platform/upstream/gcc.git
4 years agoFix typo
Eric Botcazou [Mon, 23 Sep 2019 08:28:36 +0000 (08:28 +0000)]
Fix typo

From-SVN: r276046

4 years agotrans.c (gnat_compile_time_expr_list): New variable.
Eric Botcazou [Mon, 23 Sep 2019 08:27:40 +0000 (08:27 +0000)]
trans.c (gnat_compile_time_expr_list): New variable.

* gcc-interface/trans.c (gnat_compile_time_expr_list): New variable.
(Pragma_to_gnu): Rename local variable.  Save the (first) expression
of pragma Compile_Time_{Error|Warning} for later processing.
(Compilation_Unit_to_gnu): Process the expressions saved above.

From-SVN: r276045

4 years agotrans.c (Attribute_to_gnu): Test Can_Use_Internal_Rep on the underlying type of the...
Eric Botcazou [Mon, 23 Sep 2019 08:08:08 +0000 (08:08 +0000)]
trans.c (Attribute_to_gnu): Test Can_Use_Internal_Rep on the underlying type of the node.

* gcc-interface/trans.c (Attribute_to_gnu): Test Can_Use_Internal_Rep
on the underlying type of the node.
(Call_to_gnu): Likewise with the type of the prefix.

From-SVN: r276041

4 years agodecl.c (components_to_record): Do not reorder fields in packed record types if...
Eric Botcazou [Mon, 23 Sep 2019 07:45:58 +0000 (07:45 +0000)]
decl.c (components_to_record): Do not reorder fields in packed record types if...

* gcc-interface/decl.c (components_to_record): Do not reorder fields
in packed record types if they contain fixed-size fields that cannot
be laid out in a packed manner.

From-SVN: r276036

4 years agoDaily bump.
GCC Administrator [Mon, 23 Sep 2019 00:16:24 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r276035

4 years ago[Darwin, PPC] Clean up symbol stubs code.
Iain Sandoe [Sun, 22 Sep 2019 19:24:14 +0000 (19:24 +0000)]
[Darwin, PPC] Clean up symbol stubs code.

Remove dead code for the the TARGET_LINK_STACK which is not
applicable to Darwin. Use MACHOPIC_PURE instead of a hard-wired
PIC level to determine the stub kind.

Merge common code blocks.

gcc/ChangeLog:

2019-09-22  Iain Sandoe  <iain@sandoe.co.uk>

* config/rs6000/rs6000.c (machopic_output_stub): Remove dead
code.  Merge code blocks with common conditionals. Use declared
macro instead of a magic number for PIC level.

From-SVN: r276030

4 years agoPR c++/91819 - ICE with operator++ and enum.
Marek Polacek [Sun, 22 Sep 2019 12:35:00 +0000 (12:35 +0000)]
PR c++/91819 - ICE with operator++ and enum.

* call.c (build_new_op_1): Set arg2_type.

* g++.dg/other/operator4.C: New test.

From-SVN: r276027

4 years agoDaily bump.
GCC Administrator [Sun, 22 Sep 2019 00:16:15 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r276026

4 years agoPR middle-end/91830 - Bogus -Warray-bounds on strcpy into a member
Martin Sebor [Sat, 21 Sep 2019 22:32:59 +0000 (22:32 +0000)]
PR middle-end/91830 - Bogus -Warray-bounds on strcpy into a member

PR middle-end/91830 - Bogus -Warray-bounds on strcpy into a member
of a subobject compiling binutils

gcc/ChangeLog:
* gcc/gimple-ssa-warn-restrict.c (builtin_memref::set_base_and_offset):
Simplify computation of the offset of the referenced subobject.

gcc/testsuite/ChangeLog:
* gcc/testsuite/gcc.dg/Warray-bounds-47.c: New test.

From-SVN: r276022

4 years agore PR c++/30277 (bit-field: wrong overload resolution)
Jakub Jelinek [Sat, 21 Sep 2019 21:54:38 +0000 (23:54 +0200)]
re PR c++/30277 (bit-field: wrong overload resolution)

PR c++/30277
* g++.dg/expr/bitfield14.C (struct S): Use signed long long instead
of signed long.
(foo): Use long long instead of long.

From-SVN: r276021

4 years ago[Darwin] Update machopic_legitimize_pic_address.
Iain Sandoe [Sat, 21 Sep 2019 19:48:27 +0000 (19:48 +0000)]
[Darwin] Update machopic_legitimize_pic_address.

Some changes were missed here in the transition to LRA.  The Darwin
archs are all using LRA now.

gcc/ChangeLog:

2019-09-21  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.c (machopic_legitimize_pic_address): Check
for lra not reload.

From-SVN: r276020

4 years agoDR 2345 - Jumping across initializers in init-statements and conditions.
Marek Polacek [Sat, 21 Sep 2019 13:59:29 +0000 (13:59 +0000)]
DR 2345 - Jumping across initializers in init-statements and conditions.

* g++.dg/cpp1z/init-statement10.C: New test.

From-SVN: r276019

4 years agoAvoid adding impossible copies in ira-conflicts.c:process_reg_shuffles
Richard Sandiford [Sat, 21 Sep 2019 12:57:13 +0000 (12:57 +0000)]
Avoid adding impossible copies in ira-conflicts.c:process_reg_shuffles

If an insn requires two operands to be tied, and the input operand dies
in the insn, IRA acts as though there were a copy from the input to the
output with the same execution frequency as the insn.  Allocating the
same register to the input and the output then saves the cost of a move.

If there is no such tie, but an input operand nevertheless dies
in the insn, IRA creates a similar move, but with an eighth of the
frequency.  This helps to ensure that chains of instructions reuse
registers in a natural way, rather than using arbitrarily different
registers for no reason.

This heuristic seems to work well in the vast majority of cases.
However, for SVE, the handling of untied operands ends up creating
copies between dying predicate registers and vector outputs, even though
vector and predicate registers are distinct classes and can never be
tied.  This is a particular problem because the dying predicate tends
to be the loop control predicate, which is used by most instructions
in a vector loop and so (rightly) has a very high allocation priority.
Any copies involving the loop predicate therefore tend to get processed
before copies involving only vector registers.  The end result is that
we tend to allocate the output of the last vector instruction in a loop
ahead of its natural place in the allocation order and don't benefit
from chains created between vector registers.

This patch tries to avoid the problem by not adding register shuffle
copies if there appears to be no chance that the two operands could be
allocated to the same register.

2019-09-21  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* ira-conflicts.c (can_use_same_reg_p): New function.
(process_reg_shuffles): Take an insn parameter.  Ignore cases
in which input operand op_num could seemingly never be allocated
to the same register as the destination.
(add_insn_allocno_copies): Update call to process_reg_shuffles.

gcc/testsuite/
* gcc.target/aarch64/sve/cond_convert_1.c: Remove XFAILs.
* gcc.target/aarch64/sve/cond_convert_4.c: Likewise.
* gcc.target/aarch64/sve/cond_unary_2.c: Likewise.

From-SVN: r276018

4 years agoExtend neg_const_int simplifications to other const rtxes
Richard Sandiford [Sat, 21 Sep 2019 12:56:50 +0000 (12:56 +0000)]
Extend neg_const_int simplifications to other const rtxes

This patch generalises some neg_const_int-based rtx simplifications
so that they handle all CONST_SCALAR_INTs and also CONST_POLY_INT.
This actually simplifies things a bit, since we no longer have
to treat HOST_WIDE_INT_MIN specially.

This is tested by later SVE patches.

2019-09-21  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* simplify-rtx.c (neg_const_int): Replace with...
(neg_poly_int_rtx): ...this new function.
(simplify_binary_operation_1): Extend (minus x C) -> (plus X -C)
to all CONST_SCALAR_INTs and to CONST_POLY_INT.
(simplify_plus_minus): Likewise for constant terms here.

From-SVN: r276017

4 years agoDaily bump.
GCC Administrator [Sat, 21 Sep 2019 00:16:17 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r276015

4 years agomicroblaze.h (ASM_OUTPUT_SKIP): Use HOST_WIDE_PRINT_UNSIGNED.
Jonas Pfeil [Fri, 20 Sep 2019 22:50:42 +0000 (22:50 +0000)]
microblaze.h (ASM_OUTPUT_SKIP): Use HOST_WIDE_PRINT_UNSIGNED.

* config/microblaze/microblaze.h (ASM_OUTPUT_SKIP): Use
HOST_WIDE_PRINT_UNSIGNED.

From-SVN: r276011

4 years agopa.c (pa_trampoline_init): Remove spurious extended character.
John David Anglin [Fri, 20 Sep 2019 21:47:56 +0000 (21:47 +0000)]
pa.c (pa_trampoline_init): Remove spurious extended character.

* config/pa/pa.c (pa_trampoline_init): Remove spurious extended
character.

From-SVN: r276007

4 years agore PR target/86811 (Vax port needs updating for CVE-2017-5753)
Maya Rashish [Fri, 20 Sep 2019 20:23:29 +0000 (20:23 +0000)]
re PR target/86811 (Vax port needs updating for CVE-2017-5753)

PR target/86811
* config/vax/vax.c (TARGET_HAVE_SPECULATION_SAFE_VALUE):
Define to speculation_safe_value_not_needed.

From-SVN: r276006

4 years ago[Darwin, X86, testsuite] Fix naked-1.c fail.
Iain Sandoe [Fri, 20 Sep 2019 18:52:05 +0000 (18:52 +0000)]
[Darwin, X86, testsuite] Fix naked-1.c fail.

This fails at m32 because the scan-asm is looking for an absence
of "ret". Darwin is generating the correct code for the function
but the picbase thunk has a 'ret' insn.  Fixed by making the test
use -mdynamic-no-pic for m32.

gcc/testsuite/ChangeLog:

2019-09-20  Iain Sandoe  <iain@sandoe.co.uk>

* gcc.target/i386/naked-1.c: Alter options to use non-
PIC codegen for m32 Darwin.

From-SVN: r276004

4 years agore PR fortran/78260 (ICE in gimplify_expr, at gimplify.c:11939)
Tobias Burnus [Fri, 20 Sep 2019 16:05:06 +0000 (18:05 +0200)]
re PR fortran/78260 (ICE in gimplify_expr, at gimplify.c:11939)

2019-09-20  Tobias Burnus  <tobias@codesourcery.com>

        PR fortran/78260
        * openmp.c (gfc_resolve_oacc_declare): Reject all
        non variables but accept function result variables.
        * trans-openmp.c (gfc_trans_omp_clauses): Handle
        function-result variables for remaing cases.

