gcc.git
4 years agolibsanitizer: Fix GetPcSpBp determination of SP on 32-bit Solaris/x86
Rainer Orth [Thu, 16 Jul 2020 08:38:48 +0000 (10:38 +0200)]
libsanitizer: Fix GetPcSpBp determination of SP on 32-bit Solaris/x86

The latest Solaris 11.4/x86 update uncovered a libsanitizer bug that
caused one test to FAIL for 32-bit:

+FAIL: c-c++-common/asan/null-deref-1.c   -O0  output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c   -O1  output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c   -O2  output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c   -O2 -flto  output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c -O2 -flto -flto-partition=none
output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c   -O3 -g  output pattern test
+FAIL: c-c++-common/asan/null-deref-1.c   -Os  output pattern test

I've identified the problem and the fix has just landed in upstream
llvm-project:

https://reviews.llvm.org/D83664

Tested on i386-pc-solaris2.11 and x86_64-pc-linux.gnu.

libsanitizer:
* sanitizer_common/sanitizer_linux.cpp: Cherry-pick llvm-project
revision f0e9b76c3500496f8f3ea7abe6f4bf801e3b41e7.

4 years agoDaily bump.
GCC Administrator [Thu, 16 Jul 2020 00:16:34 +0000 (00:16 +0000)]
Daily bump.

4 years agoi386: Introduce peephole2 to use flags from CMPXCHG more [PR96189]
Uros Bizjak [Wed, 15 Jul 2020 19:27:00 +0000 (21:27 +0200)]
i386: Introduce peephole2 to use flags from CMPXCHG more [PR96189]

CMPXCHG instruction sets ZF flag if the values in the destination operand
and EAX register are equal; otherwise the ZF flag is cleared and value
from destination operand is loaded to EAX. Following assembly:

        movl    %esi, %eax
        lock cmpxchgl   %edx, (%rdi)
        cmpl    %esi, %eax
        sete    %al

can be optimized by removing the unneeded comparison, since set ZF flag
signals that no update to EAX happened.

2020-15-07  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:
PR target/95355
* config/i386/sync.md
(peephole2 to remove unneded compare after CMPXCHG): New pattern.

gcc/testsuite/ChangeLog:
PR target/95355
* gcc.target/i386/pr96189.c: New test.

4 years agolibgomp.fortran/alloc-1.F90: Fix testcase for 32bit size_t
Tobias Burnus [Wed, 15 Jul 2020 15:23:04 +0000 (17:23 +0200)]
libgomp.fortran/alloc-1.F90: Fix testcase for 32bit size_t

libgomp/ChangeLog:

* testsuite/libgomp.fortran/alloc-1.F90: Use c_size_t to
avoid conversion on 32bit systems from 32bit to 64bit due
to -fdefault-integer-8.

4 years agoopenmp: Fix up loop-21.c
Jakub Jelinek [Wed, 15 Jul 2020 14:34:54 +0000 (16:34 +0200)]
openmp: Fix up loop-21.c

I've missed
+FAIL: libgomp.c/loop-21.c execution test
during testing of the recent patch.  The problem is that while
for the number of iterations computation it doesn't matter if we compute
min_inner_iterations as (m2 * first + n2 + (adjusted step) + m1 * first + n1) / step
or (m2 * last + n2 + (adjusted step) + m1 * last + n1) / step provided that
in the second case we use as factor (m1 - m2) * ostep / step rather than
(m2 - m1) * ostep / step, for the logical to actual iterator values computation
it does matter and in my hand written C implementations of all the cases (outer
vs. inner loop with increasing vs. decreasing iterator) I'm using the same computation
and it worked well for all the pseudo-random iterators testing it was doing.

It also means min_inner_iterations is misnamed, because it is not really
minimum number of inner iterations, whether the first or last outer iteration
results in the smaller or larger value of this can be (sometimes) only
determined at runtime.
So this patch also renames it to first_inner_iterations.

2020-07-15  Jakub Jelinek  <jakub@redhat.com>

PR libgomp/96198
* omp-general.h (struct omp_for_data): Rename min_inner_iterations
member to first_inner_iterations, adjust comment.
* omp-general.c (omp_extract_for_data): Adjust for the above change.
Always use n1first and n2first to compute it, rather than depending
on single_nonrect_cond_code.  Similarly, always compute factor
as (m2 - m1) * outer_step / inner_step rather than sometimes m1 - m2
depending on single_nonrect_cond_code.
* omp-expand.c (expand_omp_for_init_vars): Rename min_inner_iterations
to first_inner_iterations and min_inner_iterationsd to
first_inner_iterationsd.

4 years agoc++: refactor some parser code
Nathan Sidwell [Wed, 15 Jul 2020 14:38:35 +0000 (07:38 -0700)]
c++: refactor some parser code

cp_parser_declaration copies tokens to local variables, before inspecting
(some of) their fields.  There's no need.  Just point at them in the token
buffer -- they don't move.  Also, we never look at the second token if the
first is EOF, so no need for some kind of dummy value in that case.

gcc/cp/
* parser.c (cp_parser_declaration): Avoid copying tokens.
(cp_parser_block_declaration): RAII token pointer.

4 years ago[Ada] Cleanup code related to object overlays
Piotr Trojanek [Fri, 5 Jun 2020 11:30:54 +0000 (13:30 +0200)]
[Ada] Cleanup code related to object overlays

gcc/ada/

* sem_util.adb (Find_Overlaid_Entity): Fix style in comment.
(Note_Possible_Modification): Simplify repeated calls to Ekind.

4 years ago[Ada] Cleanup in Convert_To_Positional after previous work
Eric Botcazou [Fri, 5 Jun 2020 09:35:25 +0000 (11:35 +0200)]
[Ada] Cleanup in Convert_To_Positional after previous work

gcc/ada/

* exp_aggr.adb (Flatten): Adjust description.
(Convert_To_Positional): Remove obsolete ??? comment and use
Compile_Time_Known_Value in the final test.

4 years ago[Ada] Ongoing work for AI12-0212: container aggregates
Arnaud Charlet [Wed, 3 Jun 2020 07:42:19 +0000 (03:42 -0400)]
[Ada] Ongoing work for AI12-0212: container aggregates

gcc/ada/

* par-ch4.adb (P_Iterated_Component_Association): Extended to
recognzize the similar Iterated_Element_Association. This node
is only generated when an explicit Key_Expression is given.
Otherwise the distinction between the two iterated forms is done
during semantic analysis.
* sinfo.ads: New node N_Iterated_Element_Association, for
Ada202x container aggregates.  New field Key_Expression.
* sinfo.adb: Subprograms for new node and newn field.
* sem_aggr.adb (Resolve_Iterated_Component_Association): Handle
the case where the Iteration_Scheme is an
Iterator_Specification.
* exp_aggr.adb (Wxpand_Iterated_Component): Handle a component
with an Iterated_Component_Association, generate proper loop
using given Iterator_Specification.
* exp_util.adb (Insert_Axtions): Handle new node as other
aggregate components.
* sem.adb, sprint.adb: Handle new node.
* tbuild.adb (Make_Implicit_Loop_Statement): Handle properly a
loop with an Iterator_ specification.

4 years ago[Ada] Fix logic in Allocate_Any_Controlled
Arnaud Charlet [Thu, 4 Jun 2020 18:18:18 +0000 (14:18 -0400)]
[Ada] Fix logic in Allocate_Any_Controlled

gcc/ada/

* libgnat/s-stposu.adb (Allocate_Any_Controlled): Fix logic in
lock/unlock.

4 years ago[Ada] Target name is an object reference
Piotr Trojanek [Thu, 28 May 2020 18:28:28 +0000 (20:28 +0200)]
[Ada] Target name is an object reference

gcc/ada/

* sem_util.adb (Is_Object_Reference): Return True on
N_Target_Name.

4 years ago[Ada] Spurious error on Predicate_Failure aspect
Arnaud Charlet [Thu, 4 Jun 2020 09:09:59 +0000 (05:09 -0400)]
[Ada] Spurious error on Predicate_Failure aspect

gcc/ada/

* sem_ch13.adb (Check_Aspect_At_End_Of_Declarations): Add proper
handling of Aspect_Predicate_Failure, consistent with
Check_Aspect_At_Freeze_Point.

4 years ago[Ada] Mention -gnat2020 in error message
Arnaud Charlet [Thu, 4 Jun 2020 08:13:16 +0000 (10:13 +0200)]
[Ada] Mention -gnat2020 in error message

gcc/ada/

* sem_ch13.adb (Check_Aspect_Too_Late): Mention -gnat2020 switch
in error message.

4 years ago[Ada] Do not generate elaboration code for alignment aspect
Eric Botcazou [Thu, 4 Jun 2020 08:15:33 +0000 (10:15 +0200)]
[Ada] Do not generate elaboration code for alignment aspect

gcc/ada/

* einfo.ads (Delayed Freezing and Elaboration): Adjust description.
* freeze.adb (Freeze_Object_Declaration): Likewise.
* sem_ch3.adb (Delayed_Aspect_Present): Likewise.  Do not return
true for Alignment.
* sem_ch13.adb (Analyze_Aspect_Specifications): Do not always delay
for Alignment.  Moreover, for Alignment and various Size aspects,
do not delay if the expression is an attribute whose prefix is the
Standard package.

4 years ago[Ada] Do not generate extra copies inside initialization procedures
Eric Botcazou [Thu, 4 Jun 2020 07:56:02 +0000 (09:56 +0200)]
[Ada] Do not generate extra copies inside initialization procedures

gcc/ada/

* exp_ch6.adb (Requires_Atomic_Or_Volatile_Copy): Return false
inside an initialization procedure.

4 years ago[Ada] Guard against access to wrong fields in Is_Renaming
Ghjuvan Lacambre [Wed, 3 Jun 2020 09:09:20 +0000 (11:09 +0200)]
[Ada] Guard against access to wrong fields in Is_Renaming

gcc/ada/

* sem_util.adb (Is_Renaming): Add ekind checks.

4 years ago[Ada] Remove doc on obsolete tools
Arnaud Charlet [Wed, 3 Jun 2020 20:14:38 +0000 (16:14 -0400)]
[Ada] Remove doc on obsolete tools

gcc/ada/

* doc/gnat_ugn/gnat_utility_programs.rst: Remove doc on obsolete
tools.
* gnat_ugn.texi: Regenerate.

4 years ago[Ada] Assert failure with -gnatwr
Arnaud Charlet [Wed, 3 Jun 2020 16:27:43 +0000 (12:27 -0400)]
[Ada] Assert failure with -gnatwr

gcc/ada/

* sem_res.adb (Resolve_Type_Conversion): Protect against null
entity.  Add proper tag for -gnatwr warning.

4 years ago[Ada] Missing error on operator call
Arnaud Charlet [Wed, 3 Jun 2020 07:42:19 +0000 (03:42 -0400)]
[Ada] Missing error on operator call

gcc/ada/

* sem_ch6.adb (Analyze_Procedure_Call): Detect use of operators
in a procedure call.
* sem_util.adb: Minor edit.

