platform/upstream/gcc.git
23 months agoarm: Define with_float to hard when target name ends with hf
Christophe Lyon [Wed, 17 Aug 2022 08:06:24 +0000 (10:06 +0200)]
arm: Define with_float to hard when target name ends with hf

On arm, the --with-float= configure option is used to define include
files search path (among other things).  However, when targeting
arm-linux-gnueabihf, one would expect to automatically default to the
hard-float ABI, but this is not the case. As a consequence, GCC
bootstrap fails on an arm-linux-gnueabihf target if --with-float=hard
is not used.

This patch checks if the target name ends with 'hf' and defines
with_float to hard if not already defined.  This is achieved in
gcc/config.gcc, just before selecting the default CPU depending on the
$with_float value.

2022-08-17  Christophe Lyon  <christophe.lyon@arm.com>

gcc/
* config.gcc (arm): Define with_float to hard if target name ends
with 'hf'.

23 months agoRefactor back_threader_profitability
Richard Biener [Mon, 15 Aug 2022 13:38:09 +0000 (15:38 +0200)]
Refactor back_threader_profitability

The following refactors profitable_path_p in the backward threader,
splitting out parts that can be computed once the exit block is known,
parts that contiguously update and that can be checked allowing
for the path to be later identified as FSM with larger limits,
possibly_profitable_path_p, and final checks done when the whole
path is known, profitable_path_p.

I've removed the back_threader_profitability instance from the
back_threader class and instead instantiate it once per path
discovery.  I've kept the size compute non-incremental to simplify
the patch and not worry about unwinding.

There's key changes to previous behavior - namely we apply
the param_max_jump_thread_duplication_stmts early only when
we know the path cannot become an FSM one (multiway + thread through
latch) but make sure to elide the path query when we we didn't
yet discover that but are over this limit.  Similarly the
speed limit is now used even when we did not yet discover a
hot BB on the path.  Basically the idea is to only stop path
discovery when we know the path will never become profitable
but avoid the expensive path range query when we know it's
currently not.

I've done a few cleanups, merging functions, on the way.

* tree-ssa-threadbackward.cc
(back_threader_profitability): Split profitable_path_p
into possibly_profitable_path_p and itself, keep state
as new members.
(back_threader::m_profit): Remove.
(back_threader::find_paths): Likewise.
(back_threader::maybe_register_path): Take profitability
instance as parameter.
(back_threader::find_paths_to_names): Likewise.  Use
possibly_profitable_path_p and avoid the path range query
when the path is currently too large.
(back_threader::find_paths): Fold into ...
(back_threader::maybe_thread_block): ... this.
(get_gimple_control_stmt): Remove.
(back_threader_profitability::possibly_profitable_path_p):
Split out from profitable_path_p, do early profitability
checks.
(back_threader_profitability::profitable_path_p): Do final
profitability path after the taken edge has been determined.

23 months agoFix bug in emergency cxa pool free
Keef Aragon [Wed, 17 Aug 2022 06:45:15 +0000 (08:45 +0200)]
Fix bug in emergency cxa pool free

This probably has never actually affected anyone in practice. The normal
ABI implementation just uses malloc and only falls back to the pool on
malloc failure. But if that happens a bunch of times the freelist gets out
of order which violates some of the invariants of the freelist (as well as
the comments that follow the bug). The bug is just a comparison reversal
when traversing the freelist in the case where the pointer being returned
to the pool is after the existing freelist.

libstdc++-v3/
* libsupc++/eh_alloc.cc (pool::free): Inverse comparison.

23 months agoLoongArch: Provide fmin/fmax RTL pattern
Xi Ruoyao [Tue, 16 Aug 2022 07:34:36 +0000 (15:34 +0800)]
LoongArch: Provide fmin/fmax RTL pattern

We already had smin/smax RTL pattern using fmin/fmax instruction.  But
for smin/smax, it's unspecified what will happen if either operand is
NaN.  So we would generate calls to libc fmin/fmax functions with
-fno-finite-math-only (the default for all optimization levels expect
-Ofast).

But, LoongArch fmin/fmax instruction is IEEE-754-2008 conformant so we
can also use the instruction for fmin/fmax pattern and avoid the library
function call.

gcc/ChangeLog:

* config/loongarch/loongarch.md (fmax<mode>3): New RTL pattern.
(fmin<mode>3): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/loongarch/fmax-fmin.c: New test.

23 months agoAbstract interesting ssa-names from GORI.
Andrew MacLeod [Tue, 16 Aug 2022 17:18:37 +0000 (13:18 -0400)]
Abstract interesting ssa-names from GORI.

Provide a routine to pick out the ssa-names from interesting statements.

* gimple-range-fold.cc (gimple_range_ssa_names): New.
* gimple-range-fold.h (gimple_range_ssa_names): New prototype.
* gimple-range-gori.cc (range_def_chain::get_def_chain): Move
  code to new routine.

23 months agoDaily bump.
GCC Administrator [Wed, 17 Aug 2022 00:17:17 +0000 (00:17 +0000)]
Daily bump.

23 months agoc++: remove some xfails
Jason Merrill [Tue, 16 Aug 2022 19:19:46 +0000 (15:19 -0400)]
c++: remove some xfails

These tests are now passing.

gcc/testsuite/ChangeLog:

* g++.dg/warn/Wstringop-overflow-4.C: Only xfail for C++98.
* g++.target/i386/bfloat_cpp_typecheck.C: Remove xfail.

23 months agoc++: Fix pragma suppression of -Wc++20-compat diagnostics [PR106423]
Tom Honermann [Mon, 1 Aug 2022 18:49:00 +0000 (14:49 -0400)]
c++: Fix pragma suppression of -Wc++20-compat diagnostics [PR106423]

Gcc's '#pragma GCC diagnostic' directives are processed in "early mode"
(see handle_pragma_diagnostic_early) for the C++ frontend and, as such,
require that the target diagnostic option be enabled for the preprocessor
(see c_option_is_from_cpp_diagnostics).  This change modifies the
-Wc++20-compat option definition to register it as a preprocessor option
so that its associated diagnostics can be suppressed.  The changes also
implicitly disable the option in C++20 and later modes.  These changes
are consistent with the definition of the -Wc++11-compat option.

This support is motivated by the need to suppress the following diagnostic
otherwise issued in C++17 and earlier modes due to the char8_t typedef
present in the uchar.h header file in glibc 2.36.
  warning: identifier ‘char8_t’ is a keyword in C++20 [-Wc++20-compat]

Tests are added to validate suppression of both -Wc++11-compat and
-Wc++20-compat related diagnostics (fixes were only needed for the C++20
case).

PR c++/106423

gcc/c-family/ChangeLog:
* c-opts.cc (c_common_post_options): Disable -Wc++20-compat
diagnostics in C++20 and later.
* c.opt (Wc++20-compat): Enable hooks for the preprocessor.

gcc/cp/ChangeLog:
* parser.cc (cp_lexer_saving_tokens): Add comment regarding
diagnostic requirements.

gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/keywords2.C: New test.
* g++.dg/cpp2a/keywords2.C: New test.

libcpp/ChangeLog:
* include/cpplib.h (cpp_warning_reason): Add CPP_W_CXX20_COMPAT.
* init.cc (cpp_create_reader): Add cpp_warn_cxx20_compat.

23 months agodocs: remove link to www.bullfreeware.com from install
Martin Liska [Tue, 16 Aug 2022 16:15:55 +0000 (18:15 +0200)]
docs: remove link to bullfreeware.com from install

As mentioned at https://gcc.gnu.org/PR106637#c2, the discontinued
providing binaries.

PR target/106637

gcc/ChangeLog:

* doc/install.texi: Remove link to www.bullfreeware.com

23 months agoRISC-V: Support zfh and zfhmin extension
Kito Cheng [Mon, 4 Jul 2022 15:09:57 +0000 (23:09 +0800)]
RISC-V: Support zfh and zfhmin extension

Zfh and Zfhmin are extensions for IEEE half precision, both are ratified
in Jan. 2022[1]:

- Zfh has full set of operation like F or D for single or double precision.
- Zfhmin has only provide minimal support for half precision operation,
  like conversion, load, store and move instructions.

[1] https://github.com/riscv/riscv-isa-manual/commit/b35a54079e0da11740ce5b1e6db999d1d5172768

gcc/ChangeLog:

* common/config/riscv/riscv-common.cc (riscv_implied_info): Add
zfh and zfhmin.
(riscv_ext_version_table): Ditto.
(riscv_ext_flag_table): Ditto.
* config/riscv/riscv-opts.h (MASK_ZFHMIN): New.
(MASK_ZFH): Ditto.
(TARGET_ZFHMIN): Ditto.
(TARGET_ZFH): Ditto.
* config/riscv/riscv.cc (riscv_output_move): Handle HFmode move
for zfh and zfhmin.
(riscv_emit_float_compare): Handle HFmode.
* config/riscv/riscv.md (ANYF): Add HF.
(SOFTF): Add HF.
(load): Ditto.
(store): Ditto.
(truncsfhf2): New.
(truncdfhf2): Ditto.
(extendhfsf2): Ditto.
(extendhfdf2): Ditto.
(*movhf_hardfloat): Ditto.
(*movhf_softfloat): Make sure not ZFHMIN.
* config/riscv/riscv.opt (riscv_zf_subext): New.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/_Float16-zfh-1.c: New.
* gcc.target/riscv/_Float16-zfh-2.c: Ditto.
* gcc.target/riscv/_Float16-zfh-3.c: Ditto.
* gcc.target/riscv/_Float16-zfhmin-1.c: Ditto.
* gcc.target/riscv/_Float16-zfhmin-2.c: Ditto.
* gcc.target/riscv/_Float16-zfhmin-3.c: Ditto.
* gcc.target/riscv/arch-16.c: Ditto.
* gcc.target/riscv/arch-17.c: Ditto.
* gcc.target/riscv/predef-21.c: Ditto.
* gcc.target/riscv/predef-22.c: Ditto.

23 months agoRISC-V: Support _Float16 type.
Kito Cheng [Mon, 4 Jul 2022 12:17:47 +0000 (20:17 +0800)]
RISC-V: Support _Float16 type.

RISC-V decide use _Float16 as primary IEEE half precision type, and this
already become part of psABI, this patch has added folloing support for
_Float16:

- Soft-float support for _Float16.
- Make sure _Float16 available on C++ mode.
- Name mangling for _Float16 on C++ mode.

gcc/ChangeLog

* config/riscv/riscv-builtins.cc: include stringpool.h
(riscv_float16_type_node): New.
(riscv_init_builtin_types): Ditto.
(riscv_init_builtins): Call riscv_init_builtin_types.
* config/riscv/riscv-modes.def (HF): New.
* config/riscv/riscv.cc (riscv_output_move): Handle HFmode.
(riscv_mangle_type): New.
(riscv_scalar_mode_supported_p): Ditto.
(riscv_libgcc_floating_mode_supported_p): Ditto.
(riscv_excess_precision): Ditto.
(riscv_floatn_mode): Ditto.
(riscv_init_libfuncs): Ditto.
(TARGET_MANGLE_TYPE): Ditto.
(TARGET_SCALAR_MODE_SUPPORTED_P): Ditto.
(TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P): Ditto.
(TARGET_INIT_LIBFUNCS): Ditto.
(TARGET_C_EXCESS_PRECISION): Ditto.
(TARGET_FLOATN_MODE): Ditto.
* config/riscv/riscv.md (mode): Add HF.
(softload): Add HF.
(softstore): Ditto.
(fmt): Ditto.
(UNITMODE): Ditto.
(movhf): New.
(*movhf_softfloat): New.

libgcc/ChangeLog:

* config/riscv/sfp-machine.h (_FP_NANFRAC_H): New.
(_FP_NANFRAC_H): Ditto.
(_FP_NANSIGN_H): Ditto.
* config/riscv/t-softfp32 (softfp_extensions): Add HF related
routines.
(softfp_truncations): Ditto.
(softfp_extras): Ditto.
* config/riscv/t-softfp64 (softfp_extras): Add HF related routines.

gcc/testsuite/ChangeLog:

* g++.target/riscv/_Float16.C: New.
* gcc.target/riscv/_Float16-soft-1.c: Ditto.
* gcc.target/riscv/_Float16-soft-2.c: Ditto.
* gcc.target/riscv/_Float16-soft-3.c: Ditto.
* gcc.target/riscv/_Float16-soft-4.c: Ditto.
* gcc.target/riscv/_Float16.c: Ditto.