2019-09-20  Tobias Burnus  <tobias@codesourcery.com>

        PR fortran/78260
        * gfortran.dg/goacc/parameter.f95: Change
        dg-error as it is now detected earlier.
        * gfortran.dg/goacc/pr85701.f90: Modify to
        use a separate result variable.
        * gfortran.dg/goacc/pr78260.f90: New.
        * gfortran.dg/goacc/pr78260-2.f90: New.
        * gfortran.dg/gomp/pr78260.f90: New.
        * gfortran.dg/gomp/pr78260-2.f90: New.
        * gfortran.dg/gomp/pr78260-3.f90: New.

From-SVN: r276002

4 years agoRevert [ARM/FDPIC v6 13/24] [ARM] FDPIC: Force LSB bit for PC in Cortex-M architecture
Christophe Lyon [Fri, 20 Sep 2019 13:32:20 +0000 (15:32 +0200)]
Revert [ARM/FDPIC v6 13/24] [ARM] FDPIC: Force LSB bit for PC in Cortex-M architecture

This is causing regressions when mixing with user code compiled in ARM mode.

2019-09-20  Christophe Lyon  <christophe.lyon@st.com>

Revert:
2019-09-10  Christophe Lyon  <christophe.lyon@st.com>
Mickaël Guêné <mickael.guene@st.com>

* config/arm/unwind-arm.c (_Unwind_VRS_Set): Handle thumb-only
architecture.

From-SVN: r276001

4 years agoRestrict gnat.dg/system_info1.adb to Linux and Windows hosts
Olivier Hainque [Fri, 20 Sep 2019 12:17:20 +0000 (12:17 +0000)]
Restrict gnat.dg/system_info1.adb to Linux and Windows hosts

Where it is know to work, still covering the original test intent.

From-SVN: r275999

4 years agore PR target/91814 (ICE in elimination_costs_in_insn, at reload1.c:3549 since r274926)
Richard Biener [Fri, 20 Sep 2019 11:14:34 +0000 (11:14 +0000)]
re PR target/91814 (ICE in elimination_costs_in_insn, at reload1.c:3549 since r274926)

2019-09-20  Richard Biener  <rguenther@suse.de>
Uros Bizjak  <ubizjak@gmail.com>

PR target/91814
* config/i386/i386-features.c (gen_gpr_to_xmm_move_src): Revert
previous change.
(general_scalar_chain::convert_op): Force not suitable memory
operands to a register.

Co-Authored-By: Uros Bizjak <ubizjak@gmail.com>
From-SVN: r275998

4 years agore PR testsuite/91821 (r275928 breaks gcc.target/powerpc/sad-vectorize-2.c)
Richard Biener [Fri, 20 Sep 2019 09:54:54 +0000 (09:54 +0000)]
re PR testsuite/91821 (r275928 breaks gcc.target/powerpc/sad-vectorize-2.c)

2019-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91821
* tree-vect-loop.c (check_reduction_path): Check we can compute
reduc_idx.
(vect_is_simple_reduction): Set STMT_VINFO_REDUC_IDX.
* tree-vect-patterns.c (vect_reassociating_reduction_p): Return
operands in canonical order.
* tree-vectorizer.c (vec_info::new_stmt_vec_info): Initialize
STMT_VINFO_REDUC_IDX.
* tree-vectorizer.h (_stmt_vec_info::reduc_idx): New.
(STMT_VINFO_REDUC_IDX): Likewise.

From-SVN: r275996

4 years agore PR target/91269 (unaligned floating-point register with -mcpu=niagara4 -fcall...
Eric Botcazou [Fri, 20 Sep 2019 09:42:40 +0000 (09:42 +0000)]
re PR target/91269 (unaligned floating-point register with -mcpu=niagara4 -fcall-used-g6)

PR target/91269
* config/sparc/sparc.h (HARD_REGNO_CALLER_SAVE_MODE): Define.

From-SVN: r275994

4 years agore PR c/91815 (questionable error on type definition at file scope)
Eric Botcazou [Fri, 20 Sep 2019 09:11:20 +0000 (09:11 +0000)]
re PR c/91815 (questionable error on type definition at file scope)

PR c/91815
* c-decl.c (pushdecl): In C detect duplicate declarations across scopes
of identifiers in the external scope only for variables and functions.

From-SVN: r275992

4 years agore PR tree-optimization/91822 (FAIL: gcc.dg/pr88031.c (internal compiler error))
Richard Biener [Fri, 20 Sep 2019 08:35:59 +0000 (08:35 +0000)]
re PR tree-optimization/91822 (FAIL: gcc.dg/pr88031.c (internal compiler error))

2019-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91822
* tree-vectorizer.h (vectorizable_condition): Restore for_reduction
parameter.
* tree-vect-loop.c (vectorizable_reduction): Adjust asserts
for reduc_index in nested cycles, adjust vectorizable_condition
calls.
* tree-vect-stmts.c (vectorizable_condition): Restore for_reduction
parameter.
(vect_analyze_stmt): Adjust.
(vect_transform_stmt): Likewise.

From-SVN: r275990

4 years agore PR target/91767 (After r274953, clang-compiled xgcc segfaults during RTL pass...
Richard Biener [Fri, 20 Sep 2019 06:42:39 +0000 (06:42 +0000)]
re PR target/91767 (After r274953, clang-compiled xgcc segfaults during RTL pass: stv)

2019-09-20  Richard Biener  <rguenther@suse.de>

PR target/91767
* config/i386/i386-features.c (general_scalar_chain::convert_registers):
Ensure there's a sequence point between allocating the new register
and passing a reference to a reg via regno_reg_rtx.

From-SVN: r275989

4 years agoDaily bump.
GCC Administrator [Fri, 20 Sep 2019 00:16:15 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r275986

4 years agoNew IPA-SRA
Martin Jambor [Thu, 19 Sep 2019 22:25:04 +0000 (00:25 +0200)]
New IPA-SRA

2019-09-20  Martin Jambor  <mjambor@suse.cz>

        * coretypes.h (cgraph_edge): Declare.
        * ipa-param-manipulation.c: Rewrite.
        * ipa-param-manipulation.h: Likewise.
        * Makefile.in (GTFILES): Added ipa-param-manipulation.h and ipa-sra.c.
        (OBJS): Added ipa-sra.o.
        * cgraph.h (ipa_replace_map): Removed fields old_tree, replace_p
        and ref_p, added fields param_adjustments and performed_splits.
        (struct cgraph_clone_info): Remove ags_to_skip and
        combined_args_to_skip, new field param_adjustments.
        (cgraph_node::create_clone): Changed parameters to use
        ipa_param_adjustments.
        (cgraph_node::create_virtual_clone): Likewise.
        (cgraph_node::create_virtual_clone_with_body): Likewise.
        (tree_function_versioning): Likewise.
        (cgraph_build_function_type_skip_args): Removed.
        * cgraph.c (cgraph_edge::redirect_call_stmt_to_callee): Convert to
        using ipa_param_adjustments.
        (clone_of_p): Likewise.
        * cgraphclones.c (cgraph_build_function_type_skip_args): Removed.
        (build_function_decl_skip_args): Likewise.
        (duplicate_thunk_for_node): Adjust parameters using
        ipa_param_body_adjustments, copy param_adjustments instead of
        args_to_skip.
        (cgraph_node::create_clone): Convert to using ipa_param_adjustments.
        (cgraph_node::create_virtual_clone): Likewise.
        (cgraph_node::create_version_clone_with_body): Likewise.
        (cgraph_materialize_clone): Likewise.
        (symbol_table::materialize_all_clones): Likewise.
        * ipa-fnsummary.c (ipa_fn_summary_t::duplicate): Simplify
        ipa_replace_map check.
        * ipa-cp.c (get_replacement_map): Do not initialize removed fields.
        (initialize_node_lattices): Make aware that some parameters might have
        already been removed.
        (want_remove_some_param_p): New function.
        (create_specialized_node): Convert to using ipa_param_adjustments and
        deal with possibly pre-existing adjustments.
        * lto-cgraph.c (output_cgraph_opt_summary_p): Likewise.
        (output_node_opt_summary): Do not stream removed fields.  Stream
        parameter adjustments instead of argumetns to skip.
        (input_node_opt_summary): Likewise.
        (input_node_opt_summary): Likewise.
        * lto-section-in.c (lto_section_name): Added ipa-sra section.
        * lto-streamer.h (lto_section_type): Likewise.
        * tree-inline.h (copy_body_data): New fields killed_new_ssa_names and
param_body_adjs.
        (copy_decl_to_var): Declare.
        * tree-inline.c (update_clone_info): Do not remap old_tree.
        (remap_gimple_stmt): Use ipa_param_body_adjustments to modify gimple
        statements, walk all extra generated statements and remap their
        operands.
        (redirect_all_calls): Add killed SSA names to a hash set.
        (remap_ssa_name): Do not remap killed SSA names.
        (copy_arguments_for_versioning): Renames to copy_arguments_nochange,
        half of functionality moved to ipa_param_body_adjustments.
        (copy_decl_to_var): Make exported.
        (copy_body): Destroy killed_new_ssa_names hash set.
        (expand_call_inline): Remap performed splits.
        (update_clone_info): Likewise.
        (tree_function_versioning): Simplify tree_map processing.  Updated to
        accept ipa_param_adjustments and use ipa_param_body_adjustments.
        * omp-simd-clone.c (simd_clone_vector_of_formal_parm_types): Adjust
        for the new interface.
        (simd_clone_clauses_extract): Likewise, make args an auto_vec.
        (simd_clone_compute_base_data_type): Likewise.
        (simd_clone_init_simd_arrays): Adjust for the new interface.
        (simd_clone_adjust_argument_types): Likewise.
        (struct modify_stmt_info): Likewise.
        (ipa_simd_modify_stmt_ops): Likewise.
        (ipa_simd_modify_function_body): Likewise.
        (simd_clone_adjust): Likewise.
        * tree-sra.c: Removed IPA-SRA.  Include tree-sra.h.
        (type_internals_preclude_sra_p): Make public.
        * tree-sra.h: New file.
        * ipa-inline-transform.c (save_inline_function_body): Update to
        refelct new tree_function_versioning signature.
        * ipa-prop.c (adjust_agg_replacement_values): Use a helper from
        ipa_param_adjustments to get current parameter indices.
        (ipcp_modif_dom_walker::before_dom_children): Likewise.
        (ipcp_update_bits): Likewise.
        (ipcp_update_vr): Likewise.
        * ipa-split.c (split_function): Convert to using ipa_param_adjustments.
        * ipa-sra.c: New file.
        * multiple_target.c (create_target_clone): Update to reflet new type
        of create_version_clone_with_body.
        * trans-mem.c (ipa_tm_create_version): Update to reflect new type of
        tree_function_versioning.
(modify_function): Update to reflect new type of
tree_function_versioning.
        * params.def (PARAM_IPA_SRA_MAX_REPLACEMENTS): New.
        * passes.def: Remove old IPA-SRA and add new one.
        * tree-pass.h (make_pass_early_ipa_sra): Remove declaration.
        (make_pass_ipa_sra): Declare.
* dbgcnt.def: Remove eipa_sra.  Added ipa_sra_params and
ipa_sra_retvalues.
* doc/invoke.texi (ipa-sra-max-replacements): New.