4 years ago[Ada] Fix for possibly null ranges in 'Update and delta_aggregate
Piotr Trojanek [Tue, 2 Jun 2020 22:17:21 +0000 (00:17 +0200)]
[Ada] Fix for possibly null ranges in 'Update and delta_aggregate

gcc/ada/

* exp_spark.adb (Expand_SPARK_Delta_Or_Update): Apply scalar
range checks against the base type of an index type, not against
the index type itself.

4 years ago[Ada] Small addition and tweaks in documentation on freezing
Eric Botcazou [Wed, 3 Jun 2020 08:30:09 +0000 (10:30 +0200)]
[Ada] Small addition and tweaks in documentation on freezing

gcc/ada/

* einfo.ads (Delayed Freezing and Elaboration): Minor tweaks.
Document the discrepancy between the aspect and the non-aspect
cases for alignment settings in object declarations.

4 years ago[Ada] ACATS 4.1P [BDB4001] - 13.11.4(22-23/3) not enforced
Arnaud Charlet [Sat, 30 May 2020 18:04:33 +0000 (14:04 -0400)]
[Ada] ACATS 4.1P [BDB4001] - 13.11.4(22-23/3) not enforced

gcc/ada/

* exp_ch3.adb (Freeze_Type): Remove warning in expander,
replaced by a corresponding error in sem_ch13.adb. Replace
RTE_Available by RTU_Loaded to avoid adding unnecessary
dependencies.
* sem_ch13.adb (Associate_Storage_Pool): New procedure.
(Analyze_Attribute_Definition_Clause
[Attribute_Simple_Storage_Pool| Attribute_Storage_Pool]): Call
Associate_Storage_Pool to add proper legality checks on
subpools.

4 years ago[Ada] Mark standard containers as not in SPARK
Yannick Moy [Tue, 2 Jun 2020 12:16:25 +0000 (14:16 +0200)]
[Ada] Mark standard containers as not in SPARK

gcc/ada/

* libgnat/a-cbdlli.adb, libgnat/a-cbdlli.ads,
libgnat/a-cbhama.adb, libgnat/a-cbhama.ads,
libgnat/a-cbhase.adb, libgnat/a-cbhase.ads,
libgnat/a-cbmutr.adb, libgnat/a-cbmutr.ads,
libgnat/a-cborma.adb, libgnat/a-cborma.ads,
libgnat/a-cborse.adb, libgnat/a-cborse.ads,
libgnat/a-cbprqu.adb, libgnat/a-cbprqu.ads,
libgnat/a-cbsyqu.adb, libgnat/a-cbsyqu.ads,
libgnat/a-cdlili.adb, libgnat/a-cdlili.ads,
libgnat/a-cidlli.adb, libgnat/a-cidlli.ads,
libgnat/a-cihama.adb, libgnat/a-cihama.ads,
libgnat/a-cihase.adb, libgnat/a-cihase.ads,
libgnat/a-cimutr.adb, libgnat/a-cimutr.ads,
libgnat/a-ciorma.adb, libgnat/a-ciorma.ads,
libgnat/a-ciormu.adb, libgnat/a-ciormu.ads,
libgnat/a-ciorse.adb, libgnat/a-ciorse.ads,
libgnat/a-cohama.adb, libgnat/a-cohama.ads,
libgnat/a-cohase.adb, libgnat/a-cohase.ads,
libgnat/a-coinve.adb, libgnat/a-coinve.ads,
libgnat/a-comutr.adb, libgnat/a-comutr.ads,
libgnat/a-convec.adb, libgnat/a-convec.ads,
libgnat/a-coorma.adb, libgnat/a-coorma.ads,
libgnat/a-coormu.adb, libgnat/a-coormu.ads,
libgnat/a-coorse.adb, libgnat/a-coorse.ads: Add SPARK_Mode =>
Off.

4 years ago[Ada] Fix oversight in Delayed_Aspect_Present predicate
Eric Botcazou [Tue, 2 Jun 2020 22:08:17 +0000 (00:08 +0200)]
[Ada] Fix oversight in Delayed_Aspect_Present predicate

gcc/ada/

* sem_ch3.adb (Delayed_Aspect_Present): Fix oversight in loop.
* freeze.adb (Freeze_Object_Declaration): Use Declaration_Node
instead of Parent for the sake of consistency.

4 years ago[Ada] Wrong resolution of 'access in protected subprogram
Javier Miranda [Tue, 2 Jun 2020 19:46:10 +0000 (15:46 -0400)]
[Ada] Wrong resolution of 'access in protected subprogram

gcc/ada/

* sem_attr.adb (Resolve_Attribute): Resolve overloaded
N_Selected_Component prefix of 'Access. Required to handle
overloaded prefixed view of protected subprograms.

4 years ago[Ada] Documentation cleanups
Arnaud Charlet [Tue, 2 Jun 2020 12:40:26 +0000 (08:40 -0400)]
[Ada] Documentation cleanups

gcc/ada/

* doc/gnat_ugn/about_this_guide.rst: Remove old section and
update for Ada 202x.
* doc/gnat_ugn/getting_started_with_gnat.rst: Add a system
requirements section. Remove obsolete section and minimal
rewording on the getting started section.
* gnat_ugn.texi: Regenerate.

4 years ago[Ada] Fix slices and qualified expressions being effectively volatile
Piotr Trojanek [Sat, 30 May 2020 09:56:38 +0000 (11:56 +0200)]
[Ada] Fix slices and qualified expressions being effectively volatile

gcc/ada/

* exp_ch5.adb (Expand_Assign_Array): Use short-circuit operator
(style).
* sem_res.adb (Resolve_Indexed_Component): Fix style in comment.
* sem_util.adb (Is_Effectively_Volatile_Object): Handle slices
just like indexed components; handle qualified expressions and
type conversions lie in Is_OK_Volatile_Context.
(Is_OK_Volatile_Context): Handle qualified expressions just like
type conversions.

4 years ago[Ada] Cleanup condition for an effectively volatile array type
Piotr Trojanek [Fri, 29 May 2020 18:57:53 +0000 (20:57 +0200)]
[Ada] Cleanup condition for an effectively volatile array type

gcc/ada/

* sem_prag.adb (Atomic_Components): Simplify with Ekind_In.
(Complex_Representation): Fix type of E_Id, which just like when
for pragma Atomic_Components will hold an N_Identifier node, not
an entity.
* sem_util.adb (Is_Effectively_Volatile): Refactor to avoid
unnecessary computation.

4 years ago[Ada] Extend static functions
Arnaud Charlet [Wed, 27 May 2020 10:10:35 +0000 (06:10 -0400)]
[Ada] Extend static functions

gcc/ada/

* inline.adb, inline.ads
(Inline_Static_Expression_Function_Call): Renamed
Inline_Static_Function_Call.
* sem_ch13.adb (Analyze_Aspect_Static): Allow static intrinsic
imported functions under -gnatX.
* sem_util.ads, sem_util.adb (Is_Static_Expression_Function):
Renamed Is_Static_Function.
(Is_Static_Expression_Function_Call): Renamed
Is_Static_Function_Call.
* sem_ch6.adb, sem_elab.adb, sem_res.adb: Update calls to
Is_Static_Function*.
* sem_eval.adb (Fold_Dummy, Eval_Intrinsic_Call, Fold_Shift):
New.
(Eval_Call): Add support for intrinsic calls, code refactoring.
(Eval_Entity_Name): Code refactoring.
(Eval_Logical_Op): Update comment.
(Eval_Shift): Call Fold_Shift. Update comments.
* par-prag.adb (Par [Pragma_Extensions_Allowed]): Set
Ada_Version to Ada_Version_Type'Last to handle
Extensions_Allowed (On) consistently.
* opt.ads (Extensions_Allowed): Update documentation.
* sem_attr.adb: Update comment.
* doc/gnat_rm/implementation_defined_pragmas.rst: Update
documentation of Extensions_Allowed.
* gnat_rm.texi: Regenerate.

4 years ago[Ada] Potential access to uninitialized variable
Arnaud Charlet [Tue, 2 Jun 2020 06:57:11 +0000 (02:57 -0400)]
[Ada] Potential access to uninitialized variable

gcc/ada/

* sem_aggr.adb (Resolve_Iterated_Component_Association): Ensure
Typ is never accessed uninitialized.

4 years ago[Ada] Fix typo in "accommodate" in docs and comments
Piotr Trojanek [Fri, 8 May 2020 22:06:16 +0000 (00:06 +0200)]
[Ada] Fix typo in "accommodate" in docs and comments

gcc/ada/

* doc/gnat_rm/representation_clauses_and_pragmas.rst: Fix typo.
* gnat_rm.texi: Regenerate.
* libgnat/s-secsta.ads (Memory_Alignment): Likewise.

4 years ago[Ada] Minor reformatting of comments and some code, plus a typo correction
Gary Dismukes [Mon, 1 Jun 2020 19:22:48 +0000 (15:22 -0400)]
[Ada] Minor reformatting of comments and some code, plus a typo correction

gcc/ada/

* exp_ch6.adb: Add a comma and fix a typo (machinary =>
machinery) in comment.
* exp_aggr.adb: Reformat, fix capitalization, and add a couple
of commas in a comment. Adjust columns in several code
fragments.
* sem_aggr.adb: Reformat and add a comma in a comment.

4 years ago[Ada] Use Sloc of delay statement in timed entry call
Eric Botcazou [Mon, 1 Jun 2020 18:49:43 +0000 (20:49 +0200)]
[Ada] Use Sloc of delay statement in timed entry call

gcc/ada/

* exp_ch9.adb (Expand_N_Timed_Entry_Call): Use the Sloc of
the delay statement in the expansion.

4 years ago[Ada] Ongoing work for unnamed and named container aggregates
Ed Schonberg [Sun, 31 May 2020 20:01:41 +0000 (16:01 -0400)]
[Ada] Ongoing work for unnamed and named container aggregates

gcc/ada/

* sem_aggr.adb (Resolve_Iterated_Component_Association): New
procedure, internal to Resolve_Container_Aggregate, to complete
semantic analysis of Iterated_Component_Associations.
* exp_aggr.adb (Expand_Iterated_Component): New procedure,
internal to Expand_Container_Aggregate, to expand the construct
into an implicit loop that performs individual insertions into
the target aggregate.

4 years ago[Ada] Spurious accessibility error on allocator
Justin Squirek [Fri, 29 May 2020 19:44:50 +0000 (15:44 -0400)]
[Ada] Spurious accessibility error on allocator

gcc/ada/

* exp_ch6.adb (Make_Build_In_Place_Call_Allocator): Normalize
the associated node for internally generated objects to be like
their SOAAT counter-parts.

4 years ago[Ada] Fix potentially uninitialized variable flagged by CodePeer
Arnaud Charlet [Sat, 30 May 2020 18:01:17 +0000 (14:01 -0400)]
[Ada] Fix potentially uninitialized variable flagged by CodePeer

gcc/ada/

* libgnat/g-socket.adb (Wait_On_Socket): Fix potentially
uninitialized variable.

