]> git.ipfire.org Git - thirdparty/gcc.git/log
thirdparty/gcc.git
2 weeks agotestsuite: arm: [MVE] Relax expected code for vbicq_f [PR122223]
Christophe Lyon [Thu, 23 Oct 2025 07:59:20 +0000 (07:59 +0000)] 
testsuite: arm: [MVE] Relax expected code for vbicq_f [PR122223]

The original versions of the pr122223.c test only took into account
code generated with -mfloat-abi=hard, which uses q0.

With -mfloat-abi=softfp, this can be any Q register, so replace q0
with a suitable regex.

gcc/testsuite/ChangeLog:

PR target/122223
* gcc.target/arm/mve/intrinsics/pr122223.c: Relax expected code.

2 weeks agoSupport reduc_sbool_and_scal_m for V{QI,SI,DI}mode.
liuhongt [Wed, 15 Oct 2025 09:00:30 +0000 (02:00 -0700)] 
Support reduc_sbool_and_scal_m for V{QI,SI,DI}mode.

gcc/ChangeLog:

PR target/101639
* config/i386/sse.md
(VI_AVX): New mode iterator.
(VI_AVX_CMP): Ditto.
(ssebytemode): Add V16HI, V32QI, V16QI.
(reduc_sbool_and_scal_<mode>): New expander.
(reduc_sbool_ior_scal_<mode>): Ditto.
(reduc_sbool_xor_scal_<mode>): Ditto.
(*eq<mode>3_2_negate): New pre_reload splitter.
(*ptest<mode>_ccz): Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr101639_reduc_mask_vdi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_vqi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_vsi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_ior_vqi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_and_vqi.c: New test.

2 weeks agoSupport reduc_sbool_{and,ior,xor}_scal_m for avx512 kmask.
liuhongt [Wed, 15 Oct 2025 07:30:01 +0000 (00:30 -0700)] 
Support reduc_sbool_{and,ior,xor}_scal_m for avx512 kmask.

gcc/ChangeLog:

PR target/101639
* config/i386/sse.md
(reduc_sbool_and_scal_<mode>): New expander.
(reduc_sbool_ior_scal_<mode>): Ditto.
(reduc_sbool_xor_scal_<mode>): Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr101639_reduc_mask_di.c: New test.
* gcc.target/i386/pr101639_reduc_mask_hi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_qi.c: New test.
* gcc.target/i386/pr101639_reduc_mask_si.c: New test.

2 weeks agoDaily bump.
GCC Administrator [Thu, 23 Oct 2025 00:19:12 +0000 (00:19 +0000)] 
Daily bump.

2 weeks agox86: Use HOST_WIDE_INT_(0|M1)U to initialize unsigned HOST_WIDE_INT
H.J. Lu [Wed, 22 Oct 2025 11:46:58 +0000 (19:46 +0800)] 
x86: Use HOST_WIDE_INT_(0|M1)U to initialize unsigned HOST_WIDE_INT

Use HOST_WIDE_INT_0U, instead of 0, HOST_WIDE_INT_M1U, instead of -1, to
initialize unsigned HOST_WIDE_INT.

* config/i386/i386-expand.cc (ix86_expand_set_or_cpymem): Use
HOST_WIDE_INT_0U and HOST_WIDE_INT_M1U to initialize unsigned
HOST_WIDE_INT.

Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
2 weeks agotestsuite: Fix local labels [PR122378]
Paul-Antoine Arras [Wed, 22 Oct 2025 14:14:02 +0000 (16:14 +0200)] 
testsuite: Fix local labels [PR122378]

r16-4540-g80af807e52e4f4 exposed a bug in two testcases where the declaration of
local labels was wrongly commented out. That caused "duplicate label" errors.
Uncommenting declarations fixes it.

PR middle-end/122378

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/attrs-metadirective-2.c: Uncomment local label
declaration.
* c-c++-common/gomp/metadirective-2.c: Likewise.

2 weeks agolibstdc++: Avoid incrementing input iterators with std::prev [PR122224]
Jonathan Wakely [Tue, 21 Oct 2025 16:06:49 +0000 (17:06 +0100)] 
libstdc++: Avoid incrementing input iterators with std::prev [PR122224]

As explained in PR libstdc++/122224 we do not make it ill-formed to call
std::prev with a non-Cpp17BidirectionalIterator. Instead we just use a
runtime assertion to check the std::advance precondition that the
distance is not negative.

