|
|
the loop, or -1 if no estimate is available. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static unsigned
|
|
|
|
determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
|
2007-03-01 23:14:23 +01:00
|
|
|
unsigned ninsns, struct tree_niter_desc *desc,
|
|
|
|
HOST_WIDE_INT est_niter)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
2006-11-12 21:05:49 +01:00
|
|
|
unsigned upper_bound;
|
|
|
|
unsigned nfactor, factor, mod_constraint;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct mem_ref_group *agp;
|
|
|
|
struct mem_ref *ref;
|
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
/* First check whether the loop is not too large to unroll. We ignore
|
|
|
|
PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
|
|
|
|
from unrolling them enough to make exactly one cache line covered by each
|
|
|
|
iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
|
|
|
|
us from unrolling the loops too many times in cases where we only expect
|
|
|
|
gains from better scheduling and decreasing loop overhead, which is not
|
|
|
|
the case here. */
|
|
|
|
upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
|
2007-03-01 23:14:23 +01:00
|
|
|
|
|
|
|
/* If we unrolled the loop more times than it iterates, the unrolled version
|
|
|
|
of the loop would be never entered. */
|
|
|
|
if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
|
|
|
|
upper_bound = est_niter;
|
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
if (upper_bound <= 1)
|
2006-02-14 14:51:51 +01:00
|
|
|
return 1;
|
|
|
|
|
2006-11-12 21:05:49 +01:00
|
|
|
/* Choose the factor so that we may prefetch each cache just once,
|
|
|
|
but bound the unrolling by UPPER_BOUND. */
|
|
|
|
factor = 1;
|
2006-02-14 14:51:51 +01:00
|
|
|
for (agp = refs; agp; agp = agp->next)
|
|
|
|
for (ref = agp->refs; ref; ref = ref->next)
|
2006-11-12 21:05:49 +01:00
|
|
|
if (should_issue_prefetch_p (ref))
|
|
|
|
{
|
|
|
|
mod_constraint = ref->prefetch_mod;
|
|
|
|
nfactor = least_common_multiple (mod_constraint, factor);
|
|
|
|
if (nfactor <= upper_bound)
|
|
|
|
factor = nfactor;
|
|
|
|
}
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (!should_unroll_loop_p (loop, desc, factor))
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
return factor;
|
|
|
|
}
|
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
/* Returns the total volume of the memory references REFS, taking into account
|
|
|
|
reuses in the innermost loop and cache line size. TODO -- we should also
|
|
|
|
take into account reuses across the iterations of the loops in the loop
|
|
|
|
nest. */
|
|
|
|
|
|
|
|
static unsigned
|
|
|
|
volume_of_references (struct mem_ref_group *refs)
|
|
|
|
{
|
|
|
|
unsigned volume = 0;
|
|
|
|
struct mem_ref_group *gr;
|
|
|
|
struct mem_ref *ref;
|
|
|
|
|
|
|
|
for (gr = refs; gr; gr = gr->next)
|
|
|
|
for (ref = gr->refs; ref; ref = ref->next)
|
|
|
|
{
|
|
|
|
/* Almost always reuses another value? */
|
|
|
|
if (ref->prefetch_before != PREFETCH_ALL)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* If several iterations access the same cache line, use the size of
|
|
|
|
the line divided by this number. Otherwise, a cache line is
|
|
|
|
accessed in each iteration. TODO -- in the latter case, we should
|
|
|
|
take the size of the reference into account, rounding it up on cache
|
|
|
|
line size multiple. */
|
|
|
|
volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
|
|
|
|
}
|
|
|
|
return volume;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Returns the volume of memory references accessed across VEC iterations of
|
|
|
|
loops, whose sizes are described in the LOOP_SIZES array. N is the number
|
|
|
|
of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
|
|
|
|
|
|
|
|
static unsigned
|
|
|
|
volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; i < n; i++)
|
|
|
|
if (vec[i] != 0)
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (i == n)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
gcc_assert (vec[i] > 0);
|
|
|
|
|
|
|
|
/* We ignore the parts of the distance vector in subloops, since usually
|
|
|
|
the numbers of iterations are much smaller. */
|
|
|
|
return loop_sizes[i] * vec[i];
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
|
|
|
|
at the position corresponding to the loop of the step. N is the depth
|
|
|
|
of the considered loop nest, and, LOOP is its innermost loop. */
|
|
|
|
|
|
|
|
static void
|
|
|
|
add_subscript_strides (tree access_fn, unsigned stride,
|
|
|
|
HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
|
|
|
|
{
|
|
|
|
struct loop *aloop;
|
|
|
|
tree step;
|
|
|
|
HOST_WIDE_INT astep;
|
|
|
|
unsigned min_depth = loop_depth (loop) - n;
|
|
|
|
|
|
|
|
while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
|
|
|
|
{
|
|
|
|
aloop = get_chrec_loop (access_fn);
|
|
|
|
step = CHREC_RIGHT (access_fn);
|
|
|
|
access_fn = CHREC_LEFT (access_fn);
|
|
|
|
|
|
|
|
if ((unsigned) loop_depth (aloop) <= min_depth)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (host_integerp (step, 0))
|
|
|
|
astep = tree_low_cst (step, 0);
|
|
|
|
else
|
|
|
|
astep = L1_CACHE_LINE_SIZE;
|
|
|
|
|
|
|
|
strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
|
|
|
|
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Returns the volume of memory references accessed between two consecutive
|
|
|
|
self-reuses of the reference DR. We consider the subscripts of DR in N
|
|
|
|
loops, and LOOP_SIZES contains the volumes of accesses in each of the
|
|
|
|
loops. LOOP is the innermost loop of the current loop nest. */
|
|
|
|
|
|
|
|
static unsigned
|
|
|
|
self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
|
|
|
|
struct loop *loop)
|
|
|
|
{
|
|
|
|
tree stride, access_fn;
|
|
|
|
HOST_WIDE_INT *strides, astride;
|
|
|
|
VEC (tree, heap) *access_fns;
|
|
|
|
tree ref = DR_REF (dr);
|
|
|
|
unsigned i, ret = ~0u;
|
|
|
|
|
|
|
|
/* In the following example:
|
|
|
|
|
|
|
|
for (i = 0; i < N; i++)
|
|
|
|
for (j = 0; j < N; j++)
|
|
|
|
use (a[j][i]);
|
|
|
|
the same cache line is accessed each N steps (except if the change from
|
|
|
|
i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
|
|
|
|
we cannot rely purely on the results of the data dependence analysis.