4 years agoc++: error recovery & pragmas
Nathan Sidwell [Wed, 15 Jul 2020 13:00:10 +0000 (06:00 -0700)]
c++: error recovery & pragmas

Parser error recovery can get confused by the tokens within a deferred
pragma, as treats those as regular tokens.  This adjusts the recovery
so that the pragma is treated as a unit.  Also, the preprocessor now
ensures that we never have an EOF token inside a pragma -- the pragma
is always closed first.

gcc/cp/
* parser.c (cp_parser_skip_to_closing_parenthesis_1): Deal with
meeting a deferred pragma.
(cp_parser_skip_to_end_of_statement): Likewise.
(cp_parser_skip_to_end_of_block_or_statement): Likewise.
(cp_parser_skip_to_pragma_eol): We should never meet EOF.
(cp_parser_omp_declare_simd): Likewise.
(cp_parser_omp_declare_reduction, cp_parser_oacc_routine)
(pragma_lex): Likewise.
gcc/testsuite/
* g++.dg/parse/pragma-recovery.C: New.

4 years agolibgomp.fortran/struct-elem-map-1.f90: Add char kind=4 tests
Tobias Burnus [Wed, 15 Jul 2020 10:29:44 +0000 (12:29 +0200)]
libgomp.fortran/struct-elem-map-1.f90: Add char kind=4 tests

As the Fortran PR 95837 has been fixed, the test could be be added.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/struct-elem-map-1.f90: Remove unused
variables; add character(kind=4) tests; update TODO comment.

4 years agofix _mm512_{,mask_}cmp*_p[ds]_mask at -O0 [PR96174]
Jakub Jelinek [Wed, 15 Jul 2020 09:34:44 +0000 (11:34 +0200)]
fix _mm512_{,mask_}cmp*_p[ds]_mask at -O0 [PR96174]

The _mm512_{,mask_}cmp_p[ds]_mask and also _mm_{,mask_}cmp_s[ds]_mask
intrinsics have an argument which must have a constant passed to it
and so use an inline version only for ifdef __OPTIMIZE__ and have
a #define for -O0.  But the _mm512_{,mask_}cmp*_p[ds]_mask intrinsics
don't need a constant argument, they are essentially the first
set with the constant added to them implicitly based on the comparison
name, and so there is no #define version for them (correctly).
But their inline versions are defined in between the first and s[ds]
set and so inside of ifdef __OPTIMIZE__, which means that with -O0
they aren't defined at all.

This patch fixes that by moving those after the #ifdef __OPTIMIZE #else
use #define #endif block.

2020-07-15  Jakub Jelinek  <jakub@redhat.com>

PR target/96174
* config/i386/avx512fintrin.h (_mm512_cmpeq_pd_mask,
_mm512_mask_cmpeq_pd_mask, _mm512_cmplt_pd_mask,
_mm512_mask_cmplt_pd_mask, _mm512_cmple_pd_mask,
_mm512_mask_cmple_pd_mask, _mm512_cmpunord_pd_mask,
_mm512_mask_cmpunord_pd_mask, _mm512_cmpneq_pd_mask,
_mm512_mask_cmpneq_pd_mask, _mm512_cmpnlt_pd_mask,
_mm512_mask_cmpnlt_pd_mask, _mm512_cmpnle_pd_mask,
_mm512_mask_cmpnle_pd_mask, _mm512_cmpord_pd_mask,
_mm512_mask_cmpord_pd_mask, _mm512_cmpeq_ps_mask,
_mm512_mask_cmpeq_ps_mask, _mm512_cmplt_ps_mask,
_mm512_mask_cmplt_ps_mask, _mm512_cmple_ps_mask,
_mm512_mask_cmple_ps_mask, _mm512_cmpunord_ps_mask,
_mm512_mask_cmpunord_ps_mask, _mm512_cmpneq_ps_mask,
_mm512_mask_cmpneq_ps_mask, _mm512_cmpnlt_ps_mask,
_mm512_mask_cmpnlt_ps_mask, _mm512_cmpnle_ps_mask,
_mm512_mask_cmpnle_ps_mask, _mm512_cmpord_ps_mask,
_mm512_mask_cmpord_ps_mask): Move outside of __OPTIMIZE__ guarded
section.

* gcc.target/i386/avx512f-vcmppd-3.c: New test.
* gcc.target/i386/avx512f-vcmpps-3.c: New test.

4 years agobuiltins: Avoid useless char/short -> int promotions before atomics [PR96176]
Jakub Jelinek [Wed, 15 Jul 2020 09:26:22 +0000 (11:26 +0200)]
builtins: Avoid useless char/short -> int promotions before atomics [PR96176]

As mentioned in the PR, we generate a useless movzbl insn before lock cmpxchg.
The problem is that the builtin for the char/short cases has the arguments
promoted to int and combine gives up, because the instructions have
MEM_VOLATILE_P arguments and recog in that case doesn't recognize anything
when volatile_ok is false, and nothing afterwards optimizes the
(reg:SI a) = (zero_extend:SI (reg:QI a))
... (subreg:QI (reg:SI a) 0) ...

The following patch fixes it at expansion time, we already have a function
that is meant to undo the promotion, so this just adds the very common case
to that.

2020-07-15  Jakub Jelinek  <jakub@redhat.com>

PR target/96176
* builtins.c: Include gimple-ssa.h, tree-ssa-live.h and
tree-outof-ssa.h.
(expand_expr_force_mode): If exp is a SSA_NAME with different mode
from MODE and get_gimple_for_ssa_name is a cast from MODE, use the
cast's rhs.

* gcc.target/i386/pr96176.c: New test.

4 years agors6000: Refine RTL unroll hook for small loops
guojiufu [Wed, 15 Jul 2020 08:07:47 +0000 (16:07 +0800)]
rs6000: Refine RTL unroll hook for small loops

For very small loops (< 6 insns), it would be fine to unroll 4
times to run fast with less latency and better cache usage.  Like
below loops:
 while (i) a[--i] = NULL;   while (p < e)  *d++ = *p++;

With this patch enhances, we could see some performance improvement
for some workloads(e.g. SPEC2017).

2020-07-13  Jiufu Guo   <guojiufu@cn.ibm.com>

* config/rs6000/rs6000.c (rs6000_loop_unroll_adjust): Refine hook.

4 years agolibgomp: Add Fortran routine support for allocators
Tobias Burnus [Wed, 15 Jul 2020 06:33:20 +0000 (08:33 +0200)]
libgomp: Add Fortran routine support for allocators

libgomp/ChangeLog:

* allocator.c: Add ialias for omp_init_allocator and
omp_destroy_allocator.
* configure.ac: Set INTPTR_T_KIND.
* configure: Regenerate.
* Makefile.in: Regenerate.
* testsuite/Makefile.in: Regenerate.
* fortran.c (omp_init_allocator_, omp_destroy_allocator_,
omp_set_default_allocator_, omp_get_default_allocator_): New
functions and ialias_redirect.
* icv.c: Add ialias for omp_set_default_allocator and
omp_get_default_allocator.
* libgomp.map (OMP_5.0.1): Add omp_init_allocator_,
omp_destroy_allocator_, omp_set_default_allocator_ and
omp_get_default_allocator_.
* omp_lib.f90.in: Add allocator traits parameters, declare
allocator routines and add related kind parameters.
* omp_lib.h.in: Likewise.
* testsuite/libgomp.c-c++-common/alloc-2.c: Fix sizeof.
* testsuite/libgomp.fortran/alloc-1.F90: New test.
* testsuite/libgomp.fortran/alloc-2.F90: New test.
* testsuite/libgomp.fortran/alloc-3.F: New test.
* testsuite/libgomp.fortran/alloc-4.f90: New test.
* testsuite/libgomp.fortran/alloc-5.f90: New test.

4 years agoDaily bump.
GCC Administrator [Wed, 15 Jul 2020 00:16:35 +0000 (00:16 +0000)]
Daily bump.

4 years agoc++: Add new test [PR59978]
Marek Polacek [Tue, 14 Jul 2020 21:31:08 +0000 (17:31 -0400)]
c++: Add new test [PR59978]

Fixed in r224162.  That came without a test so adding this one.
Previously, we issued a bogus "too few arguments to function" error.

gcc/testsuite/ChangeLog:

PR c++/59978
* g++.dg/cpp0x/vt-59978.C: New test.

4 years agox86: Replace __glibc_unlikely with __builtin_expect
H.J. Lu [Tue, 14 Jul 2020 21:01:51 +0000 (14:01 -0700)]
x86: Replace __glibc_unlikely with __builtin_expect

Replace glibc specific __glibc_unlikely with __builtin_expect.

PR target/95443
* gcc.target/i386/pr95443-1.c (simple_strstr): Replace
__glibc_unlikely with __builtin_expect.

4 years agoc++: Make convert_like complain about bad ck_ref_bind again [PR95789]
Marek Polacek [Tue, 23 Jun 2020 01:26:49 +0000 (21:26 -0400)]
c++: Make convert_like complain about bad ck_ref_bind again [PR95789]

convert_like issues errors about bad_p conversions at the beginning
of the function, but in the ck_ref_bind case, it only issues them
after we've called convert_like on the next conversion.

This doesn't work as expected since r10-7096 because when we see
a conversion from/to class type in a template, we return early, thereby
missing the error, and a bad_p conversion goes by undetected.  That
made the attached test to compile even though it should not.

I had thought that I could just move the ck_ref_bind/bad_p errors
above to the rest of them, but that regressed diagnostics because
expr then wasn't converted yet by the nested convert_like_real call.

So, for bad_p conversions, do the normal processing, but still return
the IMPLICIT_CONV_EXPR to avoid introducing trees that the template
processing can't handle well.  This I achieved by adding a wrapper
function.

gcc/cp/ChangeLog:

PR c++/95789
PR c++/96104
PR c++/96179
* call.c (convert_like_real_1): Renamed from convert_like_real.
(convert_like_real): New wrapper for convert_like_real_1.

gcc/testsuite/ChangeLog:

PR c++/95789
PR c++/96104
PR c++/96179
* g++.dg/conversion/ref4.C: New test.
* g++.dg/conversion/ref5.C: New test.
* g++.dg/conversion/ref6.C: New test.

4 years agors6000: Don't call movsi_from_sf in 32 bit mode.
David Edelsohn [Tue, 14 Jul 2020 19:31:23 +0000 (15:31 -0400)]
rs6000: Don't call movsi_from_sf in 32 bit mode.

movsi_from_sf uses rldimi instruction, which will cause the compiler to ICE
in 32 bit mode.  This patch limits the recently added pattern and call to
TARGET_POWERPC64.