This allows us to support std::prev on types which model the C++20
std::bidirectional_iterator concept but do not meet the
Cpp17BidirectionalIterator requirements, e.g. iota_view's iterators.

It also allows us to support std::prev(iter, -1) which is admittedly
weird, but there's no reason it shouldn't be equivalent to
std::next(iter), which is perfectly fine to use on non-bidirectional
iterators. In other words, "reverse decrementing" is valid for
non-bidirectional iterators.

However, the current implementation of std::advance for
non-bidirectional iterators uses a loop that does `while (n--) ++i;`
which assumes that n is not negative and so will eventually reach zero.
When the assertion for the precondition is not enabled, incrementing the
iterator while n is non-zero means that using std::prev(iter) or
std::next(iter, -1) on a non-bidirectional iterator will keep
incrementing the iterator until n reaches INT_MIN, overflows, and then
keeps decrementing until it eventually reaches zero. Incrementing most
iterators that many times will cause memory safety errors long before
the integer reaches zero and terminates the loop.

This commit changes the loop to use `while (n-- > 0)` which means that
the loop doesn't execute at all if a negative n is used. We still
consider such calls to be erroneous, but when the precondition isn't
checked by an assertion, the function now has no effects. The undefined
behaviour resulting from incrementing the iterator is prevented.

libstdc++-v3/ChangeLog:

PR libstdc++/122224
* include/bits/stl_iterator_base_funcs.h (prev): Compare
distance as n > 0 instead of n != 0.
* testsuite/24_iterators/range_operations/122224.cc: New test.

Reviewed-by: Tomasz Kamiński <tkaminsk@redhat.com>
2 weeks agoMAINTAINERS: Update my contact info.
Josef Melcr [Wed, 22 Oct 2025 12:58:26 +0000 (14:58 +0200)] 
MAINTAINERS: Update my contact info.

ChangeLog:

* MAINTAINERS: Update my contact information.

Signed-off-by: Josef Melcr <jmelcr02@gmail.com>
2 weeks agoc++: Fix up RAW_DATA_CST handling in braced_list_to_string [PR122302]
Jakub Jelinek [Wed, 22 Oct 2025 11:11:52 +0000 (13:11 +0200)] 
c++: Fix up RAW_DATA_CST handling in braced_list_to_string [PR122302]

The following testcase is miscompiled, because a RAW_DATA_CST tree
node is shared by multiple CONSTRUCTORs and when the braced_list_to_string
function changes one to extend the RAW_DATA_CST over the single preceding
and single succeeding INTEGER_CST, it changes the RAW_DATA_CST in
the other CONSTRUCTOR where the elts around it are still present.

Fixed by tweaking a copy of it instead, like we handle it in other spots.

2025-10-22  Jakub Jelinek  <jakub@redhat.com>

PR c++/122302
* c-common.cc (braced_list_to_string): Call copy_node on RAW_DATA_CST
before changing RAW_DATA_POINTER and RAW_DATA_LENGTH on it.

* g++.dg/cpp0x/pr122302.C: New test.
* g++.dg/cpp/embed-27.C: New test.

2 weeks agoAArch64: Add support for boolean reductions for Adv. SIMD using SVE
Tamar Christina [Wed, 22 Oct 2025 09:52:43 +0000 (10:52 +0100)] 
AArch64: Add support for boolean reductions for Adv. SIMD using SVE

When doing boolean reductions for Adv. SIMD vectors and SVE is available
we can use SVE instructions instead of Adv. SIMD ones to do the reduction.

For instance OR-reductions are

        umaxp v3.4s, v3.4s, v3.4s
        fmov x1, d3
        cmp x1, 0
        cset w0, ne

and with SVE we generate:

        ptrue p1.b, vl16
        cmpne p1.b, p1/z, z3.b, #0
        cset w0, any

Where the ptrue is normally executed much earlier so it's not a bottleneck for
the compare.

For the remaining codegen see test vect-reduc-bool-18.c.

gcc/ChangeLog:

* config/aarch64/aarch64-simd.md (reduc_sbool_and_scal_<mode>,
reduc_sbool_ior_scal_<mode>, reduc_sbool_xor_scal_<mode>): Use SVE if
available.
* config/aarch64/aarch64-sve.md (*cmp<cmp_op><mode>_ptest): Rename ...
(@aarch64_pred_cmp<cmp_op><mode>_ptest): ... To this.
(reduc_sbool_xor_scal_<mode>): Rename ...
(@reduc_sbool_xor_scal_<mode>): ... To this.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/sve/vect-reduc-bool-10.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-11.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-12.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-13.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-14.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-15.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-16.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-17.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-18.c: New test.

