official-gcc.git
5 hours agoFortran: fix issues with class(*) assignment [PR114827]mastertrunk
Harald Anlauf [Mon, 29 Apr 2024 17:52:52 +0000 (29 19:52 +0200)]
Fortran: fix issues with class(*) assignment [PR114827]

gcc/fortran/ChangeLog:

PR fortran/114827
* trans-array.cc (gfc_alloc_allocatable_for_assignment): Take into
account _len of unlimited polymorphic entities when calculating
the effective element size for allocation size and array span.
Set _len of lhs to _len of rhs.
* trans-expr.cc (trans_class_assignment): Take into account _len
of unlimited polymorphic entities for allocation size.

gcc/testsuite/ChangeLog:

PR fortran/114827
* gfortran.dg/asan/unlimited_polymorphic_34.f90: New test.

15 hours agoUpdate libbid according to the latest Intel Decimal Floating-Point Math Library.
liuhongt [Wed, 27 Mar 2024 00:20:13 +0000 (27 08:20 +0800)]
Update libbid according to the latest Intel Decimal Floating-Point Math Library.

The Intel Decimal Floating-Point Math Library is available as open-source on Netlib[1].

[1] https://www.netlib.org/misc/intel/.

libgcc/config/libbid/ChangeLog:

* bid128_fma.c (add_and_round): Fix bug: the result
of (+5E+368)*(+10E-34)+(-10E+369) was returning
-9999999999999999999999999999999999E+336 instead of expected
result -1000000000000000000000000000000000E+337.
(bid128_ext_fma): Ditto.
(bid64qqq_fma): Ditto.
* bid128_noncomp.c: Change return type of bid128_class from
int to class_t.
* bid128_round_integral.c: Add default case to avoid compiler
warning.
* bid128_string.c (bid128_to_string): Replace 0x30 with '0'
for zero digit.
(bid128_from_string): Ditto.
* bid32_to_bid128.c (bid128_to_bid32): Fix Bug. In addition
to the INEXACT flag, the UNDERFLOW flag needs to be set (and
was not) when converting an input such as
+6931674235302037148946035460357709E+1857 to +1000000E-101
* bid32_to_bid64.c (bid64_to_bid32): fix Bug, In addition to
the INEXACT flag, the UNDERFLOW flag needs to be set (and was
not) when converting an input such as +9999999000000001E-111
to +1000000E-101. Furthermore, significant bits of NaNs are
set correctly now. For example,  0x7c00003b9aca0000 was
returning 0x7c000002 instead of 0x 7c000100.
* bid64_noncomp.c: Change return type of bid64_class from int
to class_t.
* bid64_round_integral.c (bid64_round_integral_exact): Add
default case to avoid compiler warning.
* bid64_string.c (bid64_from_string): Fix bug for rounding
up. The input string "10000000000000000" was returning
+1000000000000001E+1 instead of +1000000000000000E+1.
* bid64_to_bid128.c (bid128_to_bid64): Fix bug, in addition to
the INEXACT flag, the UNDERFLOW flag needs to be set (and was
not) when converting an input such as
+9999999999999999999999999999999999E-417 to
+1000000000000000E-398.
* bid_binarydecimal.c (bid32_to_binary64): Fix bug for
conversion between binary and bid types. For example,
0x7c0F4240 was returning 0x7FFFA12000000000 instead of
expected double precision 0x7FF8000000000000.
(binary64_to_bid32): Ditto.
(binary80_to_bid32): Ditto.
(binary128_to_bid32): Ditto.
(binary80_to_bid64): Ditto.
(binary128_to_bid64): Ditto.
* bid_conf.h (BID_HIGH_128W): New macro.
(BID_LOW_128W): Ditto.
* bid_functions.h (__ENABLE_BINARY80__): Ditto.
(ALIGN): Ditto.
* bid_inline_add.h (get_add128): Add default case to avoid compiler
warning.
* bid_internal.h (get_BID64): Ditto.
(fast_get_BID64_check_OF): Ditto.
(ALIGN): New macro.

Co-authored-by: Anderson, Cristina S <cristina.s.anderson@intel.com>
Co-authored-by: Akkas, Ahmet <ahmet.akkas@intel.com>
Co-authored-by: Cornea, Marius <marius.cornea@intel.com>
23 hours agoRemove m_nloops field from loop_versioning
Andrew Pinski [Sun, 28 Apr 2024 01:54:43 +0000 (27 18:54 -0700)]
Remove m_nloops field from loop_versioning

This is a small cleanup of loop_versioning where m_nloops
is only used in the constructor so we can remove the whole
field.

Bootstrapped and tested on x86_64-linux-gnu.

gcc/ChangeLog:

* gimple-loop-versioning.cc (loop_versioning): Remove m_nloops field.
(loop_versioning::loop_versioning): Remove initialization of
m_nloops field and move it to be a local variable.
(loop_versioning::analyze_blocks): Fix formating.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
23 hours agoDaily bump.
GCC Administrator [Sun, 5 May 2024 00:16:48 +0000 (5 00:16 +0000)]
Daily bump.

39 hours agoAdd prange entries in gimple-range-op.cc.
Aldy Hernandez [Wed, 20 Mar 2024 10:27:21 +0000 (20 11:27 +0100)]
Add prange entries in gimple-range-op.cc.

gcc/ChangeLog:

* gimple-range-op.cc (class cfn_pass_through_arg1): Add overloads
for prange operations.
(cfn_strlen): Same.

39 hours agoImplement operator_ge for prange....
Aldy Hernandez [Wed, 20 Mar 2024 10:15:03 +0000 (20 11:15 +0100)]
Implement operator_ge for prange....

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_ge::fold_range): New.
(operator_ge::op1_range): New.
(operator_ge::op2_range): New.
(operator_ge::op1_op2_relation): New.
(operator_ge::pointers_handled_p): New.

39 hours agoImplement operator_gt for prange.
Aldy Hernandez [Wed, 20 Mar 2024 10:10:03 +0000 (20 11:10 +0100)]
Implement operator_gt for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_gt::fold_range): New.
(operator_gt::op1_range): New.
(operator_gt::op2_range): New.
(operator_gt::op1_op2_relation): New.
(operator_gt::pointers_handled_p): New.

39 hours agoImplement operator_le for prange.
Aldy Hernandez [Wed, 20 Mar 2024 10:07:30 +0000 (20 11:07 +0100)]
Implement operator_le for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_le::fold_range): New.
(operator_le::op1_range): New.
(operator_le::op2_range): New.
(operator_le::op1_op2_relation): New.
(operator_le::pointers_handled_p): New.

39 hours agoImplement operator_lt for prange.
Aldy Hernandez [Wed, 20 Mar 2024 10:03:24 +0000 (20 11:03 +0100)]
Implement operator_lt for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (max_limit): New.
(min_limit): New.
(build_lt): New.
(build_le): New.
(build_gt): New.
(build_ge): New.
(operator_lt::fold_range): New.
(operator_lt::op1_range): New.
(operator_lt::op2_range): New.
(operator_lt::op1_op2_relation): New.
(operator_lt::pointers_handled_p): New.

39 hours agoImplement operator_equal for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:54:39 +0000 (20 10:54 +0100)]
Implement operator_equal for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_equal::fold_range): New.
(operator_equal::op1_range): New.
(operator_equal::op2_range): New.
(operator_equal::op1_op2_relation): New.
(operator_equal::pointers_handled_p): New.

39 hours agoImplement operator_not_equal for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:49:11 +0000 (20 10:49 +0100)]
Implement operator_not_equal for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_not_equal::fold_range): New.
(operator_not_equal::op1_range): New.
(operator_not_equal::op2_range): New.
(operator_not_equal::op1_op2_relation): New.
(operator_not_equal::pointers_handled_p): New.

39 hours agoImplement operator_bitwise_or for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:29:50 +0000 (20 10:29 +0100)]
Implement operator_bitwise_or for prange.

We seem to have a range-op entry for pointer bitwise OR that we've
inherited from the original VRP implementation, but it never gets
used.  If this is not valid gimple, we can safely remove this entry.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_bitwise_or::pointers_handled_p): New.

39 hours agoImplement operator_bitwise_and for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:23:31 +0000 (20 10:23 +0100)]
Implement operator_bitwise_and for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_bitwise_and::fold_range): New.
(operator_bitwise_and::pointers_handled_p): New.

39 hours agoImplement operator_pointer_diff for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:12:47 +0000 (20 10:12 +0100)]
Implement operator_pointer_diff for prange.

gcc/ChangeLog:

* range-op-ptr.cc
(operator_pointer_diff::op1_op2_relation_effect): New.
(operator_pointer_diff::pointers_handled_p): New.

39 hours agoImplement pointer_plus_operator for prange.
Aldy Hernandez [Wed, 20 Mar 2024 09:04:41 +0000 (20 10:04 +0100)]
Implement pointer_plus_operator for prange.

gcc/ChangeLog:

* range-op-ptr.cc (class pointer_plus_operator): Add overloaded declarations
for pointer variants.
(pointer_plus_operator::fold_range): New.
(pointer_plus_operator::op2_range): New.
(pointer_plus_operator::pointers_handled_p): New.

39 hours agoImplement operator_addr_expr for prange.
Aldy Hernandez [Wed, 20 Mar 2024 08:51:33 +0000 (20 09:51 +0100)]
Implement operator_addr_expr for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_addr_expr::op1_range): New.
(operator_addr_expr::pointers_handled_p): New.

39 hours agoImplement operator_min and operator_max for prange.
Aldy Hernandez [Wed, 20 Mar 2024 07:44:49 +0000 (20 08:44 +0100)]
Implement operator_min and operator_max for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_min::fold_range): New.
(operator_min::pointers_handled_p): New.
(operator_max::fold_range): New.
(operator_max::pointers_handled_p): New.

39 hours agoImplement operator_cast for prange.
Aldy Hernandez [Wed, 20 Mar 2024 07:04:32 +0000 (20 08:04 +0100)]
Implement operator_cast for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_cast::fold_range): New.
(operator_cast::op1_range): New.
(operator_cast::lhs_op1_relation): New.
(operator_cast::pointers_handled_p): New.

39 hours agoImplement operator_cst for prange.
Aldy Hernandez [Wed, 20 Mar 2024 06:55:57 +0000 (20 07:55 +0100)]
Implement operator_cst for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for pointer variants.
* range-op-ptr.cc (operator_cst::fold_range): New.
(operator_cst::pointers_handled_p): New.

39 hours agoImplement operator_identity for prange.
Aldy Hernandez [Wed, 20 Mar 2024 06:50:11 +0000 (20 07:50 +0100)]
Implement operator_identity for prange.

gcc/ChangeLog:

* range-op-mixed.h: Add overloaded declarations for fold_range, op1_range,
lhs_op1_relation, pointers_handled_p.
* range-op-ptr.cc (operator_identity::fold_range): New.
(operator_identity::lhs_op1_relation): New.
(operator_identity::op1_range): New.
(operator_identity::pointers_handled_p): New.