2020-07-14  David Edelsohn  <dje.gcc@gmail.com>

gcc/ChangeLog

* config/rs6000/rs6000.md (rotldi3_insert_sf): Add TARGET_POWERPC64
condition.
* config/rs6000/rs6000.c (rs6000_expand_vector_init): Add
TARGET_POWERPC64 requirement to TARGET_P8_VECTOR case.

4 years agoc++: Parser entry cleanup
Nathan Sidwell [Tue, 14 Jul 2020 17:49:03 +0000 (10:49 -0700)]
c++: Parser entry cleanup

The handling of PCH is a little trick, because we have to deal with it before
allocating memory.  I found the layering somewhat confusing.  This patch
reorganizes that, so that the stopping of PCH is done in exactly one place,
and the ordering of lexer creation relative to that is much clearer.

I also changed the error message about multiple source files as with C++20,
'modules' means something rather specific.

Other than the error message changes, no functional changes.

gcc/cp/
* parser.c (cp_lexer_alloc): Do not deal with PCH here.
(cp_lexer_new_main): Deal with PCH here.  Store the tokens directly
into the buffer.
(cp_lexer_new_from_tokens): Assert last token isn't purged either.
(cp_lexer_get_preprocessor_token): Change first arg to flags, adjust.
(cp_parser_new): Pass the lexer in, don't create it here.
(cp_parser_translation_unit): Initialize access checks here.
(cp_parser_initial_pragma): First token is provided by caller,
don't deal with PCH stopping here.  Adjust error message.
(c_parse_file): Adjust, change error message to avoid C++20 module
confusion.

4 years agolibgomp: Fix hang when profiling OpenACC programs with CUDA 9.0 nvprof
Kwok Cheung Yeung [Tue, 14 Jul 2020 17:31:35 +0000 (10:31 -0700)]
libgomp: Fix hang when profiling OpenACC programs with CUDA 9.0 nvprof

The version of nvprof in CUDA 9.0 causes a hang when used to profile an
OpenACC program.  This is because it calls acc_get_device_type from
a callback called during device initialization, which then attempts
to acquire acc_device_lock while it is already taken, resulting in
deadlock.  This works around the issue by returning acc_device_none
from acc_get_device_type without attempting to acquire the lock when
initialization has not completed yet.

2020-07-14  Tom de Vries  <tom@codesourcery.com>
    Cesar Philippidis  <cesar@codesourcery.com>
    Thomas Schwinge  <thomas@codesourcery.com>
    Kwok Cheung Yeung  <kcy@codesourcery.com>

libgomp/
* oacc-init.c (acc_init_state_lock, acc_init_state, acc_init_thread):
New variable.
(acc_init_1): Set acc_init_thread to pthread_self ().  Set
acc_init_state to initializing at the start, and to initialized at the
end.
(self_initializing_p): New function.
(acc_get_device_type): Return acc_device_none if called by thread that
is currently executing acc_init_1.
* libgomp.texi (acc_get_device_type): Update documentation.
(Implementation Status and Implementation-Defined Behavior): Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-init-2.c: New.

4 years agodemangler: don't treat lambda as a substitution candidate
Ian Lance Taylor [Fri, 10 Jul 2020 16:34:28 +0000 (09:34 -0700)]
demangler: don't treat lambda as a substitution candidate

libiberty/ChangeLog:

PR demangler/96143
* cp-demangle.c (d_lambda): Don't add substitution candidate.
* testsuite/demangle-expected: Update a few existing test cases
accordingly, and add a new test case.

4 years agodiagnostics: Support conversion of tabs to spaces [PR49973] [PR86904]
Lewis Hyatt [Tue, 14 Jul 2020 16:05:56 +0000 (12:05 -0400)]
diagnostics: Support conversion of tabs to spaces [PR49973] [PR86904]

Supports conversion of tabs to spaces when outputting diagnostics. Also
adds -fdiagnostics-column-unit and -fdiagnostics-column-origin options to
control how the column number is output, thereby resolving the two PRs.

gcc/c-family/ChangeLog:

PR other/86904
* c-indentation.c (should_warn_for_misleading_indentation): Get
global tabstop from the new source.
* c-opts.c (c_common_handle_option): Remove handling of -ftabstop, which
is now a common option.
* c.opt: Likewise.

gcc/ChangeLog:

PR preprocessor/49973
PR other/86904
* common.opt: Handle -ftabstop here instead of in c-family
options.  Add -fdiagnostics-column-unit= and
-fdiagnostics-column-origin= options.
* opts.c (common_handle_option): Handle the new options.
* diagnostic-format-json.cc (json_from_expanded_location): Add
diagnostic_context argument.  Use it to convert column numbers as per
the new options.
(json_from_location_range): Likewise.
(json_from_fixit_hint): Likewise.
(json_end_diagnostic): Pass the new context argument to helper
functions above.  Add "column-origin" field to the output.
(test_unknown_location): Add the new context argument to calls to
helper functions.
(test_bad_endpoints): Likewise.
* diagnostic-show-locus.c
(exploc_with_display_col::exploc_with_display_col): Support
tabstop parameter.
(layout_point::layout_point): Make use of class
exploc_with_display_col.
(layout_range::layout_range): Likewise.
(struct line_bounds): Clarify that the units are now always
display columns.  Rename members accordingly.  Add constructor.
(layout::print_source_line): Add support for tab expansion.
(make_range): Adapt to class layout_range changes.
(layout::maybe_add_location_range): Likewise.
(layout::layout): Adapt to class exploc_with_display_col changes.
(layout::calculate_x_offset_display): Support tabstop parameter.
(layout::print_annotation_line): Adapt to struct line_bounds changes.
(layout::print_line): Likewise.
(line_label::line_label): Add diagnostic_context argument.
(get_affected_range): Likewise.
(get_printed_columns): Likewise.
(layout::print_any_labels): Adapt to struct line_label changes.
(class correction): Add m_tabstop member.
(correction::correction): Add tabstop argument.
(correction::compute_display_cols): Use m_tabstop.
(class line_corrections): Add m_context member.
(line_corrections::line_corrections): Add diagnostic_context argument.
(line_corrections::add_hint): Use m_context to handle tabstops.
(layout::print_trailing_fixits): Adapt to class line_corrections
changes.
(test_layout_x_offset_display_utf8): Support tabstop parameter.
(test_layout_x_offset_display_tab): New selftest.
(test_one_liner_colorized_utf8): Likewise.
(test_tab_expansion): Likewise.
(test_diagnostic_show_locus_one_liner_utf8): Call the new tests.
(diagnostic_show_locus_c_tests): Likewise.
(test_overlapped_fixit_printing): Adapt to helper class and
function changes.
(test_overlapped_fixit_printing_utf8): Likewise.
(test_overlapped_fixit_printing_2): Likewise.
* diagnostic.h (enum diagnostics_column_unit): New enum.
(struct diagnostic_context): Add members for the new options.
(diagnostic_converted_column): Declare.
(json_from_expanded_location): Add new context argument.
* diagnostic.c (diagnostic_initialize): Initialize new members.
(diagnostic_converted_column): New function.
(maybe_line_and_column): Be willing to output a column of 0.
(diagnostic_get_location_text): Convert column number as per the new
options.
(diagnostic_report_current_module): Likewise.
(assert_location_text): Add origin and column_unit arguments for
testing the new functionality.
(test_diagnostic_get_location_text): Test the new functionality.
* doc/invoke.texi: Document the new options and behavior.
* input.h (location_compute_display_column): Add tabstop argument.
* input.c (location_compute_display_column): Likewise.
(test_cpp_utf8): Add selftests for tab expansion.
* tree-diagnostic-path.cc (default_tree_make_json_for_path): Pass the
new context argument to json_from_expanded_location().

libcpp/ChangeLog:

PR preprocessor/49973
PR other/86904
* include/cpplib.h (struct cpp_options):  Removed support for -ftabstop,
which is now handled by diagnostic_context.
(class cpp_display_width_computation): New class.
(cpp_byte_column_to_display_column): Add optional tabstop argument.
(cpp_display_width): Likewise.
(cpp_display_column_to_byte_column): Likewise.
* charset.c
(cpp_display_width_computation::cpp_display_width_computation): New
function.
(cpp_display_width_computation::advance_display_cols): Likewise.
(compute_next_display_width): Removed and implemented this
functionality in a new function...
(cpp_display_width_computation::process_next_codepoint): ...here.
(cpp_byte_column_to_display_column): Added tabstop argument.
Reimplemented in terms of class cpp_display_width_computation.
(cpp_display_column_to_byte_column): Likewise.
* init.c (cpp_create_reader): Remove handling of -ftabstop, which is now
handled by diagnostic_context.

gcc/testsuite/ChangeLog:

PR preprocessor/49973
PR other/86904
* c-c++-common/Wmisleading-indentation-3.c: Adjust expected output
for new defaults.
* c-c++-common/Wmisleading-indentation.c: Likewise.
* c-c++-common/diagnostic-format-json-1.c: Likewise.
* c-c++-common/diagnostic-format-json-2.c: Likewise.
* c-c++-common/diagnostic-format-json-3.c: Likewise.
* c-c++-common/diagnostic-format-json-4.c: Likewise.
* c-c++-common/diagnostic-format-json-5.c: Likewise.
* c-c++-common/missing-close-symbol.c: Likewise.
* g++.dg/diagnostic/bad-binary-ops.C: Likewise.
* g++.dg/parse/error4.C: Likewise.
* g++.old-deja/g++.brendan/crash11.C: Likewise.
* g++.old-deja/g++.pt/overload2.C: Likewise.
* g++.old-deja/g++.robertl/eb109.C: Likewise.
* gcc.dg/analyzer/malloc-paths-9.c: Likewise.
* gcc.dg/bad-binary-ops.c: Likewise.
* gcc.dg/format/branch-1.c: Likewise.
* gcc.dg/format/pr79210.c: Likewise.
* gcc.dg/plugin/diagnostic-test-expressions-1.c: Likewise.
* gcc.dg/plugin/diagnostic-test-string-literals-1.c: Likewise.
* gcc.dg/redecl-4.c: Likewise.
* gfortran.dg/diagnostic-format-json-1.F90: Likewise.
* gfortran.dg/diagnostic-format-json-2.F90: Likewise.
* gfortran.dg/diagnostic-format-json-3.F90: Likewise.
* go.dg/arrayclear.go: Add a comment explaining why adding a
comment was necessary to work around a dejagnu bug.
* c-c++-common/diagnostic-units-1.c: New test.
* c-c++-common/diagnostic-units-2.c: New test.
* c-c++-common/diagnostic-units-3.c: New test.
* c-c++-common/diagnostic-units-4.c: New test.
* c-c++-common/diagnostic-units-5.c: New test.
* c-c++-common/diagnostic-units-6.c: New test.
* c-c++-common/diagnostic-units-7.c: New test.
* c-c++-common/diagnostic-units-8.c: New test.

4 years agoc++: tree dumper
Nathan Sidwell [Tue, 14 Jul 2020 15:19:36 +0000 (08:19 -0700)]
c++: tree dumper

A couple of C++ nodes were unknown to the tree dumper.

gcc/cp/
* ptree.c (cxx_print_type): Add TYPEOF_TYPE and BASES.