2 weeks agoAArch64: Add support for boolean reductions for Adv. SIMD
Tamar Christina [Wed, 22 Oct 2025 09:51:41 +0000 (10:51 +0100)] 
AArch64: Add support for boolean reductions for Adv. SIMD

The vectorizer has learned how to do boolean reductions of masks to a C bool
for the operations OR, XOR and AND.

This implements the new optabs for Adv.SIMD.  Adv.SIMD today can already
vectorize such loops but does so through SHIFT-AND-INSERT to perform the
reductions step-wise and inorder.  As an example, an OR reduction today does:

        movi    v3.4s, 0
        ext     v5.16b, v30.16b, v3.16b, #8
        orr     v5.16b, v5.16b, v30.16b
        ext     v29.16b, v5.16b, v3.16b, #4
        orr     v29.16b, v29.16b, v5.16b
        ext     v4.16b, v29.16b, v3.16b, #2
        orr     v4.16b, v4.16b, v29.16b
        ext     v3.16b, v4.16b, v3.16b, #1
        orr     v3.16b, v3.16b, v4.16b
        fmov    w1, s3
        and     w1, w1, 1

For reducing to a boolean however we don't need the stepwise reduction and can
just look at the bit patterns. For e.g. OR we now generate:

        umaxp v3.4s, v3.4s, v3.4s
        fmov x1, d3
        cmp x1, 0
        cset w0, ne

For the remaining codegen see test vect-reduc-bool-9.c.

gcc/ChangeLog:

* config/aarch64/aarch64-simd.md (reduc_sbool_and_scal_<mode>,
reduc_sbool_ior_scal_<mode>, reduc_sbool_xor_scal_<mode>): New.
* config/aarch64/iterators.md (VALLI): New.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/vect-reduc-bool-1.c: New test.
* gcc.target/aarch64/vect-reduc-bool-2.c: New test.
* gcc.target/aarch64/vect-reduc-bool-3.c: New test.
* gcc.target/aarch64/vect-reduc-bool-4.c: New test.
* gcc.target/aarch64/vect-reduc-bool-5.c: New test.
* gcc.target/aarch64/vect-reduc-bool-6.c: New test.
* gcc.target/aarch64/vect-reduc-bool-7.c: New test.
* gcc.target/aarch64/vect-reduc-bool-8.c: New test.
* gcc.target/aarch64/vect-reduc-bool-9.c: New test.

2 weeks agoAArch64: Add support for boolean reductions for SVE
Tamar Christina [Wed, 22 Oct 2025 09:50:35 +0000 (10:50 +0100)] 
AArch64: Add support for boolean reductions for SVE

The vectorizer has learned how to do boolean reductions of masks to a C bool
for the operations OR, XOR and AND.

This implements the new optabs for SVE.

For SVE & and the | case would use the CC registers.

or_reduc:
        ptest   p0, p0.b
        cset    w0, any

and_reduc:
        ptrue   p3.b, all
        nots    p3.b, p3/z, p0.b
        cset    w0, none

and the ^ case we'd see if the number of active predicate lanes
is a multiple of two.

xor_reduc:
        ptrue   p3.b, all
        cntp    x0, p3, p0.b
        and     w0, w0, 1

gcc/ChangeLog:

* config/aarch64/aarch64-sve.md (reduc_sbool_and_scal_<mode>,
reduc_sbool_ior_scal_<mode>, reduc_sbool_xor_scal_<mode>): New.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/sve/vect-reduc-bool-1.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-2.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-3.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-4.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-5.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-6.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-7.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-8.c: New test.
* gcc.target/aarch64/sve/vect-reduc-bool-9.c: New test.

2 weeks agovect: Add support for boolean reductions for VLA
Tamar Christina [Wed, 22 Oct 2025 09:49:26 +0000 (10:49 +0100)] 
vect: Add support for boolean reductions for VLA

The support for the new boolean reduction optabs didn't quite work for VLA
because the code later on insists on the target still having a shift-and-insert
optab.