        testsuite/
        * g++.dg/ipa/pr81248.C: Adjust dg-options and dump-scan.
        * gcc.dg/ipa/ipa-sra-1.c: Likewise.
        * gcc.dg/ipa/ipa-sra-10.c: Likewise.
        * gcc.dg/ipa/ipa-sra-11.c: Likewise.
        * gcc.dg/ipa/ipa-sra-3.c: Likewise.
        * gcc.dg/ipa/ipa-sra-4.c: Likewise.
        * gcc.dg/ipa/ipa-sra-5.c: Likewise.
        * gcc.dg/ipa/ipacost-2.c: Disable ipa-sra.
        * gcc.dg/ipa/ipcp-agg-9.c: Likewise.
        * gcc.dg/ipa/pr78121.c: Adjust scan pattern.
        * gcc.dg/ipa/vrp1.c: Likewise.
        * gcc.dg/ipa/vrp2.c: Likewise.
        * gcc.dg/ipa/vrp3.c: Likewise.
        * gcc.dg/ipa/vrp7.c: Likewise.
        * gcc.dg/ipa/vrp8.c: Likewise.
        * gcc.dg/noreorder.c: use noipa attribute instead of noinline.
        * gcc.dg/ipa/20040703-wpa.c: New test.
* gcc.dg/ipa/ipa-sra-12.c: New test.
* gcc.dg/ipa/ipa-sra-13.c: Likewise.
* gcc.dg/ipa/ipa-sra-14.c: Likewise.
* gcc.dg/ipa/ipa-sra-15.c: Likewise.
* gcc.dg/ipa/ipa-sra-16.c: Likewise.
* gcc.dg/ipa/ipa-sra-17.c: Likewise.
* gcc.dg/ipa/ipa-sra-18.c: Likewise.
* gcc.dg/ipa/ipa-sra-19.c: Likewise.
* gcc.dg/ipa/ipa-sra-20.c: Likewise.
* gcc.dg/ipa/ipa-sra-21.c: Likewise.
* gcc.dg/ipa/ipa-sra-22.c: Likewise.
* gcc.dg/sso/ipa-sra-1.c: Likewise.
* g++.dg/ipa/ipa-sra-2.C: Likewise.
* g++.dg/ipa/ipa-sra-3.C: Likewise.
* gcc.dg/tree-ssa/ipa-cp-1.c: Make return value used.
* g++.dg/ipa/devirt-19.C: Add missing return, add -fipa-cp-clone
option.
* g++.dg/lto/devirt-19_0.C: Add -fipa-cp-clone option.
* gcc.dg/ipa/ipa-sra-2.c: Removed.
* gcc.dg/ipa/ipa-sra-6.c: Likewise.

From-SVN: r275982

4 years agoPR middle-end/91631 - buffer overflow into an array member of a declared object not...
Martin Sebor [Thu, 19 Sep 2019 22:15:34 +0000 (22:15 +0000)]
PR middle-end/91631 - buffer overflow into an array member of a declared object not detected

gcc/ChangeLog:

PR middle-end/91631
* builtins.c (component_size): Correct trailing array computation,
rename to component_ref_size and move...
(compute_objsize): Adjust.
* gimple-ssa-warn-restrict.c (builtin_memref::refsize): New member.
(builtin_access::strict): Do not consider mememmove.
(builtin_access::write_off): New function.
(builtin_memref::builtin_memref): Initialize refsize.
(builtin_memref::set_base_and_offset): Adjust refoff and compute
refsize.
(builtin_memref::offset_out_of_bounds): Use ooboff input values.
Handle refsize.
(builtin_access::builtin_access): Intialize dstoff to destination
refeence offset here instead of in maybe_diag_overlap.  Adjust
referencess even to unrelated objects. Adjust sizrange of bounded
string functions to reflect bound.  For strcat, adjust destination
sizrange by that of source.
(builtin_access::strcat_overlap):  Adjust offsets and sizes
to reflect the increase in destination sizrange above.
(builtin_access::overlap): Do not set dstoff here but instead
in builtin_access::builtin_access.
(check_bounds_or_overlap): Use builtin_access::write_off.
(maybe_diag_access_bounds): Add argument.  Add informational notes.
(dump_builtin_memref, dump_builtin_access): New functions.
* tree.c (component_ref_size): ...to here.
* tree.h (component_ref_size): Declare.
* tree-ssa-strlen (handle_builtin_strcat): Include the terminating
nul in the size of the source string.

gcc/testsuite/ChangeLog:

PR middle-end/91631
* /c-c++-common/Warray-bounds-3.c: Correct expected offsets.
* /c-c++-common/Warray-bounds-4.c: Same.
* gcc.dg/Warray-bounds-39.c: Remove xfails.
* gcc.dg/Warray-bounds-45.c: New test.
* gcc.dg/Warray-bounds-46.c: New test.

From-SVN: r275981

4 years agoSupport extended characters in C/C++ identifiers (PR c/67224)
Lewis Hyatt [Thu, 19 Sep 2019 19:56:11 +0000 (19:56 +0000)]
Support extended characters in C/C++ identifiers (PR c/67224)

libcpp/ChangeLog
2019-09-19  Lewis Hyatt  <lhyatt@gmail.com>

PR c/67224
* charset.c (_cpp_valid_utf8): New function to help lex UTF-8 tokens.
* internal.h (_cpp_valid_utf8): Declare.
* lex.c (forms_identifier_p): Use it to recognize UTF-8 identifiers.
(_cpp_lex_direct): Handle UTF-8 in identifiers and CPP_OTHER tokens.
Do all work in "default" case to avoid slowing down typical code paths.
Also handle $ and UCN in the default case for consistency.

gcc/Changelog
2019-09-19  Lewis Hyatt  <lhyatt@gmail.com>

PR c/67224
* doc/cpp.texi: Document support for extended characters in
identifiers.
* doc/cppopts.texi: Likewise.

gcc/testsuite/ChangeLog
2019-09-19  Lewis Hyatt  <lhyatt@gmail.com>

PR c/67224
* c-c++-common/cpp/ucnid-2011-1-utf8.c: New test.
* g++.dg/cpp/ucnid-1-utf8.C: New test.
* g++.dg/cpp/ucnid-2-utf8.C: New test.
* g++.dg/cpp/ucnid-3-utf8.C: New test.
* g++.dg/cpp/ucnid-4-utf8.C: New test.
* g++.dg/other/ucnid-1-utf8.C: New test.
* gcc.dg/cpp/ucnid-1-utf8.c: New test.
* gcc.dg/cpp/ucnid-10-utf8.c: New test.
* gcc.dg/cpp/ucnid-11-utf8.c: New test.
* gcc.dg/cpp/ucnid-12-utf8.c: New test.
* gcc.dg/cpp/ucnid-13-utf8.c: New test.
* gcc.dg/cpp/ucnid-14-utf8.c: New test.
* gcc.dg/cpp/ucnid-15-utf8.c: New test.
* gcc.dg/cpp/ucnid-2-utf8.c: New test.
* gcc.dg/cpp/ucnid-3-utf8.c: New test.
* gcc.dg/cpp/ucnid-4-utf8.c: New test.
* gcc.dg/cpp/ucnid-6-utf8.c: New test.
* gcc.dg/cpp/ucnid-7-utf8.c: New test.
* gcc.dg/cpp/ucnid-9-utf8.c: New test.
* gcc.dg/ucnid-1-utf8.c: New test.
* gcc.dg/ucnid-10-utf8.c: New test.
* gcc.dg/ucnid-11-utf8.c: New test.
* gcc.dg/ucnid-12-utf8.c: New test.
* gcc.dg/ucnid-13-utf8.c: New test.
* gcc.dg/ucnid-14-utf8.c: New test.
* gcc.dg/ucnid-15-utf8.c: New test.
* gcc.dg/ucnid-16-utf8.c: New test.
* gcc.dg/ucnid-2-utf8.c: New test.
* gcc.dg/ucnid-3-utf8.c: New test.
* gcc.dg/ucnid-4-utf8.c: New test.
* gcc.dg/ucnid-5-utf8.c: New test.
* gcc.dg/ucnid-6-utf8.c: New test.
* gcc.dg/ucnid-7-utf8.c: New test.
* gcc.dg/ucnid-8-utf8.c: New test.
* gcc.dg/ucnid-9-utf8.c: New test.

From-SVN: r275979

4 years agoHandle [[likely]] on compound-statement.
Jason Merrill [Thu, 19 Sep 2019 19:41:17 +0000 (15:41 -0400)]
Handle [[likely]] on compound-statement.

I overlooked this case when adding [[likely]] handling to
cp_parser_statement.

* parser.c (cp_parser_statement): Handle [[likely]] on
compound-statement.

From-SVN: r275978

4 years ago[Darwin, PPC, testsuite] Fix pr89313.c fail.
Iain Sandoe [Thu, 19 Sep 2019 19:39:30 +0000 (19:39 +0000)]
[Darwin, PPC, testsuite] Fix pr89313.c fail.

Darwin defines '__POWERPC__' rather than '__powerpc__' so check for
the upper case version too in order to select the correct register
name.

gcc/testsuite:

2019-09-19  Iain Sandoe  <iain@sandoe.co.uk>

* gcc.dg/pr89313.c: Test for __POWERPC__ in addition to
__powerpc__ in register name selection.

From-SVN: r275977

4 years agoRevert "Fix conversions for built-in operator overloading candidates."
Jason Merrill [Thu, 19 Sep 2019 19:39:22 +0000 (15:39 -0400)]
Revert "Fix conversions for built-in operator overloading candidates."

This reverts commit 948d5b831affef14a49f56804b01e3f1ba00cdb3.

From-SVN: r275976

4 years agotree-vect-loop.c (vect_is_slp_reduction): Remove.
Richard Biener [Thu, 19 Sep 2019 17:19:31 +0000 (17:19 +0000)]
tree-vect-loop.c (vect_is_slp_reduction): Remove.

2019-09-19  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_is_slp_reduction): Remove.
(check_reduction_path): New overload having the path as result.
(vect_is_simple_reduction): From the detected reduction
path build a SLP reduction chain if possible.