4 years agoc++: Refactor some class fns
Nathan Sidwell [Tue, 14 Jul 2020 15:11:13 +0000 (08:11 -0700)]
c++: Refactor some class fns

Storing CLASSTYPE_AS_BASE in a local variable makes some code clearer
(and textually no longer).  For some reason we store a DECL in a variable
called 'value', which is confusing.

gcc/cp/
* class.c (build_base_field_1): Cache CLASSTYPE_AS_BASE.
(build_self_reference): Rename value -> decl.
(dump_class_hierarchy_1): Cache CLASSTYPE_AS_BASE.

4 years agoaix: FAT libraries: test native compiler mode directly
David Edelsohn [Mon, 13 Jul 2020 18:43:39 +0000 (14:43 -0400)]
aix: FAT libraries: test native compiler mode directly

The FAT libraries config fragments need to know which library is native
and which is a multilib to choose the correct multilib from which to
append the additional object file or shared object file.  Testing the
top-level archive is fragile because it will fail if rebuilding.  This
patch tests the compiler preprocessing macros for the 64 bit AIX specific
__64BIT__ to determine the native mode of the compiler in MULTILIBTOP.

2020-07-14  David Edelsohn  <dje.gcc@gmail.com>

libatomic/ChangeLog

* config/t-aix: Set BITS from compiler cpp macro.

libgcc/ChangeLog

* config/rs6000/t-slibgcc-aix: Set BITS from compiler cpp macro.

libgfortran/ChangeLog

* config/t-aix: Set BITS from compiler cpp macro.

libgomp/ChangeLog

* config/t-aix: Set BITS from compiler cpp macro.

libstdc++-v3/ChangeLog

* config/os/aix/t-aix: Set BITS from compiler cpp macro.

4 years agoFix goacc/finalize-1.f tree dump-scanning for -m32
Tobias Burnus [Tue, 14 Jul 2020 14:28:37 +0000 (16:28 +0200)]
Fix goacc/finalize-1.f tree dump-scanning for -m32

gcc/testsuite/ChangeLog:

* gfortran.dg/goacc/finalize-1.f: Relax scan-tree-dump-times
pattern to work on 32bit-pointer systems.

4 years agotestsuite: adapt g++.dg/ipa/pr83667.C for AIX
David Edelsohn [Sat, 11 Jul 2020 23:42:56 +0000 (19:42 -0400)]
testsuite: adapt g++.dg/ipa/pr83667.C for AIX

The current testsuite regex looks for THUNK0.  AIX generates THUNK..0.
This patch expands the regex to allow 0 or more dots between THUNK
and the 0.

gcc/testsuite/ChangeLog

2020-07-14  David Edelsohn  <dje.gcc@gmail.com>

* g++.dg/ipa/pr83667.C: Allow 0 or more dots between THUNK and 0.

4 years agoexpr: Unbreak build of mesa [PR96194]
Jakub Jelinek [Tue, 14 Jul 2020 14:01:11 +0000 (16:01 +0200)]
expr: Unbreak build of mesa [PR96194]

> > The store to the whole of each volatile object was picked apart
> > like there had been an individual assignment to each of the
> > fields.  Reads were added as part of that; see PR for details.
> > The reads from volatile memory were a clear bug; individual
> > stores questionable.  A separate patch clarifies the docs.

This breaks building of mesa on both the trunk and 10 branch.

The problem is that the middle-end may never create temporaries of non-POD
(TREE_ADDRESSABLE) types, those can be only created when the language says
so and thus only the FE is allowed to create those.

This patch just reverts the behavior to what we used to do before for the
stores to volatile non-PODs.  Perhaps we want to do something else, but
definitely we can't create temporaries of the non-POD type.  It is up to
discussions on what should happen in those cases.

2020-07-14  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/96194
* expr.c (expand_constructor): Don't create temporary for store to
volatile MEM if exp has an addressable type.

* g++.dg/opt/pr96194.C: New test.

4 years agoc++: Improve checking of decls with trailing return type [PR95820]
Marek Polacek [Wed, 24 Jun 2020 21:39:21 +0000 (17:39 -0400)]
c++: Improve checking of decls with trailing return type [PR95820]

This is an ICE-on-invalid but I've been seeing it when reducing
various testcases, so it's more important for me than usually.

splice_late_return_type now checks that if we've seen a late return
type, the function return type was auto.  That's a fair assumption
but grokdeclarator/cdk_function wasn't giving errors for function
pointers and similar.  So we want to perform various checks not only
when funcdecl_p || inner_declarator == NULL.  But only give the
!late_return_type errors when funcdecl_p, to accept e.g.

auto (*fp)() = f;

in C++11.  Here's a diff -w to ease the review:

--- a/gcc/cp/decl.c
+++ b/gcc/cp/decl.c
@@ -12102,14 +12102,9 @@ grokdeclarator (const cp_declarator *declarator,

      /* Handle a late-specified return type.  */
      tree late_return_type = declarator->u.function.late_return_type;
-     if (funcdecl_p
- /* This is the case e.g. for
-    using T = auto () -> int.  */
- || inner_declarator == NULL)
-       {
      if (tree auto_node = type_uses_auto (type))
        {
-     if (!late_return_type)
+ if (!late_return_type && funcdecl_p)
    {
      if (current_class_type
  && LAMBDA_TYPE_P (current_class_type))
@@ -12201,7 +12196,6 @@ grokdeclarator (const cp_declarator *declarator,
      "type specifier", name);
  return error_mark_node;
        }
-       }
      type = splice_late_return_type (type, late_return_type);
      if (type == error_mark_node)
        return error_mark_node;

gcc/cp/ChangeLog:

PR c++/95820
* decl.c (grokdeclarator) <case cdk_function>: Check also
pointers/references/... to functions.

gcc/testsuite/ChangeLog:

PR c++/95820
* g++.dg/cpp1y/auto-fn58.C: New test.

4 years agocore: formatting & comment
Nathan Sidwell [Tue, 14 Jul 2020 13:08:44 +0000 (06:08 -0700)]
core: formatting & comment

One of hash_map's getters returns a pointer to T or null, The comment
said it returned T or null.  Also an unaligned pair of comments.

gcc/
* hash-map.h (hash_map::get): Note it is a pointer to value.
* incpath.h (incpath_kind): Align comments.

4 years agoc++: Comments & formatting
Nathan Sidwell [Tue, 14 Jul 2020 12:57:57 +0000 (05:57 -0700)]
c++:  Comments & formatting

I found some bad formatting and misleading or incomplete comments
during my spelunking around the c++FE.  May as well clean up trunk and
record what I noted.

gcc/cp/
* cp-tree.h: Correct some tree lang flag comments,
reformat some structure definitions.  Note some structure
sizes.  Clarify some comments.
(yyungetc): Delete.  Not been a thing for some time.
* class.c (copy_fndecl_with_name): Comment.
(check_bases_and_members): Unnecessary {}.
(layout_class_type): Comment.
* cp-tree.def (UNBOUND_CLASS_TEMPLATE): Adjust comment.
* decl.c:  Fix some formatting & whitespace issues.
(function_requirements_equivalent_p): Note why
substitutions are needed.
* decl2.c (no_linkage_error): Note that heroics about
'typedef struct { ... };' are no longer needed.
* method.c: Whitespace.
* name-lookup.c: Whitespace.
(add_decl_to_level): Reformat a line.
(print_binding_stack): Mark as DEBUG_FUNCTION.
(has_using_namespace_std_directive_p): Delete comment.
* pt.c: Whitespace
* ptree.c: Whitespace.
* rtti.c: Whitespace & comment.
* tree.c: Comment.
* typeck.c (structural_comptypes): Add comment.

4 years agocore: add tree-node comments
Nathan Sidwell [Mon, 13 Jul 2020 20:52:03 +0000 (13:52 -0700)]
core: add tree-node comments

As we've moved to 64-bit systems, the padding information has become
conditionally inaccurate.  I also hit cases where invalid tree codes
did not get flagged as invalid.

gcc/
* tree-core.h (tree_decl_with_vis, tree_function_decl):
Note additional padding on 64-bits
* tree.c (cache_integer_cst): Note why no caching of enum literals.
(get_tree_code_name): Robustify error case.

4 years agodoc: Fix some typos
Nathan Sidwell [Mon, 13 Jul 2020 20:42:52 +0000 (13:42 -0700)]
doc: Fix some typos

gty calls gt_clear*e*_cache not gt_clear_cache.  I know not why it is
named so, but at least document it correctly.  invoke.texi had a duplicate opindex.

gcc/
* doc/gty.texi: Fic gt_cleare_cache name.
* doc/invoke.texi: Remove duplicate opindex Wabi-tag.

4 years agoFortran : ICE in gfc_check_pointer_assign PR95612
Mark Eggleston [Thu, 11 Jun 2020 10:05:40 +0000 (11:05 +0100)]
Fortran  : ICE in gfc_check_pointer_assign PR95612

Output an error if the right hand value is a zero sized array or
does not have a symbol tree otherwise continue checking.

2020-07-14  Steven G. Kargl  <kargl@gcc.gnu.org>

gcc/fortran/

PR fortran/95612
* expr.c (gfc_check_pointer_assigb): Output an error if
rvalue is a zero sized array or output an error if rvalue
doesn't have a symbol tree.

2020-07-14  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/95612
* gfortran.dg/pr95612.f90: New test.

4 years ago[OpenMP, Fortran] Add structure/derived-type element mapping
Tobias Burnus [Tue, 14 Jul 2020 11:39:46 +0000 (13:39 +0200)]
[OpenMP, Fortran] Add structure/derived-type element mapping

gcc/fortran/ChangeLog:

* openmp.c (gfc_match_omp_clauses): Match also derived-type
component refs in OMP_CLAUSE_MAP.
(resolve_omp_clauses): Resolve those.
* trans-openmp.c (gfc_trans_omp_array_section, gfc_trans_omp_clauses):
Handle OpenMP structure-element mapping.
(gfc_trans_oacc_construct, gfc_trans_oacc_executable_directive,
(gfc_trans_oacc_combined_directive, gfc_trans_oacc_declare): Update
add openacc=true in gfc_trans_omp_clauses call.

gcc/testsuite/ChangeLog:

* gfortran.dg/goacc/finalize-1.f: Update dump scan pattern.
* gfortran.dg/gomp/map-1.f90: Update dg-error.
* gfortran.dg/gomp/map-2.f90: New test.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/struct-elem-map-1.f90: New test.

4 years ago[Fortran, OpenMP] Fix allocatable-components check (PR67311)
Tobias Burnus [Tue, 14 Jul 2020 10:55:53 +0000 (12:55 +0200)]
[Fortran, OpenMP] Fix allocatable-components check (PR67311)

gcc/fortran/ChangeLog:

PR fortran/67311
* trans-openmp.c (gfc_has_alloc_comps): Return false also for
pointers to arrays.

libgomp/ChangeLog:

PR fortran/67311
* testsuite/libgomp.fortran/target-map-1.f90: New test.