23 months agosoft-fp: Update soft-fp from glibc
Kito Cheng [Wed, 10 Aug 2022 13:54:24 +0000 (21:54 +0800)]
soft-fp: Update soft-fp from glibc

This patch is updating all soft-fp from glibc, most changes are
copyright years update, removing "Contributed by" lines and update URL for
license, and changes other than those update are adding conversion
function between IEEE half and 32-bit/64-bit integer, those functions are
required by RISC-V _Float16 support.

libgcc/ChangeLog:

* soft-fp/fixhfdi.c: New.
* soft-fp/fixhfsi.c: Likewise.
* soft-fp/fixunshfdi.c: Likewise.
* soft-fp/fixunshfsi.c: Likewise.
* soft-fp/floatdihf.c: Likewise.
* soft-fp/floatsihf.c: Likewise.
* soft-fp/floatundihf.c: Likewise.
* soft-fp/floatunsihf.c: Likewise.
* soft-fp/adddf3.c: Updating copyright years, removing "Contributed by"
lines and update URL for license.
* soft-fp/addsf3.c: Likewise.
* soft-fp/addtf3.c: Likewise.
* soft-fp/divdf3.c: Likewise.
* soft-fp/divsf3.c: Likewise.
* soft-fp/divtf3.c: Likewise.
* soft-fp/double.h: Likewise.
* soft-fp/eqdf2.c: Likewise.
* soft-fp/eqhf2.c: Likewise.
* soft-fp/eqsf2.c: Likewise.
* soft-fp/eqtf2.c: Likewise.
* soft-fp/extenddftf2.c: Likewise.
* soft-fp/extended.h: Likewise.
* soft-fp/extendhfdf2.c: Likewise.
* soft-fp/extendhfsf2.c: Likewise.
* soft-fp/extendhftf2.c: Likewise.
* soft-fp/extendhfxf2.c: Likewise.
* soft-fp/extendsfdf2.c: Likewise.
* soft-fp/extendsftf2.c: Likewise.
* soft-fp/extendxftf2.c: Likewise.
* soft-fp/fixdfdi.c: Likewise.
* soft-fp/fixdfsi.c: Likewise.
* soft-fp/fixdfti.c: Likewise.
* soft-fp/fixhfti.c: Likewise.
* soft-fp/fixsfdi.c: Likewise.
* soft-fp/fixsfsi.c: Likewise.
* soft-fp/fixsfti.c: Likewise.
* soft-fp/fixtfdi.c: Likewise.
* soft-fp/fixtfsi.c: Likewise.
* soft-fp/fixtfti.c: Likewise.
* soft-fp/fixunsdfdi.c: Likewise.
* soft-fp/fixunsdfsi.c: Likewise.
* soft-fp/fixunsdfti.c: Likewise.
* soft-fp/fixunshfti.c: Likewise.
* soft-fp/fixunssfdi.c: Likewise.
* soft-fp/fixunssfsi.c: Likewise.
* soft-fp/fixunssfti.c: Likewise.
* soft-fp/fixunstfdi.c: Likewise.
* soft-fp/fixunstfsi.c: Likewise.
* soft-fp/fixunstfti.c: Likewise.
* soft-fp/floatdidf.c: Likewise.
* soft-fp/floatdisf.c: Likewise.
* soft-fp/floatditf.c: Likewise.
* soft-fp/floatsidf.c: Likewise.
* soft-fp/floatsisf.c: Likewise.
* soft-fp/floatsitf.c: Likewise.
* soft-fp/floattidf.c: Likewise.
* soft-fp/floattihf.c: Likewise.
* soft-fp/floattisf.c: Likewise.
* soft-fp/floattitf.c: Likewise.
* soft-fp/floatundidf.c: Likewise.
* soft-fp/floatundisf.c: Likewise.
* soft-fp/floatunditf.c: Likewise.
* soft-fp/floatunsidf.c: Likewise.
* soft-fp/floatunsisf.c: Likewise.
* soft-fp/floatunsitf.c: Likewise.
* soft-fp/floatuntidf.c: Likewise.
* soft-fp/floatuntihf.c: Likewise.
* soft-fp/floatuntisf.c: Likewise.
* soft-fp/floatuntitf.c: Likewise.
* soft-fp/gedf2.c: Likewise.
* soft-fp/gesf2.c: Likewise.
* soft-fp/getf2.c: Likewise.
* soft-fp/half.h: Likewise.
* soft-fp/ledf2.c: Likewise.
* soft-fp/lesf2.c: Likewise.
* soft-fp/letf2.c: Likewise.
* soft-fp/muldf3.c: Likewise.
* soft-fp/mulsf3.c: Likewise.
* soft-fp/multf3.c: Likewise.
* soft-fp/negdf2.c: Likewise.
* soft-fp/negsf2.c: Likewise.
* soft-fp/negtf2.c: Likewise.
* soft-fp/op-1.h: Likewise.
* soft-fp/op-2.h: Likewise.
* soft-fp/op-4.h: Likewise.
* soft-fp/op-8.h: Likewise.
* soft-fp/op-common.h: Likewise.
* soft-fp/quad.h: Likewise.
* soft-fp/single.h: Likewise.
* soft-fp/soft-fp.h: Likewise.
* soft-fp/subdf3.c: Likewise.
* soft-fp/subsf3.c: Likewise.
* soft-fp/subtf3.c: Likewise.
* soft-fp/truncdfhf2.c: Likewise.
* soft-fp/truncdfsf2.c: Likewise.
* soft-fp/truncsfhf2.c: Likewise.
* soft-fp/trunctfdf2.c: Likewise.
* soft-fp/trunctfhf2.c: Likewise.
* soft-fp/trunctfsf2.c: Likewise.
* soft-fp/trunctfxf2.c: Likewise.
* soft-fp/truncxfhf2.c: Likewise.
* soft-fp/unorddf2.c: Likewise.
* soft-fp/unordsf2.c: Likewise.
* soft-fp/unordtf2.c: Likewise.

23 months agoStop backwards thread discovery when leaving a loop
Richard Biener [Tue, 16 Aug 2022 12:14:55 +0000 (14:14 +0200)]
Stop backwards thread discovery when leaving a loop

The backward threader copier cannot deal with the situation of
copying blocks belonging to different loops and will reject those
paths late.  The following uses this to prune path discovery,
saving on compile-time.  Note the off-loop block is still considered
as entry edge origin.

* tree-ssa-threadbackward.cc (back_threader::find_paths_to_names):
Do not walk further if we are leaving the current loop.

23 months agodriver: fix environ corruption after putenv() [PR106624]
Sergei Trofimovich [Tue, 16 Aug 2022 11:35:07 +0000 (12:35 +0100)]
driver: fix environ corruption after putenv() [PR106624]

The bug appeared afte r13-2010-g1270ccda70ca09 "Factor out
jobserver_active_p" slightly changed `putenv()` use from allocating
to non-allocating:

    -xputenv (concat ("MAKEFLAGS=", dup, NULL));
    +xputenv (jinfo.skipped_makeflags.c_str ());

`xputenv()` (and `putenv()`) don't copy strings and only store the
pointer in the `environ` global table. As a result `environ` got
corrupted as soon as `jinfo.skipped_makeflags` store got deallocated.

This started causing bootstrap crashes in `execv()` calls:

    xgcc: fatal error: cannot execute '/build/build/./prev-gcc/collect2': execv: Bad address

The change restores memory allocation for `xputenv()` argument.

gcc/

PR driver/106624
* gcc.cc (driver::detect_jobserver): Allocate storage xputenv()
argument using xstrdup().

23 months agoc++: Implement P2327R1 - De-deprecating volatile compound operations
Jakub Jelinek [Tue, 16 Aug 2022 11:15:32 +0000 (13:15 +0200)]
c++: Implement P2327R1 - De-deprecating volatile compound operations

From what I can see, this has been voted in as a DR and as it means
we warn less often than before in -std={gnu,c}++2{0,3} modes or with
-Wvolatile, I wonder if it shouldn't be backported to affected release
branches as well.

2022-08-16  Jakub Jelinek  <jakub@redhat.com>

* typeck.cc (cp_build_modify_expr): Implement
P2327R1 - De-deprecating volatile compound operations.  Don't warn
for |=, &= or ^= with volatile lhs.
* expr.cc (mark_use) <case MODIFY_EXPR>: Adjust warning wording,
leave out simple.

* g++.dg/cpp2a/volatile1.C: Adjust for de-deprecation of volatile
compound |=, &= and ^= operations.
* g++.dg/cpp2a/volatile3.C: Likewise.
* g++.dg/cpp2a/volatile5.C: Likewise.

23 months agod: Update DIP links in gdc documentation to point at upstream repository
Iain Buclaw [Tue, 16 Aug 2022 10:22:10 +0000 (12:22 +0200)]
d: Update DIP links in gdc documentation to point at upstream repository

The wiki links probably worked at some point in the distant past, but
now the official location of tracking all D Improvement Proposals is on
the upstream dlang/DIPs GitHub repository.

PR d/106638

gcc/d/ChangeLog:

* gdc.texi: Update DIP links to point at upstream dlang/DIPs
repository.

23 months agoRename imports nomenclature in path_range_query to exit_dependencies.
Aldy Hernandez [Tue, 16 Aug 2022 08:52:37 +0000 (10:52 +0200)]
Rename imports nomenclature in path_range_query to exit_dependencies.

The purpose of this change is to disambiguate the imports name with
its use in GORI.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::import_p): Rename to...
(path_range_query::exit_dependency_p): ...this.
(path_range_query::dump): Rename imports to exit dependencies.
(path_range_query::compute_ranges_in_phis): Same.
(path_range_query::compute_ranges_in_block): Same.
(path_range_query::adjust_for_non_null_uses): Same.
(path_range_query::compute_ranges): Same.
(path_range_query::compute_phi_relations): Same.
(path_range_query::add_to_imports): Rename to...
(path_range_query::add_to_exit_dependencies): ...this.
(path_range_query::compute_imports): Rename to...
(path_range_query::compute_exit_dependencies): ...this.
* gimple-range-path.h (class path_range_query): Rename imports to
exit dependencies.

23 months agoVR: mitigate -Wfinal-dtor-non-final-class clang warnings
Martin Liska [Tue, 16 Aug 2022 09:23:43 +0000 (11:23 +0200)]
VR: mitigate -Wfinal-dtor-non-final-class clang warnings

Fixes:

gcc/value-range-storage.h:129:40: warning: class with destructor marked 'final' cannot be inherited from [-Wfinal-dtor-non-final-class]
gcc/value-range-storage.h:146:36: warning: class with destructor marked 'final' cannot be inherited from [-Wfinal-dtor-non-final-class]

gcc/ChangeLog:

* value-range-storage.h (class obstack_vrange_allocator): Mark
  the class as final.
(class ggc_vrange_allocator): Likewise.

23 months agoVR: add more virtual dtors
Martin Liska [Tue, 16 Aug 2022 09:15:30 +0000 (11:15 +0200)]
VR: add more virtual dtors

Add 2 virtual destructors in order to address:

gcc/alloc-pool.h:522:5: warning: destructor called on non-final 'value_range_equiv' that has virtual functions but non-virtual destructor [-Wdelete-non-abstract-non-virtual-dtor]
gcc/ggc.h:166:3: warning: destructor called on non-final 'int_range<1>' that has virtual functions but non-virtual destructor [-Wdelete-non-abstract-non-virtual-dtor]

gcc/ChangeLog:

* value-range-equiv.h (class value_range_equiv): Add virtual
  destructor.
* value-range.h: Likewise.

23 months agomiddle-end/106630 - avoid ping-pong between extract_muldiv and match.pd
Richard Biener [Tue, 16 Aug 2022 07:43:24 +0000 (09:43 +0200)]
middle-end/106630 - avoid ping-pong between extract_muldiv and match.pd

The following avoids ping-pong between the match.pd pattern changing
(sizetype) ((a_9 + 1) * 48) to (sizetype)(a_9 + 1) * 48 and
extract_muldiv performing the reverse transform by restricting the
match.pd pattern to narrowing conversions as the comment indicates.

PR middle-end/106630
* match.pd ((T)(x * CST) -> (T)x * CST): Restrict to
narrowing conversions.

* gcc.dg/torture/pr106630.c: New testcase.

23 months agoVR: add missing override keyworks
Martin Liska [Tue, 16 Aug 2022 09:06:04 +0000 (11:06 +0200)]
VR: add missing override keyworks

Address:

gcc/value-range-equiv.h:57:8: warning: 'set_undefined' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
gcc/value-range-equiv.h:58:8: warning: 'set_varying' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]

gcc/ChangeLog:

* value-range-equiv.h (class value_range_equiv):

23 months agoanalyzer: add more final override keywords
Martin Liska [Tue, 16 Aug 2022 09:00:14 +0000 (11:00 +0200)]
analyzer: add more final override keywords

gcc/analyzer/ChangeLog:

* region-model.cc: Fix -Winconsistent-missing-override clang
  warning.
* region.h: Likewise.

23 months agoi386: add 'final' and 'override' to scalar_chain
Martin Liska [Tue, 16 Aug 2022 08:53:00 +0000 (10:53 +0200)]
i386: add 'final' and 'override' to scalar_chain

In c3ed9e0d6e96d8697e4bab994f8acbc5506240ee, David added some
"final override" and since that there are 2 new warnings that
need the same treatment:

gcc/config/i386/i386-features.h:186:8: warning: 'convert_op' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
gcc/config/i386/i386-features.h:186:8: warning: 'convert_op' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
gcc/config/i386/i386-features.h:199:8: warning: 'convert_op' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
gcc/config/i386/i386-features.h:199:8: warning: 'convert_op' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]

gcc/ChangeLog:

* config/i386/i386-features.h (class general_scalar_chain): Add
  final override for a method.
(class timode_scalar_chain): Likewise.

23 months agodocs: fix link destination
Martin Liska [Tue, 16 Aug 2022 08:04:36 +0000 (10:04 +0200)]
docs: fix link destination

gcc/fortran/ChangeLog:

* gfortran.texi: Fix link destination to a valid URL.

23 months agoAdjust max-jump-thread-paths docs
Richard Biener [Tue, 16 Aug 2022 07:51:16 +0000 (09:51 +0200)]
Adjust max-jump-thread-paths docs

The following fixes spelling and changes edge degree for number of
incoming edges.

* doc/invoke.texi (max-jump-thread-paths): Adjust.

23 months agojobserver: fix fifo mode by opening pipe in proper mode
Martin Liska [Tue, 16 Aug 2022 06:22:29 +0000 (08:22 +0200)]
jobserver: fix fifo mode by opening pipe in proper mode

The current jobserver_info relies on non-blocking FDs,
thus one the pipe in such mode.

gcc/ChangeLog:

* opts-common.cc (jobserver_info::connect): Open fifo
in non-blocking mode.