From-SVN: r275972

4 years agore PR target/91814 (ICE in elimination_costs_in_insn, at reload1.c:3549 since r274926)
Richard Biener [Thu, 19 Sep 2019 17:05:51 +0000 (17:05 +0000)]
re PR target/91814 (ICE in elimination_costs_in_insn, at reload1.c:3549 since r274926)

2019-09-19  Richard Biener  <rguenther@suse.de>

PR target/91814
* config/i386/i386-features.c (gen_gpr_to_xmm_move_src):
Force operand to a register if it isn't nonimmediate_operand.

From-SVN: r275971

4 years ago[ARM] Simplify logical DImode iterators
Wilco Dijkstra [Thu, 19 Sep 2019 16:51:08 +0000 (16:51 +0000)]
[ARM] Simplify logical DImode iterators

Further simplify the logical DImode expander using code iterator and
obtab attributes.  This avoids adding unnecessary code_attr entries.

    gcc/
* config/arm/arm.md (<logical_op>di3): Use <optab> and <CODE>.
* config/arm/iterators.md (optab): Add and, ior, xor entries.
(logical_op): Remove code attribute.
(logical_OP): Likewise.

From-SVN: r275970

4 years agoFix cast in sort_congruence_class_groups_by_decl_uid.
Martin Liska [Thu, 19 Sep 2019 15:16:15 +0000 (17:16 +0200)]
Fix cast in sort_congruence_class_groups_by_decl_uid.

2019-09-19  Martin Liska  <mliska@suse.cz>

* ipa-icf.c (sort_congruence_class_groups_by_decl_uid):
Use proper casting.

From-SVN: r275969

4 years agoaarch64: Implement -moutline-atomics
Richard Henderson [Thu, 19 Sep 2019 14:36:43 +0000 (14:36 +0000)]
aarch64: Implement -moutline-atomics

* config/aarch64/aarch64.opt (-moutline-atomics): New.
* config/aarch64/aarch64.c (aarch64_atomic_ool_func): New.
(aarch64_ool_cas_names, aarch64_ool_swp_names): New.
(aarch64_ool_ldadd_names, aarch64_ool_ldset_names): New.
(aarch64_ool_ldclr_names, aarch64_ool_ldeor_names): New.
(aarch64_expand_compare_and_swap): Honor TARGET_OUTLINE_ATOMICS.
* config/aarch64/atomics.md (atomic_exchange<ALLI>): Likewise.
(atomic_<atomic_op><ALLI>): Likewise.
(atomic_fetch_<atomic_op><ALLI>): Likewise.
(atomic_<atomic_op>_fetch<ALLI>): Likewise.
* doc/invoke.texi: Document -moutline-atomics.
testsuite/
* gcc.target/aarch64/atomic-op-acq_rel.c: Use -mno-outline-atomics.
* gcc.target/aarch64/atomic-comp-swap-release-acquire.c: Likewise.
* gcc.target/aarch64/atomic-op-acquire.c: Likewise.
* gcc.target/aarch64/atomic-op-char.c: Likewise.
* gcc.target/aarch64/atomic-op-consume.c: Likewise.
* gcc.target/aarch64/atomic-op-imm.c: Likewise.
* gcc.target/aarch64/atomic-op-int.c: Likewise.
* gcc.target/aarch64/atomic-op-long.c: Likewise.
* gcc.target/aarch64/atomic-op-relaxed.c: Likewise.
* gcc.target/aarch64/atomic-op-release.c: Likewise.
* gcc.target/aarch64/atomic-op-seq_cst.c: Likewise.
* gcc.target/aarch64/atomic-op-short.c: Likewise.
* gcc.target/aarch64/atomic_cmp_exchange_zero_reg_1.c: Likewise.
* gcc.target/aarch64/atomic_cmp_exchange_zero_strong_1.c: Likewise.
* gcc.target/aarch64/sync-comp-swap.c: Likewise.
* gcc.target/aarch64/sync-op-acquire.c: Likewise.
* gcc.target/aarch64/sync-op-full.c: Likewise.

From-SVN: r275968

4 years agoaarch64: Add out-of-line functions for LSE atomics
Richard Henderson [Thu, 19 Sep 2019 14:36:38 +0000 (14:36 +0000)]
aarch64: Add out-of-line functions for LSE atomics

This is the libgcc part of the interface -- providing the functions.
Rationale is provided at the top of libgcc/config/aarch64/lse.S.

* config/aarch64/lse-init.c: New file.
* config/aarch64/lse.S: New file.
* config/aarch64/t-lse: New file.
* config.host: Add t-lse to all aarch64 tuples.

From-SVN: r275967

4 years agoaarch64: Tidy aarch64_split_compare_and_swap
Richard Henderson [Thu, 19 Sep 2019 14:36:33 +0000 (14:36 +0000)]
aarch64: Tidy aarch64_split_compare_and_swap

With aarch64_track_speculation, we had extra code to do exactly what the
!strong_zero_p path already did.  The rest is reducing code duplication.

* config/aarch64/aarch64 (aarch64_split_compare_and_swap): Disable
strong_zero_p for aarch64_track_speculation; unify some code paths;
use aarch64_gen_compare_reg instead of open-coding.

From-SVN: r275966

4 years agoaarch64: Implement TImode compare-and-swap
Richard Henderson [Thu, 19 Sep 2019 14:36:29 +0000 (14:36 +0000)]
aarch64: Implement TImode compare-and-swap

This pattern will only be used with the __sync functions, because
we do not yet have a bare TImode atomic load.

* config/aarch64/aarch64.c (aarch64_gen_compare_reg): Add support
for NE comparison of TImode values.
(aarch64_emit_load_exclusive): Add support for TImode.
(aarch64_emit_store_exclusive): Likewise.
(aarch64_split_compare_and_swap): Disable strong_zero_p for TImode.
* config/aarch64/atomics.md (@atomic_compare_and_swap<ALLI_TI>):
Change iterator from ALLI to ALLI_TI.
(@atomic_compare_and_swap<JUST_TI>): New.
(@atomic_compare_and_swap<JUST_TI>_lse): New.
(aarch64_load_exclusive_pair): New.
(aarch64_store_exclusive_pair): New.
* config/aarch64/iterators.md (JUST_TI): New.

From-SVN: r275965

4 years agoaarch64: Extend %R for integer registers
Richard Henderson [Thu, 19 Sep 2019 14:36:24 +0000 (14:36 +0000)]
aarch64: Extend %R for integer registers

* config/aarch64/aarch64.c (aarch64_print_operand): Allow integer
registers with %R.

From-SVN: r275964

4 years agoUse post-dom info to update if/switch predicate
Feng Xue [Thu, 19 Sep 2019 14:16:01 +0000 (14:16 +0000)]
Use post-dom info to update if/switch predicate

2019-09-19  Feng Xue  <fxue@os.amperecomputing.com>

        * ipa-fnsummary.c (set_cond_stmt_execution_predicate): Do not compute
        trivial predicate for condition branch.
        (set_switch_stmt_execution_predicate): Do not compute trivial predicate
        for switch case.
        (compute_bb_predicates): Update predicate based on post-dominating
        relationship.
        (analyze_function_body): Calculate post-dominating information.

2019-09-19  Feng Xue  <fxue@os.amperecomputing.com>

        * gcc.dg/ipa/pr91089.c: Add a new function and pattern.

From-SVN: r275963

4 years agoEnforce correct COND_EXPR order for EXTRACT_LAST_REDUCTION
Richard Sandiford [Thu, 19 Sep 2019 13:33:55 +0000 (13:33 +0000)]
Enforce correct COND_EXPR order for EXTRACT_LAST_REDUCTION

For conditional reductions, the "then" value needs to be the candidate
value calculated by this iteration while the "else" value needs to be
the result carried over from previous iterations.  If the COND_EXPR
is the other way around, we need to swap it.

2019-09-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vectorizer.h (vectorizable_condition): Take an int
reduction index instead of a boolean flag.
* tree-vect-stmts.c (vectorizable_condition): Likewise.
Swap the "then" and "else" values for EXTRACT_LAST_REDUCTION
reductions if the reduction accumulator is the "then" rather
than the "else" value.
(vect_analyze_stmt): Update call accordingly.
(vect_transform_stmt): Likewise.
* tree-vect-loop.c (vectorizable_reduction): Likewise,
asserting that the index is > 0.

From-SVN: r275962

4 years agoSpeed up qsort in IPA ICF.
Martin Liska [Thu, 19 Sep 2019 13:13:21 +0000 (15:13 +0200)]
Speed up qsort in IPA ICF.

2019-09-19  Martin Liska  <mliska@suse.cz>

* ipa-icf.c (sort_sem_items_by_decl_uid): Simplify comparator.
(sort_congruence_classes_by_decl_uid): Likewise.
(sort_congruence_class_groups_by_decl_uid): Use std::pair for
easier sorting.
(sem_item_optimizer::merge_classes): Likewise.

From-SVN: r275961

4 years agore PR tree-optimization/91812 (GCC ignores volatile modifier)
Richard Biener [Thu, 19 Sep 2019 12:49:45 +0000 (12:49 +0000)]
re PR tree-optimization/91812 (GCC ignores volatile modifier)

2019-09-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91812
* tree-ssa-phiprop.c (propagate_with_phi): Do not replace
volatile loads.

* gcc.dg/torture/pr91812.c: New testcase.

From-SVN: r275960

4 years agoRework constant subreg folds and handle more variable-length cases
Richard Sandiford [Thu, 19 Sep 2019 09:04:28 +0000 (09:04 +0000)]
Rework constant subreg folds and handle more variable-length cases

This patch rewrites the way simplify_subreg handles constants.
It uses similar native_encode/native_decode routines to the
tree-level handling of VIEW_CONVERT_EXPR, meaning that we can
move between rtx constants and the target memory image of them.

The main point of this patch is to support subregs of constant-length
vectors for VLA vectors, beyond the very simple cases that were already
handled.  Many of the new tests failed before the patch for variable-
length vectors.

The boolean side is tested more by the upcoming SVE ACLE work.

2019-09-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* defaults.h (TARGET_UNIT): New macro.
(target_unit): New type.
* rtl.h (native_encode_rtx, native_decode_rtx)
(native_decode_vector_rtx, subreg_size_lsb): Declare.
(subreg_lsb_1): Turn into an inline wrapper around subreg_size_lsb.
* rtlanal.c (subreg_lsb_1): Delete.
(subreg_size_lsb): New function.
* simplify-rtx.c: Include rtx-vector-builder.h
(simplify_immed_subreg): Delete.
(native_encode_rtx, native_decode_vector_rtx, native_decode_rtx)
(simplify_const_vector_byte_offset, simplify_const_vector_subreg): New
functions.
(simplify_subreg): Use them.
(test_vector_subregs_modes, test_vector_subregs_repeating)
(test_vector_subregs_fore_back, test_vector_subregs_stepped)
(test_vector_subregs): New functions.
(test_vector_ops): Call test_vector_subregs for integer vector
modes with at least 2 elements.