This is however not needed if the target can do the reduction using the new
optabs, and the initial reduction value matches the neutral value and we
have one SLP lane while not having a reduction chain.

gcc/ChangeLog:

* tree-vect-loop.cc (vectorizable_reduction): Don't always require
IFN_VEC_SHL_INSERT when using reduc sbool optabs.

2 weeks agoaarch64: Add autoregenerated files for AArch64 options.
Srinath Parvathaneni [Wed, 22 Oct 2025 09:19:29 +0000 (10:19 +0100)] 
aarch64: Add autoregenerated files for AArch64 options.

In the previous committed patch to "add support for
menable-sysreg-checking flag", I have made changes to
config/aarch64/aarch64.opt, but missed to update the
autoregenerated files.

This patch adds the updated autoregenerated aarch64.opt.urls
changes.

gcc/ChangeLog:

* config/aarch64/aarch64.opt.urls: Regenerate.

2 weeks agotree-optimization/122364 - reduction chain with conversion
Richard Biener [Tue, 21 Oct 2025 13:23:18 +0000 (15:23 +0200)] 
tree-optimization/122364 - reduction chain with conversion

The following handles detecting of a reduction chain wrapped in a
conversion.  This does not yet try to combine operands with different
signedness, but we should now handle signed integer accumulation
to both a signed and unsigned accumulator fine.

PR tree-optimization/122364
* tree-vect-slp.cc (vect_analyze_slp_reduc_chain): Re-try
linearization on a conversion source.

* gcc.dg/vect/vect-reduc-chain-5.c: New testcase.

2 weeks agotree-optimization/122370 - ICE with reduction and masks
Richard Biener [Wed, 22 Oct 2025 08:05:01 +0000 (10:05 +0200)] 
tree-optimization/122370 - ICE with reduction and masks

The following fixes bad interaction with mask demotion to data
and the code dealing with UB on signed reductions by making sure
to also update compute_vectype when updating vectype.

PR tree-optimization/122370
* tree-vect-loop.cc (vect_create_epilog_for_reduction):
Also update compute_vectype when demoting masks to an
integer vector.

* gcc.dg/vect/vect-pr122370.c: New testcase.

2 weeks agolibstdc++: Add missing constraints to views::indices
Jonathan Wakely [Tue, 21 Oct 2025 20:21:45 +0000 (21:21 +0100)] 
libstdc++: Add missing constraints to views::indices

Calling views::indices(n) should be expression equivalent to
views::iota(decltype(n)(0), n), which means it should have the same
constraints as views::iota and be SFINAE-friendly.

libstdc++-v3/ChangeLog:

* include/std/ranges (indices::operator()): Constrain using
__can_iota_view concept.
* testsuite/std/ranges/indices/1.cc: Check SFINAE-friendliness
required by expression equivalence. Replace unused <vector>
header with <stddef.h> needed for size_t.

2 weeks agotree-optimization/122371 - ICE with reduction chain and fold-left reduction
Richard Biener [Wed, 22 Oct 2025 07:41:15 +0000 (09:41 +0200)] 
tree-optimization/122371 - ICE with reduction chain and fold-left reduction

The fold-left reduction transform relies on preserving the scalar
cycle PHI.  The following rewrites how we connect this to the
involved stmt-infos instead of relying on (the actually bogus for
reduction chain) scalar stmts in SLP nodes more than absolutely
necessary.  This also makes sure to not re-associate to form a
reduction chain when a fold-left reduction is required.

PR tree-optimization/122371
* tree-vect-loop.cc (vectorize_fold_left_reduction): Get
to the scalar def to replace via the scalar PHI backedge def.
* tree-vect-slp.cc (vect_analyze_slp_reduc_chain): Do not
re-associate to for a reduction chain if a fold-left
reduction is required.

* gcc.dg/vect/vect-pr122371.c: New testcase.

2 weeks agolibstdc++: Implement optional<T&> from P2988R12 [PR121748]
Giuseppe D'Angelo [Wed, 22 Oct 2025 07:31:46 +0000 (09:31 +0200)] 
libstdc++: Implement optional<T&> from P2988R12 [PR121748]

This patch implements optional<T&> based on the P2988R12 paper, incorporating
corrections from LWG4300, LWG4304, and LWG3467. The resolution for LWG4015
is also extended to cover optional<T&>.

We introduce _M_fwd() helper, that is equivalent to operator*(), except that
it does not check non-empty precondition. It is used in to correctly propagate
the value during move construction from optional<T&>. This is necessary because
moving an optional<T&> must not move the contained object, which is the key
distinction between *std::move(opt) and std::move(*opt).