23 months agors6000: Adjust mov optabs for opaque modes [PR103353]
Kewen.Lin [Tue, 16 Aug 2022 05:24:07 +0000 (00:24 -0500)]
rs6000: Adjust mov optabs for opaque modes [PR103353]

As PR103353 shows, we may want to continue to expand built-in
function __builtin_vsx_lxvp, even if we have already emitted
error messages about some missing required conditions.  As
shown in that PR, without one explicit mov optab on OOmode
provided, it would call emit_move_insn recursively.

So this patch is to allow the mov pattern to be generated during
expanding phase if compiler has already seen errors.

PR target/103353

gcc/ChangeLog:

* config/rs6000/mma.md (define_expand movoo): Move TARGET_MMA condition
check to preparation statements and add handlings for !TARGET_MMA.
(define_expand movxo): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/pr103353.c: New test.

23 months agovect: Don't allow vect_emulated_vector_p type in vectorizable_call [PR106322]
Kewen Lin [Tue, 16 Aug 2022 05:18:51 +0000 (00:18 -0500)]
vect: Don't allow vect_emulated_vector_p type in vectorizable_call [PR106322]

As PR106322 shows, in some cases for some vector type whose
TYPE_MODE is a scalar integral mode instead of a vector mode,
it's possible to obtain wrong target support information when
querying with the scalar integral mode.  For example, for the
test case in PR106322, on ppc64 32bit vectorizer gets vector
type "vector(2) short unsigned int" for scalar type "short
unsigned int", its mode is SImode instead of V2HImode.  The
target support querying checks umul_highpart optab with SImode
and considers it's supported, then vectorizer further generates
.MULH IFN call for that vector type.  Unfortunately it's wrong
to use SImode support for that vector type multiply highpart
here.

This patch is to teach vectorizable_call analysis not to allow
vect_emulated_vector_p type for both vectype_in and vectype_out
as Richi suggested.

PR tree-optimization/106322

gcc/ChangeLog:

* tree-vect-stmts.cc (vectorizable_call): Don't allow
vect_emulated_vector_p type for both vectype_in and vectype_out.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr106322.c: New test.
* gcc.target/powerpc/pr106322.c: New test.

23 months agoxtensa: Turn on -fsplit-wide-types-early by default
Takayuki 'January June' Suwa [Sun, 14 Aug 2022 09:29:33 +0000 (18:29 +0900)]
xtensa: Turn on -fsplit-wide-types-early by default

Since GCC10, the "subreg2" optimization pass was no longer tied to enabling
"subreg1" unless -fsplit-wide-types-early was turned on (PR88233).  However
on the Xtensa port, the lack of "subreg2" can degrade the quality of the
output code, especially for those that produce many D[FC]mode pseudos.

This patch turns on -fsplit-wide-types-early by default in order to restore
the previous behavior.

gcc/ChangeLog:

* common/config/xtensa/xtensa-common.cc
(xtensa_option_optimization_table): Add OPT_fsplit_wide_types_early
for OPT_LEVELS_ALL in order to restore pre-GCC10 behavior.

23 months agoDaily bump.
GCC Administrator [Tue, 16 Aug 2022 00:17:23 +0000 (00:17 +0000)]
Daily bump.

23 months agod: Defer compiling inline definitions until after the module has finished.
Iain Buclaw [Mon, 15 Aug 2022 17:00:43 +0000 (19:00 +0200)]
d: Defer compiling inline definitions until after the module has finished.

This is to prevent the case of when generating the methods of a struct
type, we don't accidentally emit an inline function that references it,
as the outer struct itself would still be incomplete.

gcc/d/ChangeLog:

* d-tree.h (d_defer_declaration): Declare.
* decl.cc (function_needs_inline_definition_p): Defer checking
DECL_UNINLINABLE and DECL_DECLARED_INLINE_P.
(maybe_build_decl_tree): Call d_defer_declaration instead of
build_decl_tree.
* modules.cc (deferred_inline_declarations): New variable.
(build_module_tree): Set deferred_inline_declarations and a handle
declarations pushed to it.
(d_defer_declaration): New function.

23 months agod: Fix internal compiler error: Segmentation fault at gimple-expr.cc:88
Iain Buclaw [Mon, 15 Aug 2022 15:51:03 +0000 (17:51 +0200)]
d: Fix internal compiler error: Segmentation fault at gimple-expr.cc:88

Because complex types are deprecated in the language, the new way to
expose native complex types is by defining an enum with a basetype of a
library-defined struct that is implicitly treated as-if it is native.
As casts are not implicitly added by the front-end when downcasting from
enum to its underlying type, we must insert an explicit cast during the
code generation pass.

PR d/106623

gcc/d/ChangeLog:

* d-codegen.cc (underlying_complex_expr): New function.
(d_build_call): Handle passing native complex objects as the
library-defined equivalent.
* d-tree.h (underlying_complex_expr): Declare.
* expr.cc (ExprVisitor::visit (DotVarExp *)): Call
underlying_complex_expr instead of build_vconvert.

gcc/testsuite/ChangeLog:

* gdc.dg/torture/pr106623.d: New test.

23 months agod: Build internal TypeInfo types when module name is "object"
Iain Buclaw [Mon, 15 Aug 2022 11:43:29 +0000 (13:43 +0200)]
d: Build internal TypeInfo types when module name is "object"

If for whatever reason the module declaration doesn't exist in the
object file, ensure that the internal definitions for TypeInfo and
TypeInfo_Class are still created, otherwise an ICE could occur later if
they are required for a run-time helper call.

gcc/d/ChangeLog:

* d-compiler.cc (Compiler::onParseModule): Call create_tinfo_types
when module name is object.
* typeinfo.cc (create_tinfo_types): Add guard for multiple
invocations.

23 months agod: Field names of anonymous delegates should be same as regular delegate types.
Iain Buclaw [Mon, 15 Aug 2022 07:54:15 +0000 (09:54 +0200)]
d: Field names of anonymous delegates should be same as regular delegate types.

Doesn't change anything in the code generation or ABI, but makes it
consistent with regular delegates as names would match up when
inspecting tree dumps.

gcc/d/ChangeLog:

* d-codegen.cc (build_delegate_cst): Give anonymous delegate field
names same as per ABI spec.

23 months agoanalyzer: fix direction of -Wanalyzer-out-of-bounds note [PR106626]
David Malcolm [Mon, 15 Aug 2022 18:47:03 +0000 (14:47 -0400)]
analyzer: fix direction of -Wanalyzer-out-of-bounds note [PR106626]

Fix a read/write typo.

Also, add more test coverage of -Wanalyzer-out-of-bounds to help
establish a baseline for experiments on tweaking the wording of
the warning (PR analyzer/106626).

gcc/analyzer/ChangeLog:
PR analyzer/106626
* region-model.cc (buffer_overread::emit): Fix copy&paste error in
direction of the access in the note.

gcc/testsuite/ChangeLog:
PR analyzer/106626
* gcc.dg/analyzer/out-of-bounds-read-char-arr.c: New test.
* gcc.dg/analyzer/out-of-bounds-read-int-arr.c: New test.
* gcc.dg/analyzer/out-of-bounds-write-char-arr.c: New test.
* gcc.dg/analyzer/out-of-bounds-write-int-arr.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
23 months agoanalyzer: better fix for -Wanalyzer-use-of-uninitialized-value [PR106573]
David Malcolm [Mon, 15 Aug 2022 18:47:02 +0000 (14:47 -0400)]
analyzer: better fix for -Wanalyzer-use-of-uninitialized-value [PR106573]

gcc/analyzer/ChangeLog:
PR analyzer/106573
* region-model.cc (region_model::on_call_pre): Use check_call_args
when ensuring that we call get_arg_svalue on all args.  Remove
redundant call from handling for stdio builtins.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
23 months agoCheck for undefined and varying first.
Andrew MacLeod [Mon, 15 Aug 2022 14:16:23 +0000 (10:16 -0400)]
Check for undefined and varying first.

Rearrange order in irange:set to ensure all POLY_INTs map to varying.

PR tree-optimization/106621
gcc/
* value-range.cc (irange::set): Check for POLY_INT_CST early.

gcc/testsuite/
* gcc.dg/pr106621.c

23 months agoanalyzer: fix for ICE in sm-fd.cc [PR106551]
Immad Mir [Mon, 15 Aug 2022 16:51:27 +0000 (22:21 +0530)]
analyzer: fix for ICE in sm-fd.cc [PR106551]

This patch fixes the ICE caused by valid_to_unchecked_state
in sm-fd.cc by exiting early if first argument of any "dup"
functions is invalid.

gcc/analyzer/ChangeLog:
PR analyzer/106551
* sm-fd.cc (check_for_dup): exit early if first
argument is invalid for all dup functions.

gcc/testsuite/ChangeLog:
PR analyzer/106551
* gcc.dg/analyzer/fd-dup-1.c: New testcase.

Signed-off-by: Immad Mir <mirimmad@outlook.com>
23 months agoSupport shifts and rotates by integer constants in TImode STV on x86_64.
Roger Sayle [Mon, 15 Aug 2022 16:46:38 +0000 (17:46 +0100)]
Support shifts and rotates by integer constants in TImode STV on x86_64.

This patch adds support for converting 128-bit TImode shifts and rotates
to SSE equivalents using V1TImode during the TImode STV pass.
Previously, only logical shifts by multiples of 8 were handled
(from my patch earlier this month).

As an example of the benefits, the following rotate by 32-bits:

unsigned __int128 a, b;
void rot32() { a = (b >> 32) | (b << 96); }

when compiled on x86_64 with -O2 previously generated:

        movq    b(%rip), %rax
        movq    b+8(%rip), %rdx
        movq    %rax, %rcx
        shrdq   $32, %rdx, %rax
        shrdq   $32, %rcx, %rdx
        movq    %rax, a(%rip)
        movq    %rdx, a+8(%rip)
        ret

with this patch, now generates:

        movdqa  b(%rip), %xmm0
        pshufd  $57, %xmm0, %xmm0
        movaps  %xmm0, a(%rip)
        ret

[which uses a V4SI permutation for those that don't read SSE].
This should help 128-bit cryptography codes, that interleave XORs
with rotations (but that don't use additions or subtractions).

2022-08-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-features.cc
(timode_scalar_chain::compute_convert_gain): Provide costs for
shifts and rotates.
(timode_scalar_chain::convert_insn): Handle ASHIFTRT, ROTATERT
and ROTATE just like existing ASHIFT and LSHIFTRT cases.
(timode_scalar_to_vector_candidate_p): Handle all shifts and
rotates by integer constants between 0 and 127.

gcc/testsuite/ChangeLog
* gcc.target/i386/sse4_1-stv-9.c: New test case.

23 months agoImproved gain calculation for COMPARE to 0 or -1 in TImode STV on x86_64.
Roger Sayle [Mon, 15 Aug 2022 16:43:02 +0000 (17:43 +0100)]
Improved gain calculation for COMPARE to 0 or -1 in TImode STV on x86_64.

This patch tweaks timode_scalar_chain::compute_convert_gain to provide
more accurate costs for converting TImode comparisons against zero or
minus 1 to V1TImode equivalents.

2022-08-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-features.cc
(timode_scalar_chain::compute_convert_gain): Provide gains for
comparisons against 0/-1, including "*testti" patterns.

23 months agoPR tree-optimization/64992: (B << 2) != 0 is B when B is Boolean.
Roger Sayle [Mon, 15 Aug 2022 16:39:47 +0000 (17:39 +0100)]
PR tree-optimization/64992: (B << 2) != 0 is B when B is Boolean.

This patch resolves both PR tree-optimization/64992 and PR
tree-optimization/98956 which are missed optimization enhancement
request, for which Andrew Pinski already has a proposed solution
(related to a fix for PR tree-optimization/98954).  Yesterday,
I proposed an alternate improved patch for PR98954, which although
superior in most respects, alas didn't address this case [which
doesn't include a BIT_AND_EXPR], hence this follow-up fix.

For many functions, F(B), of a (zero-one) Boolean value B, the
expression F(B) != 0 can often be simplified to just B.  Hence
"(B * 5) != 0" is B, "-B != 0" is B, "bswap(B) != 0" is B,
"(B >>r 3) != 0" is B.  These are all currently optimized by GCC,
with the strange exception of left shifts by a constant (possibly
due to the undefined/implementation defined behaviour when the
shift constant is larger than the first operand's precision).
This patch adds support for this particular case, when the shift
constant is valid.

2022-08-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
PR tree-optimization/64992
PR tree-optimization/98956
* match.pd (ne (lshift @0 @1) 0): Simplify (X << C) != 0 to X
when X is zero_one_valued_p and the shift constant C is valid.
(eq (lshift @0 @1) 0): Likewise, simplify (X << C) == 0 to !X
when X is zero_one_valued_p and the shift constant C is valid.

gcc/testsuite/ChangeLog
PR tree-optimization/64992
* gcc.dg/pr64992.c: New test case.

23 months agoPR tree-optimization/71343: Optimize (X<<C)&(Y<<C) as (X&Y)<<C.
Roger Sayle [Mon, 15 Aug 2022 16:32:26 +0000 (17:32 +0100)]
PR tree-optimization/71343: Optimize (X<<C)&(Y<<C) as (X&Y)<<C.

This patch is the first part of a solution to PR tree-optimization/71343,
a missed-optimization enhancement request where GCC fails to see that
(a<<2)+(b<<2) == a*4+b*4.

This piece is that (X<<C) op (Y<<C) can be simplified to (X op Y) << C,
for many binary operators, including AND, IOR, XOR, and (if overflow
isn't an issue) PLUS and MINUS.  Likewise, the right shifts (both logical
and arithmetic) and bit-wise logical operators can be simplified in a
similar fashion.  These all reduce the number of GIMPLE binary operations
from 3 to 2, by combining/eliminating a shift operation.