39 hours agoImplement range-op dispatch for prange.
Aldy Hernandez [Wed, 20 Mar 2024 06:19:45 +0000 (20 07:19 +0100)]
Implement range-op dispatch for prange.

This patch adds the range-op dispatch code for prange, and adds some
temporary sanity checks (for flag_checking only) to make sure we handle
all the pointer/integer variants.

In order to make sure I got all the combinations right, I started with
a clean slate, trapping on all pointer operands.  Then I added support
for each one piecemeal.  To verify the work, I added a
pointers_handled_p() helper that is implemented for each range-op
entry and returns TRUE iff the operator can handle a given combination
of pointers.  If this helper returns false, we will trap, because it
indicates an operator that was not implemented.  This is temporary
checking code, and I will rip it out once the the dust has
settled in a few days.

gcc/ChangeLog:

* range-op-mixed.h: Add using declarator for all classes.
* range-op-ptr.cc (range_operator::pointers_handled_p): New.
(range_operator::fold_range): New.
(range_operator::op1_op2_relation_effect): New.
(range_operator::op1_range): New.
(range_operator::op2_range): New.
(range_operator::op1_op2_relation): New.
(range_operator::lhs_op1_relation): New.
(range_operator::update_bitmask): New.
(class pointer_plus_operator): New.
(class operator_pointer_diff): New.
(class hybrid_min_operator): New.
(class hybrid_max_operator): New.
* range-op.cc: Add RO_PPP, RO_PPI, RO_IPP, RO_IPI, RO_PIP, RO_PII.
(range_op_handler::discriminator_fail): New.
(has_pointer_operand_p): New.
(range_op_handler::fold_range): Add pointer support.
(range_op_handler::op1_range): Same.
(range_op_handler::op2_range): Same.
(range_op_handler::lhs_op1_relation): Same.
(range_op_handler::lhs_op2_relation): Same.
(range_op_handler::op1_op2_relation): Same.
(class operator_div): Add using.
(class operator_lshift): Add using.
(class operator_rshift):Add using.
(class operator_trunc_mod):Add using.
(class operator_absu):Add using.
* range-op.h (enum range_op_dispatch_type): New.
Add extern definitions for RO_*.

39 hours agoAdd prange implementation for get_legacy_range.
Aldy Hernandez [Wed, 20 Mar 2024 05:39:48 +0000 (20 06:39 +0100)]
Add prange implementation for get_legacy_range.

gcc/ChangeLog:

* value-range.cc (get_legacy_range): New version for prange.

39 hours agoAdd hashing support for prange.
Aldy Hernandez [Wed, 20 Mar 2024 05:39:07 +0000 (20 06:39 +0100)]
Add hashing support for prange.

gcc/ChangeLog:

* value-range.cc (add_vrange): Add prange support.

39 hours agoAdd storage support for prange.
Aldy Hernandez [Wed, 20 Mar 2024 05:38:06 +0000 (20 06:38 +0100)]
Add storage support for prange.

gcc/ChangeLog:

* value-range-storage.cc (vrange_allocator::clone_varying): Add
prange support.
(vrange_allocator::clone_undefined): Same.
(vrange_storage::alloc): Same.
(vrange_storage::set_vrange): Same.
(vrange_storage::get_vrange): Same.
(vrange_storage::fits_p): Same.
(vrange_storage::equal_p): Same.
(prange_storage::alloc): New.
(prange_storage::prange_storage): New.
(prange_storage::set_prange): New.
(prange_storage::get_prange): New.
(prange_storage::equal_p): New.
(prange_storage::fits_p): New.
* value-range-storage.h (class prange_storage): Add prange support.

39 hours agoAdd streaming support for prange.
Aldy Hernandez [Wed, 20 Mar 2024 05:34:26 +0000 (20 06:34 +0100)]
Add streaming support for prange.

gcc/ChangeLog:

* data-streamer-in.cc (streamer_read_value_range): Add prange support.
* data-streamer-out.cc (streamer_write_vrange): Same.

39 hours agoImplement basic prange class.
Aldy Hernandez [Wed, 20 Mar 2024 05:25:52 +0000 (20 06:25 +0100)]
Implement basic prange class.

This provides a bare prange class with bounds and bitmasks.  It will
be a drop-in replacement for pointer ranges, so we can pull their
support from irange.  The range-op code will be contributed as a
follow-up.

The code is disabled by default, as irange::supports_p still accepts
pointers:

inline bool
irange::supports_p (const_tree type)
{
  return INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type);
}

Once the prange operators are implemented in range-ops, pointer
support will be removed from irange to activate pranges.

gcc/ChangeLog:

* value-range-pretty-print.cc (vrange_printer::visit): New.
* value-range-pretty-print.h: Declare prange visit() method.
* value-range.cc (vrange::operator=): Add prange support.
(vrange::operator==): Same.
(prange::accept): New.
(prange::set_nonnegative): New.
(prange::set): New.
(prange::contains_p): New.
(prange::singleton_p): New.
(prange::lbound): New.
(prange::ubound): New.
(prange::union_): New.
(prange::intersect): New.
(prange::operator=): New.
(prange::operator==): New.
(prange::invert): New.
(prange::verify_range): New.
(prange::update_bitmask): New.
(range_tests_misc): Use prange.
* value-range.h (enum value_range_discriminator): Add VR_PRANGE.
(class prange): New.
(Value_Range::init): Add prange support.
(Value_Range::operator=): Same.
(Value_Range::supports_type_p): Same.
(prange::prange):  New.
(prange::supports_p): New.
(prange::supports_type_p): New.
(prange::set_undefined): New.
(prange::set_varying): New.
(prange::set_nonzero): New.
(prange::set_zero): New.
(prange::contains_p): New.
(prange::zero_p): New.
(prange::nonzero_p): New.
(prange::type): New.
(prange::lower_bound): New.
(prange::upper_bound): New.
(prange::varying_compatible_p): New.
(prange::get_bitmask): New.
(prange::fits_p): New.

39 hours agoMinimal prange class showing inlining degradation to VRP.
Aldy Hernandez [Sat, 4 May 2024 04:45:18 +0000 (4 06:45 +0200)]
Minimal prange class showing inlining degradation to VRP.

There is a 2% slowdown to VRP unrelated to the work at hand.  This patch
is a skeleton implementation of prange that exhibits this degradation.  It
is meant as a place in the commit history we can return to in order to revisit
the issue.

The relevant discussion is here:

https://gcc.gnu.org/pipermail/gcc/2024-May/243898.html

gcc/ChangeLog:

* value-range.h (class prange): New.

47 hours agoDaily bump.
GCC Administrator [Sat, 4 May 2024 00:16:30 +0000 (4 00:16 +0000)]
Daily bump.

2 days agolibbacktrace: add DLLS as they are loaded
Ian Lance Taylor [Fri, 3 May 2024 22:23:23 +0000 (3 15:23 -0700)]
libbacktrace: add DLLS as they are loaded

Patch from Björn Schäpers.

* pecoff.c (struct dll_notification_data): Define.
(LDR_DLL_NOTIFICATION): New typedef.
(LDR_REGISTER_FUNCTION): New typedef.
(struct dll_notification_context): Define.
(dll_notification): New static function.
(backtrace_initialize): Register DLL notification.

2 days agoFix printing COMPOUND_EXPR in .original [PR23872]
Andrew Pinski [Wed, 17 Apr 2024 00:43:36 +0000 (16 17:43 -0700)]
Fix printing COMPOUND_EXPR in .original [PR23872]

Starting with the merge of the openmp branch into the trunk
(r0-73077-g953ff28998b59b), COMPOUND_EXPR started to be printed
as `expr; , expr` which is wrong. This was due to the wrong
conversion of dumping_stmts into `!(flags & TDF_SLIM)`. That is wrong
as we are not dumping stmts at this point (`!(flags & TDF_SLIM)` was always
true for this case as TDF_SLIM case was handled before hand). So switch it
to be always false.

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

gcc/ChangeLog:

PR middle-end/23872
* tree-pretty-print.cc (dump_generic_node <case COMPOUND_EXPR>): Fix
calls to dump_generic_node and also remove unreachable code that is testing
`flags & TDF_SLIM`.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/atomic-21.f90: Update testcase for the removal of `;`.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
2 days agoc++: initializer_list<string> and EH [PR114935]
Jason Merrill [Fri, 3 May 2024 13:52:46 +0000 (3 09:52 -0400)]
c++: initializer_list<string> and EH [PR114935]

When we initialize an array of a type with a non-trivial destructor, such as
the backing array for the initializer_list, we have a cleanup to destroy any
constructed elements if a later constructor throws.  When the array being
created is a variable, the end of that EH region naturally coincides with
the beginning of the EH region for the cleanup for the variable as a whole.

But if the array is a temporary, or a subobject of one, the array cleanup
region lasts for the rest of the full-expression, along with the normal
cleanup for the TARGET_EXPR.  As a result, when tata throws we clean it up
twice.  Before r14-1705 we avoided this by disabling the array cleanup in
split_nonconstant_init, but after that we don't go through
split_nonconstant_init, so let's handle it in cp_genericize_target_expr.

PR c++/114935

gcc/cp/ChangeLog:

* cp-gimplify.cc (cp_genericize_init): Add flags parm.
(cp_genericize_init_expr): Pass nullptr.
(cp_genericize_target_expr): Handle cleanup flags.
* typeck2.cc (build_disable_temp_cleanup): Factor out of...
(split_nonconstant_init): ...here.
* cp-tree.h (build_disable_temp_cleanup): Declare.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/initlist-eh1.C: New test.

2 days agoPR modula2/114929 extra for loop iteration count regression tests
Gaius Mulley [Fri, 3 May 2024 19:48:01 +0000 (3 20:48 +0100)]
PR modula2/114929 extra for loop iteration count regression tests

This patch introduces three more for loop tests checking the iteration
count using the CHAR and enumeration data types.

gcc/testsuite/ChangeLog:

PR modula2/114929
* gm2/pim/run/pass/testforloopchar.mod: New test.
* gm2/pim/run/pass/testforloopchar2.mod: New test.
* gm2/pim/run/pass/testforloopenum.mod: New test.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
2 days agoc++: Implement __is_pointer built-in trait
Ken Matsui [Mon, 10 Jul 2023 00:49:08 +0000 (9 17:49 -0700)]
c++: Implement __is_pointer built-in trait

This patch implements built-in trait for std::is_pointer.

gcc/cp/ChangeLog:

* cp-trait.def: Define __is_pointer.
* constraint.cc (diagnose_trait_expr): Handle CPTK_IS_POINTER.
* semantics.cc (trait_expr_value): Likewise.
(finish_trait_expr): Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/ext/has-builtin-1.C: Test existence of __is_pointer.
Arrange the order lexically around __is_pointer.
* g++.dg/ext/is_pointer.C: New test.