|
|
|
|
|
|
|
|
Instead, we compute the stride of the reference in each loop, and consider
|
|
|
|
the innermost loop in that the stride is less than cache size. */
|
|
|
|
|
|
|
|
strides = XCNEWVEC (HOST_WIDE_INT, n);
|
|
|
|
access_fns = DR_ACCESS_FNS (dr);
|
|
|
|
|
|
|
|
for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
|
|
|
|
{
|
|
|
|
/* Keep track of the reference corresponding to the subscript, so that we
|
|
|
|
know its stride. */
|
|
|
|
while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
|
|
|
|
ref = TREE_OPERAND (ref, 0);
|
2009-11-25 11:55:54 +01:00
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
if (TREE_CODE (ref) == ARRAY_REF)
|
|
|
|
{
|
|
|
|
stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
|
|
|
|
if (host_integerp (stride, 1))
|
|
|
|
astride = tree_low_cst (stride, 1);
|
|
|
|
else
|
|
|
|
astride = L1_CACHE_LINE_SIZE;
|
|
|
|
|
|
|
|
ref = TREE_OPERAND (ref, 0);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
astride = 1;
|
|
|
|
|
|
|
|
add_subscript_strides (access_fn, astride, strides, n, loop);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = n; i-- > 0; )
|
|
|
|
{
|
|
|
|
unsigned HOST_WIDE_INT s;
|
|
|
|
|
|
|
|
s = strides[i] < 0 ? -strides[i] : strides[i];
|
|
|
|
|
|
|
|
if (s < (unsigned) L1_CACHE_LINE_SIZE
|
|
|
|
&& (loop_sizes[i]
|
|
|
|
> (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
|
|
|
|
{
|
|
|
|
ret = loop_sizes[i];
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
free (strides);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Determines the distance till the first reuse of each reference in REFS
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
|
|
|
|
memory references in the loop. */
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
|
|
|
|
static void
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
|
|
|
|
bool no_other_refs)
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
{
|
|
|
|
struct loop *nest, *aloop;
|
|
|
|
VEC (data_reference_p, heap) *datarefs = NULL;
|
|
|
|
VEC (ddr_p, heap) *dependences = NULL;
|
|
|
|
struct mem_ref_group *gr;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
struct mem_ref *ref, *refb;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
VEC (loop_p, heap) *vloops = NULL;
|
|
|
|
unsigned *loop_data_size;
|
|
|
|
unsigned i, j, n;
|
|
|
|
unsigned volume, dist, adist;
|
|
|
|
HOST_WIDE_INT vol;
|
|
|
|
data_reference_p dr;
|
|
|
|
ddr_p dep;
|
|
|
|
|
|
|
|
if (loop->inner)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* Find the outermost loop of the loop nest of loop (we require that
|
|
|
|
there are no sibling loops inside the nest). */
|
|
|
|
nest = loop;
|
|
|
|
while (1)
|
|
|
|
{
|
|
|
|
aloop = loop_outer (nest);
|
|
|
|
|
|
|
|
if (aloop == current_loops->tree_root
|
|
|
|
|| aloop->inner->next)
|
|
|
|
break;
|
|
|
|
|
|
|
|
nest = aloop;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* For each loop, determine the amount of data accessed in each iteration.
|
|
|
|
We use this to estimate whether the reference is evicted from the
|
|
|
|
cache before its reuse. */
|
|
|
|
find_loop_nest (nest, &vloops);
|
|
|
|
n = VEC_length (loop_p, vloops);
|
|
|
|
loop_data_size = XNEWVEC (unsigned, n);
|
|
|
|
volume = volume_of_references (refs);
|
|
|
|
i = n;
|
|
|
|
while (i-- != 0)
|
|
|
|
{
|
|
|
|
loop_data_size[i] = volume;
|
|
|
|
/* Bound the volume by the L2 cache size, since above this bound,
|
|
|
|
all dependence distances are equivalent. */
|
|
|
|
if (volume > L2_CACHE_SIZE_BYTES)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
aloop = VEC_index (loop_p, vloops, i);
|
|
|
|
vol = estimated_loop_iterations_int (aloop, false);
|
|
|
|
if (vol < 0)
|
|
|
|
vol = expected_loop_iterations (aloop);
|
|
|
|
volume *= vol;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Prepare the references in the form suitable for data dependence
|
auto-inc-dec.c, [...]: Fix comment typos.
* auto-inc-dec.c, c-incpath.c, config/c4x/libgcc.S,
config/sh/divcost-analysis, dbgcnt.def, df-core.c,
df-problems.c, df-scan.c, df.h, dominance.c, dse.c, regstat.c,
tree-data-ref.c, tree-ssa-loop-im.c, tree-ssa-loop-prefetch.c,
tree-vect-transform.c: Fix comment typos. Follow spelling
conventions.
From-SVN: r125666
2007-06-13 04:27:04 +02:00
|
|
|
analysis. We ignore unanalyzable data references (the results
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
are used just as a heuristics to estimate temporality of the
|
|
|
|
references, hence we do not need to worry about correctness). */
|
|
|
|
for (gr = refs; gr; gr = gr->next)
|
|
|
|
for (ref = gr->refs; ref; ref = ref->next)
|
|
|
|
{
|
|
|
|
dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
|
|
|
|
|
|
|
|
if (dr)
|
|
|
|
{
|
|
|
|
ref->reuse_distance = volume;
|
|
|
|
dr->aux = ref;
|
|
|
|
VEC_safe_push (data_reference_p, heap, datarefs, dr);
|
|
|
|
}
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
else
|
|
|
|
no_other_refs = false;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
|
|
|
|
{
|
|
|
|
dist = self_reuse_distance (dr, loop_data_size, n, loop);
|
ggc.h (GGC_RESIZEVAR): New, reorder macros.
* ggc.h (GGC_RESIZEVAR): New, reorder macros.
* tracer.c (tail_duplicate): Fix for -Wc++-compat.
* tree-affine.c (aff_combination_expand, free_name_expansion):
Likewise.
* tree-cfg.c (replace_by_duplicate_decl, replace_ssa_name,
move_stmt_r, new_label_mapper): Likewise.
* tree-complex.c (cvc_lookup): Likewise.
* tree-dfa.c (create_function_ann): Likewise.
* tree-dump.c (dump_register): Likewise.
* tree-if-conv.c (tree_if_conversion, add_to_predicate_list,
find_phi_replacement_condition): Likewise.
* tree-inline.c (copy_phis_for_bb, estimate_num_insns_1,
tree_function_versioning): Likewise.
* tree-into-ssa.c (cmp_dfsnum): Likewise.
* tree-iterator.c (tsi_link_before, tsi_link_after): Likewise.
* tree-nested.c (lookup_field_for_decl, lookup_tramp_for_decl,
get_nonlocal_debug_decl, convert_nonlocal_reference,
convert_nonlocal_omp_clauses, get_local_debug_decl,
convert_local_reference, convert_local_omp_clauses,
convert_nl_goto_reference, convert_nl_goto_receiver,
convert_tramp_reference, convert_call_expr): Likewise.
* tree-outof-ssa.c (contains_tree_r): Likewise.
* tree-parloops.c (reduction_phi, initialize_reductions,
eliminate_local_variables_1, add_field_for_reduction,
add_field_for_name, create_phi_for_local_result,
create_call_for_reduction_1, create_loads_for_reductions,
create_stores_for_reduction, create_loads_and_stores_for_name):
Likewise.
* tree-phinodes.c (allocate_phi_node): Likewise.
* tree-predcom.c (order_drefs, execute_pred_commoning_cbck):
Likewise.
* tree-sra.c (sra_elt_hash, sra_elt_eq, lookup_element):
Likewise.
* tree-ssa-alias.c (get_mem_sym_stats_for): Likewise.
* tree-ssa-coalesce.c (compare_pairs): Likewise.
* tree-ssa-loop-im.c (mem_ref_in_stmt, memref_hash, memref_eq,
memref_free, gather_mem_refs_stmt, vtoe_hash, vtoe_eq, vtoe_free,
record_vop_access, get_vop_accesses, get_vop_stores): Likewise.
* tree-ssa-loop-prefetch.c (determine_loop_nest_reuse): Likewise.
* tree-ssa-sccvn.c (VN_INFO_GET, free_phi, free_reference,
vn_nary_op_insert): Likewise.
* tree-ssa.c (redirect_edge_var_map_add,
redirect_edge_var_map_clear, redirect_edge_var_map_dup): Likewise.
* tree-vectorizer.c (vectorize_loops): Likewise.
* tree.c (make_node_stat, copy_node_stat, build_int_cst_wide,
build_fixed, build_real, make_tree_binfo_stat, make_tree_vec_stat,
tree_cons_stat, build1_stat, build_variant_type_copy,
decl_init_priority_lookup, decl_fini_priority_lookup,
decl_priority_info, decl_restrict_base_lookup,
decl_restrict_base_insert, decl_debug_expr_lookup,
decl_debug_expr_insert, decl_value_expr_lookup,
decl_value_expr_insert, type_hash_eq, type_hash_lookup,
type_hash_add, get_file_function_name, tree_check_failed,
tree_not_check_failed, tree_range_check_failed,
omp_clause_range_check_failed, build_omp_clause,
build_vl_exp_stat): Likewise.
* value-prof.c (gimple_histogram_value,
gimple_duplicate_stmt_histograms): Likewise.