2022-08-15  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
PR tree-optimization/71343
* match.pd (op (lshift @0 @1) (lshift @2 @1)): Optimize the
expression (X<<C) + (Y<<C) to (X+Y)<<C for multiple operators.
(op (rshift @0 @1) (rshift @2 @1)): Likewise, simplify (X>>C)^(Y>>C)
to (X^Y)>>C for binary logical operators, AND, IOR and XOR.

gcc/testsuite/ChangeLog
PR tree-optimization/71343
* gcc.dg/pr71343-1.c: New test case.

23 months agoc++: Fix module line no testcase
Nathan Sidwell [Mon, 15 Aug 2022 14:19:36 +0000 (07:19 -0700)]
c++: Fix module line no testcase

Not all systems have the same injected headers, leading to line
location table differences that are immaterial to the test.  Fix the
regexp more robustly.

gcc/testsuite/
* g++.dg/modules/loc-prune-4.C: Adjust regexp

23 months agoc++: Extend -Wpessimizing-move for class prvalues [PR106276]
Marek Polacek [Fri, 22 Jul 2022 22:10:30 +0000 (18:10 -0400)]
c++: Extend -Wpessimizing-move for class prvalues [PR106276]

We already have a warning that warns about pessimizing std::move
in a return statement, when it prevents the NRVO:

  T fn()
  {
    T t;
    return std::move (t); // warning \o/
  }

However, the warning doesn't warn when what we are returning is a class
prvalue, that is, when std::move prevents the RVO:

  T fn()
  {
    T t;
    return std::move (T{}); // no warning :-(
  }

This came up recently in GCC:
<https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598177.html>.

This patch fixes that.  I would like to extend the warning further, so
that it warns in more contexts, e.g.:

  T t = std::move(T());

or

  void foo (T);
  foo (std::move(T()));

PR c++/106276

gcc/cp/ChangeLog:

* typeck.cc (can_do_rvo_p): New.
(maybe_warn_pessimizing_move): Warn when moving a temporary object
in a return statement prevents copy elision.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/Wpessimizing-move7.C: New test.

23 months agoSimplify range_on_path_entry
Richard Biener [Mon, 15 Aug 2022 11:03:54 +0000 (13:03 +0200)]
Simplify range_on_path_entry

I've noticed that range_on_path_entry does mightly complicated things
that don't make sense to me and the commentary might just be
out of date.  For the sake of it I replaced it with range_on_entry
and statistics show we thread _more_ jumps with that, so better
not do magic there.

* gimple-range-path.cc (range_on_path_entry): Just
call range_on_entry.

23 months agoi386 PIE: testsuite: cope with default pie on ia32
Alexandre Oliva [Mon, 15 Aug 2022 12:21:33 +0000 (09:21 -0300)]
i386 PIE: testsuite: cope with default pie on ia32

This patch continues the effort of cleaning up the testsuite for
--enable-default-pie; the focus herein is mostly 32-bit x86.

As much as I tried to avoid it, most of the changes to the testsuite
simply disable PIC/PIE, for reasons I'm going to detail below.

static-cdtor1.C gets new patterns to match PIE output.  Some
avx512fp16 tests change only in register allocation, because of the
register used to hold the GOT base address.  Interrupt tests changed
in this regard as well, but here it also affected register saving and
restoring.

The previous patch modified cet-sjlj tests, mentioning a single regexp
covering PIC and nonPIC got incorrect match counts.  I found out that
adding ?: to parenthesized subpatterns avoids miscounting matches.
Other tests that count certain kinds of insns needed adjustment over
insns in get_pc_thunk, extra loads from the GOT, or extra adds to
compute addresses.  In one case, namely stack-check-12, it is nonPIC
that had extra insns, that PIC gets rid of, or rather, pushing and
popping the PIC register obviates the dummy push and matching pop used
for stack probing in nonpic.

pr95126 tests were supposed to optimize loads into known constants,
but the @GOTOFF addresses prevent that for reasons I have not
investigated, but that would be clearly desirable, so I've XFAILed
these.  pr95852 is another case of missed optimization: sibcalls are
not possible when the PIC register needs to be set up for the call,
which prevents the expected constant propagation to the return block;
I have adjusted the codegen expectations of these tests.

As for tests that disable PIE...  Some are judgment calls, that fail
for similar reasons as tests described above, but I chose not to
adjust their expectations; others are just not possible with PIC, or
not worth the effort of adjusting.

anon[14].C check for no global or comdat symbols, respectively, but
-fPIE outputs get_pc_thunk, as global hidden comdat.
initlist-const1.C wants .rodata and checks for no .data, but PIC
outputs constant data that needs relocations in .data.rel.ro.local.
no-stack-protector-attr-3.C and stackprotectexplicit2.C count
stack_check_fail matches; -fPIE calls stack_check_fail_local instead,
which matches the pattern, but this symbol is also marked as .hidden,
so the match count needs to be adjusted.

pr71694.C checks for no movl, but get_pc_thunk contains one.
pr102892-1.c is a missed optimization, ivopts creates an induction
variable because the array address can't be part of an indexing base
address with PIE, and that ends up stopping a load from being resolved
to a constant as expected.  sibcall-11.c needs @PLT for the call,
which requires the PIC register, which makes sibcalling impossible.
builtin-self.c, in turn, expects no calls, but PIC calls get_pc_thunk.

avx* vector tests that had PIE disabled were affected in that the need
for GOT-based addressing modes changed instruction selection in ways
that deviated from the expectations of the tests.  Ditto other vector
tests: pr100865*, pr101796-1, pr101846, pr101989-broadcast-1, and
pr102021, pr54855-[37], and pr90773-17.

pr15184* tests need a PIC register to access global variables, which
affects register allocation, so the patterns would have to be
adjusted.  pr27971 can't use the expected addressing mode to
dereference the array with PIC, so it ends up selecting an indexed
addressing mode, obviating the expected separate shift insn.

pr70263-2 is another case that implicitly expects a sibcall,
impossible because of the need for the PIC register; without a
sibcall, the expected REG_EQUIV for the reuse of the stack slot of an
incoming argument does not occur.  pr78035 duplicates the final
compare in both then and else blocks with PIE, which deviates from the
expected cmp count.  pr81736-[57] test for no frame pointer, but the
PIC register assignment to a call-saved register forces a frame; the
former ends up not using the PIC register, but it's only optimized out
after committing to a stack frame to preserve it.  pr85620-6 also
expects a tail call in a situation that is impossible on ia32 PIC.

pr85667-6 doesn't expect the movl in get_pc_thunk.  pr93492-5 tests
-mfentry, not available with PIC on ia32.  pr96539 expects a
tail-call, to avoid copying a large-ish struct argument, but the call
requires the PIC register, so no tail-call.  stack-prot-sym.c expects
a nonpic addressing mode.

for  gcc/testsuite/ChangeLog

* g++.dg/abi/anon1.C: Disable pie on ia32.
* g++.dg/abi/anon4.C: Likewise.
* g++.dg/cpp0x/initlist-const1.C: Likewise.
* g++.dg/no-stack-protector-attr-3.C: Likewise.
* g++.dg/stackprotectexplicit2.C: Likewise.
* g++.dg/pr71694.C: Likewise.
* gcc.dg/pr102892-1.c: Likewise.
* gcc.dg/sibcall-11.c: Likewise.
* gcc.dg/torture/builtin-self.c: Likewise.
* gcc.target/i386/avx2-dest-false-dep-for-glc.c: Likewise.
* gcc.target/i386/avx512bf16-cvtsbh2ss-1.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-1.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-3.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-5.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-7.c: Likewise.
* gcc.target/i386/avx512fp16-broadcast-1.c: Likewise.
* gcc.target/i386/avx512fp16-pr101846.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-1.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-3.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-5.c: Likewise.
* gcc.target/i386/pr100865-2.c: Likewise.
* gcc.target/i386/pr100865-3.c: Likewise.
* gcc.target/i386/pr100865-4a.c: Likewise.
* gcc.target/i386/pr100865-4b.c: Likewise.
* gcc.target/i386/pr100865-5a.c: Likewise.
* gcc.target/i386/pr100865-5b.c: Likewise.
* gcc.target/i386/pr100865-6a.c: Likewise.
* gcc.target/i386/pr100865-6b.c: Likewise.
* gcc.target/i386/pr100865-6c.c: Likewise.
* gcc.target/i386/pr100865-7b.c: Likewise.
* gcc.target/i386/pr101796-1.c: Likewise.
* gcc.target/i386/pr101846-2.c: Likewise.
* gcc.target/i386/pr101989-broadcast-1.c: Likewise.
* gcc.target/i386/pr102021.c: Likewise.
* gcc.target/i386/pr90773-17.c: Likewise.
* gcc.target/i386/pr54855-3.c: Likewise.
* gcc.target/i386/pr54855-7.c: Likewise.
* gcc.target/i386/pr15184-1.c: Likewise.
* gcc.target/i386/pr15184-2.c: Likewise.
* gcc.target/i386/pr27971.c: Likewise.
* gcc.target/i386/pr70263-2.c: Likewise.
* gcc.target/i386/pr78035.c: Likewise.
* gcc.target/i386/pr81736-5.c: Likewise.
* gcc.target/i386/pr81736-7.c: Likewise.
* gcc.target/i386/pr85620-6.c: Likewise.
* gcc.target/i386/pr85667-6.c: Likewise.
* gcc.target/i386/pr93492-5.c: Likewise.
* gcc.target/i386/pr96539.c: Likewise.
PR target/81708 (%gs:my_guard)
* gcc.target/i386/stack-prot-sym.c: Likewise.
* g++.dg/init/static-cdtor1.C: Add alternate patterns for PIC.
* gcc.target/i386/avx512fp16-vcvtsh2si-1a.c: Extend patterns
for PIC/PIE register allocation.
* gcc.target/i386/pr100704-3.c: Likewise.
* gcc.target/i386/avx512fp16-vcvtsh2usi-1a.c: Likewise.
* gcc.target/i386/avx512fp16-vcvttsh2si-1a.c: Likewise.
* gcc.target/i386/avx512fp16-vcvttsh2usi-1a.c: Likewise.
* gcc.target/i386/avx512fp16-vmovsh-1a.c: Likewise.
* gcc.target/i386/interrupt-11.c: Likewise, allowing for
preservation of the PIC register.
* gcc.target/i386/interrupt-12.c: Likewise.
* gcc.target/i386/interrupt-13.c: Likewise.
* gcc.target/i386/interrupt-15.c: Likewise.
* gcc.target/i386/interrupt-16.c: Likewise.
* gcc.target/i386/interrupt-17.c: Likewise.
* gcc.target/i386/interrupt-8.c: Likewise.
* gcc.target/i386/cet-sjlj-6a.c: Combine patterns from
previous change.
* gcc.target/i386/cet-sjlj-6b.c: Likewise.
* gcc.target/i386/pad-10.c: Accept insns in get_pc_thunk.
* gcc.target/i386/pr70321.c: Likewise.
* gcc.target/i386/pr81563.c: Likewise.
* gcc.target/i386/pr84278.c: Likewise.
* gcc.target/i386/pr90773-2.c: Likewise, plus extra loads from
the GOT.
* gcc.target/i386/pr90773-3.c: Likewise.
* gcc.target/i386/pr94913-2.c: Accept additional PIC insns.
* gcc.target/i386/stack-check-17.c: Likewise.
* gcc.target/i386/stack-check-12.c: Do not require dummy stack
probing obviated with PIC.
* gcc.target/i386/pr95126-m32-1.c: Expect missed optimization
with PIC.
* gcc.target/i386/pr95126-m32-2.c: Likewise.
* gcc.target/i386/pr95852-2.c: Accept different optimization
with PIC.
* gcc.target/i386/pr95852-4.c: Likewise.

23 months agoifcvt: Fix up noce_convert_multiple_sets [PR106590]
Jakub Jelinek [Mon, 15 Aug 2022 11:56:57 +0000 (13:56 +0200)]
ifcvt: Fix up noce_convert_multiple_sets [PR106590]

The following testcase is miscompiled on x86_64-linux.
The problem is in the noce_convert_multiple_sets optimization.
We essentially have:
if (g == 1)
  {
    g = 1;
    f = 23;
  }
else
  {
    g = 2;
    f = 20;
  }
and for each insn try to create a conditional move sequence.
There is code to detect overlap with the regs used in the condition
and the destinations, so we actually try to construct:
tmp_g = g == 1 ? 1 : 2;
f = g == 1 ? 23 : 20;
g = tmp_g;
which is fine.  But, we actually try to create two different
conditional move sequences in each case, seq1 with the whole
(eq (reg/v:HI 82 [ g ]) (const_int 1 [0x1]))
condition and seq2 with cc_cmp
(eq (reg:CCZ 17 flags) (const_int 0 [0]))
to rely on the earlier present comparison.  In each case, we
compare the rtx costs and choose the cheaper sequence (seq1 if both
have the same cost).
The problem is that with the skylake tuning,
tmp_g = g == 1 ? 1 : 2;
is actually expanded as
tmp_g = (g == 1) + 1;
in seq1 (which clobbers (reg 17 flags)) and as a cmov in seq2
(which doesn't).  The tuning says both have the same cost, so we
pick seq1.  Next we check sequences for
f = g == 1 ? 23 : 20; and here the seq2 cmov is cheaper, but it
uses (reg 17 flags) which has been clobbered earlier.

The following patch fixes that by detecting if we in the chosen
sequence clobber some register mentioned in cc_cmp or rev_cc_cmp,
and if yes, arranges for only seq1 (i.e. sequences that emit the
comparison itself) to be used after that.

2022-08-15  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/106590
* ifcvt.cc (check_for_cc_cmp_clobbers): New function.
(noce_convert_multiple_sets_1): If SEQ sets or clobbers any regs
mentioned in cc_cmp or rev_cc_cmp, don't consider seq2 for any
further conditional moves.

* gcc.dg/torture/pr106590.c: New test.