From-SVN: r275959

4 years ago[Ada] Emit DW_AT_GNU_bias with -fgnat-encodings=gdb
Tom Tromey [Thu, 19 Sep 2019 08:14:57 +0000 (08:14 +0000)]
[Ada] Emit DW_AT_GNU_bias with -fgnat-encodings=gdb

Emit DW_AT_GNU_bias with -fgnat-encodings=gdb.  gdb implements this,
but not the encoded variant.

2019-09-19  Tom Tromey  <tromey@adacore.com>

gcc/ada/

* gcc-interface/misc.c (gnat_get_type_bias): Return the bias
when -fgnat-encodings=gdb.

gcc/testsuite/

* gnat.dg/bias1.adb: New testcase.

From-SVN: r275958

4 years ago[Ada] Accept concatentation arguments to pragma Annotate
Steve Baird [Thu, 19 Sep 2019 08:14:52 +0000 (08:14 +0000)]
[Ada] Accept concatentation arguments to pragma Annotate

In cases where pragma Annotate accepts a string literal as an argument,
we now also accept a concatenation of string literals.

2019-09-19  Steve Baird  <baird@adacore.com>

gcc/ada/

* sem_prag.adb (Preferred_String_Type): A new function. Given an
expression, determines whether the preference rules defined for
the third-and-later arguments of pragma Annotate suffice to
determine the type of the expression. If so, then the preferred
type is returned; if not then Empty is returned. Handles
concatenations.
(Analyze_Pragma): Replace previous code, which dealt only with
string literals, with calls to the new Preferred_String_Type
function, which also handles concatenations.
* doc/gnat_rm/implementation_defined_pragmas.rst: Update
documentation for pragma Annotate.
* gnat_rm.texi: Regenerate.

gcc/testsuite/

* gnat.dg/annotation1.adb: New testcase.

From-SVN: r275957

4 years ago[Ada] Get rid of useless temporary for slice in overaligned record type
Eric Botcazou [Thu, 19 Sep 2019 08:14:47 +0000 (08:14 +0000)]
[Ada] Get rid of useless temporary for slice in overaligned record type

This fixes a recent code quality regression for targets that do not
require the strict alignment of memory accesses: the compiler would
generate a useless temporary for a slice of an array component in an
overaligned record type.

Running these commands:

  gcc -c p.adb -gnatws -gnatD
  grep loop p.adb.dg

On the following sources:

procedure P (N : Positive) is

  type Rec1 is record
    I : Integer;
  end record;

  type Arr is array (Positive range <>) of Rec1;

  type Rec2 is record
    A : Arr (1 .. 128);
  end record;
  for Rec2'Alignment use 8;

  procedure Proc (A : Arr) is
  begin
    null;
  end;

  R : Rec2;

begin
  Proc (R.A (1 .. N));
end;

Should execute silently.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_util.adb (Is_Possibly_Unaligned_Slice): Do not return true
on pure alignment considerations if the target does not require
the strict alignment of memory accesses.

From-SVN: r275956

4 years ago[Ada] Sem_Ch12: add a comment to indicate future work
Eric Botcazou [Thu, 19 Sep 2019 08:14:42 +0000 (08:14 +0000)]
[Ada] Sem_Ch12: add a comment to indicate future work

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* sem_ch12.adb (Check_Private_View): Add a comment to indicate
future work.

From-SVN: r275955

4 years ago[Ada] Fix copy operation with private discriminated record type
Eric Botcazou [Thu, 19 Sep 2019 08:14:38 +0000 (08:14 +0000)]
[Ada] Fix copy operation with private discriminated record type

This prevents the object code from reading too many bytes from the
source for a copy operation involving a private discriminated record
type with default discriminants and generated for the assignment of an
aggregate to a variable or the initialization of a constant.

The front-end already knows that it needs to convert the operation
involving the aggregate into individual assignments if the type of the
aggregate has mutable components, but it would not do so if this type is
private, which does not change anything for code generation.

Running these commands:

  gnatmake -q p -g -fsanitize=address
  p

On the following sources:

with Q; use Q;

procedure P is

   type Rec is record
      A : T;
   end record;

   C : constant Rec := Rec'(A => Default_T);

begin
   null;
end;

package Q is

   type T is private;

   Default_T : constant T;

private

   A : constant := 170;
   B : constant := 8192;

   type A_Index is range 1 .. A;
   type B_Index is range 1 .. B;

   type A_Array is array (A_Index) of Boolean;
   type B_Array is array (B_Index) of Boolean;

   type Data_Type is (A_Type, B_Type);

   type T (Discriminant : Data_Type := A_Type) is record
      case Discriminant is
         when A_Type =>
            Field_A : A_Array;
         when B_Type =>
            Field_B : B_Array;
      end case;
   end record;

   Default_T : constant T :=
     T'(Discriminant => A_Type, Field_A => (others => True));

end Q;

Should execute silently.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_aggr.adb (Has_Mutable_Components): Look at the underlying
type of components to find out whether they are mutable.

From-SVN: r275954

4 years ago[Ada] Fix bogus "too late" error with nested generics and inlining
Eric Botcazou [Thu, 19 Sep 2019 08:14:28 +0000 (08:14 +0000)]
[Ada] Fix bogus "too late" error with nested generics and inlining

This prevents the compiler from issuing a bogus error about a constant
whose full declaration appears too late, if it is declared in a nested
generic package and instantiated in another nested instantiation, when
the instantiations are done in a unit withed from the main unit and
containing an inlined subprogram, and cross-unit inlining is enabled.

It turns out that, under these very peculiar conditions, the compiler
ends up instantiating the body of the generic package twice, which leads
to various semantic errors, in particular for declarations of constants.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* sem_ch12.adb (Instantiate_Package_Body): Check that the body
has not already been instantiated when the body of the parent
was being loaded.

gcc/testsuite/

* gnat.dg/inline21.adb, gnat.dg/inline21_g.ads,
gnat.dg/inline21_h.adb, gnat.dg/inline21_h.ads,
gnat.dg/inline21_q.ads: New testcase.

From-SVN: r275953

4 years ago[Ada] Fix bogus visibility error with nested generics and inlining
Eric Botcazou [Thu, 19 Sep 2019 08:14:23 +0000 (08:14 +0000)]
[Ada] Fix bogus visibility error with nested generics and inlining

This prevents the compiler from issuing a bogus error about the
visibility of an operator in an instantiation of a nested generic
package which is itself used as an actual of an instantiation of another
generic package, when the instantiations are done in a unit withed from
the main unit and containing an inlined subprogram, and cross-unit
inlining is enabled.

In most cases, the compiler does not check the visibility of operators
in an instantiation context because this has already been done when the
generic package has been analyzed. However, there are exceptions like
the actuals of an instantiation of a generic child unit which is done
as a compilation unit and the In_Instance predicate has a special check
for these cases.

This check would incorrectly trigger here and needs to be tightened.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* sem_util.adb (In_Instance): Test whether the current unit has
been analyzed instead of being on the scope stack to detect the
case of actuals of an instantiation of a generic child unit done
as a compilation unit.

gcc/testsuite/

* gnat.dg/inline20.adb, gnat.dg/inline20_g.adb,
gnat.dg/inline20_g.ads, gnat.dg/inline20_h.ads,
gnat.dg/inline20_i.ads, gnat.dg/inline20_q-io.ads,
gnat.dg/inline20_q.ads, gnat.dg/inline20_r.ads: New testcase.

From-SVN: r275952

4 years ago[Ada] New routine GNAT.Sockets.Create_Socket_Pair
Dmitriy Anisimkov [Thu, 19 Sep 2019 08:14:18 +0000 (08:14 +0000)]
[Ada] New routine GNAT.Sockets.Create_Socket_Pair

New routine to create 2 connected sockets. This routine is analog of the
UNIX system call socketpair. On UNIX platforms it is implemented on the
base of socketpair. On other platforms it is implemented by conecting
network sockets over loopback interface.

2019-09-19  Dmitriy Anisimkov  <anisimko@adacore.com>

gcc/ada/

* libgnat/g-socket.ads, libgnat/g-socket.adb
(Create_Socket_Pair): New routine.
* libgnat/g-socthi.ads (OS_Has_Socketpair): Boolean constant.
(C_Socketpair): New imported routine.
* libgnat/g-socthi__mingw.ads, libgnat/g-socthi__vxworks.ads
(Default_Socket_Pair_Family): New constant.
(C_Socketpair): New routine.
* libgnat/g-socthi__mingw.adb, libgnat/g-socthi__vxworks.adb
(C_Socketpair): Is separated in anouther file.
* libgnat/g-sthcso.adb (C_Socketpair): Non UNIX implementation.
* libgnat/g-stsifd__sockets.adb: Reuse C_Socketpair.

From-SVN: r275951

4 years ago[Ada] Use declared type for deciding on SPARK pointer rules
Yannick Moy [Thu, 19 Sep 2019 08:14:13 +0000 (08:14 +0000)]
[Ada] Use declared type for deciding on SPARK pointer rules

A constant of pointer type is considered as mutable in SPARK, according
to SPARK RM 3.10, but this should be based on the declared type of the
constant instead of its underlying type.

There is no impact on compilation hence no test.

2019-09-19  Yannick Moy  <moy@adacore.com>

gcc/ada/

* sem_prag.adb (Analyze_Depends_In_Decl_Part): Simplify previous
test.

From-SVN: r275950

4 years ago[Ada] Spurious visibility error in generic child unit
Ed Schonberg [Thu, 19 Sep 2019 08:14:08 +0000 (08:14 +0000)]
[Ada] Spurious visibility error in generic child unit

2019-09-19  Ed Schonberg  <schonberg@adacore.com>

gcc/ada/

* sem_ch7.adb (Install_Parent_Private_Declarations): If a
generic child unit is instantiated within a sibling child unit,
the analysis of its private part requires that the private part
of the ancestor be available, as is done when the context
includes an instance of the ancestor.

gcc/testsuite/

* gnat.dg/generic2-child.ads, gnat.dg/generic2-io_any.adb,
gnat.dg/generic2-io_any.ads, gnat.dg/generic2.ads: New testcase.

From-SVN: r275949

4 years ago[Ada] Disable inlining of traversal function in GNATprove
Yannick Moy [Thu, 19 Sep 2019 08:14:03 +0000 (08:14 +0000)]
[Ada] Disable inlining of traversal function in GNATprove

Traversal functions as defined in SPARK RM 3.10 should not be inlined
for analysis in GNATprove, as this changes the ownership behavior.
Disable the inlining performed in GNATprove on functions which could be
interpreted as such.