The implementation deviates from the standard by providing a separate std::swap
overload for std::optional<T&>, which simplifies preserving the resolution of
LWG2766.

This introduces a few changes to make_optional behavior (see included test):
* some previously valid uses of make_optional<T>({...}) (where T is not a
  reference type) now become ill-formed (see optional/make_optional_neg.cc).
* make_optional<T&>(t) and make_optional<const T&>(ct), where decltype(t) is T&,
  and decltype(ct) is const T& now produce optional<T&> and optional<const T&>
  respectively, instead of optional<T>.
* a few other uses of make_optional<R> with reference type R are now ill-formed.

PR libstdc++/121748

libstdc++-v3/ChangeLog:

* include/bits/version.def: Bump value for optional,
* include/bits/version.h: Regenerate.
* include/std/optional (std::__is_valid_contained_type_for_optional):
Define.
(std::optional<T>): Use __is_valid_contained_type_for_optional.
(optional<T>(const optional<_Up>&), optional<T>(optional<_Up>&&))
(optional<T>::operator=(const optional<_Up>&))
(optional<T>::operator=(optional<_Up>&&)): Replacex._M_get() with
x._M_fwd(), and std::move(x._M_get()) with std::move(x)._M_fwd().
(optional<T>::and_then): Remove uncessary remove_cvref_t.
(optional<T>::_M_fwd): Define.
(std::optional<T&>): Define new partial specialization.
(std::swap(std::optional<T&>, std::optional<T&>)): Define.
(std::make_optional(_Tp&&)): Add non-type template parameter.
(std::make_optional): Use parenthesis to constructor optional.
(std::hash<optional<T>>): Add comment.
* testsuite/20_util/optional/make_optional-2.cc: Guarded not longer
working example.
* testsuite/20_util/optional/relops/constrained.cc: Expand test to
cover optionals of reference.
* testsuite/20_util/optional/requirements.cc: Ammend for
optional<T&>.
* testsuite/20_util/optional/requirements_neg.cc: Likewise.
* testsuite/20_util/optional/version.cc: Test new value of
__cpp_lib_optional.
* testsuite/20_util/optional/make_optional_neg.cc: New test.
* testsuite/20_util/optional/monadic/ref_neg.cc: New test.
* testsuite/20_util/optional/ref/access.cc: New test.
* testsuite/20_util/optional/ref/assign.cc: New test.
* testsuite/20_util/optional/ref/cons.cc: New test.
* testsuite/20_util/optional/ref/internal_traits.cc: New test.
* testsuite/20_util/optional/ref/make_optional/1.cc: New test.
* testsuite/20_util/optional/ref/make_optional/from_args_neg.cc:
New test.
* testsuite/20_util/optional/ref/make_optional/from_lvalue_neg.cc:
New test.
* testsuite/20_util/optional/ref/make_optional/from_rvalue_neg.cc:
New test.
* testsuite/20_util/optional/ref/monadic.cc: New test.
* testsuite/20_util/optional/ref/relops.cc: New test.

Reviewed-by: Jonathan Wakely <jwakely@redhat.com>
Co-authored-by: Tomasz Kamiński <tkaminsk@redhat.com>
2 weeks agolibstdc++: Add comparison operators between tuple<> and array<T, 0> [PR119721]
Osama Abdelkader [Mon, 20 Oct 2025 15:16:52 +0000 (18:16 +0300)] 
libstdc++: Add comparison operators between tuple<> and array<T, 0> [PR119721]

This fixes the C++23 compliance issue where std::tuple<> cannot be compared
with other empty tuple-like types such as std::array<T, 0>.

The operators correctly allow comparison with array<T, 0> even when T is not
comparable, because empty tuple-like types don't compare element values.

PR libstdc++/119721

libstdc++-v3/ChangeLog:

* include/std/tuple (tuple<>::operator==, tuple<>::operator<=>):
Define.
* testsuite/23_containers/tuple/comparison_operators/119721.cc:
New test.