Signed-off-by: Ken Matsui <kmatsui@gcc.gnu.org>
Reviewed-by: Jason Merrill <jason@redhat.com>
2 days agoc++: Implement __is_volatile built-in trait
Ken Matsui [Wed, 22 Mar 2023 23:26:25 +0000 (22 16:26 -0700)]
c++: Implement __is_volatile built-in trait

This patch implements built-in trait for std::is_volatile.

gcc/cp/ChangeLog:

* cp-trait.def: Define __is_volatile.
* constraint.cc (diagnose_trait_expr): Handle CPTK_IS_VOLATILE.
* semantics.cc (trait_expr_value): Likewise.
(finish_trait_expr): Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/ext/has-builtin-1.C: Test existence of __is_volatile.
* g++.dg/ext/is_volatile.C: New test.

Signed-off-by: Ken Matsui <kmatsui@gcc.gnu.org>
Reviewed-by: Jason Merrill <jason@redhat.com>
2 days agoc++: Implement __is_const built-in trait
Ken Matsui [Tue, 21 Mar 2023 16:10:52 +0000 (21 09:10 -0700)]
c++: Implement __is_const built-in trait

This patch implements built-in trait for std::is_const.

gcc/cp/ChangeLog:

* cp-trait.def: Define __is_const.
* constraint.cc (diagnose_trait_expr): Handle CPTK_IS_CONST.
* semantics.cc (trait_expr_value): Likewise.
(finish_trait_expr): Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/ext/has-builtin-1.C: Test existence of __is_const.
* g++.dg/ext/is_const.C: New test.

Signed-off-by: Ken Matsui <kmatsui@gcc.gnu.org>
Reviewed-by: Jason Merrill <jason@redhat.com>
2 days agoRISC-V: miscll comment fixes [NFC]
Vineet Gupta [Tue, 1 Mar 2022 11:45:19 +0000 (1 03:45 -0800)]
RISC-V: miscll comment fixes [NFC]

gcc/ChangeLog:
* config/riscv/riscv.cc: Comment updates.
* config/riscv/riscv.h: Ditto.

Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
2 days agodocs: rtl: document GET_MODE_INNER
Vineet Gupta [Thu, 17 Feb 2022 21:49:45 +0000 (17 13:49 -0800)]
docs: rtl: document GET_MODE_INNER

gcc/ChangeLog
* doc/rtl.texi: Add entry for GET_MODE_INNER.

Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
2 days agotestsuite: fix analyzer C++ failures on Solaris [PR111475]
David Malcolm [Fri, 3 May 2024 13:05:29 +0000 (3 09:05 -0400)]
testsuite: fix analyzer C++ failures on Solaris [PR111475]

As part of PR analyzer/96395, these patches moved testcases from
gcc.dg/analyzer to c-c++-common/analyzer:
r14-3503-g55f6a7d949abc7
r14-3823-g50b5199cff6908
r14-6564-gae034b9106fbdd

Unfortunately this led to numerous g++ testsuite failures on Solaris,
tracked as PR analyzer/111475.

Almost all of the failures are due to standard library differences where
including a C standard library on C++ e.g. <stdlib.h> leads to the plain
symbols referencing the symbols "std::" via a "using" declaration,
whereas I had written the code expecting them to use symbols in the root
namespace.

The analyzer has special-case handling of many functions by name.
This patch generalizes such handling to also match against functions
in "std::" for all of the cases I found in the testsuite (via manual
inspection of the preprocessed test cases against Solaris headers).
This fixes cases where the analyzer was failing to "know about" the
behavior of such functions.

Other such failures are due to "std::" prefixes appearing in names of
functions in the output, leading to mismatches against expected output.
The patch adds regexes to some cases, and moves some other cases back
from c-c++-common to gcc.dg where the dg-multiline syntax isn't
expressive enough.

Various "fd-*.c" failures relate to Solaris's socket-handling functions
not being marked with "noexcept", where due to PR analyzer/97111 we
mishandle the exception-handling edges in the CFG, leading to leak
false positives.  The patch works around this by adding -fno-exceptions
to these cases, pending a proper fix for PR analyzer/97111.

gcc/analyzer/ChangeLog:
PR analyzer/111475
* analyzer.cc (is_special_named_call_p): Add "look_in_std" param.
(is_std_function_p): Make non-static.
* analyzer.h (is_special_named_call_p): Add optional "look_in_std"
param.
(is_std_function_p): New decl.
* engine.cc (stmt_requires_new_enode_p): Look for both "signal"
and "std::signal".
* kf.cc (register_known_functions): Add various "std::" copies
of the known functions.
* known-function-manager.cc
(known_function_manager::~known_function_manager): Clean up
m_std_ns_map_id_to_kf.
(known_function_manager::add_std_ns): New.
(known_function_manager::get_match): Also look for known "std::"
functions.
(known_function_manager::get_by_identifier_in_std_ns): New.
* known-function-manager.h
(known_function_manager::add_std_ns): New decl.
(known_function_manager::get_by_identifier_in_std_ns): New decl.
(known_function_manager::m_std_ns_map_id_to_kf): New field.
* sm-file.cc (register_known_file_functions): Add various "std::"
copies of the known functions.
* sm-malloc.cc (malloc_state_machine::on_stmt): Handle
"std::realloc".
* sm-signal.cc (signal_unsafe_p): Consider "std::" copies of the
functions as also being async-signal-unsafe.
(signal_state_machine::on_stmt): Consider "std::signal".

gcc/testsuite/ChangeLog:
PR analyzer/111475
* c-c++-common/analyzer/fd-glibc-byte-stream-socket.c: Add
-fno-exceptions for now.
* c-c++-common/analyzer/fd-manpage-getaddrinfo-client.c: Likewise.
* c-c++-common/analyzer/fd-mappage-getaddrinfo-server.c: Rename to...
* c-c++-common/analyzer/fd-manpage-getaddrinfo-server.c: ...this, and
add -fno-exceptions for now.
* c-c++-common/analyzer/fd-socket-meaning.c: Add -fno-exceptions
for now.
* c-c++-common/analyzer/fd-symbolic-socket.c: Likewise.
* c-c++-common/analyzer/flexible-array-member-1.c: Use regexp to
handle C vs C++ differences in spelling of function name, which
could have a "std::" prefix on some targets.
* c-c++-common/analyzer/pr106539.c: Likewise.
* c-c++-common/analyzer/malloc-ipa-8-unchecked.c: Move back to...
* gcc.dg/analyzer/malloc-ipa-8-unchecked.c: ...here, dropping
attempt to generalize output for C vs C++.
* c-c++-common/analyzer/signal-4a.c: Move back to...
* gcc.dg/analyzer/signal-4a.c: ...here, dropping attempt to
generalize output for C vs C++.
* c-c++-common/analyzer/signal-4b.c: Move back to...
* gcc.dg/analyzer/signal-4b.c: ...here, dropping attempt to
generalize output for C vs C++.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
2 days agoAdd default bitmap obstack allocation check
Richard Biener [Fri, 5 Apr 2024 09:02:46 +0000 (5 11:02 +0200)]
Add default bitmap obstack allocation check

The following adds a check that the global bitmap obstack is initialized
when allocating a bitmap from it.

* bitmap.cc (bitmap_alloc): When using the global bitmap obstack
assert that is initialized.

2 days agoSilence two instances of -Wcalloc-transposed-args
Peter Damianov [Fri, 3 May 2024 08:02:35 +0000 (3 01:02 -0700)]
Silence two instances of -Wcalloc-transposed-args

libgcc/
* libgcov-util.c (tag_counters): Swap order of arguments to xcalloc.
(topen_to_memory_representation): Likewise.

Signed-off-by: Peter Damianov <peter0x44@disroot.org>
2 days agolibstdc++: Update powerpc-linux-gnu baseline_symbols
Andreas Schwab [Sat, 3 Feb 2024 13:00:41 +0000 (3 14:00 +0100)]
libstdc++: Update powerpc-linux-gnu baseline_symbols

* config/abi/post/powerpc-linux-gnu/baseline_symbols.txt: Update.
* config/abi/post/powerpc64-linux-gnu/32/baseline_symbols.txt: Update.

2 days agoAvoid changing type in the type_hash_canon hash
Richard Biener [Fri, 3 May 2024 09:48:07 +0000 (3 11:48 +0200)]
Avoid changing type in the type_hash_canon hash

When building a type and type_hash_canon returns an existing type
avoid changing it, in particular its TYPE_CANONICAL.

PR middle-end/114931
* tree.cc (build_array_type_1): Return early when type_hash_canon
returned an older existing type.
(build_function_type): Likewise.
(build_method_type_directly): Likewise.
(build_offset_type): Likewise.

2 days agocfgrtl: Fix MEM_EXPR update in duplicate_insn_chain [PR114924]
Alex Coplan [Fri, 3 May 2024 08:23:59 +0000 (3 09:23 +0100)]
cfgrtl: Fix MEM_EXPR update in duplicate_insn_chain [PR114924]

The PR shows that when cfgrtl.cc:duplicate_insn_chain attempts to
update the MR_DEPENDENCE_CLIQUE information for a MEM_EXPR we can end up
accidentally dropping (e.g.) an ARRAY_REF from the MEM_EXPR and end up
replacing it with the underlying MEM_REF.  This leads to an
inconsistency in the MEM_EXPR information, and could lead to wrong code.

While the walk down to the MEM_REF is necessary to update
MR_DEPENDENCE_CLIQUE, we should use the outer tree expression for the
MEM_EXPR.  This patch does that.

gcc/ChangeLog:

PR rtl-optimization/114924
* cfgrtl.cc (duplicate_insn_chain): When updating MEM_EXPRs,
don't strip (e.g.) ARRAY_REFs from the final MEM_EXPR.

2 days agotree-inline: Add __builtin_stack_{save,restore} pair about inline calls with calls...
Jakub Jelinek [Fri, 3 May 2024 07:44:30 +0000 (3 09:44 +0200)]
tree-inline: Add __builtin_stack_{save,restore} pair about inline calls with calls to alloca [PR113596]

The following patch adds save_NNN = __builtin_stack_save (); ...
__builtin_stack_restore (save_NNN);
pair around inline calls which call alloca (alloca calls because of
VLA vars are ignored in that decision).
The patch doesn't change anything on whether we try to inline such calls or
not, it just fixes the behavior when we inline them despite those checks.
The stack save/restore restores the behavior that alloca acquired regions
are freed at the end of the containing call.

2024-05-03  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/113596
* tree-inline.cc (expand_call_inline): Emit __builtin_stack_save
and __builtin_stack_restore calls around inlined functions which
call alloca.

* gcc.dg/pr113596.c: New test.
* gcc.dg/tree-ssa/pr113596.c: New test.