23 months agox86: Enable __bf16 type for TARGET_SSE2 and above
konglin1 [Mon, 15 Aug 2022 06:39:22 +0000 (14:39 +0800)]
x86: Enable __bf16 type for TARGET_SSE2 and above

gcc/ChangeLog:

* config/i386/i386-builtin-types.def (BFLOAT16): New primitive type.
* config/i386/i386-builtins.cc : Support __bf16 type for i386 backend.
(ix86_register_bf16_builtin_type): New function.
(ix86_bf16_type_node): New.
(ix86_bf16_ptr_type_node): Ditto.
(ix86_init_builtin_types): Add ix86_register_bf16_builtin_type function call.
* config/i386/i386-modes.def (FLOAT_MODE): Add BFmode.
(ADJUST_FLOAT_FORMAT): Ditto.
* config/i386/i386.cc (classify_argument): Handle BFmode.
(construct_container): Ditto.
(function_value_32): Return __bf16 by %xmm0.
(function_value_64): Return __bf16 by SSE register.
(ix86_output_ssemov): Handle BFmode.
(ix86_legitimate_constant_p): Disable BFmode constant double.
(ix86_secondary_reload): Require gpr as intermediate register
to store __bf16 from sse register when sse4 is not available.
(ix86_scalar_mode_supported_p): Enable __bf16 under sse2.
(ix86_mangle_type): Add manlging for __bf16 type.
(ix86_invalid_conversion): New function for target hook.
(ix86_invalid_unary_op): Ditto.
(ix86_invalid_binary_op): Ditto.
(TARGET_INVALID_CONVERSION): New define for target hook.
(TARGET_INVALID_UNARY_OP): Ditto.
(TARGET_INVALID_BINARY_OP): Ditto.
* config/i386/i386.h (host_detect_local_cpu): Add BFmode.
* config/i386/i386.md ("mode"): Add BFmode.
(MODE_SIZE): Ditto.
(X87MODEFH): Ditto.
(HFBF): Add new define_mode_iterator.
(*pushhf_rex64): Change for BFmode.
(*push<mode>_rex64): Ditto.
(*pushhf): Ditto.
(*push<mode>): Ditto.
(MODESH): Ditto.
(hfbfconstf): Add new define_mode_attr.
(*mov<mode>_internal): Add BFmode.

gcc/testsuite/ChangeLog:

* g++.target/i386/bfloat_cpp_typecheck.C: New test.
* gcc.target/i386/bfloat16-1.c: Ditto.
* gcc.target/i386/sse2-bfloat16-1.c: Ditto.
* gcc.target/i386/sse2-bfloat16-2.c: Ditto.
* gcc.target/i386/sse2-bfloat16-scalar-typecheck.c: Ditto.

23 months agoDaily bump.
GCC Administrator [Mon, 15 Aug 2022 00:16:42 +0000 (00:16 +0000)]
Daily bump.

23 months agoDaily bump.
GCC Administrator [Sun, 14 Aug 2022 00:16:56 +0000 (00:16 +0000)]
Daily bump.

23 months agoMove V1TI shift/rotate lowering from expand to pre-reload split on x86_64.
Roger Sayle [Sat, 13 Aug 2022 13:52:45 +0000 (14:52 +0100)]
Move V1TI shift/rotate lowering from expand to pre-reload split on x86_64.

This patch moves the lowering of 128-bit V1TImode shifts and rotations by
constant bit counts to sequences of SSE operations from the RTL expansion
pass to the pre-reload split pass.  Postponing this splitting of shifts
and rotates enables (will enable) the TImode equivalents of these operations/
instructions to be considered as candidates by the (TImode) STV pass.
Technically, this patch changes the existing expanders to continue to
lower shifts by variable amounts, but constant operands become RTL
instructions, specified by define_insn_and_split that are triggered by
x86_pre_reload_split.  The one minor complication is that logical shifts
by multiples of eight, don't get split, but are handled by existing insn
patterns, such as sse2_ashlv1ti3 and sse2_lshrv1ti3.  There should be no
changes in generated code with this patch, which just adjusts the pass
in which transformations get applied.

2022-08-13  Roger Sayle  <roger@nextmovesoftware.com>
    Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog
* config/i386/predicates.md (const_0_to_255_not_mul_8_operand):
New predicate for values between 0/1 and 255, not multiples of 8.
* config/i386/sse.md (ashlv1ti3): Delay lowering of logical left
shifts by constant bit counts.
(*ashlvti3_internal): New define_insn_and_split that lowers
logical left shifts by constant bit counts, that aren't multiples
of 8, before reload.
(lshrv1ti3): Delay lowering of logical right shifts by constant.
(*lshrv1ti3_internal): New define_insn_and_split that lowers
logical right shifts by constant bit counts, that aren't multiples
of 8, before reload.
(ashrv1ti3):: Delay lowering of arithmetic right shifts by
constant bit counts.
(*ashrv1ti3_internal): New define_insn_and_split that lowers
arithmetic right shifts by constant bit counts before reload.
(rotlv1ti3): Delay lowering of rotate left by constant.
(*rotlv1ti3_internal): New define_insn_and_split that lowers
rotate left by constant bits counts before reload.
(rotrv1ti3): Delay lowering of rotate right by constant.
(*rotrv1ti3_internal): New define_insn_and_split that lowers
rotate right by constant bits counts before reload.

23 months agotestsuite: Disable out-of-bounds checker in analyzer/torture/pr93451.c
Tim Lange [Sat, 13 Aug 2022 11:31:45 +0000 (13:31 +0200)]
testsuite: Disable out-of-bounds checker in analyzer/torture/pr93451.c

This patch disables Wanalyzer-out-of-bounds for analyzer/torture/pr93451.c
and makes the test case pass when compiled with -m32.

The emitted warning is a true positive but only occurs if
sizeof (long int) is less than sizeof (double). I've already discussed a
similar case with Dave in the context of pr96764.c and we came to the
conclusion that we just disable the checker in such cases.

Committed under the "obvious fix" rule.

gcc/testsuite/ChangeLog:

* gcc.dg/analyzer/torture/pr93451.c:
Disable Wanalyzer-out-of-bounds.

23 months agoDaily bump.
GCC Administrator [Sat, 13 Aug 2022 00:16:54 +0000 (00:16 +0000)]
Daily bump.

23 months ago[Committed] arm: Document +no options for Cortex-M55 CPU.
Srinath Parvathaneni [Fri, 12 Aug 2022 15:29:10 +0000 (16:29 +0100)]
[Committed] arm: Document +no options for Cortex-M55 CPU.

This patch documents the following options for Arm Cortex-M55 CPU
under -mcpu= list.

+nomve.fp (disables MVE single precision floating point instructions)
+nomve (disables MVE integer and single precision floating point instructions)
+nodsp (disables dsp, MVE integer and single precision floating point instructions)
+nofp (disables floating point instructions)
Committed as obvious to master.

gcc/ChangeLog:

2022-08-12  Srinath Parvathaneni  <srinath.parvathaneni@arm.com>

* doc/invoke.texi (Arm Options): Document -mcpu=cortex-m55 options.

23 months agoFix invalid devirtualization when combining final keyword and anonymous types
Jan Hubicka [Fri, 12 Aug 2022 14:25:28 +0000 (16:25 +0200)]
Fix invalid devirtualization when combining final keyword and anonymous types

this patch fixes a wrong code issue where we incorrectly devirtualize to
__builtin_unreachable.  The problem occurs in combination of anonymous
namespaces and final keyword used on methods.  We do two optimizations here
 1) when reacing final method we cut the search for possible new targets
 2) if the type is anonymous we detect whether it is ever instatiated by
    looking if its vtable is referred to.
Now this goes wrong when thre is an anonymous type with final method that
is not instantiated while its derived type is.  So if 1 triggers we need
to make 2 to look for vtables of all derived types as done by this patch.

Bootstrpaped/regtested x86_64-linux

Honza

gcc/ChangeLog:

2022-08-10  Jan Hubicka  <hubicka@ucw.cz>

PR middle-end/106057
* ipa-devirt.cc (type_or_derived_type_possibly_instantiated_p): New
function.
(possible_polymorphic_call_targets): Use it.

gcc/testsuite/ChangeLog:

2022-08-10  Jan Hubicka  <hubicka@ucw.cz>

PR middle-end/106057
* g++.dg/tree-ssa/pr101839.C: New test.

23 months agoImprove comment for tree_niter_desc.{control,bound,cmp}
Andrew Carlotti [Fri, 12 Aug 2022 13:36:22 +0000 (14:36 +0100)]
Improve comment for tree_niter_desc.{control,bound,cmp}

Fix typos and explain ERROR_MARK usage.

gcc/ChangeLog:

* tree-ssa-loop.h: Improve comment

23 months agophiopt: Remove unnecessary checks from spaceship_replacement [PR106506]
Jakub Jelinek [Fri, 12 Aug 2022 11:40:43 +0000 (13:40 +0200)]
phiopt: Remove unnecessary checks from spaceship_replacement [PR106506]

Those 2 checks were just me trying to be extra careful, the
(phires & 1) == phires and variants it is folded to of course make only sense
for the -1/0/1/2 result spaceship, for -1/0/1 one can just use comparisons of
phires.  We only floating point spaceship if nans aren't honored, so the
2 case is ignored, and if it is, with Aldy's changes we can simplify the
2 case away from the phi but the (phires & 1) == phires stayed.  It is safe
to treat the phires comparison as phires >= 0 even then.

2022-08-12  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/106506
* tree-ssa-phiopt.cc (spaceship_replacement): Don't punt for
is_cast or orig_use_lhs cases if phi_bb has 3 predecessors.

* g++.dg/opt/pr94589-2.C: New test.

23 months agotree-optimization/106593 - fix ICE with backward threading
Richard Biener [Fri, 12 Aug 2022 10:20:13 +0000 (12:20 +0200)]
tree-optimization/106593 - fix ICE with backward threading

With the last re-org I failed to make sure to not add SSA names
nor supported by ranger into m_imports which then triggers an
ICE in range_on_path_entry because range_of_expr returns false.

PR tree-optimization/106593
* tree-ssa-threadbackward.cc (back_threader::find_paths):
If the imports from the conditional do not satisfy
gimple_range_ssa_p don't try to thread anything.

23 months agosve: Fix fcmuo combine patterns [PR106524]
Tamar Christina [Fri, 12 Aug 2022 11:28:41 +0000 (12:28 +0100)]
sve: Fix fcmuo combine patterns [PR106524]

There's no encoding for fcmuo with zero.  This restricts the combine patterns
from accepting zero registers.

gcc/ChangeLog:

PR target/106524
* config/aarch64/aarch64-sve.md (*fcmuo<mode>_nor_combine,
*fcmuo<mode>_bic_combine): Don't accept comparisons against zero.

gcc/testsuite/ChangeLog:

PR target/106524
* gcc.target/aarch64/sve/pr106524.c: New test.

23 months agoanalyzer: out-of-bounds checker [PR106000]
Tim Lange [Fri, 12 Aug 2022 08:27:16 +0000 (10:27 +0200)]
analyzer: out-of-bounds checker [PR106000]

This patch adds an experimental out-of-bounds checker to the analyzer.

The checker was tested on coreutils, curl, httpd and openssh. It is mostly
accurate but does produce false-positives on yacc-generated files and
sometimes when the analyzer misses an invariant. These cases will be
documented in bugzilla.
Regression-tested on Linux x86-64, further ran the analyzer tests with
the -m32 option.

2022-08-11  Tim Lange  <mail@tim-lange.me>

gcc/analyzer/ChangeLog:

PR analyzer/106000
* analyzer.opt: Add Wanalyzer-out-of-bounds.
* region-model.cc (class out_of_bounds): Diagnostics base class
for all out-of-bounds diagnostics.
(class past_the_end): Base class derived from out_of_bounds for
the buffer_overflow and buffer_overread diagnostics.
(class buffer_overflow): Buffer overflow diagnostics.
(class buffer_overread): Buffer overread diagnostics.
(class buffer_underflow): Buffer underflow diagnostics.
(class buffer_underread): Buffer overread diagnostics.
(region_model::check_region_bounds): New function to check region
bounds for out-of-bounds accesses.
(region_model::check_region_access):
Add call to check_region_bounds.
(region_model::get_representative_tree): New function that accepts
a region instead of an svalue.
* region-model.h (class region_model):
Add region_model::check_region_bounds.
* region.cc (region::symbolic_p): New predicate.
(offset_region::get_byte_size_sval): Only return the remaining
byte size on offset_regions.
* region.h: Add region::symbolic_p.
* store.cc (byte_range::intersects_p):
Add new function equivalent to bit_range::intersects_p.
(byte_range::exceeds_p): New function.
(byte_range::falls_short_of_p): New function.
* store.h (struct byte_range): Add byte_range::intersects_p,
byte_range::exceeds_p and byte_range::falls_short_of_p.

gcc/ChangeLog:

PR analyzer/106000
* doc/invoke.texi: Add Wanalyzer-out-of-bounds.

gcc/testsuite/ChangeLog:

PR analyzer/106000
* g++.dg/analyzer/pr100244.C: Disable out-of-bounds warning.
* gcc.dg/analyzer/allocation-size-3.c:
Disable out-of-bounds warning.
* gcc.dg/analyzer/memcpy-2.c: Disable out-of-bounds warning.
* gcc.dg/analyzer/pr101962.c: Add dg-warning.
* gcc.dg/analyzer/pr96764.c: Disable out-of-bounds warning.
* gcc.dg/analyzer/pr97029.c:
Add dummy buffer to prevent an out-of-bounds warning.
* gcc.dg/analyzer/realloc-5.c: Add dg-warning.
* gcc.dg/analyzer/test-setjmp.h:
Add dummy buffer to prevent an out-of-bounds warning.
* gcc.dg/analyzer/zlib-3.c: Add dg-bogus.
* g++.dg/analyzer/out-of-bounds-placement-new.C: New test.
* gcc.dg/analyzer/out-of-bounds-1.c: New test.
* gcc.dg/analyzer/out-of-bounds-2.c: New test.
* gcc.dg/analyzer/out-of-bounds-3.c: New test.
* gcc.dg/analyzer/out-of-bounds-container_of.c: New test.
* gcc.dg/analyzer/out-of-bounds-coreutils.c: New test.
* gcc.dg/analyzer/out-of-bounds-curl.c: New test.