There is no impact on compilation and thus no test.

2019-09-19  Yannick Moy  <moy@adacore.com>

gcc/ada/

* inline.adb (Can_Be_Inlined_In_GNATprove_Mode): Add special
case for traversal functions.

From-SVN: r275948

4 years ago[Ada] Allow constants of access type in Global contracts
Yannick Moy [Thu, 19 Sep 2019 08:13:58 +0000 (08:13 +0000)]
[Ada] Allow constants of access type in Global contracts

Now that SPARK supports access types, global constants of access type
may appear as outputs of a subprogram, with the meaning that the
underlying memory can be modified (see SPARK RM 3.10).

2019-09-19  Yannick Moy  <moy@adacore.com>

gcc/ada/

* sem_prag.adb (Analyze_Global_In_Decl_Part): Do not issue an
error when a constant of an access type is used as output in a
Global contract.
(Analyze_Depends_In_Decl_Part): Do not issue an error when a
constant of an access type is used as output in a Depends
contract.

gcc/testsuite/

* gnat.dg/global2.adb, gnat.dg/global2.ads: New testcase.

From-SVN: r275947

4 years ago[Ada] Exp_Attr: remove obsolete comment
Arnaud Charlet [Thu, 19 Sep 2019 08:13:52 +0000 (08:13 +0000)]
[Ada] Exp_Attr: remove obsolete comment

2019-09-19  Arnaud Charlet  <charlet@adacore.com>

gcc/ada/

* exp_attr.adb: Remove obsolete comment.

From-SVN: r275946

4 years ago[Ada] Fix run-time segfault with derived access-to-subprogram type
Eric Botcazou [Thu, 19 Sep 2019 08:13:48 +0000 (08:13 +0000)]
[Ada] Fix run-time segfault with derived access-to-subprogram type

This fixes a segfault at run time for the call to a local subprogram
through an access value if the type of this access value is derived
from an initial access-to-subprogram type and the access value was
originally obtained with the initial type.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* sem_ch3.adb (Build_Derived_Access_Type): If this is an access-
to-subprogram type, copy Can_Use_Internal_Rep from the parent.

gcc/testsuite/

* gnat.dg/access9.adb: New testcase.

From-SVN: r275945

4 years ago[Ada] Move SPARK borrow-checker to gnat2why codebase
Yannick Moy [Thu, 19 Sep 2019 08:13:43 +0000 (08:13 +0000)]
[Ada] Move SPARK borrow-checker to gnat2why codebase

Unit sem_spark was implementing the borrow-checker for the support of
ownership pointers in SPARK. It has been moved to gnat2why codebase to
facilitate its evolution and allow the more powerful flow analysis to
provide its results for better analysis on pointers.

2019-09-19  Yannick Moy  <moy@adacore.com>

gcc/ada/

* gcc-interface/Make-lang.in: Remove references to sem_spark.
* sem_spark.adb, sem_spark.ads: Remove unit.

From-SVN: r275944

4 years ago[Ada] Implement Machine_Rounding attribute in line when possible
Eric Botcazou [Thu, 19 Sep 2019 08:13:38 +0000 (08:13 +0000)]
[Ada] Implement Machine_Rounding attribute in line when possible

GNAT implements Machine_Rounding as an alias for Rounding but, whereas
the implementation of the latter is in line when possible, that of the
former is always out of line, which is not aligned with the intent of
the Ada RM.

This changes the compiler to using for Machine_Rounding the same in line
implementation as Rounding when possible.

Running these commands:

  gcc -c f.adb -gnatD
  grep system f.adb.dg

On the following sources:

function F (Val : Float) return Integer is
begin
  return Integer (Float'Machine_Rounding (Val));
end;

Should execute silently.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_attr.adb (Is_Inline_Floating_Point_Attribute): Treat
Machine_Rounding as an alias for Rounding.
* sem_res.adb (Simplify_Type_Conversion): Likewise.

From-SVN: r275943

4 years ago[Ada] Suppress GNAT FE up-level reference transformation for GNAT-LLVM
Gary Dismukes [Thu, 19 Sep 2019 08:13:34 +0000 (08:13 +0000)]
[Ada] Suppress GNAT FE up-level reference transformation for GNAT-LLVM

In the case of GNAT-LLVM, the GNAT FE no longer does expansion of
up-level references identified by the subprogram unnesting machinery
into activation record references. This is now only done by the FE when
generating C code. This expansion is already taken care of by the
gnat-llvm middle phase, so there's no benefit to also doing it in the
front end.

2019-09-19  Gary Dismukes  <dismukes@adacore.com>

gcc/ada/

* exp_unst.adb (Unnest_Subprogram): Bypass the transformation of
up-level references unless Opt.Generate_C_Code is enabled.

From-SVN: r275942

4 years ago[Ada] Streamline comparison for equality of 2-element arrays
Eric Botcazou [Thu, 19 Sep 2019 08:13:29 +0000 (08:13 +0000)]
[Ada] Streamline comparison for equality of 2-element arrays

In the general case, the comparison for equality of array objects is
implemented by a local function that contains, among other things, a
loop running over the elements, comparing them one by one and exiting
as soon as an element is not the same in the two array objects.

For the specific case of constrained 2-element arrays, this is rather
heavy and unnecessarily obfuscates the control flow of the program,
so this change implements a simple conjunction of comparisons for it.

Running these commands:

  gcc -c p.ads -O -gnatD
  grep loop p.ads.dg

On the following sources:

package P is

  type Rec is record
    Re : Float;
    Im : Float;
  end record;

  type Arr is array (1 .. 2) of Rec;

  function Equal (A, B : Arr) return Boolean is (A = B);

end P;

Should execute silently.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_ch4.adb (Expand_Array_Equality): If optimization is
enabled, generate a simple conjunction of comparisons for the
specific case of constrained 1-dimensional 2-element arrays.
Fix formatting.

From-SVN: r275941

4 years ago[Ada] Remove duplicated routines for getting homonym number
Piotr Trojanek [Thu, 19 Sep 2019 08:13:25 +0000 (08:13 +0000)]
[Ada] Remove duplicated routines for getting homonym number

Routines Homonym_Number and Get_Homonym_Number were exactly the same,
except for minor style differences. Keep the one in Exp_Util; remove the
one in Exp_Dbug. No test attached, because semantics is unaffected.

2019-09-19  Piotr Trojanek  <trojanek@adacore.com>

gcc/ada/

* exp_dbug.ads, exp_dbug.adb (Get_Homonym_Number): Remove.
(Append_Homonym_Number): Use Homonym_Number instead of
Get_Homonym_Number.
* exp_util.ads, exp_util.adb (Homonym_Number): Mirror style of
the removed Get_Homonym_Number routine, i.e. initialize local
objects at declaration and refine the type of result.
* sem_util.adb (Add_Homonym_Suffix): Use Homonym_Number instead
of Get_Homonym_Number.

From-SVN: r275940

4 years ago[Ada] Crash on predicate in full view in a generic unit
Ed Schonberg [Thu, 19 Sep 2019 08:13:20 +0000 (08:13 +0000)]
[Ada] Crash on predicate in full view in a generic unit

This patch fixes a compiler abort on a dynamic predicate applied to the
full view of a type in a generic package declaration, when the
expression for the predicate is a conditionql expression that contains
references to components of the full view of the type.

2019-09-19  Ed Schonberg  <schonberg@adacore.com>

gcc/ada/

* sem_ch13.adb (Check_Aspect_At_End_Of_Declarations): Simplify
handling of expressions in predicates when the context is a
generic unit.

gcc/testsuite/

* gnat.dg/predicate14.adb, gnat.dg/predicate14.ads: New
testcase.

From-SVN: r275939

4 years ago[Ada] Memory leak with 'Range of a function call in a loop
Bob Duff [Thu, 19 Sep 2019 08:13:15 +0000 (08:13 +0000)]
[Ada] Memory leak with 'Range of a function call in a loop

If a for loop starts with "for X in F (...)'Range loop", where F is a
function returning an unconstrained array, then memory is leaked. This
patch fixes that bug.

Running these commands:

  gnatmake -q -f main.adb
  main

On the following sources:

with Text_IO; use Text_IO;
package P is

   function Get_Objects return String;

end P;

package body P is
   function Get_Objects return String is
   begin
      return "xyzzy";
   end Get_Objects;

end P;

with Text_IO; use Text_IO;
pragma Warnings (Off, "an internal GNAT unit");
with System.Secondary_Stack;
pragma Warnings (On, "an internal GNAT unit");
with P; use P;

procedure Main is
   Max_Iterations : constant Integer := 1_000;

   procedure Leak_Call is
   begin
      for Id in Get_Objects'Range loop
         null;
      end loop;
   end Leak_Call;

   procedure SS_Info is new System.Secondary_Stack.SS_Info
(Text_IO.Put_Line);

begin
   for Iteration in 1 .. Max_Iterations loop
      Leak_Call;
   end loop;

   SS_Info;

end Main;

Should produce the following output:

  Secondary Stack information:
    Total size              :  10240 bytes
    Current allocated space :  0 bytes
    Number of Chunks        :  1
    Default size of Chunks  :  10240

2019-09-19  Bob Duff  <duff@adacore.com>

gcc/ada/

* sem_attr.adb (Resolve_Attribute): Make sure the secondary
stack is properly managed in the case of a 'Range attribute in a
loop.

From-SVN: r275938

4 years ago[Ada] Propagate documentation to formal bounded sets
Raphael Amiard [Thu, 19 Sep 2019 08:13:10 +0000 (08:13 +0000)]
[Ada] Propagate documentation to formal bounded sets

2019-09-19  Raphael Amiard  <amiard@adacore.com>

gcc/ada/

* libgnat/a-cfhase.ads (Set): Add comments to public primitives.

From-SVN: r275937

4 years ago[Ada] Add comments wrt. deallocation of bounded sets/maps
Raphael Amiard [Thu, 19 Sep 2019 08:13:06 +0000 (08:13 +0000)]
[Ada] Add comments wrt. deallocation of bounded sets/maps

2019-09-19  Raphael Amiard  <amiard@adacore.com>

gcc/ada/

* libgnat/a-cbhama.ads, libgnat/a-cbhase.ads,
libgnat/a-chtgop.ads (Clear): Refine comments

From-SVN: r275936

4 years ago[Ada] Fix spurious type mismatch failure on nested instantiations
Eric Botcazou [Thu, 19 Sep 2019 08:13:01 +0000 (08:13 +0000)]
[Ada] Fix spurious type mismatch failure on nested instantiations

This fixes a spurious type mismatch failure reported between formal and
actual of a call to a subprogram that comes from the instantiation of a
child generic unit that itself contains an instantiation of a slibling
child generic unit, when the parent is itself a generic unit with
private part. The regression was introduced by a recent change made to
clear the Is_Generic_Actual_Type on the implicit full view built when a
generic package is instantiated on a private type.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* sem_ch12.adb (Restore_Private_Views): Comment out new code
that clear the Is_Generic_Actual_Type also on the full view.

gcc/testsuite/

* gnat.dg/generic_inst13.adb,
gnat.dg/generic_inst13_pkg-nested_g.ads,
gnat.dg/generic_inst13_pkg-ops_g.ads,
gnat.dg/generic_inst13_pkg.ads: New testcase.