* var-tracking.c (attrs_list_insert, attrs_list_copy,
unshare_variable, variable_union_info_cmp_pos, variable_union,
dataflow_set_different_1, dataflow_set_different_2,
vt_find_locations, variable_was_changed, set_variable_part,
emit_notes_for_differences_1, emit_notes_for_differences_2):
Likewise.
* varasm.c (prefix_name, emutls_decl, section_entry_eq,
section_entry_hash, object_block_entry_eq,
object_block_entry_hash, create_block_symbol,
initialize_cold_section_name, default_function_rodata_section,
strip_reg_name, set_user_assembler_name, const_desc_eq,
build_constant_desc, output_constant_def, lookup_constant_def,
const_desc_rtx_hash, const_desc_rtx_eq, const_rtx_hash_1,
create_constant_pool, force_const_mem, compute_reloc_for_rtx_1,
default_internal_label): Likewise.
* varray.c (varray_init, varray_grow): Likewise.
* vec.c (vec_gc_o_reserve_1, vec_heap_o_reserve_1): Likewise.
From-SVN: r136992
2008-06-20 20:34:07 +02:00
|
|
|
ref = (struct mem_ref *) dr->aux;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
if (ref->reuse_distance > dist)
|
|
|
|
ref->reuse_distance = dist;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
|
|
|
if (no_other_refs)
|
|
|
|
ref->independent_p = true;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
compute_all_dependences (datarefs, &dependences, vloops, true);
|
|
|
|
|
|
|
|
for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
|
|
|
|
{
|
|
|
|
if (DDR_ARE_DEPENDENT (dep) == chrec_known)
|
|
|
|
continue;
|
|
|
|
|
ggc.h (GGC_RESIZEVAR): New, reorder macros.
* ggc.h (GGC_RESIZEVAR): New, reorder macros.
* tracer.c (tail_duplicate): Fix for -Wc++-compat.
* tree-affine.c (aff_combination_expand, free_name_expansion):
Likewise.
* tree-cfg.c (replace_by_duplicate_decl, replace_ssa_name,
move_stmt_r, new_label_mapper): Likewise.
* tree-complex.c (cvc_lookup): Likewise.
* tree-dfa.c (create_function_ann): Likewise.
* tree-dump.c (dump_register): Likewise.
* tree-if-conv.c (tree_if_conversion, add_to_predicate_list,
find_phi_replacement_condition): Likewise.
* tree-inline.c (copy_phis_for_bb, estimate_num_insns_1,
tree_function_versioning): Likewise.
* tree-into-ssa.c (cmp_dfsnum): Likewise.
* tree-iterator.c (tsi_link_before, tsi_link_after): Likewise.
* tree-nested.c (lookup_field_for_decl, lookup_tramp_for_decl,
get_nonlocal_debug_decl, convert_nonlocal_reference,
convert_nonlocal_omp_clauses, get_local_debug_decl,
convert_local_reference, convert_local_omp_clauses,
convert_nl_goto_reference, convert_nl_goto_receiver,
convert_tramp_reference, convert_call_expr): Likewise.
* tree-outof-ssa.c (contains_tree_r): Likewise.
* tree-parloops.c (reduction_phi, initialize_reductions,
eliminate_local_variables_1, add_field_for_reduction,
add_field_for_name, create_phi_for_local_result,
create_call_for_reduction_1, create_loads_for_reductions,
create_stores_for_reduction, create_loads_and_stores_for_name):
Likewise.
* tree-phinodes.c (allocate_phi_node): Likewise.
* tree-predcom.c (order_drefs, execute_pred_commoning_cbck):
Likewise.
* tree-sra.c (sra_elt_hash, sra_elt_eq, lookup_element):
Likewise.
* tree-ssa-alias.c (get_mem_sym_stats_for): Likewise.
* tree-ssa-coalesce.c (compare_pairs): Likewise.
* tree-ssa-loop-im.c (mem_ref_in_stmt, memref_hash, memref_eq,
memref_free, gather_mem_refs_stmt, vtoe_hash, vtoe_eq, vtoe_free,
record_vop_access, get_vop_accesses, get_vop_stores): Likewise.
* tree-ssa-loop-prefetch.c (determine_loop_nest_reuse): Likewise.
* tree-ssa-sccvn.c (VN_INFO_GET, free_phi, free_reference,
vn_nary_op_insert): Likewise.
* tree-ssa.c (redirect_edge_var_map_add,
redirect_edge_var_map_clear, redirect_edge_var_map_dup): Likewise.
* tree-vectorizer.c (vectorize_loops): Likewise.
* tree.c (make_node_stat, copy_node_stat, build_int_cst_wide,
build_fixed, build_real, make_tree_binfo_stat, make_tree_vec_stat,
tree_cons_stat, build1_stat, build_variant_type_copy,
decl_init_priority_lookup, decl_fini_priority_lookup,
decl_priority_info, decl_restrict_base_lookup,
decl_restrict_base_insert, decl_debug_expr_lookup,
decl_debug_expr_insert, decl_value_expr_lookup,
decl_value_expr_insert, type_hash_eq, type_hash_lookup,
type_hash_add, get_file_function_name, tree_check_failed,
tree_not_check_failed, tree_range_check_failed,
omp_clause_range_check_failed, build_omp_clause,
build_vl_exp_stat): Likewise.
* value-prof.c (gimple_histogram_value,
gimple_duplicate_stmt_histograms): Likewise.
* var-tracking.c (attrs_list_insert, attrs_list_copy,
unshare_variable, variable_union_info_cmp_pos, variable_union,
dataflow_set_different_1, dataflow_set_different_2,
vt_find_locations, variable_was_changed, set_variable_part,
emit_notes_for_differences_1, emit_notes_for_differences_2):
Likewise.
* varasm.c (prefix_name, emutls_decl, section_entry_eq,
section_entry_hash, object_block_entry_eq,
object_block_entry_hash, create_block_symbol,
initialize_cold_section_name, default_function_rodata_section,
strip_reg_name, set_user_assembler_name, const_desc_eq,
build_constant_desc, output_constant_def, lookup_constant_def,
const_desc_rtx_hash, const_desc_rtx_eq, const_rtx_hash_1,
create_constant_pool, force_const_mem, compute_reloc_for_rtx_1,
default_internal_label): Likewise.
* varray.c (varray_init, varray_grow): Likewise.
* vec.c (vec_gc_o_reserve_1, vec_heap_o_reserve_1): Likewise.
From-SVN: r136992
2008-06-20 20:34:07 +02:00
|
|
|
ref = (struct mem_ref *) DDR_A (dep)->aux;
|
|
|
|
refb = (struct mem_ref *) DDR_B (dep)->aux;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
|
|
|
|
|| DDR_NUM_DIST_VECTS (dep) == 0)
|
|
|
|
{
|
auto-inc-dec.c, [...]: Fix comment typos.
* auto-inc-dec.c, c-incpath.c, config/c4x/libgcc.S,
config/sh/divcost-analysis, dbgcnt.def, df-core.c,
df-problems.c, df-scan.c, df.h, dominance.c, dse.c, regstat.c,
tree-data-ref.c, tree-ssa-loop-im.c, tree-ssa-loop-prefetch.c,
tree-vect-transform.c: Fix comment typos. Follow spelling
conventions.
From-SVN: r125666
2007-06-13 04:27:04 +02:00
|
|
|
/* If the dependence cannot be analyzed, assume that there might be
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
a reuse. */
|
|
|
|
dist = 0;
|
2009-11-25 11:55:54 +01:00
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
ref->independent_p = false;
|
|
|
|
refb->independent_p = false;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
auto-inc-dec.c, [...]: Fix comment typos.
* auto-inc-dec.c, c-incpath.c, config/c4x/libgcc.S,
config/sh/divcost-analysis, dbgcnt.def, df-core.c,
df-problems.c, df-scan.c, df.h, dominance.c, dse.c, regstat.c,
tree-data-ref.c, tree-ssa-loop-im.c, tree-ssa-loop-prefetch.c,
tree-vect-transform.c: Fix comment typos. Follow spelling
conventions.