4 years agoopenmp: Adjust outer bounds of non-rect loops
Jakub Jelinek [Tue, 14 Jul 2020 08:31:59 +0000 (10:31 +0200)]
openmp: Adjust outer bounds of non-rect loops

In loops like:
  #pragma omp parallel for collapse(2)
  for (i = -4; i < 8; i++)
    for (j = 3 * i; j > 2 * i; j--)
for some outer loop iterations there are no inner loop iterations at all,
the condition is false.  In order to use Summæ Potestate to count number
of iterations or to transform the logical iteration number to actual
iterator values using quadratic non-equation root discovery the outer
iterator range needs to be adjusted, such that the inner loop has at least
one iteration for each of the outer loop iterator value in the reduced
range.  Sometimes this adjustment is done at the start of the range,
at other times at the end.

This patch implements it during the compile time number of loop computation
(if all expressions are compile time constants).

2020-07-14  Jakub Jelinek  <jakub@redhat.com>

* omp-general.h (struct omp_for_data): Add adjn1 member.
* omp-general.c (omp_extract_for_data): For non-rect loop, punt on
count computing if n1, n2 or step are not INTEGER_CST earlier.
Narrow the outer iterator range if needed so that non-rect loop
has at least one iteration for each outer range iteration.  Compute
adjn1.
* omp-expand.c (expand_omp_for_init_vars): Use adjn1 if non-NULL
instead of the outer loop's n1.

* testsuite/libgomp.c/loop-21.c: New test.

4 years agoLTO: pick up -fcf-protection flag for the link step
Matthias Klose [Tue, 14 Jul 2020 08:12:08 +0000 (10:12 +0200)]
LTO: pick up -fcf-protection flag for the link step

2020-07-14  Matthias Klose  <doko@ubuntu.com>

PR lto/95604
* lto-wrapper.c (merge_and_complain): Add decoded options as parameter,
error on different values for -fcf-protection.
(append_compiler_options): Pass -fcf-protection option.
(find_and_merge_options): Add decoded options as parameter,
pass decoded_options to merge_and_complain.
(run_gcc): Pass decoded options to find_and_merge_options.
* lto-opts.c (lto_write_options): Pass -fcf-protection option.

4 years agoFortran : Implicitly type parameter causes an invalid error PR96038
Mark Eggleston [Mon, 6 Jul 2020 06:14:59 +0000 (07:14 +0100)]
Fortran  : Implicitly type parameter causes an invalid error PR96038

If a paramter to declared and initialised before its type is
declared a bogus error is output at the type declaration
idicating that initialisation is missing.

2020-07-14  Steven G. Kargl  <kargl@gcc.gnu.org>

gcc/fortran/

PR fortran/96038
* decl.c (add_init_expr_sym):  For a symbol that is a
parameter accept an initialisation if it does not have a
value otherwise output a error and reject.

2020-07-14  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/96038
* gfortran.dg/pr96038.f90: New test.

4 years agoDaily bump.
GCC Administrator [Tue, 14 Jul 2020 00:16:30 +0000 (00:16 +0000)]
Daily bump.

4 years agoMerge sibcall_local32/64
Alan Modra [Mon, 13 Jul 2020 23:49:43 +0000 (09:19 +0930)]
Merge sibcall_local32/64

This patch consolidates four insn patterns into two.

* config/rs6000/rs6000.md (sibcall_local): Merge sibcall_local32
and sibcall_local64.
(sibcall_value_local): Similarly.

4 years agors6000: clean up testsuite power10_hw check
Aaron Sawdey [Mon, 13 Jul 2020 21:19:33 +0000 (16:19 -0500)]
rs6000: clean up testsuite power10_hw check

Because the check for power10_hw is not called
check_effective_target_power10_hw, it needs to be looked
for by is-effective-target-keyword. Also reorder things
in is-effective-target to put power10_hw with the other
ppc stuff.

gcc/testsuite/

* lib/target-supports.exp (is-effective-target):
Reorder to put powerpc stuff together.
(is-effective-target-keyword): Add power10_hw.

4 years agotestsuite: scan-lang-dump-times & scan-lang-dump-not
Nathan Sidwell [Mon, 13 Jul 2020 20:37:46 +0000 (13:37 -0700)]
testsuite: scan-lang-dump-times & scan-lang-dump-not

Sigh, last week's success at not breaking things failed with an
incorrect 'fix' this morning.  Let's reduce my confusion by making
lib/scanlang.exp the same on trunk as modules.

gcc/testsuite/
* lib/scanlang.exp (scan-lang-dump): Fix breakage.
(scan-lang-dump-times, scan-lang-dump-not): New.

4 years agoRemove long-gone cxxmain.c from makefiles
Nathan Sidwell [Mon, 13 Jul 2020 20:17:20 +0000 (13:17 -0700)]
Remove long-gone cxxmain.c from makefiles

I discovered we were attempting to delete some no-longer generated
files.

gcc/
* Makefile.in (distclean): Remove long gone cxxmain.c
gcc/cp/
* Make-lang.in (c++.disclean): Likewise.

4 years agors6000: add effective-target test ppc_mma_hw
Aaron Sawdey [Fri, 10 Jul 2020 22:46:17 +0000 (17:46 -0500)]
rs6000: add effective-target test ppc_mma_hw

Add a test for dejagnu to determine if execution of MMA instructions is
supported in the test environment. Add an execution test to make sure
that __builtin_cpu_supports("mma") is true if we can execute MMA
instructions.

gcc/testsuite/

* lib/target-supports.exp (check_ppc_mma_hw_available):
New function.
(is-effective-target): Add ppc_mma_hw.
(is-effective-target-keyword): Add ppc_mma_hw.
* gcc.target/powerpc/mma-supported.c: New file.
* gcc.target/powerpc/mma-single-test.c: Require ppc_mma_hw.
* gcc.target/powerpc/mma-double-test.c: Require ppc_mma_hw.

4 years agox86: Pass a copy of the string length to cmpstrnqi
H.J. Lu [Mon, 13 Jul 2020 17:33:47 +0000 (10:33 -0700)]
x86: Pass a copy of the string length to cmpstrnqi

cmpstrnsi expander may pass the actual string length directly to cmpstrnqi
patterns.  For cmpstrnsi, one of the strings must be a constant and
expand_builtin_strncmp rewrites the length argument to be the minimum of
the const string length and the actual string length.  But it is not the
case for cmpmemsi.  Pass a copy of the string length to cmpstrnqi patterns
to avoid changing the actual string length by cmpstrnqi patterns.

gcc/

PR target/95443
* config/i386/i386.md (cmpstrnsi): Pass a copy of the string
length to cmpstrnqi patterns.

gcc/testsuite/

PR target/95443
* gcc.target/i386/pr95443-1.c: New test.
* gcc.target/i386/pr95443-2.c: Likewise.

4 years agoc++: Add test [PR95288]
Marek Polacek [Mon, 13 Jul 2020 17:05:06 +0000 (13:05 -0400)]
c++: Add test [PR95288]

Somewhat improved by r11-2064, though we still generate junk that seems
redundant.  But at least it says
error: expected ‘}’ before ‘.’ token

PR c++/95288
* g++.dg/diagnostic/enum2.C: New test.

4 years agoipa-fnsummary: Fix ICE with switch predicates [PR96130]
Jakub Jelinek [Mon, 13 Jul 2020 16:25:53 +0000 (18:25 +0200)]
ipa-fnsummary: Fix ICE with switch predicates [PR96130]

The following testcase ICEs since r10-3199.
There is a switch with default label, where the controlling expression has
range just 0..7 and there are case labels for all those 8 values, but
nothing has yet optimized away the default.
Since r10-3199, set_switch_stmt_execution_predicate sets the switch to
default label's edge's predicate to a false predicate and then
compute_bb_predicates propagates the predicates through the cfg, but false
predicates aren't really added.  The caller of compute_bb_predicates
in one place handles NULL bb->aux as false predicate:
      if (fbi.info)
{
  if (bb->aux)
    bb_predicate = *(predicate *) bb->aux;
  else
    bb_predicate = false;
}
      else
bb_predicate = true;
but then in two further spots that the patch below is changing
it assumes bb->aux must be non-NULL.  Those two spots are guarded by a
condition that is only true if fbi.info is non-NULL, so I think the right
fix is to treat NULL aux as false predicate in those spots too.

2020-07-13  Jakub Jelinek  <jakub@redhat.com>

PR ipa/96130
* ipa-fnsummary.c (analyze_function_body): Treat NULL bb->aux
as false predicate.

* gcc.dg/torture/pr96130.c: New test.

4 years agoFortran : accepts pointer initialization of DT dummy args PR45337
Mark Eggleston [Wed, 10 Jun 2020 06:22:50 +0000 (07:22 +0100)]
Fortran  : accepts pointer initialization of DT dummy args PR45337

Initialisation of a variable results in an implicit save attribute
being added to the variable.  The save attribute is not allowed for
variables with the dummy attribute set.  Initialisation should be
rejected for dummy variables.

2020-07-13  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/fortran/

PR fortran/45337
* resolve.c (resolve_fl_variable): Remove type and intent
checks from the check for dummy.

2020-07-13  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/45337
* gfortran.dg/pr45337_1.f90: New test.
* gfortran.dg/pr45337_2.f90: New test.

4 years agoc++: Fix tentative parsing of enum-specifier [PR96077]
Marek Polacek [Fri, 10 Jul 2020 00:44:05 +0000 (20:44 -0400)]
c++: Fix tentative parsing of enum-specifier [PR96077]

Here's an interesting issue: in this code a ) is missing:

  enum { E = (2 } e;

but we compile the code anyway, and E is set to 0 in build_enumerator,
which is sneaky.

The problem is that cp_parser_enum_specifier parses tentatively, because
when we see the enum keyword, we don't know yet if we'll find an
enum-specifier, opaque-enum-declaration, or elaborated-enum-specifier.

In this test when we call cp_parser_enumerator_list we're still parsing
tentatively, and as a consequence, parens.require_close (parser) in
cp_parser_primary_expression doesn't report any errors.  But we only go
on to parse the enumerator-list after we've seen a {, at which point we
might as well commit -- we know we're dealing with an enum-specifier.

gcc/cp/ChangeLog:

PR c++/96077
* parser.c (cp_parser_enum_specifier): Commit to tentative parse
after we've seen an opening brace.

gcc/testsuite/ChangeLog:

PR c++/96077
* g++.dg/parse/enum14.C: New test.

4 years agotestsuite: Fix scan-lang-dump-not
Nathan Sidwell [Mon, 13 Jul 2020 15:04:26 +0000 (08:04 -0700)]
testsuite: Fix scan-lang-dump-not

turned out scan-lang-dump-not was broken in the 3 argument case -- I'd missed a
necessary empty arg.  Fixed thusly.

gcc/testsuite/
* lib/scanlang.exp (scan-lang-dump-not): Fix 3-arg case.