23 months agoanalyzer: consider that realloc could shrink the buffer [PR106539]
Tim Lange [Fri, 12 Aug 2022 08:26:14 +0000 (10:26 +0200)]
analyzer: consider that realloc could shrink the buffer [PR106539]

This patch adds the "shrinks buffer" case to the success_with_move
modelling of realloc.

Regression-tested on Linux x86-64, further ran the analyzer tests with
the -m32 option.

2022-08-11  Tim Lange  <mail@tim-lange.me>

gcc/analyzer/ChangeLog:

PR analyzer/106539
* region-model-impl-calls.cc (region_model::impl_call_realloc):
Use the result of get_copied_size as the size for the
sized_regions in realloc.
(success_with_move::get_copied_size): New function.

gcc/testsuite/ChangeLog:

PR analyzer/106539
* gcc.dg/analyzer/pr106539.c: New test.
* gcc.dg/analyzer/realloc-5.c: New test.

23 months ago[AARCH64] Remove reference to MD_INCLUDES
Andrew Pinski [Fri, 12 Aug 2022 00:26:17 +0000 (00:26 +0000)]
[AARCH64] Remove reference to MD_INCLUDES

The comment reference to MD_INCLUDES is not needed
as it is auto generated for long time now even before
aarch64 target was added.

MD_INCLUDES has been auto generated since r0-64489.
Note some targets still manually set MD_INCLUDES and
I suspect those can be changed but I don't have access
to those targets.

Committed as obvious.

Thanks,
Andrew Pinski

gcc/ChangeLog:

* config/aarch64/aarch64.md: Remove comment
about MD_INCLUDES as it is out of date and not needed.

23 months agoDaily bump.
GCC Administrator [Fri, 12 Aug 2022 00:17:02 +0000 (00:17 +0000)]
Daily bump.

23 months agotestsuite: fd-4.c redefines mode_t on AIX.
David Edelsohn [Thu, 11 Aug 2022 20:26:37 +0000 (16:26 -0400)]
testsuite: fd-4.c redefines mode_t on AIX.

AIX stdio.h includes sys/types.h, which defines mode_t.  The
analyzer/fd-4.c testcase provides a definition of mode_t for creat()
call, which conflicts with the AIX definition.  This patch defines an
AIX macro to prevent multiple-definition of the type.

gcc/testsuite/ChangeLog:

* gcc.dg/analyzer/fd-4.c: Define _MODE_T on AIX.

23 months agotestcase: Fix AIX testsuite failures
David Edelsohn [Wed, 10 Aug 2022 13:22:28 +0000 (09:22 -0400)]
testcase: Fix AIX testsuite failures

Recent testsuite additions trip over AIX-specific features.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/initlist-const1.C: XFAIL on AIX.

23 months agoanalyzer: fix ICE casued by dup2 in sm-fd.cc[PR106551]
Immad Mir [Thu, 11 Aug 2022 16:15:54 +0000 (21:45 +0530)]
analyzer: fix ICE casued by dup2 in sm-fd.cc[PR106551]

This patch fixes the ICE caused by valid_to_unchecked_state,
at analyzer/sm-fd.cc by handling the m_start state in
check_for_dup.

Tested lightly on x86_64.

gcc/analyzer/ChangeLog:
PR analyzer/106551
* sm-fd.cc (check_for_dup): handle the m_start
state when transitioning the state of LHS
of dup, dup2 and dup3 call.

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/fd-dup-1.c: New testcases.
* gcc.dg/analyzer/fd-uninit-1.c: Remove bogus
warning.
Signed-off-by: Immad Mir <mirimmad@outlook.com>
23 months agoc-family: Honor -Wno-init-self for cv-qual vars [PR102633]
Marek Polacek [Tue, 26 Jul 2022 17:55:58 +0000 (13:55 -0400)]
c-family: Honor -Wno-init-self for cv-qual vars [PR102633]

Since r11-5188-g32934a4f45a721, we drop qualifiers during l-to-r
conversion by creating a NOP_EXPR.  For e.g.

  const int i = i;

that means that the DECL_INITIAL is '(int) i' and not 'i' anymore.
Consequently, we don't suppress_warning here:

711     case DECL_EXPR:
715       if (VAR_P (DECL_EXPR_DECL (*expr_p))
716           && !DECL_EXTERNAL (DECL_EXPR_DECL (*expr_p))
717           && !TREE_STATIC (DECL_EXPR_DECL (*expr_p))
718           && (DECL_INITIAL (DECL_EXPR_DECL (*expr_p)) == DECL_EXPR_DECL (*expr_p))
719           && !warn_init_self)
720         suppress_warning (DECL_EXPR_DECL (*expr_p), OPT_Winit_self);

because of the check on line 718 -- (int) i is not i.  So -Wno-init-self
doesn't disable the warning as it's supposed to.

The following patch fixes it by moving the suppress_warning call from
c_gimplify_expr to the front ends, at points where we haven't created
the NOP_EXPR yet.

PR middle-end/102633

gcc/c-family/ChangeLog:

* c-gimplify.cc (c_gimplify_expr) <case DECL_EXPR>: Don't call
suppress_warning here.

gcc/c/ChangeLog:

* c-parser.cc (c_parser_initializer): Add new tree parameter.  Use it.
Call suppress_warning.
(c_parser_declaration_or_fndef): Pass d down to c_parser_initializer.
(c_parser_omp_declare_reduction): Pass omp_priv down to
c_parser_initializer.

gcc/cp/ChangeLog:

* decl.cc (cp_finish_decl): Call suppress_warning.

gcc/testsuite/ChangeLog:

* c-c++-common/Winit-self1.c: New test.
* c-c++-common/Winit-self2.c: New test.

23 months agoTame path_range_query::compute_imports
Richard Biener [Thu, 11 Aug 2022 11:28:44 +0000 (13:28 +0200)]
Tame path_range_query::compute_imports

This avoids going BBs outside of the path when adding def chains
to the set of imports.  It also syncs the code with
range_def_chain::get_def_chain to not miss out on some imports
this function would identify.

* gimple-range-path.cc (path_range_query::compute_imports):
Restrict walking SSA defs to blocks inside the path.  Track
the same operands as range_def_chain::get_def_chain does.

23 months agotree-optimization/106514 - revisit m_import compute in backward threading
Richard Biener [Tue, 9 Aug 2022 11:37:26 +0000 (13:37 +0200)]
tree-optimization/106514 - revisit m_import compute in backward threading

This revisits how we compute imports later used for the ranger path
query during backwards threading.  The compute_imports function
of the path solver ends up pulling the SSA def chain of regular
stmts without limit and since it starts with just the gori imports
of the path exit it misses some interesting names to translate
during path discovery.  In fact with a still empty path this
compute_imports function looks like not the correct tool.

The following instead implements what it does during the path discovery
and since we add the exit block we seed the initial imports and
interesting names from just the exit conditional.  When we then
process interesting names (aka imports we did not yet see the definition
of) we prune local defs but add their uses in a similar way as
compute_imports would have done.

compute_imports also is lacking in its walking of the def chain
compared to range_def_chain::get_def_chain which for example
handles &_1->x specially through range_op_handler and
gimple_range_operand1, so the code copies this.  A fix for
compute_imports will be done separately, also fixing the unbound
walk there.

The patch also properly unwinds m_imports during the path discovery
backtracking and from a debugging session I have verified the two
sets evolve as expected now while previously behaving slightly erratic.

Fortunately the m_imports set now also is shrunken significantly for
the PR69592 testcase (aka PR106514) so that there's overall speedup
when increasing --param max-jump-thread-duplication-stmts as
15 -> 30 -> 60 -> 120 from 1s -> 2s -> 13s -> 27s to with the patch
1s -> 2s -> 4s -> 8s.

This runs into a latent issue in X which doesn't seem to expect
any PHI nodes with a constant argument on an edge inside the path.
But we now have those as interesting, for example for the ICEing
g++.dg/torture/pr100925.C which just has sth like

  if (i)
    x = 1;
  else
    x = 5;
  if (x == 1)
    ...

where we now have the path from if (i) to if (x) and the PHI for x
in the set of imports to consider for resolving x == 1 which IMHO
looks exactly like what we want.  The path_range_query::ssa_range_in_phi
papers over the issue and drops the range to varying instead of
crashing.  I didn't want to mess with this any further in this patch
(but I couldn't resist replacing the loop over PHI args with
PHI_ARG_DEF_FROM_EDGE, so mind the re-indenting).

PR tree-optimization/106514
* tree-ssa-threadbackward.cc (back_threader::find_paths_to_names):
Compute and unwind both m_imports and interesting on the fly during
path discovery.
(back_threader::find_paths): Compute the original m_imports
from just the SSA uses of the exit conditional.  Drop
handling single_succ_to_potentially_threadable_block.
* gimple-range-path.cc (path_range_query::ssa_range_in_phi): Handle
constant PHI arguments without crashing.  Use PHI_ARG_DEF_FROM_EDGE.

* gcc.dg/tree-ssa/ssa-thread-19.c: Un-XFAIL.
* gcc.dg/tree-ssa/ssa-thread-20.c: New testcase.

23 months agotestsuite: Fix up pr106243* tests on i686-linux [PR106243]
Jakub Jelinek [Thu, 11 Aug 2022 08:29:06 +0000 (10:29 +0200)]
testsuite: Fix up pr106243* tests on i686-linux [PR106243]

These 2 tests were FAILing on i686-linux or e.g. with
--target_board=unix/-m32/-mno-sse on x86_64-linux due to
-Wpsabi warnings.

2022-08-11  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/106243
* gcc.dg/pr106243.c: Add -Wno-psabi to dg-options.
* gcc.dg/pr106243-1.c: Likewise.

23 months agotestsuite: Fix up pr104992* tests on i686-linux [PR104992]
Jakub Jelinek [Thu, 11 Aug 2022 08:26:32 +0000 (10:26 +0200)]
testsuite: Fix up pr104992* tests on i686-linux [PR104992]

These 2 tests were FAILing on i686-linux or e.g. with
--target_board=unix/-m32/-mno-sse on x86_64-linux due to
-Wpsabi warnings and also because dg-options in the latter
test has been ignored due to missing space, so even -O2
wasn't passed at all.

2022-08-11  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/104992
* gcc.dg/pr104992.c: Add -Wno-psabi to dg-options.
* g++.dg/pr104992-1.C: Likewise.  Add space between " and } in
dg-options.

23 months agoFix path query compute_imports for external path
Richard Biener [Wed, 10 Aug 2022 10:54:05 +0000 (12:54 +0200)]
Fix path query compute_imports for external path

The following fixes the use of compute_imports from the backwards
threader which ends up accessing stale m_path from a previous
threading attempt.  The fix is to pass in the path explicitely
(and not the exit), and initializing it with the exit around this
call from the backwards threader.  That unfortunately exposed that
we rely on this broken behavior as the new testcase shows.  The
missed threading can be restored by registering all relations
from conditions on the path during solving, for the testcase the
particular important case is for relations provided by the path
entry conditional.

I've verified that the GORI query for imported ranges on edges
is not restricted this way.

This regresses the new ssa-thread-19.c testcase which is exactly
a case for the other patch re-doing how we compute imports since
this misses imports for defs that are not on the dominating path
from the exit.

That's one of the cases this regresses (it also progresses a few
due to more or the correct relations added).  Overall it
reduces the number of threads from 98649 to 98620 on my set of
cc1files.  I think it's a reasonable intermediate step to find
a stable, less random ground to compare stats to.

* gimple-range-path.h (path_range_query::compute_imports):
Take path as argument, not the exit block.
* gimple-range-path.cc (path_range_query::compute_imports):
Likewise, and adjust, avoiding possibly stale m_path.
(path_range_query::compute_outgoing_relations): Register
relations for all conditionals.
* tree-ssa-threadbackward.cc (back_threader::find_paths):
Adjust.

* gcc.dg/tree-ssa/ssa-thread-18.c: New testcase.
* gcc.dg/tree-ssa/ssa-thread-19.c: Likewise, but XFAILed.

23 months agors6000: Simplify some code with rs6000_builtin_is_supported
Kewen Lin [Wed, 10 Aug 2022 08:50:45 +0000 (03:50 -0500)]
rs6000: Simplify some code with rs6000_builtin_is_supported

In function rs6000_init_builtins, there is a oversight that
in one target debugging hunk with TARGET_DEBUG_BUILTIN we
missed to handle enum bif_enable ENB_CELL.  It's easy to
fix it by adding another if case.  But considering the long
term maintainability, this patch updates it with the existing
function rs6000_builtin_is_supported, which centralizes the
related conditions for different enum bif_enable, we only
need to update that function once some condition needs to
be changed later.  This also simplifies another usage in
function rs6000_expand_builtin.

gcc/ChangeLog:

* config/rs6000/rs6000-builtin.cc (rs6000_init_builtins): Fix the
oversight on ENB_CELL by simplifying with rs6000_builtin_is_supported.
(rs6000_expand_builtin): Simplify with rs6000_builtin_is_supported.