From-SVN: r125666
2007-06-13 04:27:04 +02:00
|
|
|
/* The distance vectors are normalized to be always lexicographically
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
positive, hence we cannot tell just from them whether DDR_A comes
|
|
|
|
before DDR_B or vice versa. However, it is not important,
|
|
|
|
anyway -- if DDR_A is close to DDR_B, then it is either reused in
|
|
|
|
DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
|
|
|
|
in cache (and marking it as nontemporal would not affect
|
|
|
|
anything). */
|
|
|
|
|
|
|
|
dist = volume;
|
|
|
|
for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
|
|
|
|
{
|
|
|
|
adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
|
|
|
|
loop_data_size, n);
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
/* If this is a dependence in the innermost loop (i.e., the
|
|
|
|
distances in all superloops are zero) and it is not
|
|
|
|
the trivial self-dependence with distance zero, record that
|
|
|
|
the references are not completely independent. */
|
|
|
|
if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
|
|
|
|
&& (ref != refb
|
|
|
|
|| DDR_DIST_VECT (dep, j)[n-1] != 0))
|
|
|
|
{
|
|
|
|
ref->independent_p = false;
|
|
|
|
refb->independent_p = false;
|
|
|
|
}
|
|
|
|
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
/* Ignore accesses closer than
|
|
|
|
L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
|
|
|
so that we use nontemporal prefetches e.g. if single memory
|
|
|
|
location is accessed several times in a single iteration of
|
|
|
|
the loop. */
|
|
|
|
if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (adist < dist)
|
|
|
|
dist = adist;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ref->reuse_distance > dist)
|
|
|
|
ref->reuse_distance = dist;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
if (refb->reuse_distance > dist)
|
|
|
|
refb->reuse_distance = dist;
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
free_dependence_relations (dependences);
|
|
|
|
free_data_refs (datarefs);
|
|
|
|
free (loop_data_size);
|
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
{
|
|
|
|
fprintf (dump_file, "Reuse distances:\n");
|
|
|
|
for (gr = refs; gr; gr = gr->next)
|
|
|
|
for (ref = gr->refs; ref; ref = ref->next)
|
|
|
|
fprintf (dump_file, " ref %p distance %u\n",
|
|
|
|
(void *) ref, ref->reuse_distance);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Determine whether or not the trip count to ahead ratio is too small based
|
|
|
|
on prefitablility consideration.
|
2009-06-08 18:00:13 +02:00
|
|
|
AHEAD: the iteration ahead distance,
|
2010-07-10 01:08:55 +02:00
|
|
|
EST_NITER: the estimated trip count. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter)
|
|
|
|
{
|
|
|
|
/* Assume trip count to ahead ratio is big enough if the trip count could not
|
|
|
|
be estimated at compile time. */
|
|
|
|
if (est_niter < 0)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
|
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file,
|
|
|
|
"Not prefetching -- loop estimated to roll only %d times\n",
|
|
|
|
(int) est_niter);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Determine whether or not the number of memory references in the loop is
|
|
|
|
reasonable based on the profitablity and compilation time considerations.
|
2009-06-08 18:00:13 +02:00
|
|
|
NINSNS: estimated number of instructions in the loop,
|
|
|
|
MEM_REF_COUNT: total number of memory references in the loop. */
|
|
|
|
|
2009-11-25 11:55:54 +01:00
|
|
|
static bool
|
2010-07-10 01:08:55 +02:00
|
|
|
mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count)
|
2009-06-08 18:00:13 +02:00
|
|
|
{
|
2010-07-10 01:08:55 +02:00
|
|
|
int insn_to_mem_ratio;
|
2009-06-08 18:00:13 +02:00
|
|
|
|
|
|
|
if (mem_ref_count == 0)
|
|
|
|
return false;
|
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Miss rate computation (is_miss_rate_acceptable) and dependence analysis
|
|
|
|
(compute_all_dependences) have high costs based on quadratic complexity.
|
|
|
|
To avoid huge compilation time, we give up prefetching if mem_ref_count
|
|
|
|
is too large. */
|
|
|
|
if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP)
|
|
|
|
return false;
|
|
|
|
|
2009-11-25 11:55:54 +01:00
|
|
|
/* Prefetching improves performance by overlapping cache missing
|
|
|
|
memory accesses with CPU operations. If the loop does not have
|
|
|
|
enough CPU operations to overlap with memory operations, prefetching
|
|
|
|
won't give a significant benefit. One approximate way of checking
|
|
|
|
this is to require the ratio of instructions to memory references to
|
2009-06-08 18:00:13 +02:00
|
|
|
be above a certain limit. This approximation works well in practice.
|
|
|
|
TODO: Implement a more precise computation by estimating the time
|
|
|
|
for each CPU or memory op in the loop. Time estimates for memory ops
|
|
|
|
should account for cache misses. */
|
2009-11-25 11:55:54 +01:00
|
|
|
insn_to_mem_ratio = ninsns / mem_ref_count;
|
2009-06-08 18:00:13 +02:00
|
|
|
|
|
|
|
if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
|
2010-05-07 18:15:45 +02:00
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file,
|
|
|
|
"Not prefetching -- instruction to memory reference ratio (%d) too small\n",
|
|
|
|
insn_to_mem_ratio);
|
|
|
|
return false;
|
|
|
|
}
|
2009-06-08 18:00:13 +02:00
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Determine whether or not the instruction to prefetch ratio in the loop is
|
|
|
|
too small based on the profitablity consideration.
|
|
|
|
NINSNS: estimated number of instructions in the loop,
|
|
|
|
PREFETCH_COUNT: an estimate of the number of prefetches,
|
|
|
|
UNROLL_FACTOR: the factor to unroll the loop if prefetching. */
|
|
|
|
|
|
|
|
static bool
|
|
|
|
insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count,
|
|
|
|
unsigned unroll_factor)
|
|
|
|
{
|
|
|
|
int insn_to_prefetch_ratio;
|
|
|
|
|
2010-05-18 00:34:03 +02:00
|
|
|
/* Prefetching most likely causes performance degradation when the instruction
|
|
|
|
to prefetch ratio is too small. Too many prefetch instructions in a loop
|
|
|
|
may reduce the I-cache performance.
|
2010-05-07 18:15:52 +02:00
|
|
|
(unroll_factor * ninsns) is used to estimate the number of instructions in
|
|
|
|
the unrolled loop. This implementation is a bit simplistic -- the number
|
|
|
|
of issued prefetch instructions is also affected by unrolling. So,
|
|
|
|
prefetch_mod and the unroll factor should be taken into account when
|
|
|
|
determining prefetch_count. Also, the number of insns of the unrolled
|
|
|
|
loop will usually be significantly smaller than the number of insns of the
|
|
|
|
original loop * unroll_factor (at least the induction variable increases
|
|
|
|
and the exit branches will get eliminated), so it might be better to use
|
|
|
|
tree_estimate_loop_size + estimated_unrolled_size. */
|
2010-05-18 00:34:03 +02:00
|
|
|
insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count;
|
|
|
|
if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO)
|
2009-06-08 18:00:13 +02:00
|
|
|
{
|
2010-05-18 00:34:03 +02:00
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file,
|
|
|
|
"Not prefetching -- instruction to prefetch ratio (%d) too small\n",
|
|
|
|
insn_to_prefetch_ratio);
|
2010-07-10 01:08:55 +02:00
|
|
|
return true;
|
2009-06-08 18:00:13 +02:00
|
|
|
}
|
2009-11-25 11:55:54 +01:00
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
return false;
|
2009-06-08 18:00:13 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Issue prefetch instructions for array references in LOOP. Returns
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
true if the LOOP was unrolled. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
static bool
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
loop_prefetch_arrays (struct loop *loop)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
struct mem_ref_group *refs;
|
2007-03-01 23:14:23 +01:00
|
|
|
unsigned ahead, ninsns, time, unroll_factor;
|
|
|
|
HOST_WIDE_INT est_niter;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct tree_niter_desc desc;
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
bool unrolled = false, no_other_refs;
|
2009-06-08 18:00:13 +02:00
|
|
|
unsigned prefetch_count;
|
|
|
|
unsigned mem_ref_count;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
loop-unswitch.c (unswitch_single_loop): Use optimize_loop_for_speed_p.