From-SVN: r275935

4 years ago[Ada] Infinite loop with concatenation and aspect
Bob Duff [Thu, 19 Sep 2019 08:12:56 +0000 (08:12 +0000)]
[Ada] Infinite loop with concatenation and aspect

This patch fixes a bug where an array object initialized with a
concatenation, and that has an aspect_specification for Alignment,
causes the compiler goes into an infinite loop.

2019-09-19  Bob Duff  <duff@adacore.com>

gcc/ada/

* exp_ch3.adb (Rewrite_As_Renaming): Return False if there are
any aspect specifications, because otherwise Insert_Actions
blows up.

gcc/testsuite/

* gnat.dg/concat3.adb: New testcase.

From-SVN: r275934

4 years ago[Ada] Fix fallout of previous change for bit-packed arrays
Eric Botcazou [Thu, 19 Sep 2019 08:12:52 +0000 (08:12 +0000)]
[Ada] Fix fallout of previous change for bit-packed arrays

This fixes a regression introduced by the previous change that improved
the handling of explicit by-reference mechanism. For the very specific
case of a component of a bit-packed array, the front-end still needs to
insert a copy around the call because this is where the rewriting into
the sequence of mask-and-shifts is done for the code generator.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_ch6.adb (Add_Simple_Call_By_Copy_Code): Add
Bit_Packed_Array parameter and documet it. Always insert a copy
if it is set True.
(Expand_Actuals): Adjust the calls to
Add_Simple_Call_By_Copy_Code.

gcc/testsuite/

* gnat.dg/pack26.adb: New testcase.

From-SVN: r275933

4 years ago[Ada] gnatxref: infinite loop on symbols not found
Bob Duff [Thu, 19 Sep 2019 08:12:47 +0000 (08:12 +0000)]
[Ada] gnatxref: infinite loop on symbols not found

This patch fixes a bug in which if a symbol is not found, gnatxref can
sometimes enter an infinite loop. No impact on compilation.

2019-09-19  Bob Duff  <duff@adacore.com>

gcc/ada/

* xref_lib.adb (Get_Symbol_Name): If we reach EOF in the first
loop without finding the symbol, return "???". Otherwise, it's
an infinite loop.
(Parse_EOL): Assert that we're not already at EOF.  Remove
processing of LF/CR -- there are no operating systems that use
that.

From-SVN: r275932

4 years ago[Ada] Improve handling of explicit by-reference mechanism
Eric Botcazou [Thu, 19 Sep 2019 08:12:39 +0000 (08:12 +0000)]
[Ada] Improve handling of explicit by-reference mechanism

This improves the handling of an explicit by-reference passing mechanism
specified by means of the GNAT pragma Export_Function.  This device sort
of circumvents the rules of the language for the by-reference passing
mechanism and it's then up to the programmer to ensure that the actual
parameter is addressable; if it is not, the compiler will generate a
temporary around the call, thus effectively passing the actual by copy.

It turns out that the compiler was too conservative when determining
whether the actual parameter is addressable, in particular if it's a
component of a record type subject to a representation clause.

The change effectively moves this computation from the front-end to the
back-end, which has much more information on the layout and alignment
of types and thus can be less conservative.

2019-09-19  Eric Botcazou  <ebotcazou@adacore.com>

gcc/ada/

* exp_ch6.adb (Is_Legal_Copy): Also return false for an aliased
formal and a formal passed by reference in convention Ada.  Add
missing guard to the existing test on Is_Valued_Procedure.

From-SVN: r275931

4 years ago[Ada] Rtsfind: improve comment on RTE_Available
Bob Duff [Thu, 19 Sep 2019 08:12:34 +0000 (08:12 +0000)]
[Ada] Rtsfind: improve comment on RTE_Available

2019-09-19  Bob Duff  <duff@adacore.com>

gcc/ada/

* rtsfind.ads (RTE_Available): Improve comment.

From-SVN: r275930

4 years agotree-parloops.c (parloops_is_slp_reduction): Do not set LOOP_VINFO_OPERANDS_SWAPPED.
Richard Biener [Thu, 19 Sep 2019 06:08:33 +0000 (06:08 +0000)]
tree-parloops.c (parloops_is_slp_reduction): Do not set LOOP_VINFO_OPERANDS_SWAPPED.

2019-09-19  Richard Biener  <rguenther@suse.de>

* tree-parloops.c (parloops_is_slp_reduction): Do not set
LOOP_VINFO_OPERANDS_SWAPPED.
(parloops_is_simple_reduction): Likewise.
* tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Do not
initialize operands_swapped.
(_loop_vec_info::~_loop_vec_info): Do not re-canonicalize stmts.
(vect_is_slp_reduction): Do not swap operands.
* tree-vectorizer.h (_loop_vec_info::operands_swapped): Remove.
(LOOP_VINFO_OPERANDS_SWAPPED): Likewise.

From-SVN: r275928

4 years agoExtend pass rpad to handle avx512f vcvtusi2ss vcvtusi2ss 538.imagick_r improved...
Hongtao Liu [Thu, 19 Sep 2019 01:21:39 +0000 (01:21 +0000)]
Extend pass rpad to handle avx512f vcvtusi2ss vcvtusi2ss 538.imagick_r improved...

Extend pass rpad to handle avx512f vcvtusi2ss vcvtusi2ss
538.imagick_r improved by 4% with single copy run on SKYLAKE workstation.

gcc/
* config/i386/i386.md
(*floatuns<SWI48:mode><MODEF:mode>2_avx512):
Add avx_partial_xmm_update.

gcc/testsuie
* gcc.target/i386/pr87007-3.c: New test.

From-SVN: r275926

4 years agoRISC-V: Fix more splitters accidentally calling gen_reg_rtx.
Jim Wilson [Thu, 19 Sep 2019 01:19:25 +0000 (01:19 +0000)]
RISC-V: Fix more splitters accidentally calling gen_reg_rtx.

PR target/91683
* config/riscv/riscv-protos.h (riscv_split_symbol): New bool parameter.
(riscv_move_integer): Likewise.
* config/riscv/riscv.c (riscv_split_integer): Pass FALSE for new
riscv_move_integer arg.
(riscv_legitimize_move): Likewise.
(riscv_force_temporary): New parameter in_splitter.  Don't call
force_reg if true.
(riscv_unspec_offset_high): Pass FALSE for new riscv_force_temporary
arg.
(riscv_add_offset): Likewise.
(riscv_split_symbol): New parameter in_splitter.  Pass to
riscv_force_temporary.
(riscv_legitimize_address): Pass FALSE for new riscv_split_symbol
arg.
(riscv_move_integer): New parameter in_splitter.  New local
can_create_psuedo.  Don't call riscv_split_integer or force_reg when
in_splitter TRUE.
(riscv_legitimize_const_move): Pass FALSE for new riscv_move_integer,
riscv_split_symbol, and riscv_force_temporary args.
* config/riscv/riscv.md (low<mode>+1): Pass TRUE for new
riscv_move_integer arg.
(low<mode>+2): Pass TRUE for new riscv_split_symbol arg.

From-SVN: r275925

4 years agoDaily bump.
GCC Administrator [Thu, 19 Sep 2019 00:16:15 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r275924

4 years agolibgo: support gollvm build on arm64 linux
Ian Lance Taylor [Thu, 19 Sep 2019 00:03:25 +0000 (00:03 +0000)]
libgo: support gollvm build on arm64 linux

    This CL serves as part of an initial change for enabling gollvm
    building on arm64 linux, the rest of the change will be covered by
    another one to the gollvm repo.

    Incorporate type definition of 'uint128' to 'runtime' and 'syscall'
    packges, the change is not specific to arm64 linux but made available
    for all platforms.

    Verified by building and unit-testing gollvm on linux x86-64 and arm64.

    Verified by building and checking gccgo on linux x86-64 and arm64.

    Fixes golang/go#33711

    Change-Id: I4720c7d810cfd4ef720962fb4104c5641b2459c0

From-SVN: r275919

4 years ago[ARM] Add logical DImode expanders
Wilco Dijkstra [Wed, 18 Sep 2019 19:52:09 +0000 (19:52 +0000)]
[ARM] Add logical DImode expanders

We currently use default mid-end expanders for logical DImode operations.
These split operations without first splitting off complex immediates or
memory operands.  The resulting expansions are non-optimal and allow for
fewer LDRD/STRD opportunities.  So add back explicit expanders which ensure
memory operands and immediates are handled more efficiently.

    gcc/
PR target/91738
* config/arm/arm.md (<logical_op>di3): Expand explicitly.
(one_cmpldi2): Likewise.
* config/arm/arm.c (const_ok_for_dimode_op): Return true if one
of the constant parts is simple.
* config/arm/iterators.md (LOGICAL): Add new code iterator.
(logical_op): Add new code attribute.
(logical_OP): Likewise.
* config/arm/predicates.md (arm_anddi_operand): Add predicate.
(arm_iordi_operand): Add predicate.
(arm_xordi_operand): Add predicate.

From-SVN: r275907

4 years agoi386: Restore Skylake SImode hard register store cost
H.J. Lu [Wed, 18 Sep 2019 19:50:45 +0000 (19:50 +0000)]
i386: Restore Skylake SImode hard register store cost

On Skylake, we should move integer register to SSE register without
going through memory.  This patch restores Skylake SImode hard register
store cost to 6.

gcc/

PR target/90878
* config/i386/x86-tune-costs.h (skylake_cost): Restore SImode
hard register store cost to 6.

gcc/testsuite/

PR target/90878
* gcc.target/i386/pr90878.c: New test.

From-SVN: r275906

4 years agoi386: Increase Skylake SImode pseudo register store cost
H.J. Lu [Wed, 18 Sep 2019 19:49:19 +0000 (19:49 +0000)]
i386: Increase Skylake SImode pseudo register store cost

On Skylake, SImode store cost isn't less than half cost of 128-bit vector
store.  This patch increases Skylake SImode pseudo register store cost to
make it the same as QImode and HImode.

gcc/

PR target/91446
* config/i386/x86-tune-costs.h (skylake_cost): Increase SImode
pseudo register store cost from 3 to 6 to make it the same as
QImode and HImode.

gcc/testsuite/

PR target/91446
* gcc.target/i386/pr91446.c: New test.