4 years agotree-optimization/96163 - fix placement issue with SLP and vectors
Richard Biener [Mon, 13 Jul 2020 10:41:35 +0000 (12:41 +0200)]
tree-optimization/96163 - fix placement issue with SLP and vectors

This avoids placing stmts beyond the vectorizer region begin which
confuses vect_stmt_dominates_stmt_p.

2020-07-13  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96163
* tree-vect-slp.c (vect_schedule_slp_instance): Put new stmts
at least after region begin.

* g++.dg/vect/pr96163.cc: New testcase.

4 years agoFortran : ICE in gfc_find_array_ref(): No ref found PR95981
Mark Eggleston [Tue, 30 Jun 2020 09:15:05 +0000 (10:15 +0100)]
Fortran  :  ICE in gfc_find_array_ref(): No ref found PR95981

When looking for an array reference allow NULL references.  If
no array reference is found dim_rank_check should return false.

2020-07-13  Steven G. Kargl  <kargl@gcc.gnu.org>

gcc/fortran/

PR fortran/95981
* check.c (dim_rank_check): Allow NULL references in call to
gfc_find_array_ref and return false if no reference is found.

2020-07-13  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/95981
* gfortran.dg/pr95981.f90: New test.

4 years agoaarch64: Add missing ACLE support for PAC-RET
Szabolcs Nagy [Thu, 2 Jul 2020 15:11:04 +0000 (16:11 +0100)]
aarch64: Add missing ACLE support for PAC-RET

Define the __ARM_FEATURE_PAC_DEFAULT feature test
macro when PAC-RET branch protection is enabled.

2020-07-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

gcc/ChangeLog:

* config/aarch64/aarch64-c.c (aarch64_update_cpp_builtins): Add
__ARM_FEATURE_PAC_DEFAULT support.

4 years agodoc: Clarify __builtin_return_address [PR94891]
Szabolcs Nagy [Thu, 28 May 2020 09:28:30 +0000 (10:28 +0100)]
doc: Clarify __builtin_return_address [PR94891]

The expected semantics and valid usage of __builtin_return_address is
not clear since it exposes implementation internals that are normally
not meaningful to portable c code.

This documentation change tries to clarify the semantics in case the
return address is stored in a mangled form. This affects AArch64 when
pointer authentication is used for the return address signing (i.e.
-mbranch-protection=pac-ret).

2020-07-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

gcc/ChangeLog:

PR target/94891
* doc/extend.texi: Update the text for  __builtin_return_address.

4 years agolibgcc: fix the handling of return address mangling [PR94891]
Szabolcs Nagy [Thu, 4 Jun 2020 08:33:35 +0000 (09:33 +0100)]
libgcc: fix the handling of return address mangling [PR94891]

Mangling, currently only used on AArch64 for return address signing,
is an internal representation that should not be exposed via

  __builtin_return_address return value,
  __builtin_eh_return handler argument,
  _Unwind_DebugHook handler argument.

Note that a mangled address might not even fit into a void *, e.g.
with AArch64 ilp32 ABI the return address is stored as 64bit, so
the mangled return address cannot be accessed via _Unwind_GetPtr.

This patch changes the unwinder hooks as follows:

MD_POST_EXTRACT_ROOT_ADDR is removed: root address comes from
__builtin_return_address which is not mangled.

MD_POST_EXTRACT_FRAME_ADDR is renamed to MD_DEMANGLE_RETURN_ADDR,
it now operates on _Unwind_Word instead of void *, so the hook
should work when return address signing is enabled on AArch64 ilp32.
(But for that __builtin_aarch64_autia1716 should be fixed to operate
on 64bit input instead of a void *.)

MD_POST_FROB_EH_HANDLER_ADDR is removed: it is the responsibility of
__builtin_eh_return to do the mangling if necessary.

2020-07-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

libgcc/ChangeLog:

PR target/94891
* config/aarch64/aarch64-unwind.h (MD_POST_EXTRACT_ROOT_ADDR): Remove.
(MD_POST_FROB_EH_HANDLER_ADDR): Remove.
(MD_POST_EXTRACT_FRAME_ADDR): Rename to ...
(MD_DEMANGLE_RETURN_ADDR): This.
(aarch64_post_extract_frame_addr): Rename to ...
(aarch64_demangle_return_addr): This.
(aarch64_post_frob_eh_handler_addr): Remove.
* unwind-dw2.c (uw_update_context): Demangle return address.
(uw_frob_return_addr): Remove.

4 years agoaarch64: fix __builtin_eh_return with pac-ret [PR94891]
Szabolcs Nagy [Thu, 4 Jun 2020 12:42:16 +0000 (13:42 +0100)]
aarch64: fix __builtin_eh_return with pac-ret [PR94891]

Currently __builtin_eh_return takes a signed return address, which can
cause ABI and API issues: 1) pointer representation problems if the
address is passed around before eh return, 2) the source code needs
pac-ret specific changes and needs to know if pac-ret is used in the
current frame, 3) signed address may not be representible as void *
(with ilp32 abi).

Using address signing to protect eh return is ineffective because the
instruction sequence in the unwinder that starts from the address
signing and ends with a ret can be used as a return to anywhere gadget.
Using indirect branch istead of ret with bti j landing pads at the
target can reduce the potential of such gadget, which also implies
that __builtin_eh_return should not take a signed address.

This is a big hammer fix to the ABI and API issues: it turns pac-ret
off for the caller completely (not just on the eh return path).  To
harden the caller against ROP attacks, it should use indirect branch
instead of ret, this is not attempted so the patch remains small and
backportable.

2020-07-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

gcc/ChangeLog:

PR target/94891
* config/aarch64/aarch64.c (aarch64_return_address_signing_enabled):
Disable return address signing if __builtin_eh_return is used.

gcc/testsuite/ChangeLog:

PR target/94891
* gcc.target/aarch64/return_address_sign_1.c: Update test.
* gcc.target/aarch64/return_address_sign_b_1.c: Likewise.

4 years agoaarch64: fix return address access with pac [PR94891][PR94791]
Szabolcs Nagy [Tue, 2 Jun 2020 15:44:41 +0000 (16:44 +0100)]
aarch64: fix return address access with pac [PR94891][PR94791]

This is a big hammer fix for __builtin_return_address (PR target/94891)
returning signed addresses (sometimes, depending on wether lr happens
to be signed or not at the time of call which depends on optimizations),
and similarly -pg may pass signed return address to _mcount
(PR target/94791).

At the time of return address expansion we don't know if it's signed or
not so it is done unconditionally.

2020-07-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

gcc/ChangeLog:

PR target/94891
PR target/94791
* config/aarch64/aarch64-protos.h (aarch64_return_addr_rtx): Declare.
* config/aarch64/aarch64.c (aarch64_return_addr_rtx): New.
(aarch64_return_addr): Use aarch64_return_addr_rtx.
* config/aarch64/aarch64.h (PROFILE_HOOK): Likewise.

4 years agoipa-devirt: Fix crash in obj_type_ref_class [PR95114]
Richard Sandiford [Mon, 13 Jul 2020 12:47:39 +0000 (13:47 +0100)]
ipa-devirt: Fix crash in obj_type_ref_class [PR95114]

The testcase has failed since r9-5035, because obj_type_ref_class
tries to look up an ODR type when no ODR type information is
available.  (The information was available earlier in the
compilation, but was freed during pass_ipa_free_lang_data.)
We then crash dereferencing the null get_odr_type result.

The test passes with -O2.  However, it fails again if -fdump-tree-all
is used, since obj_type_ref_class is called indirectly from the
dump routines.

Other code creates ODR type entries on the fly by passing “true”
as the insert parameter.  But obj_type_ref_class can't do that
unconditionally, since it should have no side-effects when used
from the dumping code.

Following a suggestion from Honza, this patch adds parameters
to say whether the routines are being called from dump routines
and uses those to derive the insert parameter.

gcc/
PR middle-end/95114
* tree.h (virtual_method_call_p): Add a default-false parameter
that indicates whether the function is being called from dump
routines.
(obj_type_ref_class): Likewise.
* tree.c (virtual_method_call_p): Likewise.
* ipa-devirt.c (obj_type_ref_class): Likewise.  Lazily add ODR
type information for the type when the parameter is false.
* tree-pretty-print.c (dump_generic_node): Update calls to
virtual_method_call_p and obj_type_ref_class accordingly.

gcc/testsuite/
PR middle-end/95114
* g++.target/aarch64/pr95114.C: New test.

4 years agolibstdc++: Fix istream::ignore exit conditions (PR 94749, PR 96161)
Jonathan Wakely [Mon, 13 Jul 2020 09:26:39 +0000 (10:26 +0100)]
libstdc++: Fix istream::ignore exit conditions (PR 94749, PR 96161)

My previous fix for PR 94749 did fix the reported case, so that the next
character is not discarded if it happens to equal the delimiter when __n
characters have already been read. But it introduced a new bug, which is
that the delimiter character would *not* be discarded if the number of
characters discarded is numeric_limits<streamsize>::max() or more before
reaching the delimiter.

The new bug happens because I changed the code to check _M_gcount < __n.
But when __n == numeric_limits<streamsize>::max() that is false, and so
we don't discard the delimiter. It's not sufficient to check for the
delimiter when the __large_ignore condition is true, because there's an
edge case where the delimiter is reached when _M_gcount == __n and so
we break out of the loop without setting __large_ignore.

PR 96161 is a similar bug to the original PR 94749 report, where eofbit
is set after discarding __n characters if there happen to be no more
characters in the stream.

This patch fixes both cases (and the regression) by checking different
conditions for the __n == max case and the __n < max case. For the
former case, we know that we must have either reached the delimiter or
EOF, and the value of _M_gcount doesn't matter (except to avoid integer
overflow). For the latter case we need to check _M_gcount first and only
set eofbit or discard the delimiter if it didn't reach __n. For the
latter case overflow can't happen because _M_gcount <= __n < max.

libstdc++-v3/ChangeLog:

PR libstdc++/94749
PR libstdc++/96161
* include/bits/istream.tcc (basic_istream::ignore(streamsize))
[n == max]: Check overflow conditions on _M_gcount. Rely on
the fact that either EOF or the delimiter was reached.
[n < max]: Check _M_gcount < n before checking for EOF or
delimiter.
(basic_istream::ignore(streamsize, char_type): Likewise.
* src/c++98/compatibility.cc (istream::ignore(streamsize))
(wistream::ignore(streamsize)): Likewise.
* src/c++98/istream.cc (istream::ignore(streamsize, char_type))
(wistream::ignore(streamsize, char_type)): Likewise.
* testsuite/27_io/basic_istream/ignore/char/94749.cc: Check that
delimiter is discarded if the number of characters ignored
doesn't fit in streamsize.
* testsuite/27_io/basic_istream/ignore/wchar_t/94749.cc:
Likewise.
* testsuite/27_io/basic_istream/ignore/char/96161.cc: New test.
* testsuite/27_io/basic_istream/ignore/wchar_t/96161.cc: New test.