23 months agors6000: Remove stale rs6000_global_entry_point_needed_p
Kewen Lin [Wed, 10 Aug 2022 08:50:17 +0000 (03:50 -0500)]
rs6000: Remove stale rs6000_global_entry_point_needed_p

r10-631 had renamed rs6000_global_entry_point_needed_p to
rs6000_global_entry_point_prologue_needed_p.  This is to
remove the stale function declaration.

gcc/ChangeLog:

* config/rs6000/rs6000-internal.h (rs6000_global_entry_point_needed_p):
Remove function declaration.

23 months agoDaily bump.
GCC Administrator [Thu, 11 Aug 2022 00:16:46 +0000 (00:16 +0000)]
Daily bump.

23 months agotree-optimization/106513 - fix mistake in bswap symbolic number shifts
Richard Biener [Wed, 10 Aug 2022 13:45:22 +0000 (15:45 +0200)]
tree-optimization/106513 - fix mistake in bswap symbolic number shifts

This fixes a mistake in typing a local variable in the symbolic
shift routine.

PR tree-optimization/106513
* gimple-ssa-store-merging.cc (do_shift_rotate): Use uint64_t
for head_marker.

* gcc.dg/torture/pr106513.c: New testcase.

23 months agolto: respect jobserver in parallel WPA streaming
Martin Liska [Tue, 9 Aug 2022 11:59:39 +0000 (13:59 +0200)]
lto: respect jobserver in parallel WPA streaming

PR lto/106328

gcc/ChangeLog:

* opts-jobserver.h (struct jobserver_info): Add pipefd.
(jobserver_info::connect): New.
(jobserver_info::disconnect): Likewise.
(jobserver_info::get_token): Likewise.
(jobserver_info::return_token): Likewise.
* opts-common.cc: Implement the new functions.

gcc/lto/ChangeLog:

* lto.cc (wait_for_child): Decrement nruns once a process
finishes.
(stream_out_partitions): Use job server if active.
(do_whole_program_analysis): Likewise.

23 months agolto: support --jobserver-style=fifo for recent GNU make
Martin Liska [Tue, 9 Aug 2022 11:59:36 +0000 (13:59 +0200)]
lto: support --jobserver-style=fifo for recent GNU make

gcc/ChangeLog:

* opts-jobserver.h: Add one member.
* opts-common.cc (jobserver_info::jobserver_info): Parse FIFO
format of --jobserver-auth.

23 months agoFactor out jobserver_active_p.
Martin Liska [Tue, 9 Aug 2022 11:59:32 +0000 (13:59 +0200)]
Factor out jobserver_active_p.

gcc/ChangeLog:

* gcc.cc (driver::detect_jobserver): Remove and move to
jobserver.h.
* lto-wrapper.cc (jobserver_active_p): Likewise.
(run_gcc): Likewise.
* opts-jobserver.h: New file.
* opts-common.cc (jobserver_info::jobserver_info): New function.

23 months ago[Committed] PR other/106575: Use "signed char" in new fold-eqandshift-4.c
Roger Sayle [Wed, 10 Aug 2022 06:05:33 +0000 (07:05 +0100)]
[Committed] PR other/106575: Use "signed char" in new fold-eqandshift-4.c

My recently added testcase gcc.dg/fold-eqandshift-4.c, incorrectly assumed
that "char" was "signed char", and hence fails on powerpc64 where this
isn't the case.  Fixed by making "signed char" explicit where needed in
this test.  Committed as obvious.

2022-08-10  Roger Sayle  <roger@nextmovesoftware.com>

gcc/testsuite/ChangeLog
PR other/106575
* gcc.dg/fold-eqandshift-4.c: Use "signed char" explicitly.

23 months agoDaily bump.
GCC Administrator [Wed, 10 Aug 2022 00:16:43 +0000 (00:16 +0000)]
Daily bump.

23 months agoanalyzer: fix missing -Wanalyzer-use-of-uninitialized-value on special-cased function...
David Malcolm [Tue, 9 Aug 2022 23:58:54 +0000 (19:58 -0400)]
analyzer: fix missing -Wanalyzer-use-of-uninitialized-value on special-cased functions [PR106573]

We were missing checks for uninitialized params on calls to functions
that the analyzer has hardcoded knowledge of - both for those that are
handled just by state machines, and for those that are handled in
region-model-impl-calls.cc (for those arguments for which the svalue
wasn't accessed in handling the call).

Fixed thusly.

gcc/analyzer/ChangeLog:
PR analyzer/106573
* region-model.cc (region_model::on_call_pre): Ensure that we call
get_arg_svalue on all arguments.

gcc/testsuite/ChangeLog:
PR analyzer/106573
* gcc.dg/analyzer/error-uninit.c: New test.
* gcc.dg/analyzer/fd-uninit-1.c: New test.
* gcc.dg/analyzer/file-uninit-1.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
23 months agoUse PTEST to perform AND in TImode STV of (A & B) != 0 on x86_64.
Roger Sayle [Tue, 9 Aug 2022 17:59:55 +0000 (18:59 +0100)]
Use PTEST to perform AND in TImode STV of (A & B) != 0 on x86_64.

This x86_64 backend patch allows TImode STV to take advantage of the
fact that the PTEST instruction performs an AND operation.  Previously
PTEST was (mostly) used for comparison against zero, by using the same
operands.  The benefits are demonstrated by the new test case:

__int128 a,b;
int foo()
{
  return (a & b) != 0;
}

Currently with -O2 -msse4 we generate:

        movdqa  a(%rip), %xmm0
        pand    b(%rip), %xmm0
        xorl    %eax, %eax
        ptest   %xmm0, %xmm0
        setne   %al
        ret

with this patch we now generate:

        movdqa  a(%rip), %xmm0
        xorl    %eax, %eax
        ptest   b(%rip), %xmm0
        setne   %al
        ret

Technically, the magic happens using new define_insn_and_split patterns.
Using two patterns allows this transformation to performed independently
of whether TImode STV is run before or after combine.  The one tricky
case is that immediate constant operands of the AND behave slightly
differently between TImode and V1TImode: All V1TImode immediate operands
becomes loads, but for TImode only values that are not hilo_operands
need to be loaded.  Hence the new *testti_doubleword accepts any
general_operand, but internally during split calls force_reg whenever
the second operand is not x86_64_hilo_general_operand.  This required
(benefits from) some tweaks to TImode STV to support CONST_WIDE_INT in
more places, using CONST_SCALAR_INT_P instead of just CONST_INT_P.

2022-08-09  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-features.cc (scalar_chain::convert_compare):
Create new pseudos only when/if needed.  Add support for TEST,
i.e. (COMPARE (AND x y) (const_int 0)), using UNSPEC_PTEST.
When broadcasting V2DImode and V4SImode use new pseudo register.
(timode_scalar_chain::convert_op): Do nothing if operand is
already V1TImode.  Avoid generating useless SUBREG conversions,
i.e. (SUBREG:V1TImode (REG:V1TImode) 0).  Handle CONST_WIDE_INT
in addition to CONST_INT by using CONST_SCALAR_INT_P.
(convertible_comparison_p): Use CONST_SCALAR_INT_P to match both
CONST_WIDE_INT and CONST_INT.  Recognize new *testti_doubleword
pattern as an STV candidate.
(timode_scalar_to_vector_candidate_p): Allow CONST_SCALAR_INT_P
operands in binary logic operations.

* config/i386/i386.cc (ix86_rtx_costs) <case UNSPEC>: Add costs
for UNSPEC_PTEST; a PTEST that performs an AND has the same cost
as regular PTEST, i.e. cost->sse_op.

* config/i386/i386.md (*testti_doubleword): New pre-reload
define_insn_and_split that recognizes comparison of TI mode AND
against zero.
* config/i386/sse.md (*ptest<mode>_and): New pre-reload
define_insn_and_split that recognizes UNSPEC_PTEST of identical
AND operands.

gcc/testsuite/ChangeLog
* gcc.target/i386/sse4_1-stv-8.c: New test case.

23 months agomiddle-end: Optimize ((X >> C1) & C2) != C3 for more cases.
Roger Sayle [Tue, 9 Aug 2022 17:54:43 +0000 (18:54 +0100)]
middle-end: Optimize ((X >> C1) & C2) != C3 for more cases.

Following my middle-end patch for PR tree-optimization/94026, I'd promised
Jeff Law that I'd clean up the dead-code in fold-const.cc now that these
optimizations are handled in match.pd.  Alas, I discovered things aren't
quite that simple, as the transformations I'd added avoided cases where
C2 overlapped with the new bits introduced by the shift, but the original
code handled any value of C2 provided that it had a single-bit set (under
the condition that C3 was always zero).

This patch upgrades the transformations supported by match.pd to cover
any values of C2 and C3, provided that C1 is a valid bit shift constant,
for all three shift types (logical right, arithmetic right and left).
This then makes the code in fold-const.cc fully redundant, and adds
support for some new (corner) cases not previously handled.  If the
constant C1 is valid for the type's precision, the shift is now always
eliminated (with C2 and C3 possibly updated to test the sign bit).

Interestingly, the fold-const.cc code that I'm now deleting was originally
added by me back in 2006 to resolve PR middle-end/21137.  I've confirmed
that those testcase(s) remain resolved with this patch (and I'll close
21137 in Bugzilla).  This patch also implements most (but not all) of the
examples mentioned in PR tree-optimization/98954, for which I have some
follow-up patches.

2022-08-09  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
PR middle-end/21137
PR tree-optimization/98954
* fold-const.cc (fold_binary_loc): Remove optimizations to
optimize ((X >> C1) & C2) ==/!= 0.
* match.pd (cmp (bit_and (lshift @0 @1) @2) @3): Remove wi::ctz
check, and handle all values of INTEGER_CSTs @2 and @3.
(cmp (bit_and (rshift @0 @1) @2) @3): Likewise, remove wi::clz
checks, and handle all values of INTEGER_CSTs @2 and @3.

gcc/testsuite/ChangeLog
PR middle-end/21137
PR tree-optimization/98954
* gcc.dg/fold-eqandshift-4.c: New test case.

23 months agolibgccjit.h: Uncomment macro definition for testing gcc_jit_context_new_bitcast support
Vibhav Pant [Tue, 9 Aug 2022 15:30:18 +0000 (11:30 -0400)]
libgccjit.h: Uncomment macro definition for testing gcc_jit_context_new_bitcast support

The macro definition for LIBGCCJIT_HAVE_gcc_jit_context_new_bitcast
was earlier located in the documentation comment for
gcc_jit_context_new_bitcast, making it unavailable to code that
consumed libgccjit.h. This commit moves the definition out of the
comment, making it effective.

gcc/jit/ChangeLog:
* libgccjit.h (LIBGCCJIT_HAVE_gcc_jit_context_new_bitcast): Move
definition out of comment.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
23 months agodocs: add notes on which functions -fanalyzer has hardcoded knowledge of
David Malcolm [Tue, 9 Aug 2022 15:30:18 +0000 (11:30 -0400)]
docs: add notes on which functions -fanalyzer has hardcoded knowledge of

gcc/ChangeLog:
* doc/invoke.texi (Static Analyzer Options): Add notes on which
functions the analyzer has hardcoded knowledge of.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
23 months agod: Fix undefined reference to pragma(inline) symbol (PR106563)
Iain Buclaw [Tue, 9 Aug 2022 10:48:14 +0000 (12:48 +0200)]
d: Fix undefined reference to pragma(inline) symbol (PR106563)

Functions that are declared `pragma(inline)' should be treated as if
they are defined in every translation unit they are referenced from,
regardless of visibility protection.  Ensure they always get
DECL_ONE_ONLY linkage, and start emitting them into other modules that
import them.

PR d/106563

gcc/d/ChangeLog:

* decl.cc (DeclVisitor::visit (FuncDeclaration *)): Set semanticRun
before generating its symbol.
(function_defined_in_root_p): New function.
(function_needs_inline_definition_p): New function.
(maybe_build_decl_tree): New function.
(get_symbol_decl): Call maybe_build_decl_tree before returning symbol.
(start_function): Use function_defined_in_root_p instead of inline
test for locally defined symbols.
(set_linkage_for_decl): Check for inline functions before private or
protected symbols.

gcc/testsuite/ChangeLog:

* gdc.dg/torture/torture.exp (srcdir): New proc.
* gdc.dg/torture/imports/pr106563math.d: New test.
* gdc.dg/torture/imports/pr106563regex.d: New test.
* gdc.dg/torture/imports/pr106563uni.d: New test.
* gdc.dg/torture/pr106563.d: New test.

23 months agoamdgcn: Vector procedure call ABI
Andrew Stubbs [Fri, 15 Jul 2022 08:47:36 +0000 (09:47 +0100)]
amdgcn: Vector procedure call ABI

Adjust the (unofficial) procedure calling ABI such that vector arguments are
passed in vector registers, not on the stack.  Scalar arguments continue to
be passed in scalar registers, making a total of 12 argument registers.

The return value is also moved to a vector register (even for scalars; it
would be possible to retain the scalar location, using untyped_call, but
there's no obvious advantage in doing so).

After this change the ABI is as follows:

s0-s13  : Reserved for kernel launch parameters.
s14-s15 : Frame pointer.
s16-s17 : Stack pointer.
s18-s19 : Link register.
s20-s21 : Exec Save.
s22-s23 : CC Save.
s24-s25 : Scalar arguments.          NO LONGER RETURN VALUE.
s26-s29 : Additional scalar arguments (makes 6 total).
s30-s31 : Static Chain.
v0      : Prologue/epilogue scratch.
v1      : Constant 0, 1, 2, 3, 4, ... 63.
v2-v7   : Prologue/epilogue scratch.
v8-v9   : Return value & vector arguments.              NEW.
v10-v13 : Additional vector arguments (makes 6 total).  NEW.

gcc/ChangeLog:

* config/gcn/gcn.cc (gcn_function_value): Allow vector return values.
(num_arg_regs): Allow vector arguments.
(gcn_function_arg): Likewise.
(gcn_function_arg_advance): Likewise.
(gcn_arg_partial_bytes): Likewise.
(gcn_return_in_memory): Likewise.
(gcn_expand_epilogue): Get return value from v8.
* config/gcn/gcn.h (RETURN_VALUE_REG): Set to v8.
(FIRST_PARM_REG): USE FIRST_SGPR_REG for clarity.
(FIRST_VPARM_REG): New.
(FUNCTION_ARG_REGNO_P): Allow vector parameters.
(struct gcn_args): Add vnum field.
(LIBCALL_VALUE): All vector return values.
* config/gcn/gcn.md (gcn_call_value): Add vector constraints.
(gcn_call_value_indirect): Likewise.