2 days agotree-optimization/114921 - _Float16 -> __bf16 isn't noop
Richard Biener [Thu, 2 May 2024 11:55:15 +0000 (2 13:55 +0200)]
tree-optimization/114921 - _Float16 -> __bf16 isn't noop

The vectorizer handles a _Float16 to __bf16 conversion through
vectorizable_assignment, thinking it's a noop.  The following
fixes this by requiring the same vector component mode when
checking for CONVERT_EXPR_CODE_P, being stricter than for
VIEW_CONVERT_EXPR.

PR tree-optimization/114921
* tree-vect-stmts.cc (vectorizable_assignment): Require
same vector component modes for input and output for
CONVERT_EXPR_CODE_P.

2 days agoc++: remove lookup_template_class's entering_scope flag
Patrick Palka [Fri, 3 May 2024 01:14:30 +0000 (2 21:14 -0400)]
c++: remove lookup_template_class's entering_scope flag

lookup_template_class's entering_scope flag controls whether to prefer
returning the primary template type A<T> instead of the corresponding
implicit instantiation A<T>.  When we want to set this flag as part of
substitution, we need to use tsubst_aggr_type which also takes this flag
as a parameter.  But having this separate entry point to type substitution
turned out to be subtly problematic because it doesn't reuse typedefs
like tsubst does, which r13-4729-gbe124477b38a71 fixed in a way that
respects the flag after the fact, by adjusting the entering_scope=false
result of lookup_template_class as if entering_scope=true was passed.

But if that's possible then it means lookup_template_class's
entering_scope flag is not necessary after all -- we can just do the
after-the-fact adjustment everywhere that we currently pass
entering_scope=true to it and tsubst_aggr_type.

To that end, this patch replaces this flag with an adjustment function
adjust_type_for_entering_scope, to be used whereever we currently need
the entering_scope=true behavior.  In turn we can get rid of
tsubst_aggr_type since the only reason we needed this entry point
was to be able to pass entering_scope=true to lookup_template_class.

gcc/cp/ChangeLog:

* coroutines.cc (instantiate_coro_traits): Adjust call to
lookup_template_class.
(instantiate_coro_handle_for_promise_type): Likewise.
* cp-tree.h (adjust_type_for_entering_scope): Declare.
(lookup_template_class): Adjust declaration.
* decl.cc (make_typename_type): Adjust call to
lookup_template_class. Likewise.
(get_tuple_size): Likewise.
(get_tuple_element_type): Likewise.
* pt.cc (adjust_type_for_entering_scope): Define.
(tsubst_entering_scope): Define.
(lookup_template_class): Remove entering_scope parameter.
Replace tsubst_aggr_type call with tsubst_entering_scope.
(tsubst_aggr_type): Remove.
(tsubst_aggr_type_1): Inline into tsubst.
(tsubst_function_decl): Replace tsubst_aggr_type call
with tsubst_entering_scope.
(tsubst_template_decl): Likewise.
(tsubst_decl): Likewise.
(tsubst) <case RECORD_TYPE, UNION_TYPE, ENUMERAL_TYPE>:
Inlined from tsubst_aggr_type_1.
<case BOUND_TEMPLATE_TEMPLATE_PARM>: Adjust calls to
lookup_template_class.
<case TYPENAME_TYPE>: Replace tsubst_aggr_type call with
tsubst_entering_scope.
<case UNBOUND_CLASS_TEMPLATE>: Likewise.
Increment processing_template_decl when substituting the
context.
(tsubst_expr) <case FIELD_DECL>: Replace tsubst_aggr_type
call with tsubst_entering_scope.
<case TEMPLATE_DECL>: Likewise.
(instantiate_template): Likewise.
(resolve_typename_type): Adjust lookup_template_class call
and call adjust_type_for_entering_scope afterward.
(listify): Adjust lookup_template_class call.
(alias_ctad_tweaks): Likewise.
* semantics.cc (finish_template_type): Adjust lookup_template_class
call and maybe call adjust_type_for_entering_scope afterward.

Reviewed-by: Jason Merrill <jason@redhat.com>
2 days agoPR modula2/114929 for loop fails to iterate down to zero when using a cardinal type
Gaius Mulley [Fri, 3 May 2024 00:22:10 +0000 (3 01:22 +0100)]
PR modula2/114929 for loop fails to iterate down to zero when using a cardinal type

There is a bug in the for loop control code which is exposed when an
unsigned type is used in the iterator variable.  See
gm2/pim/run/pass/testforloopzero[234].mod.  The bug is in the
calculation of the last iterator value.  The bug fix is to avoid using
negative expressions when calculating the last iterator value with a
negative step value.  This patch detects if e1, e2, step value are all
constant, in which case the ztype is used internally and there is no
overflow.  If the last iterator value is held in a variable then it
uses a different method to calculate the last iterator depending upon
the sign of the step value.

gcc/m2/ChangeLog:

PR modula2/114929
* gm2-compiler/M2LangDump.mod (GenQualidentSymString): Add
missing return result into identstr.
* gm2-compiler/M2Quads.mod (ForLoopLastIteratorVariable): New
procedure.
(ForLoopLastIteratorConstant): Ditto.
(ForLoopLastIterator): Ditto.
(BuildForToByDo): Remove LastIterator calculation and call
ForLoopLastIterator instead.
(FinalValue): Replace with ...
(LastIterator): ... this.

gcc/testsuite/ChangeLog:

PR modula2/114929
* gm2/pim/run/pass/testforloopzero.mod: New test.
* gm2/pim/run/pass/testforloopzero2.mod: New test.
* gm2/pim/run/pass/testforloopzero3.mod: New test.
* gm2/pim/run/pass/testforloopzero4.mod: New test.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
2 days agoDaily bump.
GCC Administrator [Fri, 3 May 2024 00:17:26 +0000 (3 00:17 +0000)]
Daily bump.

3 days ago[committed][RISC-V] Fix nearbyint failure on rv32 and formatting nits
Jeff Law [Thu, 2 May 2024 23:13:12 +0000 (2 17:13 -0600)]
[committed][RISC-V] Fix nearbyint failure on rv32 and formatting nits

The CI system tripped an execution failure for rv32 with the ceil/round patch.

The fundamental problem is the FP->INT step in these sequences requires the
input size to match the output size.  The output size was based on rv32/rv64.
Meaning that we'd try to do DF->SI->DF.

That doesn't preserve the semantics we want in at least two ways.

The net is we can't use this trick for DF values on rv32.  While inside the
code I realized we had a similar problem for HF modes.  HF modes we can support
only for Zfa.  So I fixed that proactively.

The CI system also pointed out various formatting nits.  I think this fixes all
but one overly long line.

Note I could have factored the TARGET_ZFA test.  But I think as-written it's
clearer what the desired cases to transform are.

gcc/
* config/riscv/riscv.md (<round_pattern><ANYF:mode>2): Adjust
condition to match what can be properly implemented.  Fix various
formatting issues.
(l<round_pattern><ANYF:mode>si2_sext): Fix formatting

3 days agolibgfortran: Fix up the autoreconf warnings
Francois-Xavier Coudert [Thu, 2 May 2024 21:03:37 +0000 (2 23:03 +0200)]
libgfortran: Fix up the autoreconf warnings

This means using sub-dirs and amending some of the recipes accordingly.

libgfortran/ChangeLog:

* Makefile.am: Use sub-dirs, amend recipies accordingly.
* Makefile.in: Regenerate.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
3 days ago[RFA][RISC-V] Improve constant synthesis for constants with 2 bits set
Jeff Law [Thu, 2 May 2024 20:06:22 +0000 (2 14:06 -0600)]
[RFA][RISC-V] Improve constant synthesis for constants with 2 bits set

In doing some preparation work for using zbkb's pack instructions for constant
synthesis I figured it would be wise to get a sense of how well our constant
synthesis is actually working and address any clear issues.

So the first glaring inefficiency is in our handling of constants with a small
number of bits set.  Let's start with just two bits set.   There are 2016
distinct constants in that space (rv64).  With Zbs enabled the absolute worst
we should ever do is two instructions (bseti+bseti).  Yet we have 503 cases
where we're generating 3+ instructions when there's just two bits set in the
constant.  A constant like 0x8000000000001000 generates 4 instructions!

This patch adds bseti (and indirectly binvi if we needed it) as a first class
citizen for constant synthesis.  There's two components to this change.

First, we can't generate an IOR with a constant like (1 << 45) as an operand.
The IOR/XOR define_insn is in riscv.md.  The constant argument for those
patterns must match an arith_operand which means its not really usable for
generating bseti directly in the cases we care about (at least one of the bits
will be in the 32..63 range and thus won't match arith_operand).

We have a few things we could do.  One would be to extend the existing pattern
to incorporate bseti cases.  But I suspect folks like the separation of the
base architecture (riscv.md) from the Zb* extensions (bitmanip.md).  We could
also try to generate the RTL for bseti
directly, bypassing gen_fmt_ee (which forces undesirable constants into registers based on the predicate of the appropriate define_insn). Neither of these seemed particularly appealing to me.

So what I've done instead is to make ior/xor a define_expand and have the
expander allow a wider set of constant operands when Zbs is enabled.  That
allows us to keep the bulk of Zb* support inside bitmanip.md and continue to
use gen_fmt_ee in the constant synthesis paths.

Note the code generation in this case is designed to first set as many bits as
we can with lui, then with addi since those can both set multiple bits at a
time.  If there are any residual bits left to set we can emit bseti
instructions up to the current cost ceiling.

This results in fixing all of the 503 2-bit set cases where we emitted too many
instructions.  It also significantly helps other scenarios with more bits set.

The testcase I'm including verifies the number of instructions we generate for
the full set of 2016 possible cases.  Obviously this won't be possible as we
increase the number of bits (there are something like 48k cases with just 3
bits set).

gcc/

* config/riscv/predicates.md (arith_or_zbs_operand): New predicate.
* config/riscv/riscv.cc (riscv_build_integer_one): Use bseti to set
single bits when profitable.
* config/riscv/riscv.md (*<optab><mode>3): Renamed with '*' prefix.
(<optab><mode>3): New expander for IOR/XOR.

gcc/testsuite
* gcc.target/riscv/synthesis-1.c: New test.

3 days agoRegenerate gcc.pot
Joseph Myers [Thu, 2 May 2024 19:43:15 +0000 (2 19:43 +0000)]
Regenerate gcc.pot

* gcc.pot: Regenerate.

3 days agoRISC-V: Add testcase for pr114734
Patrick O'Neill [Tue, 30 Apr 2024 20:26:45 +0000 (30 13:26 -0700)]
RISC-V: Add testcase for pr114734

gcc/testsuite/ChangeLog:

PR middle-end/114734

* gcc.target/riscv/rvv/autovec/pr114734.c: New test.