* loop-unswitch.c (unswitch_single_loop): Use optimize_loop_for_speed_p.
* tree-ssa-threadupdate.c (mark_threaded_blocks): Use optimize_function_for_size_p.
* tracer.c (ignore_bb_p): Use optimize_bb_for_size_p.
* postreload-gcse.c (eliminate_partially_redundant_load): Use optimize_bb_for_size_p.
* value-prof.c (gimple_divmod_fixed_value_transform,
gimple_mod_pow2_value_transform, gimple_mod_subtract_transform,
gimple_stringops_transform): Use optimize_bb_for_size_p.
* ipa-cp.c (ipcp_insert_stage): Use optimize_function_for_size_p.
* final.c (compute_alignments): Use optimize_function_for_size_p.
* builtins.c (fold_builtin_cabs): Use optimize_function_for_speed_p.
(fold_builtin_strcpy, fold_builtin_fputs): Use
optimize_function_for_size_p.
* fold-const.c (tree_swap_operands_p): Use optimize_function_for_size_p.
* recog.c (relax_delay_slots): Likewise.
* tree-ssa-math-opts.c (replace_reciprocal): Use optimize_bb_for_speed_p.
(execute_cse_reciprocals): Use optimize_bb_for_size_p.
* ipa-inline.c (cgraph_decide_recursive_inlining): Use
optimize_function_for_size_p.
(cgraph_decide_inlining_of_small_function): Use
optimize_function_for_size_p.
* global.c (find_reg): Use optimize_function_for_size_p.
* opts.c (decode_options): Do not clear flag_tree_ch, flag_inline_functions,
flag_unswitch_loops, flag_unroll_loops, flag_unroll_all_loops and
flag_prefetch_loop_arrays. Those can work it out from profile.
* tree-ssa-loop-ivcanon.c (tree_unroll_loops_completely): Use
optimize_loop_for_speed_p.
* predict.c (optimize_bb_for_size_p, optimize_bb_for_speed_p): Constify
argument.
(optimize_loop_nest_for_size_p, optimize_loop_nest_for_speed_p): New.
* tree-parloops.c (parallelize_loops): Use optimize_loop_for_size_p.
* tree-eh.c (decide_copy_try_finally): Use optimize_function_for_size_p.
* local-alloc.c (block_alloc): Pass BB pointer.
(find_free_reg): Add BB pointer, use optimize_bb_for_size_p.
* gcse.c (gcse_main): Use optimize_function_for_size_p.
* loop-unroll.c (decide_unrolling_and_peeling): Use optimize_loop_for_size_p.
(decide_peel_completely): Likewise.
* tree-vect-analyze.c (vect_mark_for_runtime_alias_test): Use
optimize_loop_for_size_p.
(vect_enhance_data_refs_alignment): Likewise.
* tree-ssa-coalesce.c (coalesce_cost): Add optimize_for_size argument.
(coalesce_cost_bb, coalesce_cost_edge, create_outofssa_var_map): Update call.
* cfgcleanup.c (outgoing_edges_match): Use optimize_bb_for_speed_p.
(try_crossjump_bb): Use optimize_bb_for_size_p.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays): Use
optimize_loop_for_speed_p.
* bb-reorder.c (find_traces_1_round): Likewise.
(copy_bb): Use optimize_bb_for_speed_p.
(duplicate_computed_gotos): Likewise.
* basic-block.h (optimize_loop_nest_for_size_p,
optimize_loop_nest_for_speed_p): New.
* stmt.c (expand_case): Use optimize_insn_for_size_p.
From-SVN: r139760
2008-08-29 12:35:57 +02:00
|
|
|
if (optimize_loop_nest_for_size_p (loop))
|
2007-04-15 05:36:25 +02:00
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, " ignored (cold area)\n");
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* FIXME: the time should be weighted by the probabilities of the blocks in
|
|
|
|
the loop body. */
|
|
|
|
time = tree_num_loop_insns (loop, &eni_time_weights);
|
|
|
|
if (time == 0)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
ahead = (PREFETCH_LATENCY + time - 1) / time;
|
|
|
|
est_niter = estimated_loop_iterations_int (loop, false);
|
|
|
|
|
|
|
|
/* Prefetching is not likely to be profitable if the trip count to ahead
|
|
|
|
ratio is too small. */
|
|
|
|
if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
ninsns = tree_num_loop_insns (loop, &eni_size_weights);
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Step 1: gather the memory references. */
|
2009-06-08 18:00:13 +02:00
|
|
|
refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Give up prefetching if the number of memory references in the
|
|
|
|
loop is not reasonable based on profitablity and compilation time
|
|
|
|
considerations. */
|
|
|
|
if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count))
|
|
|
|
goto fail;
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Step 2: estimate the reuse effects. */
|
|
|
|
prune_by_reuse (refs);
|
|
|
|
|
2010-06-14 22:51:26 +02:00
|
|
|
if (nothing_to_prefetch_p (refs))
|
2006-02-14 14:51:51 +01:00
|
|
|
goto fail;
|
|
|
|
|
tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-data-ref.c (dr_analyze_alias): Handle case smt is NULL.
* tree-predcom.c (mark_virtual_ops_for_renaming): Exported.
* tree-ssa-loop-prefetch.c: Include optabs.h.
(FENCE_FOLLOWING_MOVNT): New macro.
(struct mem_ref): Add independent_p and storent_p fields.
(record_ref): Initalize the new fields.
(gather_memory_references_ref): Return true if the reference
could be analysed.
(gather_memory_references): Check whether all memory accesses
in loop were recorded.
(should_issue_prefetch_p): Return false for nontemporal stores.
(nontemporal_store_p, mark_nontemporal_store, emit_mfence_after_loop,
may_use_storent_in_loop_p, mark_nontemporal_stores): New functions.
(determine_loop_nest_reuse): Detect independent memory references.
(loop_prefetch_arrays): Call mark_nontemporal_stores.
* tree-flow.h (mark_virtual_ops_for_renaming): Declare.
* Makefile.in (tree-ssa-loop-prefetch.o): Add OPTABS_H dependency.
* config/i386/i386.h (x86_mfence): Declare.
(FENCE_FOLLOWING_MOVNT): Return x86_mfence.
* config/i386/i386.c (x86_mfence): New variable.
(ix86_init_mmx_sse_builtins): Initialize x86_mfence.
* tree-pretty-print.c (dump_generic_node): Mark nontemporal stores.
* optabs.c (init_optabs): Initialize storent_optab.
* optabs.h (enum optab_index): Add OTI_storent.
(storent_optab): Declare.
* genopinit.c (optabs): Add initialization for storent_optab.
* tree.h (MOVE_NONTEMPORAL): New macro.
* expr.c (expand_assignment, store_expr, store_constructor_field,
store_constructor, store_field, expand_expr_real_1): Propagate
nontemporality of the expanded store.
(emit_storent_insn): New function.
* expr.h (expand_assignment, store_expr): Declaration changed.
* function.c (assign_parm_setup_reg): Pass false as nontemporality
to expand_assignment.
* stmt.c (expand_asm_expr): Ditto.
* calls.c (initialize_argument_information): Pass false as
nontemporality to store_expr.
* config/i386/sse.md (storentv4sf, storentv2df, storentv2di,
storentsi): New.
* gcc.dg/tree-ssa/prefetch-7.c: New test.