23 months agoautopar TLC
Richard Biener [Tue, 2 Aug 2022 11:46:28 +0000 (13:46 +0200)]
autopar TLC

The following removes all excessive update_ssa calls from OMP
expansion, thereby rewriting the atomic load and store cases to
GIMPLE code generation.  I don't think autopar ever exercises the
atomics code though.

There's not much test coverage overall so I've built SPEC 2k17
with -floop-parallelize-all -ftree-parallelize-loops=2 with and
without LTO (and otherwise -Ofast plus -march=haswell) without
fallout.

If there's any fallout it's not OK to update SSA form for
each and every OMP stmt lowered.

* omp-expand.cc (expand_omp_atomic_load): Emit GIMPLE
directly.  Avoid update_ssa when in SSA form.
(expand_omp_atomic_store): Likewise.
(expand_omp_atomic_fetch_op): Avoid update_ssa when in SSA
form.
(expand_omp_atomic_pipeline): Likewise.
(expand_omp_atomic_mutex): Likewise.
* tree-parloops.cc (gen_parallel_loop): Use
TODO_update_ssa_no_phi after loop_version.

23 months agoRemove --param max-fsm-thread-length
Richard Biener [Mon, 8 Aug 2022 12:04:43 +0000 (14:04 +0200)]
Remove --param max-fsm-thread-length

This removes max-fsm-thread-length which is obsoleted by
max-jump-thread-paths.

* doc/invoke.texi (max-fsm-thread-length): Remove.
* params.opt (max-fsm-thread-length): Likewise.
* tree-ssa-threadbackward.cc
(back_threader_profitability::profitable_path_p): Do not
check max-fsm-thread-length.

23 months agotree-optimization/106514 - add --param max-jump-thread-paths
Richard Biener [Mon, 8 Aug 2022 10:20:04 +0000 (12:20 +0200)]
tree-optimization/106514 - add --param max-jump-thread-paths

The following adds a limit for the exponential greedy search of
the backwards jump threader.  The idea is to limit the search
space in a way that the paths considered are the same if the search
were in BFS order rather than DFS.  In particular it stops considering
incoming edges into a block if the product of the in-degrees of
blocks on the path exceeds the specified limit.

When considering the low stmt copying limit of 7 (or 1 in the size
optimize case) this means the degenerate case with maximum search
space is a sequence of conditions with no actual code

  B1
   |\
   | empty
   |/
  B2
   |\
   ...
  Bn
   |\

GIMPLE_CONDs are costed 2, an equivalent GIMPLE_SWITCH already 4, so
we reach 7 already with 3 middle conditions (B1 and Bn do not count).
The search space would be 2^4 == 16 to reach this.  The FSM threads
historically allowed for a thread length of 10 but is really looking
for a single multiway branch threaded across the backedge.  I've
chosen the default of the new parameter to 64 which effectively
limits the outdegree of the switch statement (the cases reaching the
backedge) to that number (divided by 2 until I add some special
pruning for FSM threads due to the loop header indegree).  The
testcase ssa-dom-thread-7.c requires 56 at the moment (as said,
some special FSM thread pruning of considered edges would bring
it down to half of that), but we now get one more threading
and quite some more in later threadfull.  This testcase seems to
be difficult to check for expected transforms.

The new testcases add the degenerate case we currently thread
(without deciding whether that's a good idea ...) plus one with
an approripate limit that should prevent the threading.

This obsoletes the mentioned --param max-fsm-thread-length but
I am not removing it as part of this patch.  When the search
space is limited the thread stmt size limit effectively provides
max-fsm-thread-length.

The param with its default does not help PR106514 enough to unleash
path searching with the higher FSM stmt count limit.

PR tree-optimization/106514
* params.opt (max-jump-thread-paths): New.
* doc/invoke.texi (max-jump-thread-paths): Document.
* tree-ssa-threadbackward.cc (back_threader::find_paths_to_names):
Honor max-jump-thread-paths, take overall_path argument.
(back_threader::find_paths): Pass 1 as initial overall_path.

* gcc.dg/tree-ssa/ssa-thread-16.c: New testcase.
* gcc.dg/tree-ssa/ssa-thread-17.c: Likewise.
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Adjust.

23 months agoOpenMP: Fix folding with simd's linear clause [PR106492]
Tobias Burnus [Tue, 9 Aug 2022 05:57:40 +0000 (07:57 +0200)]
OpenMP: Fix folding with simd's linear clause [PR106492]

gcc/ChangeLog:

PR middle-end/106492
* omp-low.cc (lower_rec_input_clauses): Add missing folding
to data type of linear-clause list item.

gcc/testsuite/ChangeLog:

PR middle-end/106492
* g++.dg/gomp/pr106492.C: New test.

23 months agoDaily bump.
GCC Administrator [Tue, 9 Aug 2022 00:16:47 +0000 (00:16 +0000)]
Daily bump.

23 months agoEvaluate condition arguments with the correct type.
Andrew MacLeod [Mon, 8 Aug 2022 19:13:51 +0000 (15:13 -0400)]
Evaluate condition arguments with the correct type.

Processing of a cond_expr requires that a range of the correct type for the
operands of the cond_expr is passed in.

PR tree-optimization/106556
gcc/
* gimple-range-gori.cc (gori_compute::condexpr_adjust): Use the
  type of the cond_expr operands being evaluted.

gcc/testsuite/
* gfortran.dg/pr106556.f90: New.

23 months agopreprocessor/106426: Treat u8 character literals as unsigned in char8_t modes.
Tom Honermann [Tue, 2 Aug 2022 18:36:02 +0000 (14:36 -0400)]
preprocessor/106426: Treat u8 character literals as unsigned in char8_t modes.

This patch corrects handling of UTF-8 character literals in preprocessing
directives so that they are treated as unsigned types in char8_t enabled
C++ modes (C++17 with -fchar8_t or C++20 without -fno-char8_t). Previously,
UTF-8 character literals were always treated as having the same type as
ordinary character literals (signed or unsigned dependent on target or use
of the -fsigned-char or -funsigned char options).

PR preprocessor/106426

gcc/c-family/ChangeLog:
* c-opts.cc (c_common_post_options): Assign cpp_opts->unsigned_utf8char
subject to -fchar8_t, -fsigned-char, and/or -funsigned-char.

gcc/testsuite/ChangeLog:
* g++.dg/ext/char8_t-char-literal-1.C: Check signedness of u8 literals.
* g++.dg/ext/char8_t-char-literal-2.C: Check signedness of u8 literals.

libcpp/ChangeLog:
* charset.cc (narrow_str_to_charconst): Set signedness of CPP_UTF8CHAR
literals based on unsigned_utf8char.
* include/cpplib.h (cpp_options): Add unsigned_utf8char.
* init.cc (cpp_create_reader): Initialize unsigned_utf8char.

23 months agoC: Implement C2X N2653 char8_t and UTF-8 string literal changes
Tom Honermann [Tue, 2 Aug 2022 18:36:01 +0000 (14:36 -0400)]
C: Implement C2X N2653 char8_t and UTF-8 string literal changes

This patch implements the core language and compiler dependent library
changes adopted for C2X via WG14 N2653.  The changes include:
- Change of type for UTF-8 string literals from array of const char to
  array of const char8_t (unsigned char).
- A new atomic_char8_t typedef.
- A new ATOMIC_CHAR8_T_LOCK_FREE macro defined in terms of the existing
  __GCC_ATOMIC_CHAR8_T_LOCK_FREE predefined macro.

gcc/ChangeLog:

* ginclude/stdatomic.h (atomic_char8_t,
ATOMIC_CHAR8_T_LOCK_FREE): New typedef and macro.

gcc/c/ChangeLog:

* c-parser.cc (c_parser_string_literal): Use char8_t as the type
of CPP_UTF8STRING when char8_t support is enabled.
* c-typeck.cc (digest_init): Allow initialization of an array
of character type by a string literal with type array of
char8_t.

gcc/c-family/ChangeLog:

* c-lex.cc (lex_string, lex_charconst): Use char8_t as the type
of CPP_UTF8CHAR and CPP_UTF8STRING when char8_t support is
enabled.
* c-opts.cc (c_common_post_options): Set flag_char8_t if
targeting C2x.

gcc/testsuite/ChangeLog:
* gcc.dg/atomic/c2x-stdatomic-lockfree-char8_t.c: New test.
* gcc.dg/atomic/gnu2x-stdatomic-lockfree-char8_t.c: New test.
* gcc.dg/c11-utf8str-type.c: New test.
* gcc.dg/c17-utf8str-type.c: New test.
* gcc.dg/c2x-utf8str-type.c: New test.
* gcc.dg/c2x-utf8str.c: New test.
* gcc.dg/gnu2x-utf8str-type.c: New test.
* gcc.dg/gnu2x-utf8str.c: New test.

23 months agod: Fix ICE in in add_stack_var, at cfgexpand.cc:476
Iain Buclaw [Mon, 8 Aug 2022 13:17:47 +0000 (15:17 +0200)]
d: Fix ICE in in add_stack_var, at cfgexpand.cc:476

The type that triggers the ICE never got completed by the semantic
analysis pass.  Checking for size forces it to be done, or issue a
compile-time error.

PR d/106555

gcc/d/ChangeLog:

* d-target.cc (Target::isReturnOnStack): Check for return type size.

gcc/testsuite/ChangeLog:

* gdc.dg/imports/pr106555.d: New test.
* gdc.dg/pr106555.d: New test.

23 months agolibstdc++: [_GLIBCXX_DEBUG] Do not consider detached iterators as value-initialized
François Dumont [Thu, 28 Jan 2021 21:23:28 +0000 (22:23 +0100)]
libstdc++: [_GLIBCXX_DEBUG] Do not consider detached iterators as value-initialized

An attach iterator has its _M_version set to something != 0, the container version. This
value shall be preserved when detaching it so that the iterator does not look like a
value-initialized one.

libstdc++-v3/ChangeLog:

* include/debug/formatter.h (__singular_value_init): New _Iterator_state enum entry.
(_Parameter<>(const _Safe_iterator<>&, const char*, _Is_iterator)): Check if iterator
parameter is value-initialized.
(_Parameter<>(const _Safe_local_iterator<>&, const char*, _Is_iterator)): Likewise.
* include/debug/safe_iterator.h (_Safe_iterator<>::_M_value_initialized()): New. Adapt
checks.
* include/debug/safe_local_iterator.h (_Safe_local_iterator<>::_M_value_initialized()): New.
Adapt checks.
* src/c++11/debug.cc (_Safe_iterator_base::_M_reset): Do not reset _M_version.
(print_field(PrintContext&, const _Parameter&, const char*)): Adapt state_names.
* testsuite/23_containers/deque/debug/iterator1_neg.cc: New test.
* testsuite/23_containers/deque/debug/iterator2_neg.cc: New test.
* testsuite/23_containers/forward_list/debug/iterator1_neg.cc: New test.
* testsuite/23_containers/forward_list/debug/iterator2_neg.cc: New test.
* testsuite/23_containers/forward_list/debug/iterator3_neg.cc: New test.

23 months agoFix middle-end/103645: empty struct store not removed when using compound literal
Andrew Pinski [Tue, 21 Dec 2021 04:27:33 +0000 (20:27 -0800)]
Fix middle-end/103645: empty struct store not removed when using compound literal

For compound literals empty struct stores are not removed as they go down a
different path of the gimplifier; trying to optimize the init constructor.
This fixes the problem by not adding the gimple assignment at the end
of gimplify_init_constructor if it was an empty type.

Note this updates gcc.dg/pr87052.c where we had:
const char d[0] = { };
And was expecting a store to d but after this, there is no store
as the decl's type is zero in size.

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

gcc/ChangeLog:

PR middle-end/103645
* gimplify.cc (gimplify_init_constructor): Don't build/add
gimple assignment of an empty type.

gcc/testsuite/ChangeLog:
* gcc.dg/pr87052.c: Update d var to expect nothing.

23 months agoAArch32: Fix 128-bit sequential consistency atomic operations.
Tamar Christina [Mon, 8 Aug 2022 13:37:42 +0000 (14:37 +0100)]
AArch32: Fix 128-bit sequential consistency atomic operations.

Similar to AArch64 the Arm implementation of 128-bit atomics is broken.

For 128-bit atomics we rely on pthread barriers to correct guard the address
in the pointer to get correct memory ordering.  However for 128-bit atomics the
address under the lock is different from the original pointer.

This means that one of the values under the atomic operation is not protected
properly and so we fail during when the user has requested sequential
consistency as there's no barrier to enforce this requirement.

As such users have resorted to adding an

#ifdef GCC
<emit barrier>
#endif

around the use of these atomics.

This corrects the issue by issuing a barrier only when __ATOMIC_SEQ_CST was
requested.  I have hand verified that the barriers are inserted
for atomic seq cst.

libatomic/ChangeLog:

PR target/102218
* config/arm/host-config.h (pre_seq_barrier, post_seq_barrier,
pre_post_seq_barrier): Require barrier on __ATOMIC_SEQ_CST.