Signed-off-by: Patrick O'Neill <patrick@rivosinc.com>
3 days ago[committed] [RISC-V] Don't run new rounding tests on newlib risc-v targets
Jeff Law [Thu, 2 May 2024 14:42:32 +0000 (2 08:42 -0600)]
[committed] [RISC-V] Don't run new rounding tests on newlib risc-v targets

The new round_32.c and round_64.c tests depend on the optimizers to recognize
the conversions feeding the floor/ceil calls and convert them into ceilf,
floorf and the like.

Those transformations only occur when the target indicates the C library has
the appropriate routines (fnclass == function_c99_misc).  While newlib has
these routines, they are not exposed as available to the compiler and thus the
transformation the tests depend on do not happen. Naturally the scan-tests then
fail.

gcc/testsuite
* gcc.target/riscv/round_32.c: Add require-effective-target glibc.
* gcc.target/riscv/round_64.c: Likewise.

3 days agoc++: Clear is_unbraced_* when parsing declaration_seq_opt [PR114917]
Nathaniel Shead [Thu, 2 May 2024 02:55:24 +0000 (2 12:55 +1000)]
c++: Clear is_unbraced_* when parsing declaration_seq_opt [PR114917]

Currently we incorrectly retain "in_unbraced_linkage_specification_p"
and "in_unbraced_export_declaration_p" when parsing a (braced)
declaration-seq.  This patch ensures that we clear these flags before
parsing the toplevel declarations.

Strictly speaking we don't need to save and restore the flags around the
parsing because there's currently no way to provide new declarations
within the unbraced context after the closing brace, but this patch does
it anyway in case this ever changes and for consistency with other
places that these flags are adjusted.

PR c++/114917

gcc/cp/ChangeLog:

* parser.cc (cp_parser_declaration_seq_opt): Clear
parser->in_unbraced_* flags when parsing toplevel declarations.

gcc/testsuite/ChangeLog:

* g++.dg/modules/export-5_a.C: New test.
* g++.dg/modules/export-5_b.C: New test.
* g++.dg/parse/linkage4.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
3 days agomodula2: Regenerate libgm2 Makefile.ins using correct include order
Gaius Mulley [Thu, 2 May 2024 13:38:51 +0000 (2 14:38 +0100)]
modula2: Regenerate libgm2 Makefile.ins using correct include order

Regenerated libgm2/Makefile.in (and subdir Makefile.in) using
aclocal -I .. -I ../config (or autoreconf).

libgm2/ChangeLog:

* Makefile.in: Regenerate.
* libm2cor/Makefile.in: Ditto.
* libm2iso/Makefile.in: Ditto.
* libm2log/Makefile.in: Ditto.
* libm2min/Makefile.in: Ditto.
* libm2pim/Makefile.in: Ditto.
* aclocal.m4: Ditto.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
3 days agoImprove SLP dump and graph
Richard Biener [Wed, 27 Mar 2024 09:53:11 +0000 (27 10:53 +0100)]
Improve SLP dump and graph

The following notes which lanes are considered live and adds an overload
to produce a graphviz graph for multiple entries into an SLP graph.

* tree-vect-slp.cc (vect_print_slp_tree): Mark live lanes.
(dot_slp_tree): New overload for multiple entries.

3 days agoObjective-C, NeXT, v2: Correct a regression in code-gen.
Iain Sandoe [Tue, 30 Apr 2024 14:11:56 +0000 (30 15:11 +0100)]
Objective-C, NeXT, v2: Correct a regression in code-gen.

There have been several changes in the ABI of Objective-C which
depend on the OS version targetted.  In this case Protocols and
LabelProtocols should be made weak/hidden/extern from macOS 10.7
however there was a mistake in the code causing this to occur
from macOS 10.6.  Fixed thus.

gcc/objc/ChangeLog:

* objc-next-runtime-abi-02.cc (WEAK_PROTOCOLS_AFTER): New.
(next_runtime_abi_02_protocol_decl): Use WEAK_PROTOCOLS_AFTER
to determine this ABI change.
(build_v2_protocol_list_address_table): Likewise.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
3 days agoPR modula2/113836 gm2 does not dump gimple or quadruples to a file
Gaius Mulley [Thu, 2 May 2024 12:16:07 +0000 (2 13:16 +0100)]
PR modula2/113836 gm2 does not dump gimple or quadruples to a file

This patch completes the implementation of dumping the intermediate forms
to file.  It implements the filtering on symbol rules.  Filtering can be
performed through the full text name (given to the GCC tree) or qualified
modula-2 symbol or filename:qualident.

gcc/ChangeLog:

PR modula2/113836
* doc/gm2.texi (Compiler options): Add -fm2-debug-trace=,
-fm2-dump, -fm2-dump-decl=, -fm2-dump-gimple=, -fm2-dump-quad=
and -fm2-dump-filter=.

gcc/m2/ChangeLog:

PR modula2/113836
* gm2-compiler/M2AsmUtil.def: Remove export qualified and
unused import.
* gm2-compiler/M2LangDump.mod (AddRuleTextDump): New procedure.
(AddRuleScopeQualidentDump): Add warning check against unmatched
rule.
(GenQualidentSymString): New procedure function.
(IdentQualidentMatch): New procedure function.
(IsRuleFilenameMatch): New procedure function.
(CheckRuleMatch): New procedure function.
(AddRuleFilenameDump): New procedure function.
* gm2-gcc/m2misc.cc (m2misc_warning_m2_dump_filter): New function.
* gm2-gcc/m2misc.def (warning_m2_dump_filter): New procedure.
* gm2-gcc/m2misc.h (m2misc_warning_m2_dump_filter): New prototype.
* gm2-gcc/m2pp.cc (VERBOSE_TYPE_DESC): New define.
(m2pp_identifier): Define out verbose type info.
(m2pp_constructor): Define out verbose type info.
(m2pp_assignment): Define out verbose type info.
* gm2-lang.cc (ENABLE_M2DUMP_ALL): Remove.
* lang.opt (fm2-dump): Add.
(fm2-dump-decl=): Add.
(fm2-dump-gimple=): Add.
(fm2-dump-quad=): Add.
(fm2-dump-filter=): Add.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
3 days agofix single argument static_assert
Marc Poulhiès [Thu, 2 May 2024 10:23:36 +0000 (2 12:23 +0200)]
fix single argument static_assert

Single argument static_assert is C++17 only.

gcc/ChangeLog:

* value-range.h: fix static_assert to use 2 arguments.

3 days agolto-wrapper: Truncate files using -truncate driver option [PR110710]
Peter Damianov [Sun, 28 Apr 2024 23:16:12 +0000 (28 16:16 -0700)]
lto-wrapper: Truncate files using -truncate driver option [PR110710]

This commit changes the Makefiles generated by lto-wrapper to no longer use
the "mv" and "touch" shell commands. These don't exist on Windows, so when the
Makefile attempts to call them, it results in errors like:
The system cannot find the file specified.

This problem only manifested when calling gcc from cmd.exe, and having no
sh.exe present on the PATH. The Windows port of GNU Make searches the PATH for
an sh.exe, and uses it if present.

I have tested this in environments with and without sh.exe on the PATH and
confirmed it works as expected.

Signed-off-by: Peter Damianov <peter0x44@disroot.org>
PR lto/110710
* lto-wrapper.cc (run_gcc): Instead of truncating a processed
ltrans input from the Makefile use the new -truncate option
to accomplish the same.

3 days agoDriver: Add new -truncate option
Peter Damianov [Sun, 28 Apr 2024 23:16:11 +0000 (28 16:16 -0700)]
Driver: Add new -truncate option

This commit adds a new option to the driver that truncates one file after
linking.

Tested likeso:

$ gcc hello.c -c
$ du -h hello.o
4.0K  hello.o
$ gcc hello.o -truncate hello.o
$ ./a.out
Hello world
$ du -h hello.o
$ 0   hello.o

$ gcc hello.o -truncate
gcc: error: missing filename after '-truncate'

The motivation for adding this is PR110710. It is used by lto-wrapper to
truncate files in a shell-independent manner.

Signed-off-by: Peter Damianov <peter0x44@disroot.org>
PR lto/110710
* common.opt (truncate): New internal option.
* gcc.cc (totruncate_file): New global.
(driver_handle_option): Handle -truncate <file>.
(driver::final_actions): Truncate the file indicated.

3 days agolibgomp: Add gfx90c, 1036 and 1103 declare variant tests
Jakub Jelinek [Thu, 2 May 2024 09:56:16 +0000 (2 11:56 +0200)]
libgomp: Add gfx90c, 1036 and 1103 declare variant tests

Recently -march=gfx{90c,1036,1103} support has been added, but corresponding
changes weren't done in the testsuite.

The following patch adds that.

Tested on x86_64-linux (with fiji and gfx1103 devices; had to use
OMP_DEFAULT_DEVICE=1 there, fiji doesn't really work due to LLVM dropping
support, but we still list those as offloading devices).

2024-05-02  Jakub Jelinek  <jakub@redhat.com>

* testsuite/libgomp.c/declare-variant-4.h (gfx90c, gfx1036, gfx1103):
New functions.
(f): Add #pragma omp declare variant directives for those.
* testsuite/libgomp.c/declare-variant-4-gfx90c.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1036.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1103.c: New test.

3 days agoc++: Implement C++26 P2573R2 - = delete("should have a reason"); [PR114458]
Jakub Jelinek [Thu, 2 May 2024 07:34:31 +0000 (2 09:34 +0200)]
c++: Implement C++26 P2573R2 - = delete("should have a reason"); [PR114458]

The following patch implements the C++26 P2573R2
= delete("should have a reason"); paper.
I've tried to avoid increasing compile time memory for it when it isn't
used (e.g. by adding a new lang_decl tree member), so the reason is
represented as STRING_CST in DECL_INITIAL (which normally is for
DECL_DELETED_FN error_mark_node) and to differentiate this delete("reason")
initializer from some bogus attempt to initialize a function with "reason"
using the RID_DELETE identifier as TREE_TYPE of the STRING_CST, as nothing
needs to care about the type of the reason.  If preferred it could
be say TREE_LIST with the reason STRING_CST and RID_DELETE identifier or
something similar instead, but that would need more compile time memory when
it is used.

2024-05-02  Jakub Jelinek  <jakub@redhat.com>

PR c++/114458
gcc/c-family/
* c-cppbuiltin.cc (c_cpp_builtins): Predefine
__cpp_deleted_function=202403L for C++26.
gcc/cp/ChangeLog
* parser.cc (cp_parser_pure_specifier): Implement C++26 P2573R2
- = delete("should have a reason");.  Parse deleted-function-body.
* decl.cc (duplicate_decls): Copy DECL_INITIAL from DECL_DELETED_FN
olddecl to newdecl if it is a STRING_CST.
(cp_finish_decl): Handle deleted init with a reason.
* decl2.cc: Include "escaped_string.h".
(grokfield): Handle deleted init with a reason.
(mark_used): Emit DECL_DELETED_FN reason in the message if any.
* cp-tree.h (DECL_DELETED_FN): Document representation of
= delete("reason") on a DECL.
gcc/testsuite/
* g++.dg/cpp26/feat-cxx26.C (__cpp_deleted_function): Add test.
* g++.dg/cpp26/delete-reason1.C: New test.
* g++.dg/cpp26/delete-reason2.C: New test.
* g++.dg/parse/error65.C (f1): Adjust expected diagnostics.