From-SVN: r125604
2007-06-10 22:39:22 +02:00
|
|
|
determine_loop_nest_reuse (loop, refs, no_other_refs);
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Step 3: determine unroll factor. */
|
2007-03-01 23:14:23 +01:00
|
|
|
unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
|
|
|
|
est_niter);
|
2010-06-14 22:51:26 +02:00
|
|
|
|
|
|
|
/* Estimate prefetch count for the unrolled loop. */
|
|
|
|
prefetch_count = estimate_prefetch_count (refs, unroll_factor);
|
|
|
|
if (prefetch_count == 0)
|
|
|
|
goto fail;
|
|
|
|
|
2007-03-01 23:14:23 +01:00
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
2009-11-25 11:55:54 +01:00
|
|
|
fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
|
2009-06-09 19:00:31 +02:00
|
|
|
HOST_WIDE_INT_PRINT_DEC "\n"
|
2009-11-25 11:55:54 +01:00
|
|
|
"insn count %d, mem ref count %d, prefetch count %d\n",
|
|
|
|
ahead, unroll_factor, est_niter,
|
|
|
|
ninsns, mem_ref_count, prefetch_count);
|
2009-06-08 18:00:13 +02:00
|
|
|
|
2010-07-10 01:08:55 +02:00
|
|
|
/* Prefetching is not likely to be profitable if the instruction to prefetch
|
|
|
|
ratio is too small. */
|
|
|
|
if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count,
|
|
|
|
unroll_factor))
|
2009-06-08 18:00:13 +02:00
|
|
|
goto fail;
|
|
|
|
|
|
|
|
mark_nontemporal_stores (loop, refs);
|
2007-03-01 23:14:23 +01:00
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
/* Step 4: what to prefetch? */
|
|
|
|
if (!schedule_prefetches (refs, unroll_factor, ahead))
|
|
|
|
goto fail;
|
|
|
|
|
|
|
|
/* Step 5: unroll the loop. TODO -- peeling of first and last few
|
|
|
|
iterations so that we do not issue superfluous prefetches. */
|
|
|
|
if (unroll_factor != 1)
|
|
|
|
{
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
tree_unroll_loop (loop, unroll_factor,
|
2006-02-14 14:51:51 +01:00
|
|
|
single_dom_exit (loop), &desc);
|
|
|
|
unrolled = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Step 6: issue the prefetches. */
|
|
|
|
issue_prefetches (refs, unroll_factor, ahead);
|
|
|
|
|
|
|
|
fail:
|
|
|
|
release_mem_refs (refs);
|
|
|
|
return unrolled;
|
|
|
|
}
|
|
|
|
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
/* Issue prefetch instructions for array references in loops. */
|
2006-02-14 14:51:51 +01:00
|
|
|
|
2006-09-06 08:13:22 +02:00
|
|
|
unsigned int
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
tree_ssa_prefetch_arrays (void)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
tree-vrp.c (adjust_range_with_scev): Use get_chrec_loop.
* tree-vrp.c (adjust_range_with_scev): Use get_chrec_loop.
* loop-unswitch.c (unswitch_loops): Use FOR_EACH_LOOP.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree_ssa_unswitch_loops (tree_ssa_unswitch_loops): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-loop-ivcanon.c (canonicalize_induction_variables,
tree_unroll_loops_completely): Ditto.
* predict.c (predict_loops): Ditto.
* tree-if-conv.c (main_tree_if_conversion): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling): Ditto.
* cfgloopmanip.c (unloop): Use delete_loop.
(place_new_loop): Access larray vector instead of parray.
(create_preheaders, force_single_succ_latches,
fix_loop_structure): Use FOR_EACH_LOOP and delete_loop..
* loop-doloop.c (doloop_optimize_loops): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-cfg.c (replace_uses_by): Ditto.
* tree-ssa-loop-prefetch.c (tree_ssa_prefetch_arrays): Ditto.
* tree-chrec.h (CHREC_VAR, CHREC_LEFT, CHREC_RIGHT, CHREC_VARIABLE):
Moved to ...
* tree.h (CHREC_VAR, CHREC_LEFT, CHREC_RIGHT, CHREC_VARIABLE):
... here.
* tree-scalar-evolution.c (chrec_contains_symbols_defined_in_loop,
compute_overall_effect_of_inner_loop, chrec_is_positive): Use
get_loop and get_chrec_loop.
(number_of_iterations_for_all_loops): Use number_of_loops.
(scev_initialize, scev_reset, scev_const_prop): Use FOR_EACH_LOOP.
* tree-scalar-evolution.h (get_chrec_loop): New inline function.
* cfgloopanal.c (mark_irreducible_loops): Use number_of_loops,
and FOR_EACH_LOOP.
* tree-chrec.c (evolution_function_is_invariant_rec_p,
chrec_convert_1): Use get_loop and get_chrec_loop.
* loop-init.c (loop_optimizer_init): Use number_of_loops.
(loop_optimizer_init): Use FOR_EACH_LOOP.
* tree-vectorizer.c (vect_loops_num): Removed.
(vectorize_loops): Store number of loops locally. Use
FOR_EACH_LOOP and get_loop.
* tree-vectorizer.h (vect_loops_num): Removed.
* tree-data-ref.c (get_number_of_iters_for_loop): Use get_loop.
(find_data_references_in_loop): Do not set parallel_p.
* tree-data-ref.h: Do not declare VEC(loop_p).
* cfgloop.c (flow_loops_dump, mark_single_exit_loops,
verify_loop_structure): Use FOR_EACH_LOOP.
(flow_loops_free): Use FOR_EACH_LOOP, free larray vector.
(initialize_loops_parallel_p): Removed.
(flow_loops_find): Push the loops into a vector.
(delete_loop): New function.
(cancel_loop): Use delete_loop.
* cfgloop.h: Declare VEC(loop_p).
(struct loop): Remove parallel_p field.
(struct loops): Replace num and parray field by larray vector.
Remove shared_headers field.
(delete_loop): Declare.
(get_loop, get_loops, number_of_loops, fel_next, fel_init,
FOR_EACH_LOOP): New.
* doc/loop.tex: Document new accessor functions.
From-SVN: r119713
2006-12-10 23:17:15 +01:00
|
|
|
loop_iterator li;
|
2006-02-14 14:51:51 +01:00
|
|
|
struct loop *loop;
|
|
|
|
bool unrolled = false;
|
2006-09-06 08:13:22 +02:00
|
|
|
int todo_flags = 0;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (!HAVE_prefetch
|
|
|
|
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
|
|
|
|
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
|
|
|
|
of processor costs and i486 does not have prefetch, but
|
|
|
|
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|
|
|
|
|| PREFETCH_BLOCK == 0)
|
2006-09-06 08:13:22 +02:00
|
|
|
return 0;
|
2006-02-14 14:51:51 +01:00
|
|
|
|
params.c (set_param_value): Initialize the "set" field.
* params.c (set_param_value): Initialize the "set" field.
* params.h (struct param_info): Add "set" field.
(PARAM_SET_P): New macro.
(PREFETCH_LATENCY, SIMULTANEOUS_PREFETCHES, L1_CACHE_SIZE,
L1_CACHE_LINE_SIZE): New macros.
* toplev.c (DEFPARAM): Initialize the "set" field.
* tree-ssa-loop-prefetch.c (PREFETCH_LATENCY,
SIMULTANEOUS_PREFETCHES): Removed.
(PREFETCH_BLOCK): Use L1_CACHE_LINE_SIZE.
(tree_ssa_prefetch_arrays): Dump the values of the parameters.
* config/sparc/sparc.c: Include params.h.
(sparc_override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sparc/sparc.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.c: Include params.h.
(k8_cost): Change default value for SIMULTANEOUS_PREFETCHES.
(override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sh/sh.h (SIMULTANEOUS_PREFETCHES): Removed.
(OPTIMIZATION_OPTIONS): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.c (ia64_optimization_options): Set
SIMULTANEOUS_PREFETCHES and L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.h (SIMULTANEOUS_PREFETCHES, PREFETCH_BLOCK):
Removed.
* params.def (PARAM_PREFETCH_LATENCY, PARAM_SIMULTANEOUS_PREFETCHES,
PARAM_L1_CACHE_SIZE, PARAM_L1_CACHE_LINE_SIZE): New params.