Reviewed-by: Jonathan Wakely <jwakely@redhat.com>
Reviewed-by: Tomasz Kamiński <tkaminsk@redhat.com>
Signed-off-by: Osama Abdelkader <osama.abdelkader@gmail.com>
2 weeks agotree-optimization/122365 - deal with bool SLP reductions
Richard Biener [Tue, 21 Oct 2025 13:40:41 +0000 (15:40 +0200)] 
tree-optimization/122365 - deal with bool SLP reductions

I hadn't thought of these but at least added an assert which now
tripped.  Fixed thus.  There's also a latent issue with AVX512
mask types.  The by-pieces reduction code used the wrong element
sizes.

PR tree-optimization/122365
* tree-vect-loop.cc (vect_create_epilog_for_reduction):
Convert all inputs.  Use the proper vector element sizes
for the elementwise reduction.

* gcc.dg/vect/vect-reduc-bool-9.c: New testcase.

2 weeks agoInitial Nova Lake Support
Haochen Jiang [Wed, 22 Oct 2025 03:04:07 +0000 (11:04 +0800)] 
Initial Nova Lake Support

This patch will add initial support for Nova Lake according to Intel
ISE.

gcc/ChangeLog:

* common/config/i386/cpuinfo.h
(get_intel_cpu): Handle Nova Lake.
* common/config/i386/i386-common.cc (processor_name):
Add Nova Lake.
(processor_alias_table): Ditto.
* common/config/i386/i386-cpuinfo.h (enum processor_types):
Add INTEL_COREI7_NOVALAKE.
* config.gcc: Add -march=novalake.
* config/i386/driver-i386.cc (host_detect_local_cpu): Handle
novalake.
* config/i386/i386-c.cc (ix86_target_macros_internal): Ditto.
* config/i386/i386-options.cc (processor_cost_table): Ditto.
(m_NOVALAKE): New.
(m_CORE_HYBRID): Add novalake.
* config/i386/i386.h (enum processor_type): Ditto.
* doc/extend.texi: Ditto.
* doc/invoke.texi: Ditto.

gcc/testsuite/ChangeLog:

* g++.target/i386/mv16.C: Ditto.
* gcc.target/i386/funcspec-56.inc: Handle new march.

2 weeks agoi386: Correct cpu codename value for unknown model number
Haochen Jiang [Tue, 21 Oct 2025 03:21:45 +0000 (11:21 +0800)] 
i386: Correct cpu codename value for unknown model number

There are several changes for features enabled on cpus. r16-1666 disabled
CLDEMOTE on clients. r16-2224 removed Key locker since Panther Lake and
Clearwater forest. r16-4436 disabled PREFETCHI on Panther Lake.

The patches caused the current return guess value not aligned for
host_detect_local_cpu meeting the unknown model number. Correct the
logic according to the features enabled.

This patch will also backport to GCC14 and GCC15.

gcc/ChangeLog:

* config/i386/driver-i386.cc (host_detect_local_cpu): Correct
the logic for unknown model number cpu guess value.

2 weeks agoSimplify avx512 vector integer comparison when 2 operands are known equal
liuhongt [Mon, 20 Oct 2025 08:42:32 +0000 (01:42 -0700)] 
Simplify avx512 vector integer comparison when 2 operands are known equal

For comparison NEQ/LT/NLE, it's simplified to 0.
For comparison LE/EQ/NLT, it's simplied to (1u << nelt) - 1
gcc/ChangeLog:

PR target/122320
* config/i386/sse.md (*<avx512>_cmp<mode>3_dup_op): New define_insn_and_split.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr122320-mask16.c: New test.
* gcc.target/i386/pr122320-mask2.c: New test.
* gcc.target/i386/pr122320-mask32.c: New test.
* gcc.target/i386/pr122320-mask4.c: New test.
* gcc.target/i386/pr122320-mask64.c: New test.
* gcc.target/i386/pr122320-mask8.c: New test.

2 weeks agolibgccjit: Add _Float16, _Float32, _Float64 and __float128 support for jit
Antoni Boucher [Fri, 19 Sep 2025 12:44:04 +0000 (08:44 -0400)] 
libgccjit: Add _Float16, _Float32, _Float64 and __float128 support for jit

gcc/ChangeLog:

* config/i386/i386-jit.cc: Mark new float types as supported.

gcc/jit/ChangeLog:

* docs/topics/types.rst: Document new types.
* dummy-frontend.cc: Support new types in tree_type_to_jit_type.
* jit-common.h: Update NUM_GCC_JIT_TYPES.
* jit-playback.cc: Support new types in get_tree_node_for_type.
* jit-recording.cc: Support new types.
* libgccjit.h (gcc_jit_types): Add new types.

gcc/testsuite/ChangeLog:

* jit.dg/all-non-failing-tests.h: Mention new test.
* jit.dg/test-sized-float.c: New test.