3 days agoMake graph dumps use graphviz format
Richard Biener [Tue, 30 Apr 2024 11:01:43 +0000 (30 13:01 +0200)]
Make graph dumps use graphviz format

SLP build eventually uses graphds graphs, the following makes its
dump use graphviz format so you can easily visualize it.

* graphds.cc (dump_graph): Dump in graphviz format.

3 days agoRemove live-info global bitmap
Richard Biener [Tue, 9 Apr 2024 09:28:47 +0000 (9 11:28 +0200)]
Remove live-info global bitmap

The following removes the unused tree_live_info_d->global bitmap.

* tree-ssa-live.h (tree_live_info_d::global): Remove.
(partition_is_global): Likewise.
(make_live_on_entry): Do not set bit in global.
* tree-ssa-live.cc (new_tree_live_info): Do not allocate
global bitmap.
(delete_tree_live_info): Do not release it.
(set_var_live_on_entry): Do not set bits in it.

3 days agos390: testsuite: Fix risbg-ll-2.c
Stefan Schulze Frielinghaus [Thu, 2 May 2024 06:43:50 +0000 (2 08:43 +0200)]
s390: testsuite: Fix risbg-ll-2.c

Starting with r14-2047-gd0e891406b16dc we see through subregs which
means for f10 in risbg-ll-2.c we do not end up with rosbg_si_noshift but
rather rosbg_di_noshift which materializes in slightly different start
index.  This saves us an extend.

gcc/testsuite/ChangeLog:

* gcc.target/s390/risbg-ll-2.c: Fix start offset for rosbg of
f10.

3 days agoc++: Don't emit unused GMF partial specializations [PR114630]
Nathaniel Shead [Thu, 11 Apr 2024 09:15:35 +0000 (11 19:15 +1000)]
c++: Don't emit unused GMF partial specializations [PR114630]

The change in r14-8408 to also emit partial specializations in the
global module fragment caused the regression in the linked PR; this
patch fixes this by restricting emitted GM partial specializations to
those that are actually used.

PR c++/114630

gcc/cp/ChangeLog:

* module.cc (depset::hash::add_partial_entities): Mark GM
specializations as unreached.
(depset::hash::find_dependencies): Also reach entities in the
DECL_TEMPLATE_SPECIALIZATIONS list.

gcc/testsuite/ChangeLog:

* g++.dg/modules/partial-3.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
3 days agos390: testsuite: Fix zero_bits_compound-1.c
Stefan Schulze Frielinghaus [Thu, 2 May 2024 06:39:32 +0000 (2 08:39 +0200)]
s390: testsuite: Fix zero_bits_compound-1.c

Starting with r12-2731-g96146e61cd7aee we do not generate code like

_5 = (unsigned int) c_2(D);
i_6 = _5 << 8;
_7 = _5 << 20;
i_8 = i_6 | _7;

anymore but instead

_5 = (unsigned int) c_2(D);
_3 = _5 * 1048832;

which leads finally to slightly different assembly code where we
previously ended up for z10 or newer with

lr      %r1,%r2
sll     %r1,8
rosbg   %r1,%r2,32,43,20
llgfr   %r2,%r1
br      %r14

and now

lr      %r1,%r2
sll     %r1,12
ar      %r2,%r1
risbg   %r2,%r2,35,128+55,8
br      %r14

The zero-extend materializes via risbg for which the pattern contains an
"and" which is why the test fails.  Thus, instead of scanning for RTL
expressions rather scan for assembler instructions for s390.

gcc/testsuite/ChangeLog:

* gcc.dg/zero_bits_compound-1.c: Fix for s390.

3 days agomiddle-end/114579 - speed up add_scope_conflicts
Richard Biener [Thu, 4 Apr 2024 12:00:10 +0000 (4 14:00 +0200)]
middle-end/114579 - speed up add_scope_conflicts

The following speeds up stack variable conflict detection by recognizing
that the all-to-all conflict recording is only necessary for CFG merges
as it's the unioning of the live variable sets that doesn't come with
explicit mentions we record conflicts for.

If we employ this optimization we have to make sure to perform the
all-to-all conflict recording for all CFG merges even those into
empty blocks where we might previously have skipped this.

I have reworded the comment before the all-to-all conflict recording
since it seemed to be confusing and missing the point - but maybe I
am also missing something here.

Nevertheless for the testcase in the PR the compile-time spend in
add_scope_conflicts at -O1 drops from previously 67s (39%) to 10s (9%).

PR middle-end/114579
* cfgexpand.cc (add_scope_conflicts_1): Record all-to-all
conflicts only when there's a CFG merge but for all CFG merges.

3 days agoc++: Implement modules ABI for vtable emissions
Nathaniel Shead [Tue, 16 Apr 2024 12:50:26 +0000 (16 22:50 +1000)]
c++: Implement modules ABI for vtable emissions

This patch implements the changes described in
https://github.com/itanium-cxx-abi/cxx-abi/pull/171.

One restriction that is lifted in the ABI that hasn't been updated here
is that the ABI no longer requires unique vtables to be emitted with
vague linkage.  I haven't changed this behaviour for this patch, but in
the future we could look into changing the relevant target hook
('class_data_always_comdat') to default to 'false'.  But the current
behaviour is more forgiving to changes in key function identification.

Since the ABI for vtables attached to named modules no longer depends on
key methods, this also resolves the issue described in PR c++/105224.

PR c++/105224

gcc/cp/ChangeLog:

* class.cc (finish_struct_1): Also push classes attached to a
module into the 'keyed_classes' list.
* decl.cc (record_key_method_defined): Don't push classes
attached to a named module into the 'keyed_classes' list.
* module.cc (trees_in::read_class_def): Likewise.
* decl2.cc (import_export_class): Uniquely emit vtables for
non-template classes attached to a named module.
(vtables_uniquely_emitted): New function.
(import_export_decl): Update comments. Update with knowledge
about new kinds of uniquely emitted vtables.

gcc/testsuite/ChangeLog:

* g++.dg/modules/virt-2_a.C: Update linkage requirements.
* g++.dg/modules/virt-2_b.C: Likewise.
* g++.dg/modules/virt-2_c.C: Likewise.
* g++.dg/modules/virt-4_a.C: New test.
* g++.dg/modules/virt-4_b.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Jason Merrill <jason@redhat.com>
3 days agoDaily bump.
GCC Administrator [Thu, 2 May 2024 00:17:09 +0000 (2 00:17 +0000)]
Daily bump.

4 days agodoc: Describe limitations re Ada, D, and Go on FreeBSD
Gerald Pfeifer [Wed, 1 May 2024 22:45:52 +0000 (2 00:45 +0200)]
doc: Describe limitations re Ada, D, and Go on FreeBSD

gcc:
PR target/69374
PR target/112959
* doc/install.texi (Specific) <*-*-freebsd*>: The Ada and D
run-time libraries are broken on i386 which also can affect
64-bit builds. Go is broken.

4 days agoc++: drop in-charge for dtors without vbases
Jason Merrill [Tue, 12 Dec 2023 23:07:28 +0000 (12 18:07 -0500)]
c++: drop in-charge for dtors without vbases

Constructors and destructors use the in-charge parameter to decide whether
they're responsible for recursing into virtual bases.  Historically all
destructors had this parameter in order to also distinguish the deleting
destructor.  But r151673 in GCC 4.5 changed the deleting destructor to just
call the complete destructor and then operator delete, making the in-charge
parameter no longer relevant for destructors in classes without virtual
bases.  Having it causes confusion in constexpr evaluation, which assumes
that clones will have the same parameters as the cloned 'tor.

gcc/cp/ChangeLog:

* cp-tree.h (base_ctor_identifier): Adjust comment.
* call.cc (in_charge_arg_for_name): Abort on deleting dtor.
* decl2.cc (maybe_retrofit_in_chrg): Don't add it for
destructors without vbases, either.
* constexpr.cc (cxx_eval_call_expression): Remove workaround.

gcc/testsuite/ChangeLog:

* g++.dg/debug/dwarf2/array-3.C: No more 'int' for in-chrg parm.
* g++.dg/debug/dwarf2/array-4.C: Likewise.

4 days ago[committed] [RISC-V] Trivial pattern cleanup
Jeff Law [Wed, 1 May 2024 18:43:37 +0000 (1 12:43 -0600)]
[committed] [RISC-V] Trivial pattern cleanup

As I was reviewing and cleaning up some internal work, I noticed a particular
idiom being used elsewhere in the RISC-V backend.

Specifically the use of explicit subregs when an adjustment to the
match_operand would be sufficient.

Let's take this example from the and-not splitter:

>  (define_split
>    [(set (match_operand:X 0 "register_operand")
>         (and:X (not:X (lshiftrt:X (match_operand:X 1 "register_operand")
>                                   (subreg:QI (match_operand:X 2 "register_operand") 0)))
>                (const_int 1)))]

Note the explicit subreg.  We can instead use a match_operand with QImode.
This ever-so-slightly simplifies the machine description.

It also means that if we have a QImode object lying around (say we loaded it
from memory in QImode), we can use it directly rather than first extending it
to X, then truncing to QI.  So we end up with simpler RTL and in rare cases
improve the code we generate.

When used in a define_split or define_insn_and_split we need to make suitable
adjustments to the split RTL.

Bootstrapped a while back.  Just re-tested with a cross.

gcc/
* config/riscv/bitmanip.md (splitter to use w-form division): Remove
explicit subregs.
(zero extended bitfield extraction): Similarly.
* config/riscv/thead.md (*th_memidx_operand): Similarly.

4 days ago[committed] [RISC-V] Fix detection of store pair fusion cases
Jeff Law [Wed, 1 May 2024 17:28:41 +0000 (1 11:28 -0600)]
[committed] [RISC-V] Fix detection of store pair fusion cases

We've got the ability to count the number of store pair fusions happening in
the front-end of the pipeline.  When comparing some code from last year vs the
current trunk we saw a fairly dramatic drop.

The problem is the store pair fusion detection code was actively harmful due to
a minor bug in checking offsets.   So instead of pairing up 8 byte stores such
as sp+0 with sp+8, it tried to pair up sp+8 and sp+16.

Given uarch sensitivity I didn't try to pull together a testcase.  But we could
certainly see the undesirable behavior in benchmarks as simplistic as dhrystone
up through spec2017.