4 years agoopenacc: Don't strip TO_PSET/POINTER for enter/exit data
Julian Brown [Fri, 5 Jun 2020 21:46:41 +0000 (14:46 -0700)]
openacc: Don't strip TO_PSET/POINTER for enter/exit data

OpenACC 2.6 specifies that the array descriptor (when present) must be
copied to the target before attaching pointers in Fortran. This patch
reverses the stripping of GOMP_MAP_TO_PSET and GOMP_MAP_POINTER that
was introduced by the "OpenACC reference count overhaul" patch.

2020-07-10  Julian Brown  <julian@codesourcery.com>
    Thomas Schwinge  <thomas@codesourcery.com>

gcc/
* gimplify.c (gimplify_scan_omp_clauses): Do not strip
GOMP_MAP_TO_PSET/GOMP_MAP_POINTER for OpenACC enter/exit data
directives (see also PR92929).

gcc/testsuite/
* gfortran.dg/goacc/finalize-1.f: Update expected dump output.

libgomp/
* testsuite/libgomp.oacc-fortran/dynamic-pointer-1.f90: New test.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
4 years agofix global variable alignment for testcase gcc.dg/torture/pr96133.c
Richard Biener [Mon, 13 Jul 2020 09:41:16 +0000 (11:41 +0200)]
fix global variable alignment for testcase gcc.dg/torture/pr96133.c

The testcase was errorneously accessing the global variable via a
type that might require bigger alignment than provided.  Fix that
via an appropriate attribute.

2020-07-13  Richard Biener  <rguenther@suse.de>

PR testsuite/96180
* gcc.dg/torture/pr96133.c: Align global variable.

4 years agomiddle-end: Remove truly_noop_truncation check from convert.c
Roger Sayle [Mon, 13 Jul 2020 08:49:34 +0000 (09:49 +0100)]
middle-end: Remove truly_noop_truncation check from convert.c

This patch eliminates a check of targetm.truly_noop_truncation from
the early middle-end, where the gimple/generic being generated by
GCC's front-ends is being inappropriately influenced by the target's
TRULY_NOOP_TRUNCATION.  The (recent) intention of TRULY_NOOP_TRUNCATION
is to indicate that a backend requires explicit truncation instructions
rather than using SUBREGs to perform truncations.  A long standing
(and probably unintentional) side-effect has been that this setting
also controls whether the middle-end narrows integer operations at
the tree-level.  Understandably, GCC and its testsuite assume that
GIMPLE and GENERIC behave consistently across platforms, and alas
defining TRULY_NOOP_TRUNCATION away from the default triggers several
regressions (including gcc.dg/fold-rotate-1.c).

2020-07-13  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* convert.c (convert_to_integer_1): Narrow integer operations
even on targets that require explicit truncation instructions.

4 years agocris: Add new pass eliminating compares after delay-slot-filling
Hans-Peter Nilsson [Sun, 12 Jul 2020 16:41:25 +0000 (18:41 +0200)]
cris: Add new pass eliminating compares after delay-slot-filling

Delayed-branch-slot-filling a.k.a. reorg or dbr, often causes
opportunities for more compare-elimination than were visible for
the cmpelim pass.  With cc0, these were caught by the
elimination pass run in "final", thus the missed opportunities
is a regression.  A simple reorg-aware pass run just after reorg
handles most of them, if not all.  I chose to keep the "mach2"
pass identifier string I copy-pasted from the SPARC port instead
of inventing one like "postdbr_cmpelim".  Note the gap in numbers
in the test-case file names.

gcc:
PR target/93372
* config/cris/cris-passes.def: New file.
* config/cris/t-cris (PASSES_EXTRA): Add cris-passes.def.
* config/cris/cris.c: Add infrastructure bits and pass execute
function cris_postdbr_cmpelim.
* config/cris/cris-protos.h (make_pass_cris_postdbr_cmpelim): Declare.

gcc/testsuite:
* gcc.target/cris/pr93372-44.c, gcc.target/cris/pr93372-46.c: New.

4 years agocris: Remove config/cris/t-cris gt-cris.h cargo
Hans-Peter Nilsson [Fri, 10 Jul 2020 02:44:49 +0000 (04:44 +0200)]
cris: Remove config/cris/t-cris gt-cris.h cargo

Getting tired of:

make[1]: Entering directory 'x/gccobj/gcc'
Makefile:2682: warning: overriding recipe for target 'gt-cris.h'
xx/gcc/gcc/config/cris/t-cris:29: warning: ignoring old recipe for target 'gt-cris.h'

I'm just going to assume it is just stale cruft no longer (if
ever) needed since nothing else but sh/t-sh has it, and the
commit log shows just (x prepended to avoid commit-log parsing
confusion):
x    Merge from pch-branch up to tag pch-commit-20020603.
x
x    From-SVN: r54232

Building "works better"; the related warning is gone.

This effectively empties the t-cris file, but stuff will be
added soon enough that it's kept around.

gcc:
* config/cris/t-cris: Remove gt-cris.h-related excessive cargo.

4 years agocris: Use addi.b for additions where flags aren't inspected
Hans-Peter Nilsson [Wed, 8 Jul 2020 21:59:12 +0000 (23:59 +0200)]
cris: Use addi.b for additions where flags aren't inspected

Comparing to the cc0 version of the CRIS port, I ran a few
microbenchmarks, for example gcc.c-torture/execute/arith-rand.c,
where there's sometimes an addition between an operation of
interest and the test on the result.

Unfortunately this patch doesn't remedy all the performance
regression for that program.  But, this patch by itself helps
and makes sense to commit separately: lots of addi.b in
previously empty delay-slots, with functions shortened by one or
a few insns, in libgcc.  I had an experience with the
reload-related caveat of % on constraints, which is "fixed"
documentationwise since long (soon 15 years ago;
be3914df4cc8/r105517).  I removed an even older related FIXME.

gcc:
PR target/93372
* config/cris/cris.md ("*add<mode>3_addi"): New splitter.
("*addi_b_<mode>"): New pattern.
("*addsi3<setnz>"): Remove stale %-related comment.

gcc/testsuite:
PR target/93372
* gcc.target/cris/pr93372-45.c: New test.

4 years agocris: Correct output templates in define_subst patterns.
Hans-Peter Nilsson [Sun, 12 Jul 2020 19:58:58 +0000 (21:58 +0200)]
cris: Correct output templates in define_subst patterns.

Whoops.  This little gem had the effect of making the output
operand (0) constraints disappear but not the input operand (1)
constraints for define_subst:ed patterns, probably because
there's another (match_dup 1) in the output template (not
investigated).

That went surprisingly unnoticed until I added a pass leaning
just a little bit harder on the define_subst:ed patterns and
then only by the libgfortran library generating assembly with
nominally incorrect syntax.  (There was a move to a special
register from a general register, and it incorrectly matched a
pattern affecting condition codes.)

gcc:
* config/cris/cris.md ("setnz_subst", "setnz_subst", "setcc_subst"):
Use match_dup in output template, not match_operand.

4 years agomake var-tracking iteration consistent
Richard Biener [Fri, 10 Jul 2020 09:58:41 +0000 (11:58 +0200)]
make var-tracking iteration consistent

This eliminates the visited bitmap and makes whether a to be processed
block goes to the next or the current iteration only depend on its
position in RPO order rather than on whether it was visited in the
current iteration.  As optimization single-BB iteration is processed
immediately.

2020-07-10  Richard Biener  <rguenther@suse.de>

* var-tracking.c (bb_heap_node_t): Remove unused typedef.
(vt_find_locations): Eliminate visited bitmap in favor of
RPO order check.  Dump statistics about the number of
local BB dataflow computes.

4 years agoPR94600: fix volatile access to the whole of a compound object.
Hans-Peter Nilsson [Sun, 5 Jul 2020 18:50:52 +0000 (20:50 +0200)]
PR94600: fix volatile access to the whole of a compound object.

The store to the whole of each volatile object was picked apart
like there had been an individual assignment to each of the
fields.  Reads were added as part of that; see PR for details.
The reads from volatile memory were a clear bug; individual
stores questionable.  A separate patch clarifies the docs.

gcc:

2020-07-09  Richard Biener  <rguenther@suse.de>

PR middle-end/94600
* expr.c (expand_constructor): Make a temporary also if we're
storing to volatile memory.

gcc/testsuite:

2020-07-09  Hans-Peter Nilsson  <hp@axis.com>

PR middle-end/94600
* gcc.dg/pr94600-1.c, gcc.dg/pr94600-2.c, gcc.dg/pr94600-3.c,
gcc.dg/pr94600-4.c, gcc.dg/pr94600-5.c, gcc.dg/pr94600-6.c,
gcc.dg/pr94600-7.c, gcc.dg/pr94600-8.c: New tests.

4 years agors6000: Define define_insn_and_split to split unspec sldi+or to rldimi
Xionghu Luo [Mon, 13 Jul 2020 01:22:56 +0000 (20:22 -0500)]
rs6000: Define define_insn_and_split to split unspec sldi+or to rldimi

Combine pass could recognize the pattern defined and split it in split1,
this patch could optimize:

21: r130:DI=r133:DI<<0x20
11: {r129:DI=zero_extend(unspec[[r145:DI]] 87);clobber scratch;}
22: r134:DI=r130:DI|r129:DI

to

21: {r149:DI=zero_extend(unspec[[r145:DI]] 87);clobber scratch;}
22: r134:DI=r149:DI&0xffffffff|r133:DI<<0x20

rldimi is generated instead of sldi+or.

gcc/ChangeLog:

2020-07-13  Xionghu Luo  <luoxhu@linux.ibm.com>

* config/rs6000/rs6000.md (rotl_unspec): New
define_insn_and_split.

gcc/testsuite/ChangeLog:

2020-07-13  Xionghu Luo  <luoxhu@linux.ibm.com>

* gcc.target/powerpc/vector_float.c: New test.

4 years agors6000: Init V4SF vector without converting SP to DP
Xionghu Luo [Mon, 13 Jul 2020 01:21:05 +0000 (20:21 -0500)]
rs6000: Init V4SF vector without converting SP to DP

Move V4SF to V4SI, init vector like V4SI and move to V4SF back.
Better instruction sequence could be generated on Power9:

lfs + xxpermdi + xvcvdpsp + vmrgew
=>
lwz + (sldi + or) + mtvsrdd

With the patch followed, it could be continue optimized to:

lwz + rldimi + mtvsrdd

The point is to use lwz to avoid converting the single-precision to
double-precision upon load, pack four 32-bit data into one 128-bit
register directly.

gcc/ChangeLog:

2020-07-13  Xionghu Luo  <luoxhu@linux.ibm.com>

* config/rs6000/rs6000.c (rs6000_expand_vector_init):
Move V4SF to V4SI, init vector like V4SI and move to V4SF back.