2 weeks agoDaily bump.
GCC Administrator [Wed, 22 Oct 2025 00:19:22 +0000 (00:19 +0000)] 
Daily bump.

2 weeks agolibgccjit: Fix error on Power architectures caused by wrong jit_target_objs
Antoni Boucher [Thu, 9 Oct 2025 20:51:42 +0000 (16:51 -0400)] 
libgccjit: Fix error on Power architectures caused by wrong jit_target_objs

gcc/ChangeLog:
* config.gcc (jit_target_objs): Don't set this variable since
the object files don't exist.

2 weeks agoc2y: Allow unspecified arrays in generic association.
Martin Uecker [Sun, 12 Oct 2025 18:00:22 +0000 (20:00 +0200)] 
c2y: Allow unspecified arrays in generic association.

To allow unspecified arrays in generic association add a new
declaration context GENERIC_ASSOC for grokdeclarator and new
function grokgenassoc to be used by the parser.  The error
about unspecified array is moved from build_array_declarator
to grokdeclarator to be able to check for this.

gcc/c/ChangeLog:
* c-decl.cc (build_array_declarator): Remove error.
(grokgenassoc): New function.
(grokdeclarator): Add error.
* c-parser.cc (c_parser_generic_selection): Use grokgenassoc.
* c-tree.h (grokgenassoc): Add prototype.

gcc/testsuite/ChangeLog:
* gcc.dg/c2y-generic-6.c: New test.
* gcc.dg/c2y-generic-7.c: New test.

2 weeks agoc++: Implement C++23 P2674R1 - A trait for implicit lifetime types
Jakub Jelinek [Tue, 21 Oct 2025 17:17:11 +0000 (19:17 +0200)] 
c++: Implement C++23 P2674R1 - A trait for implicit lifetime types

The following patch attempts to implement the compiler side of the
C++23 P2674R1 paper.  As mentioned in the paper, since CWG2605
the trait isn't really implementable purely on the library side.

Because it is implemented completely on the compiler side, it
just uses SCALAR_TYPE_P and so can e.g. accept __int128 even in
-std=c++23 mode, even when std::is_scalar_v<__int128> is false in
that case.  And as an extention it (like Clang) accepts _Complex
types and vector types.
I must say I'm quite surprised that any array types are considered
implicit-lifetime, even if their element type is not, but perhaps
there is some reason for that.
Because std::is_array_v<int[0]> is false, it returns false for that
as well, dunno if that shouldn't be changed for implicit-lifetime.
It accepts also VLAs.

The library part has been split into a separate patch still pending
review; committing it now so that reflection can use it in its
std::meta::is_implicit_lifetime_type implementation.

2025-10-21  Jakub Jelinek  <jakub@redhat.com>

gcc/cp/
* cp-tree.h: Implement C++23 P2674R1 - A trait for implicit lifetime
types.
(implicit_lifetime_type_p): Declare.
* tree.cc (implicit_lifetime_type_p): New function.
* cp-trait.def (IS_IMPLICIT_LIFETIME): New unary trait.
* semantics.cc (trait_expr_value): Handle CPTK_IS_IMPLICIT_LIFETIME.
(finish_trait_expr): Likewise.
* constraint.cc (diagnose_trait_expr): Likewise.
gcc/testsuite/
* g++.dg/ext/is_implicit_lifetime.C: New test.

2 weeks agoarm: testsuite: [MVE] Fix expected code for vadcq_m and vsbcq_m [PR122189]
Christophe Lyon [Mon, 20 Oct 2025 14:31:21 +0000 (14:31 +0000)] 
arm: testsuite: [MVE] Fix expected code for vadcq_m and vsbcq_m [PR122189]

The original versions of these tests only took into account code
generated with -mfloat-abi=hard.

Depending on how the toolchain is configured, arm_v8_1m_mve may use
-mfloat-abi-softfp, which generates a different instructions order.

Depending on the -mtune setting, the order can also vary, so the patch
adds -fno-schedule-insns -fno-schedule-insns2 to avoid such
maintenance issues.