* doc/invoke.texi: Document new params.
From-SVN: r118728
2006-11-12 20:17:02 +01:00
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
{
|
|
|
|
fprintf (dump_file, "Prefetching parameters:\n");
|
|
|
|
fprintf (dump_file, " simultaneous prefetches: %d\n",
|
|
|
|
SIMULTANEOUS_PREFETCHES);
|
|
|
|
fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
|
|
|
|
fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
|
invoke.texi (l1-cache-size): Update documentation.
* doc/invoke.texi (l1-cache-size): Update documentation.
(l2-cache-size): Document.
* params.h (L2_CACHE_SIZE): New macro.
* tree-ssa-loop-prefetch.c (L1_CACHE_SIZE_BYTES): Reflect
that L1_CACHE_SIZE is in kB now.
(L2_CACHE_SIZE_BYTES): New macro.
(tree_ssa_prefetch_arrays): Show size in kB.
* config/i386/i386.h (struct processor_costs): Add l1_cache_size
and l2_cache_size fields.
* config/i386/driver-i386.c (describe_cache): Detect cache size
in kB.
* config/i386/i386.c (size_cost, i386_cost, i486_cost,pentium_cost,
pentiumpro_cost, geode_cost, k6_cost, athlon_cost, k8_cost,
amdfam10_cost, pentium4_cost, nocona_cost, core2_cost,
generic64_cost, generic32_cost): Add l1_cache_size and l2_cache_size.
(override_options): Set l1-cache-size and l2-cache-size to default
values if not specified otherwise.
* params.def (PARAM_L1_CACHE_SIZE): Change to set in kB.
(PARAM_L2_CACHE_SIZE): New.
From-SVN: r127117
2007-08-01 12:57:03 +02:00
|
|
|
fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
|
|
|
|
L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
|
tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.
(SET_DR_MISALIGNMENT): New.
* tree-vect-analyze.c (vect_compute_data_ref_alignment,
vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
Use SET_DR_MISALIGNMENT.
* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
pointer.
* tree-data-ref.c (create_data_ref, compute_all_dependences,
find_loop_nest): Export.
* tree-data-ref.h (struct data_reference): Change aux field to pointer.
(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
New macros.
(struct mem_ref): Add field reuse_distance.
(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
Initialize reuse_distance field.
(issue_prefetch_ref): Select temporality of prefetch according to
reuse_distance.
(volume_of_references, volume_of_dist_vector, add_subscript_strides,
self_reuse_distance, determine_loop_nest_reuse): New functions.
(loop_prefetch_arrays): Call determine_loop_nest_reuse.
(tree_ssa_prefetch_arrays): Dump L2 cache size.
* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
dependency.
* gcc.dg/tree-ssa/prefetch-6.c: New test.
From-SVN: r125172
2007-05-29 23:55:47 +02:00
|
|
|
fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
|
2009-11-25 11:55:54 +01:00
|
|
|
fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
|
|
|
|
fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
|
2009-06-08 18:00:13 +02:00
|
|
|
MIN_INSN_TO_PREFETCH_RATIO);
|
2009-11-25 11:55:54 +01:00
|
|
|
fprintf (dump_file, " min insn-to-mem ratio: %d \n",
|
2009-06-08 18:00:13 +02:00
|
|
|
PREFETCH_MIN_INSN_TO_MEM_RATIO);
|
params.c (set_param_value): Initialize the "set" field.
* params.c (set_param_value): Initialize the "set" field.
* params.h (struct param_info): Add "set" field.
(PARAM_SET_P): New macro.
(PREFETCH_LATENCY, SIMULTANEOUS_PREFETCHES, L1_CACHE_SIZE,
L1_CACHE_LINE_SIZE): New macros.
* toplev.c (DEFPARAM): Initialize the "set" field.
* tree-ssa-loop-prefetch.c (PREFETCH_LATENCY,
SIMULTANEOUS_PREFETCHES): Removed.
(PREFETCH_BLOCK): Use L1_CACHE_LINE_SIZE.
(tree_ssa_prefetch_arrays): Dump the values of the parameters.
* config/sparc/sparc.c: Include params.h.
(sparc_override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sparc/sparc.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.h (PREFETCH_BLOCK, SIMULTANEOUS_PREFETCHES):
Removed.
* config/i386/i386.c: Include params.h.
(k8_cost): Change default value for SIMULTANEOUS_PREFETCHES.
(override_options): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/sh/sh.h (SIMULTANEOUS_PREFETCHES): Removed.
(OPTIMIZATION_OPTIONS): Set SIMULTANEOUS_PREFETCHES and
L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.c (ia64_optimization_options): Set
SIMULTANEOUS_PREFETCHES and L1_CACHE_LINE_SIZE parameters.
* config/ia64/ia64.h (SIMULTANEOUS_PREFETCHES, PREFETCH_BLOCK):
Removed.
* params.def (PARAM_PREFETCH_LATENCY, PARAM_SIMULTANEOUS_PREFETCHES,
PARAM_L1_CACHE_SIZE, PARAM_L1_CACHE_LINE_SIZE): New params.
* doc/invoke.texi: Document new params.
From-SVN: r118728
2006-11-12 20:17:02 +01:00
|
|
|
fprintf (dump_file, "\n");
|
|
|
|
}
|
|
|
|
|
2006-02-14 14:51:51 +01:00
|
|
|
initialize_original_copy_tables ();
|
|
|
|
|
|
|
|
if (!built_in_decls[BUILT_IN_PREFETCH])
|
|
|
|
{
|
2010-07-16 19:51:56 +02:00
|
|
|
tree type = build_function_type_list (void_type_node,
|
|
|
|
const_ptr_type_node, NULL_TREE);
|
2006-10-23 21:47:58 +02:00
|
|
|
tree decl = add_builtin_function ("__builtin_prefetch", type,
|
|
|
|
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
|
|
|
|
NULL, NULL_TREE);
|
2006-02-14 14:51:51 +01:00
|
|
|
DECL_IS_NOVOPS (decl) = true;
|
|
|
|
built_in_decls[BUILT_IN_PREFETCH] = decl;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* We assume that size of cache line is a power of two, so verify this
|
|
|
|
here. */
|
|
|
|
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
|
|
|
|
|
tree-vrp.c (adjust_range_with_scev): Use get_chrec_loop.
* tree-vrp.c (adjust_range_with_scev): Use get_chrec_loop.
* loop-unswitch.c (unswitch_loops): Use FOR_EACH_LOOP.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree_ssa_unswitch_loops (tree_ssa_unswitch_loops): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-loop-ivcanon.c (canonicalize_induction_variables,
tree_unroll_loops_completely): Ditto.
* predict.c (predict_loops): Ditto.
* tree-if-conv.c (main_tree_if_conversion): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling): Ditto.
* cfgloopmanip.c (unloop): Use delete_loop.
(place_new_loop): Access larray vector instead of parray.
(create_preheaders, force_single_succ_latches,
fix_loop_structure): Use FOR_EACH_LOOP and delete_loop..
* loop-doloop.c (doloop_optimize_loops): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-cfg.c (replace_uses_by): Ditto.
* tree-ssa-loop-prefetch.c (tree_ssa_prefetch_arrays): Ditto.
* tree-chrec.h (CHREC_VAR, CHREC_LEFT, CHREC_RIGHT, CHREC_VARIABLE):
Moved to ...
* tree.h (CHREC_VAR, CHREC_LEFT, CHREC_RIGHT, CHREC_VARIABLE):
... here.
* tree-scalar-evolution.c (chrec_contains_symbols_defined_in_loop,
compute_overall_effect_of_inner_loop, chrec_is_positive): Use
get_loop and get_chrec_loop.
(number_of_iterations_for_all_loops): Use number_of_loops.
(scev_initialize, scev_reset, scev_const_prop): Use FOR_EACH_LOOP.
* tree-scalar-evolution.h (get_chrec_loop): New inline function.