Anyway, bootstrapped a while back.  Also verified through our performance
counters that store pair fusion rates are back up.  Regression tested with
crosses a few minutes ago.

gcc/
* config/riscv/riscv.cc (riscv_macro_fusion_pair_p): Break out
tests for easier debugging in store pair fusion case.  Fix offset
check in same.

4 days agolibstdc++: Guard uses of is_pointer_interconvertible_v [PR114891]
Jonathan Wakely [Tue, 30 Apr 2024 08:48:00 +0000 (30 09:48 +0100)]
libstdc++: Guard uses of is_pointer_interconvertible_v [PR114891]

This type trait isn't supported by Clang 18. It's only used in static
assertions, so they can just be omitted if the trait isn't available.

libstdc++-v3/ChangeLog:

PR libstdc++/114891
* include/std/generator: Check feature test macro before using
is_pointer_interconvertible_v.

4 days agoc++: const void* memchr [PR113706]
Jason Merrill [Mon, 12 Feb 2024 23:24:00 +0000 (12 18:24 -0500)]
c++: const void* memchr [PR113706]

The C++ standard specifies that the <string.h> functions have const and
non-const overloads, unlike C's single function with const argument and
non-const return.  Many systems don't actually implement this, but only add
an overload with non-const argument, so both end up having non-const return.
Solaris <string.h> does what the standard says, but we were penalizing it by
not recognizing the const overload as the built-in memchr.

PR c++/113706

gcc/cp/ChangeLog:

* decl.cc (decls_match): Handle memchr return type being
const-qualified.

gcc/testsuite/ChangeLog:

* g++.dg/opt/const-builtin1.C: New test.
* c-c++-common/pr103798-2.c: Remove xfail.

4 days agodoc: FreeBSD no longer has a GNU toolchain in base
Gerald Pfeifer [Wed, 1 May 2024 14:23:08 +0000 (1 16:23 +0200)]
doc: FreeBSD no longer has a GNU toolchain in base

gcc:
PR target/69374
PR target/112959
* doc/install.texi (Specific) <*-*-freebsd*>: No longer refer
to GCC or binutils in base. Recommend bootstrap using binutils.

4 days agodoc: Remove old details on libunwind for ia64-*-hpux*
Gerald Pfeifer [Wed, 1 May 2024 09:18:19 +0000 (1 11:18 +0200)]
doc: Remove old details on libunwind for ia64-*-hpux*

gcc:
PR target/69374
* doc/install.texi (Specific) <ia64-*-hpux*>: Remove details
on libunwind for GCC 3.4 and earlier.

4 days agoReduce startup costs for Value_Range.
Aldy Hernandez [Tue, 30 Apr 2024 17:39:00 +0000 (30 19:39 +0200)]
Reduce startup costs for Value_Range.

Value_Range is our polymorphic temporary that can hold any range.  It
is used for type agnostic code where it isn't known ahead of time,
what the type of the range will be (irange, france, etc).  Currently,
there is a temporary of each type in the object, which means we need
to construct each range for every temporary.  This isn't scaling
well now that prange is about to add yet another range type.

This patch removes each range, opting to use in-place new for a byte
buffer sufficiently large to hold ranges of any type.  It reduces the
memory footprint by 14% for every Value_Range temporary (from 792 to
680 bytes), and we are guaranteed it will never again grow as we add
more range types (strings, complex numbers, etc).

Surprisingly, it improves VRP performance by 6.61% and overall
compilation by 0.44%, which is a lot more than we bargained for
when we started working on prange performance.

There is a slight change in semantics for Value_Range.  The default
constructor does not initialize the object at all.  It must be
manually initialized with either Value_Range::set_type(), or by
assigning a range to it.  This means that IPA's m_known_value_ranges
must be initialized at allocation, instead of depending on the empty
constructor to initialize it to VR_UNDEFINED for unsupported_range.

I have taken the time to properly document both the class, and each
method.  If anything isn't clear, please let me know so I can adjust it
accordingly.

gcc/ChangeLog:

* ipa-fnsummary.cc (evaluate_properties_for_edge): Initialize Value_Range's.
* value-range.h (class Value_Range): Add a buffer and remove
m_irange and m_frange.
(Value_Range::Value_Range): Call init.
(Value_Range::set_type): Same.
(Value_Range::init): Use in place new to initialize buffer.
(Value_Range::operator=): Tidy.

4 days agoCleanups to unsupported_range.
Aldy Hernandez [Tue, 30 Apr 2024 16:54:11 +0000 (30 18:54 +0200)]
Cleanups to unsupported_range.

Here are some cleanups to unsupported_range so the assignment operator
takes an unsupported_range and behaves like the other ranges.  This
makes subsequent cleanups easier.

gcc/ChangeLog:

* value-range.cc (unsupported_range::union_): Cast vrange to
unsupported_range.
(unsupported_range::intersect): Same.
(unsupported_range::operator=): Make argument an unsupported_range.
* value-range.h: New constructor.

4 days agoc++: Propagate hidden flag on decls from partitions
Nathaniel Shead [Tue, 9 Apr 2024 11:52:38 +0000 (9 21:52 +1000)]
c++: Propagate hidden flag on decls from partitions

While working on some other fixes I noticed that the partition handling
code used the wrong flag to propagate OVL_HIDDEN_P on exported bindings
from partitions. This patch fixes that, and renames the flag to be
clearer.

gcc/cp/ChangeLog:

* name-lookup.cc (walk_module_binding): Use the
partition-specific hidden flag instead of the top-level
decl_hidden.

gcc/testsuite/ChangeLog:

* g++.dg/modules/using-16_a.C: New test.
* g++.dg/modules/using-16_b.C: New test.
* g++.dg/modules/using-16_c.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
4 days agoc++: Propagate using decls from partitions [PR114868]
Nathaniel Shead [Tue, 9 Apr 2024 11:49:58 +0000 (9 21:49 +1000)]
c++: Propagate using decls from partitions [PR114868]

The modules code currently neglects to set OVL_USING_P on the dependency
created for a using-decl, which causes it not to remember that the
OVL_EXPORT_P flag had been set on it when emitted from the primary
interface unit. This patch ensures that it occurs.

PR c++/114868

gcc/cp/ChangeLog:

* module.cc (depset::hash::add_binding_entity): Propagate
OVL_USING_P for using-declarations.

gcc/testsuite/ChangeLog:

* g++.dg/modules/using-15_a.C: New test.
* g++.dg/modules/using-15_b.C: New test.
* g++.dg/modules/using-15_c.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
4 days agoc++: Implement P2615 'Meaningful Exports' [PR107688]
Nathaniel Shead [Mon, 4 Mar 2024 12:58:30 +0000 (4 23:58 +1100)]
c++: Implement P2615 'Meaningful Exports' [PR107688]

This clarifies which kinds of declarations may and may not be exported
in various contexts. The patch additionally fixes up some small issues
that were clarified by the paper.

Most of the changes are with regards to export-declarations, which are
applied for all standards modes that we support '-fmodules-ts' for.
However there are also a couple of changes made to linkage specifiers
('extern "C"'); I've applied these as since C++20, to line up with when
modules were actually introduced.

PR c++/107688

gcc/cp/ChangeLog:

* name-lookup.cc (push_namespace): Error when exporting
namespace with internal linkage.
* parser.h (struct cp_parser): Add new flag
'in_unbraced_export_declaration_p'.
* parser.cc (cp_debug_parser): Print the new flag.
(cp_parser_new): Initialise the new flag.
(cp_parser_module_export): Set the new flag.
(cp_parser_class_specifier): Clear and restore the new flag.
(cp_parser_import_declaration): Imports can now appear directly
in a linkage specification.
(cp_parser_declaration): Categorise declarations as "name" or
"special"; error on the later in contexts where the former is
required.
(cp_parser_class_head): Error when exporting a partial
specialisation.

gcc/testsuite/ChangeLog:

* g++.dg/modules/contracts-1_a.C: Avoid now-illegal syntax.
* g++.dg/modules/contracts-2_a.C: Likewise.
* g++.dg/modules/contracts-3_a.C: Likewise.
* g++.dg/modules/contracts-4_a.C: Likewise.
* g++.dg/modules/lang-1_c.C: Clarify now-legal syntax.
* g++.dg/modules/pr101582-1.C: Remove now-legal XFAILS.
* g++.dg/template/crash71.C: Update error messages.
* g++.dg/cpp2a/linkage-spec1.C: New test.
* g++.dg/modules/export-3.C: New test.
* g++.dg/modules/export-4_a.C: New test.
* g++.dg/modules/export-4_b.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
4 days agoDaily bump.
GCC Administrator [Wed, 1 May 2024 00:17:56 +0000 (1 00:17 +0000)]
Daily bump.

5 days agoRemove incorrect asserts.
Andrew MacLeod [Tue, 13 Feb 2024 15:07:11 +0000 (13 10:07 -0500)]
Remove incorrect asserts.

Gimple_range_op handles builtin functions, and a couple of asserts that
are in place are incorrect in this context, so just remove them.

* gimple-range-op.cc (gimple_range_op_handler::calc_op1): Don't
assert that here are less than 3 operands.
(gimple_range_op_handler::maybe_builtin_call): Simply return if
there is no type for the function call.

5 days agoAdd range_on_entry/exit to value_query API.
Andrew MacLeod [Wed, 13 Mar 2024 18:18:48 +0000 (13 14:18 -0400)]
Add range_on_entry/exit to value_query API.

Add range_on_entry and range_on_exit to the value_query API.  These will
also work with generic trees like range_of_expr does.

* gimple-range.cc (gimple_ranger::range_on_entry): Adjust for new
API and support non-SSA expressions.
(gimple_ranger::range_on_exit): Ditto.
* gimple-range.h (range_on_entry, range_on_exit): Adjust API.
* value-query.cc (range_query::range_on_entry): New.
(range_query::range_on_exit): New.
(range_query::value_on_entry): New.
(range_query::value_on_exit): New.
(range_query::invoke_range_of_expr): New.
(range_query::get_tree_range): Allow stmt, on_entry or on_exit
range queries.
SSA_NAMES should invoke range_of_expr if possible.
* value-query.h (class range_query): Adjust prototypes.

5 days agoInvoke range_of_stmt on ssa_names with no context.
Andrew MacLeod [Wed, 13 Mar 2024 18:13:28 +0000 (13 14:13 -0400)]
Invoke range_of_stmt on ssa_names with no context.

Evalaute ssa-names when range_of_expr is called with no context statement
by calling range_of_stmt to get an initial value.

* gimple-range.cc (gimple_ranger::range_of_expr): Call range_of_stmt
when there is no context stmt.

5 days agoFix ranger when called from SCEV.
Andrew MacLeod [Wed, 13 Mar 2024 18:10:41 +0000 (13 14:10 -0400)]
Fix ranger when called from SCEV.