From-SVN: r275905

4 years ago[ARM] Cleanup 64-bit multiplies
Wilco Dijkstra [Wed, 18 Sep 2019 18:33:30 +0000 (18:33 +0000)]
[ARM] Cleanup 64-bit multiplies

Cleanup 64-bit multiplies.  Combine the expanders using iterators.
Merge the signed/unsigned multiplies as well as the pre-Armv6 and Armv6
variants.  Split DImode operands early into parallel sets inside the
MULL/MLAL instructions - this improves register allocation and avoids
subreg issues due to other DImode operations splitting early.

    gcc/
* config/arm/arm.md (maddsidi4): Remove expander.
(mulsidi3adddi): Remove pattern.
(mulsidi3adddi_v6): Likewise.
(mulsidi3_nov6): Likewise.
(mulsidi3_v6): Likewise.
(umulsidi3): Remove expander.
(umulsidi3_nov6): Remove pattern.
(umulsidi3_v6): Likewise.
(umulsidi3adddi): Likewise.
(umulsidi3adddi_v6): Likewise.
(<Us>mulsidi3): Add combined expander.
(<Us>maddsidi4): Likewise.
(<US>mull): Add combined umull and smull pattern.
(<US>mlal): Likewise.
* config/arm/iterators.md (Us): Add new iterator.

From-SVN: r275901

4 years ago[ARM] Cleanup highpart multiply patterns
Wilco Dijkstra [Wed, 18 Sep 2019 18:22:55 +0000 (18:22 +0000)]
[ARM] Cleanup highpart multiply patterns

Cleanup the various highpart multiply patterns using iterators.
As a result the signed and unsigned variants and the pre-Armv6
multiply operand constraints are all handled in a single pattern
and simple expander.

    gcc/
* config/arm/arm.md (smulsi3_highpart): Use <US> and <SE> iterators.
(smulsi3_highpart_nov6): Remove pattern.
(smulsi3_highpart_v6): Likewise.
(umulsi3_highpart): Likewise.
(umulsi3_highpart_nov6): Likewise.
(umulsi3_highpart_v6): Likewise.
(<US>mull_high): Add new combined multiply pattern.

From-SVN: r275899

4 years agotree-vect-loop.c (vect_is_simple_reduction): Remove operand swapping.
Richard Biener [Wed, 18 Sep 2019 18:12:38 +0000 (18:12 +0000)]
tree-vect-loop.c (vect_is_simple_reduction): Remove operand swapping.

2019-09-18  Richard Biener  <rguenther@suse.de>

* tree-vect-loop.c (vect_is_simple_reduction): Remove operand
swapping.
(vectorize_fold_left_reduction): Remove assert.
(vectorizable_reduction): Also expect COND_EXPR non-reduction
operand in position 2.  Remove assert.

From-SVN: r275898

4 years ago[ARM] Cleanup multiply patterns
Wilco Dijkstra [Wed, 18 Sep 2019 18:11:24 +0000 (18:11 +0000)]
[ARM] Cleanup multiply patterns

Cleanup the 32-bit multiply patterns.  Merge the pre-Armv6 with the Armv6
patterns, remove useless alternatives and order the accumulator operands
to prefer MLA Ra, Rb, Rc, Ra whenever feasible.

    gcc/
* config/arm/arm.md (arm_mulsi3): Remove pattern.
(arm_mulsi3_v6): Likewise.
(mulsi3addsi_v6): Likewise.
(mulsi3subsi): Likewise.
(mul): Add new multiply pattern.
(mla): Likewise.
(mls): Likewise.

From-SVN: r275897

4 years agotree-parloops.c (report_ploop_op): Copy from report_vect_op.
Richard Biener [Wed, 18 Sep 2019 18:07:06 +0000 (18:07 +0000)]
tree-parloops.c (report_ploop_op): Copy from report_vect_op.

2019-09-18  Richard Biener  <rguenther@suse.de>

* tree-parloops.c (report_ploop_op): Copy from report_vect_op.
(parloops_valid_reduction_input_p): Copy from
valid_reduction_input_p.
(parloops_is_slp_reduction): Copy from vect_is_slp_reduction.
(parloops_needs_fold_left_reduction_p): Copy from
needs_fold_left_reduction_p.
(parloops_is_simple_reduction): Copy from
vect_is_simple_reduction.
(parloops_force_simple_reduction): Copy from
vect_force_simple_reduction.
(gather_scalar_reductions): Adjust.
* tree-vect-loop.c (vect_force_simple_reduction): Make static.
* tree-vectorizer.h (vect_force_simple_reduction): Remove.

From-SVN: r275896

4 years agoTweak clang-format configuration.
Martin Liska [Wed, 18 Sep 2019 13:32:25 +0000 (15:32 +0200)]
Tweak clang-format configuration.

2019-09-18  Martin Liska  <mliska@suse.cz>

* clang-format: Tweak configuration based on new
options offered.

From-SVN: r275875

4 years agotree-vectorizer.h (get_initial_def_for_reduction): Remove.
Richard Biener [Wed, 18 Sep 2019 12:43:08 +0000 (12:43 +0000)]
tree-vectorizer.h (get_initial_def_for_reduction): Remove.

2019-09-18  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (get_initial_def_for_reduction): Remove.
* tree-vect-loop.c (get_initial_def_for_reduction): Make
static.
(vect_create_epilog_for_reduction): Remove dead code.

From-SVN: r275874

4 years agoMake assemble_real generate canonical CONST_INTs
Richard Sandiford [Wed, 18 Sep 2019 11:37:02 +0000 (11:37 +0000)]
Make assemble_real generate canonical CONST_INTs

assemble_real used GEN_INT to create integers directly from the
longs returned by real_to_target.  assemble_integer then went on
to interpret the const_ints as though they had the mode corresponding
to the accompanying size parameter:

      imode = mode_for_size (size * BITS_PER_UNIT, mclass, 0).require ();

      for (i = 0; i < size; i += subsize)
{
  rtx partial = simplify_subreg (omode, x, imode, i);

But in the assemble_real case, X might not be canonical for IMODE.

If the interface to assemble_integer is supposed to allow outputting
(say) the low 4 bytes of a DImode integer, then the simplify_subreg
above is wrong.  But if the number of bytes passed to assemble_integer
is supposed to be the number of bytes that the integer actually contains,
assemble_real is wrong.

This patch takes the latter interpretation and makes assemble_real
generate const_ints that are canonical for the number of bytes passed.

The flip_storage_order handling assumes that each long is a full
SImode, which e.g. excludes BITS_PER_UNIT != 8 and float formats
whose memory size is not a multiple of 32 bits (which includes
HFmode at least).  The patch therefore leaves that code alone.
If interpreting each integer as SImode is correct, the const_ints
that it generates are also correct.

2019-09-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* varasm.c (assemble_real): Generate canonical const_ints.

From-SVN: r275873

4 years agore PR lto/91763 (go.go-torture/execute/printnil.go FAILs)
Richard Biener [Wed, 18 Sep 2019 11:28:20 +0000 (11:28 +0000)]
re PR lto/91763 (go.go-torture/execute/printnil.go FAILs)

2019-09-18  Richard Biener  <rguenther@suse.de>

PR lto/91763
* lto-streamer-in.c (input_eh_regions): Move EH init to
lto_materialize_function.
* tree-streamer-in.c (lto_input_ts_function_decl_tree_pointers):
Likewise.

lto/
* lto.c (lto_materialize_function): Initialize EH by looking
at the function personality and flag_exceptions setting.

From-SVN: r275872

4 years agoMake get_value_for_expr check for INTEGER_CSTs
Richard Sandiford [Wed, 18 Sep 2019 09:44:06 +0000 (09:44 +0000)]
Make get_value_for_expr check for INTEGER_CSTs

CONSTANT lattice values are symbolic constants rather than
compile-time constants, so among other things can be POLY_INT_CSTs.
This patch fixes a case in which we assumed all CONSTANTs were either
ADDR_EXPRs or INTEGER_CSTs.

This is tested by later SVE patches.

2019-09-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-ssa-ccp.c (get_value_for_expr): Check whether CONSTANTs
are INTEGER_CSTs.

From-SVN: r275871

4 years agoDon't treat variable-length vectors as VLAs during gimplification
Richard Sandiford [Wed, 18 Sep 2019 09:43:42 +0000 (09:43 +0000)]
Don't treat variable-length vectors as VLAs during gimplification

Source-level SVE vectors should be gimplified in the same way
as normal fixed-length vectors rather than as VLAs.

This is tested by later SVE patches.

2019-09-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* gimplify.c (gimplify_decl_expr): Use poly_int_tree_p instead
of checking specifically for INTEGER_CST.

From-SVN: r275870

4 years agoHandle variable-length vectors in compute_record_mode
Richard Sandiford [Wed, 18 Sep 2019 09:43:29 +0000 (09:43 +0000)]
Handle variable-length vectors in compute_record_mode

This patch makes compute_record_mode handle SVE vectors in the
same way as it would handle fixed-length vectors.  There should
be no change in behaviour for other targets.

This is needed for the SVE equivalent of arm_neon.h types like
int8x8x2_t (i.e. a pair of int8x8_ts).

2019-09-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* stor-layout.c (compute_record_mode): Operate on poly_uint64
sizes instead of uhwi sizes.

From-SVN: r275869

4 years agoTwo more POLY_INT cases for dwarf2out.c
Richard Sandiford [Wed, 18 Sep 2019 09:43:10 +0000 (09:43 +0000)]
Two more POLY_INT cases for dwarf2out.c

loc_list_for_tree_1 and add_const_value_attribute currently ICE
on POLY_INTs.  loc_list_for_tree_1 can do something sensible but
add_const_value_attribute has to punt, since the constant there
needs to be a link-time rather than load-time or run-time constant.

This is tested by later SVE patches.

2019-09-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* dwarf2out.c (loc_list_from_tree_1): Handle POLY_INT_CST.
(add_const_value_attribute): Handle CONST_POLY_INT.

From-SVN: r275868

4 years agoCome up with debug counter for store-merging.
Martin Liska [Wed, 18 Sep 2019 09:03:14 +0000 (11:03 +0200)]
Come up with debug counter for store-merging.

2019-09-18  Martin Liska  <mliska@suse.cz>

* dbgcnt.def (store_merging): New counter.
* gimple-ssa-store-merging.c (imm_store_chain_info::output_merged_stores):
Use it in store merging.

From-SVN: r275867