* cfgloopanal.c (mark_irreducible_loops): Use number_of_loops,
and FOR_EACH_LOOP.
* tree-chrec.c (evolution_function_is_invariant_rec_p,
chrec_convert_1): Use get_loop and get_chrec_loop.
* loop-init.c (loop_optimizer_init): Use number_of_loops.
(loop_optimizer_init): Use FOR_EACH_LOOP.
* tree-vectorizer.c (vect_loops_num): Removed.
(vectorize_loops): Store number of loops locally. Use
FOR_EACH_LOOP and get_loop.
* tree-vectorizer.h (vect_loops_num): Removed.
* tree-data-ref.c (get_number_of_iters_for_loop): Use get_loop.
(find_data_references_in_loop): Do not set parallel_p.
* tree-data-ref.h: Do not declare VEC(loop_p).
* cfgloop.c (flow_loops_dump, mark_single_exit_loops,
verify_loop_structure): Use FOR_EACH_LOOP.
(flow_loops_free): Use FOR_EACH_LOOP, free larray vector.
(initialize_loops_parallel_p): Removed.
(flow_loops_find): Push the loops into a vector.
(delete_loop): New function.
(cancel_loop): Use delete_loop.
* cfgloop.h: Declare VEC(loop_p).
(struct loop): Remove parallel_p field.
(struct loops): Replace num and parray field by larray vector.
Remove shared_headers field.
(delete_loop): Declare.
(get_loop, get_loops, number_of_loops, fel_next, fel_init,
FOR_EACH_LOOP): New.
* doc/loop.tex: Document new accessor functions.
From-SVN: r119713
2006-12-10 23:17:15 +01:00
|
|
|
FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
|
2006-02-14 14:51:51 +01:00
|
|
|
{
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, "Processing loop %d:\n", loop->num);
|
|
|
|
|
tree-vrp.c (execute_vrp): Do not pass loops structure through arguments.
* tree-vrp.c (execute_vrp): Do not pass loops structure through
arguments.
* loop-unswitch.c (unswitch_loops, unswitch_single_loop,
unswitch_loop): Ditto.
* tree-loop-linear.c (linear_transform_loops): Ditto.
* tree-ssa-loop-im.c (determine_lsm, tree_ssa_lim_initialize,
tree_ssa_lim): Ditto.
* tree-ssa-loop-niter.c (estimate_numbers_of_iterations,
free_numbers_of_iterations_estimates): Ditto.
* tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops,
tree_unswitch_single_loop, tree_unswitch_loop): Ditto.
* cfgloopmanip.c (fix_bb_placement, fix_bb_placements,
remove_path, add_loop, loopify, unloop, fix_loop_placements,
place_new_loop, duplicate_loop, duplicate_subloops,
update_single_exit_for_duplicated_loops, copy_loops_to,
duplicate_loop_to_header_edge, create_preheaders,
force_single_succ_latches, loop_version, fix_loop_structure):
Ditto.
* tree-ssa-loop-manip.c (tree_duplicate_loop_to_header_edge,
tree_unroll_loop): Ditto.
* tree-ssa-loop-ch.c (copy_loop_headers): Ditto.
* tree-scalar-evolution.c (select_loops_exit_conditions,
scev_initialize, scev_analysis): Ditto.
* tree-scalar-evolution.h (scev_initialize): Ditto.
* cfghooks.c (cfg_hook_duplicate_loop_to_header_edge): Ditto.
* cfgloopanal.c (mark_irreducible_loops, mark_loop_exit_edges): Ditto.
* tree-ssa-loop-ivopts.c (tree_ssa_iv_optimize): Ditto.
* modulo-sched.c (sms_schedule): Ditto.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Ditto.
* loop-init.c (loop_optimizer_init, rtl_move_loop_invariants,
rtl_unswitch, rtl_unroll_and_peel_loops, rtl_doloop): Ditto.
* ifcvt.c (if_convert): Ditto.
* tree-ssa-loop-ivcanon.c (try_unroll_loop_completely,
canonicalize_loop_induction_variables,
canonicalize_induction_variables,
tree_unroll_loops_completely, remove_empty_loops): Ditto.
* tree-ssa-loop.c (tree_ssa_loop_init, tree_ssa_loop_im,
tree_ssa_loop_unswitch, tree_vectorize, tree_linear_transform,
tree_ssa_loop_ivcanon, tree_ssa_empty_loop, tree_ssa_loop_bounds,
tree_complete_unroll, tree_ssa_loop_prefetch, tree_ssa_loop_ivopts,
tree_ssa_loop_done): Ditto.
* predict.c (predict_loops, tree_estimate_probability, estimate_loops,
estimate_bb_frequencies): Ditto.
* tree-vectorizer.c (slpeel_tree_duplicate_loop_to_edge_cfg,
slpeel_tree_peel_loop_to_edge, vectorize_loops): Ditto.
* loop-unroll.c (unroll_and_peel_loops, peel_loops_completely,
decide_unrolling_and_peeling, peel_loop_completely,
unroll_loop_constant_iterations, unroll_loop_runtime_iterations,
peel_loop_simple, unroll_loop_stupid): Ditto.
* loop-doloop.c (doloop_optimize_loops): Ditto.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Ditto.
* loop-invariant.c (move_loop_invariants): Ditto.
* tree-ssa-dce.c (tree_ssa_dce_loop): Ditto.
* tree-ssa-loop-prefetch.c (loop_prefetch_arrays,
tree_ssa_prefetch_arrays): Ditto.
* lambda-code.c (gcc_loopnest_to_lambda_loopnest, perfect_nestify):
Ditto.
* tree-vect-transform.c (vect_do_peeling_for_loop_bound,
vect_do_peeling_for_alignment, vect_transform_loop): Ditto.
* cfgloop.c (flow_loops_cfg_dump, flow_loops_dump,
mark_single_exit_loops, cancel_loop, cancel_loop_tree,
verify_loop_structure): Ditto.
* tree-flow.h (vectorize_loops, tree_ssa_lim, tree_ssa_unswitch_loops,
canonicalize_induction_variables, tree_unroll_loops_completely,
tree_ssa_prefetch_arrays, remove_empty_loops, tree_ssa_iv_optimize,
estimate_numbers_of_iterations, free_numbers_of_iterations_estimates,
tree_duplicate_loop_to_header_edge, tree_ssa_loop_version,
tree_unroll_loop, linear_transform_loops): Declaration changed.
* basic-block.h: Remove declaration of struct loops.
* cfghooks.h (struct cfg_hooks): Change type of
cfg_hook_duplicate_loop_to_header_edge.
(cfg_hook_duplicate_loop_to_header_edge): Declaration changed.
* tree-vectorizer.h (slpeel_tree_peel_loop_to_edge,
vect_transform_loop): Declaration changed.
* lambda.h (gcc_loopnest_to_lambda_loopnest): Declaration changed.
* cfgloop.h (flow_loops_dump, fix_loop_structure,
mark_irreducible_loops, mark_single_exit_loops, mark_loop_exit_edges,
cancel_loop_tree, create_preheaders, force_single_succ_latches,
verify_loop_structure, duplicate_loop, duplicate_loop_to_header_edge,
loopify, loop_version, remove_path, unswitch_loops,
unroll_and_peel_loops, doloop_optimize_loops, move_loop_invariants):
Declaration changed.
From-SVN: r119189
2006-11-25 11:34:13 +01:00
|
|
|
unrolled |= loop_prefetch_arrays (loop);
|
2006-02-14 14:51:51 +01:00
|
|
|
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
|
|
fprintf (dump_file, "\n\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (unrolled)
|
|
|
|
{
|
|
|
|
scev_reset ();
|
2006-09-06 08:13:22 +02:00
|
|
|
todo_flags |= TODO_cleanup_cfg;
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
free_original_copy_tables ();
|
2006-09-06 08:13:22 +02:00
|
|
|
return todo_flags;
|
2006-02-14 14:51:51 +01:00
|
|
|
}
|