Do not pre-evaluate PHIs in the cache, and allow fill_block_cache to
be re-entrant.  This allows SCEV to call into ranger with a context
and not produce cycles or loops.

* gimple-range-cache.cc (ranger_cache::get_global_range): Do not
pre-evaluate PHI nodes from the cache.
(ranger_cache::fill_block_cache): Make re-entrant.

5 days agoRemove wrapper around gimple_range_global.
Andrew MacLeod [Tue, 20 Feb 2024 17:27:51 +0000 (20 12:27 -0500)]
Remove wrapper around gimple_range_global.

Now that legacy EVRP has been removed, we can remove the wrapper which
prevented us from using global names before inlining except under
some specific conditions.  See discussion:
 https://gcc.gnu.org/pipermail/gcc-patches/2021-June/571709.html

* value-query.cc (get_range_global): Rename to gimple_range_global.
(gimple_range_global): Remove wrapper function.
(global_range_query::range_of_expr): Call gimple_range_global.

5 days agoFix the build: error message `quote`
Andrew Pinski [Tue, 30 Apr 2024 16:44:52 +0000 (30 09:44 -0700)]
Fix the build: error message `quote`

The problem here is the quote mark is for English's
possessiveness rather than a quote but the error message
format detection is too simple so it warns which causes
-Werror to fail.

Committed as obvious after a quick build.

gcc/ChangeLog:

* tree-cfg.cc (verify_gimple_assign): Remove quote
mark to shut up the warning.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoPHIOPT: Value-replacement check undef
Andrew Pinski [Mon, 29 Apr 2024 03:21:02 +0000 (28 20:21 -0700)]
PHIOPT: Value-replacement check undef

While moving value replacement part of PHIOPT over
to use match-and-simplify, I ran into the case where
we would have an undef use that was conditional become
unconditional. This prevents that. I can't remember at this
point what the testcase was though.

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

gcc/ChangeLog:

* tree-ssa-phiopt.cc (value_replacement): Reject undef variables
so they don't become unconditional used.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoPHI-OPT: speed up value_replacement slightly
Andrew Pinski [Sun, 28 Apr 2024 01:54:45 +0000 (27 18:54 -0700)]
PHI-OPT: speed up value_replacement slightly

This adds a few early outs to value_replacement that I noticed
while rewriting this to use match-and-simplify but could be committed
seperately.
* virtual operands won't change so return early for them
* special case `A ? B : B` as that is already just `B`

Also moves the check for NE/EQ earlier as calculating empty_or_with_defined_p
is an IR walk for a BB and that might be big.

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

gcc/ChangeLog:

* tree-ssa-phiopt.cc (value_replacement): Move check for
NE/EQ earlier.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoMATCH: change single_non_singleton_phi_for_edges for singleton phis
Andrew Pinski [Sun, 28 Apr 2024 01:54:44 +0000 (27 18:54 -0700)]
MATCH: change single_non_singleton_phi_for_edges for singleton phis

I noticed that single_non_singleton_phi_for_edges could
return a phi whos entry are all the same for the edge.
This happens only if there was a single phis in the first place.
Also gimple_seq_singleton_p walks the sequence to see if it the one
element in the sequence so there is removing that check actually
reduces the number of pointer walks needed.

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

gcc/ChangeLog:

* tree-ssa-phiopt.cc (single_non_singleton_phi_for_edges):
Remove the special case of gimple_seq_singleton_p.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoRemove support for nontemporal stores with ssa_names on lhs [PR112976]
Andrew Pinski [Wed, 17 Apr 2024 21:30:06 +0000 (17 14:30 -0700)]
Remove support for nontemporal stores with ssa_names on lhs [PR112976]

When cfgexpand was changed to support expanding from tuple gimple
(r0-95521-g28ed065ef9f345), the code was added to support
doing nontemporal stores with LHS of a SSA_NAME but that will
never be a nontemporal store.
This patch removes that and asserts that expanding with a LHS
of a SSA_NAME is not a nontemporal store.

Bootstrapped and tested on x86_64-linux-gnu.

gcc/ChangeLog:

PR middle-end/112976
* cfgexpand.cc (expand_gimple_stmt_1): Remove
support for expanding nontemporal "moves" with
ssa names on the LHS.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoAdd verification of gimple_assign_nontemporal_move_p [PR112976]
Andrew Pinski [Wed, 17 Apr 2024 21:12:17 +0000 (17 14:12 -0700)]
Add verification of gimple_assign_nontemporal_move_p [PR112976]

Currently the middle-end only knows how to support temporal stores
(the undocumented storent optab) so let's verify that the only time
we set nontemporal_move on an assign is if the the lhs is not a
gimple reg.

Bootstrapped and tested on x86_64-linux-gnu no regressions.

gcc/ChangeLog:

PR middle-end/112976
* tree-cfg.cc (verify_gimple_assign): Verify that
nontmporal moves are stores.
* gimple.h (struct gimple): Note that only
nontemporal stores are supported.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
5 days agoThis is almost exclusively Jivan's work. His original post:
Jivan Hakobyan [Tue, 30 Apr 2024 15:44:02 +0000 (30 09:44 -0600)]
This is almost exclusively Jivan's work.  His original post:

> https://www.mail-archive.com/gcc-patches@gcc.gnu.org/msg336483.html

This patch is primarily meant to improve the code we generate for FP rounding
such as ceil/floor.  It also addresses some unnecessary sign extensions in the
same areas.

RISC-V's FP conversions have a bit of undesirable behavior that make them
non-suitable as-is for ceil/floor and other related functions. These
deficiencies are addressed in the Zfa extension, but there's no reason not to
pick up a nice improvement when we can.

Basically we can still use the basic FP conversions for floor/ceil and friends
when we don't care about inexact exceptions by checking for the special cases
first, then emitting the conversion when the special cases don't apply.  That's
still much faster than calling into glibc.

The redundant sign extensions are eliminated using the same trick Jivan added
last year, just in a few more places ;-)

This eliminates roughly 10% of the dynamic instruction count for imagick.  But
more importantly it's about a 17% performance improvement for that workload
within spec.

This has been bootstrapped as well as regression tested in a cross environment.
It's also successfully built & run specint/specfp correctly.

Pushing to the trunk and the coordination branch momentarily.

gcc/
* config/riscv/iterators.md (fix_ops, fix_uns): New iterators.
(RINT, rint_pattern, rint_rm): Remove unused iterators.
* config/riscv/riscv-protos.h (get_fp_rounding_coefficient): Prototype.
* config/riscv/riscv-v.cc (get_fp_rounding_coefficient): Externalize.
external linkage.
* config/riscv/riscv.md (UNSPEC_LROUND): Remove.
(fix_trunc<ANYF:mode><GPR:mode>2): Replace with ...
(<fix_uns>_trunc<ANYF:mode>si2): New expander & associated insn.
(<fix_uns>_trunc<ANYF:mode>si2_ext): New insn.
(<fix_uns>_trunc<ANYF:mode>di2): Likewise.
(l<rint_pattern><ANYF:mode><GPR:mode>2): Replace with ...
(lrint<ANYF:mode>si2): New expander and associated insn.
(lrint<ANYF:mode>si2_ext, lrint<ANYF:mode>di2): New insns.
(<round_pattern><ANYF:mode>2): Replace with....
(l<round_pattern><ANYF:mode>si2): New expander and associated insn.
(l<round_pattern><ANYF:mode>si2_sext): New insn.
(l<round_pattern><ANYF:mode>di2): Likewise.
(<round_pattern><ANYF:mode>2): New expander.

gcc/testsuite/
* gcc.target/riscv/fix.c: New test.
* gcc.target/riscv/round.c: New test.
* gcc.target/riscv/round_32.c: New test.
* gcc.target/riscv/round_64.c: New test.

5 days agoChange int_range<2> to infinite precision.
Aldy Hernandez [Tue, 30 Apr 2024 08:36:58 +0000 (30 10:36 +0200)]
Change int_range<2> to infinite precision.

In my previous change I mistakenly changed Value_Range to
int_range<2>.  The former has "infinite" precision for integer ranges,
whereas int_range<2> has two sub-ranges.  This should have been
int_range_max.

gcc/ChangeLog:

* gimple-ssa-warn-access.cc (check_nul_terminated_array): Change
int_range<2> to int_range_max.
(memmodel_to_uhwi): Same.
* tree-ssa-loop-niter.cc (refine_value_range_using_guard): Same.
(determine_value_range): Same.
(infer_loop_bounds_from_signedness): Same.
(scev_var_range_cant_overflow): Same.

5 days agotestsuite: gm2: Remove timeout overrides [PR114886]
Rainer Orth [Tue, 30 Apr 2024 11:49:28 +0000 (30 13:49 +0200)]
testsuite: gm2: Remove timeout overrides [PR114886]

A large number of gm2 tests are timing out even on current Solaris/SPARC
systems.  As detailed in the PR, the problem is that the gm2 testsuite
artificially lowers many timeouts way below the DejaGnu default of 300
seconds, often as short as 10 seconds.  The problem lies both in the
values (they may be appropriate for some targets, but too low for
others, especially under high load) and the fact that it uses absolute
values, overriding e.g. settings from a build-wide site.exp.

Therefore this patch removes all those overrides, restoring the
defaults.

Tested on sparc-sun-solaris2.11 (where all the previous timeouts are
gone) and i386-pc-solaris2.11.

2024-04-29  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

gcc/testsuite:
PR modula2/114886
* lib/gm2.exp: Don't load timeout-dg.exp.
Don't set gm2_previous_timeout.
Don't call dg-timeout.
(gm2_push_timeout, gm2_pop_timeout): Remove.
(gm2_init): Don't call dg-timeout.
* lib/gm2-torture.exp: Don't load timeout-dg.exp.
Don't set gm2_previous_timeout.
Don't call dg-timeout.
(gm2_push_timeout, gm2_pop_timeout): Remove.

* gm2/coroutines/pim/run/pass/coroutines-pim-run-pass.exp: Don't
load timeout-dg.exp.
Don't call gm2_push_timeout, gm2_pop_timeout.
* gm2/examples/map/pass/examples-map-pass.exp: Don't call
gm2_push_timeout, gm2_pop_timeout.
* gm2/iso/run/pass/iso-run-pass.exp: Don't load timeout-dg.exp.
Don't call gm2_push_timeout, gm2_pop_timeout.
* gm2/pimlib/base/run/pass/pimlib-base-run-pass.exp: Don't load
timeout-dg.exp.
Don't call gm2_push_timeout, gm2_pop_timeout.
* gm2/projects/iso/run/pass/halma/projects-iso-run-pass-halma.exp:
Don't call gm2_push_timeout, gm2_pop_timeout.
* gm2/switches/whole-program/pass/run/switches-whole-program-pass-run.exp:
Don't load timeout-dg.exp.
Don't call gm2_push_timeout, gm2_pop_timeout.