In particular, this fixes the failures with:
 -mthumb -march=armv7e-m+fp.dp -mtune=cortex-m7 -mfloat-abi=hard -mfpu=auto
 -mthumb -march=armv6s-m -mtune=cortex-m0 -mfloat-abi=soft -mfpu=auto

gcc/testsuite/ChangeLog:

PR target/122189
* gcc.target/arm/mve/intrinsics/vadcq_m_s32.c
* gcc.target/arm/mve/intrinsics/vadcq_m_u32.c
* gcc.target/arm/mve/intrinsics/vsbcq_m_s32.c
* gcc.target/arm/mve/intrinsics/vsbcq_m_u32.c

2 weeks agoOpenMP: Handle non-executable directives in intervening code [PR120180,PR122306]
Paul-Antoine Arras [Thu, 16 Oct 2025 16:22:08 +0000 (17:22 +0100)] 
OpenMP: Handle non-executable directives in intervening code [PR120180,PR122306]

OpenMP 6 permits non-executable directives in intervening code; this commit adds
support for a sensible subset, namely metadirectives, nothing, assume, and
'error at(compilation)'.
Also handle the special case where a metadirective can be resolved at parse time
to 'omp nothing'.
This fixes a build issue that affects 10 out 12 SPECaccel benchmarks.

Co-authored by: Tobias Burnus <tburnus@baylibre.com>

PR c/120180
PR fortran/122306

gcc/c/ChangeLog:

* c-parser.cc (c_parser_pragma): Accept a subset of non-executable
OpenMP directives in intervening code.
(c_parser_omp_error): Reject 'error at(execution)' in intervening code.
(c_parser_omp_metadirective): Return early if only one selector matches
and it resolves to 'omp nothing'.

gcc/cp/ChangeLog:

* parser.cc (cp_parser_omp_metadirective): Return early if only one
selector matches and it resolves to 'omp nothing'.
(cp_parser_omp_error): Reject 'error at(execution)' in intervening code.
(cp_parser_pragma): Accept a subset of non-executable OpenMP directives
as intervening code.

gcc/fortran/ChangeLog:

* gfortran.h (enum gfc_exec_op): Add EXEC_OMP_FIRST_OPENMP_EXEC and
EXEC_OMP_LAST_OPENMP_EXEC.
* openmp.cc (gfc_match_omp_context_selector): Remove static. Remove
checks on score. Add cleanup. Remove checks on trait properties.
(gfc_match_omp_context_selector_specification): Remove static. Adjust
calls to gfc_match_omp_context_selector.
(gfc_match_omp_declare_variant): Adjust call to
gfc_match_omp_context_selector_specification.
(match_omp_metadirective): Likewise.
(icode_code_error_callback): Reject all statements except
'assume' and 'metadirective'.
(gfc_resolve_omp_context_selector): New function.
(resolve_omp_metadirective): Skip metadirectives which context selectors
can be statically resolved to false. Replace metadirective by its body
if only 'nothing' remains.
(gfc_resolve_omp_declare): Call gfc_resolve_omp_context_selector for
each variant.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/imperfect1.c: Adjust dg-error.
* c-c++-common/gomp/imperfect4.c: Likewise.
* c-c++-common/gomp/pr120180.c: Move to...
* c-c++-common/gomp/pr120180-1.c: ...here. Remove dg-error.
* g++.dg/gomp/attrs-imperfect1.C: Adjust dg-error.
* g++.dg/gomp/attrs-imperfect4.C: Likewise.
* gfortran.dg/gomp/declare-variant-2.f90: Adjust dg-error.
* gfortran.dg/gomp/declare-variant-20.f90: Likewise.
* c-c++-common/gomp/pr120180-2.c: New test.
* g++.dg/gomp/pr120180-1.C: New test.
* gfortran.dg/gomp/pr120180-1.f90: New test.
* gfortran.dg/gomp/pr120180-2.f90: New test.
* gfortran.dg/gomp/pr122306-1.f90: New file.
* gfortran.dg/gomp/pr122306-2.f90: New file.

2 weeks agox86_64: Start TImode STV chains from zero-extension or *concatditi.
Roger Sayle [Tue, 21 Oct 2025 12:14:58 +0000 (13:14 +0100)] 
x86_64: Start TImode STV chains from zero-extension or *concatditi.

Currently x86_64's TImode STV pass has the restriction that candidate
chains must start with a TImode load from memory.  This patch improves
the functionality of STV to allow zero-extensions and construction of
TImode pseudos from two DImode values (i.e. *concatditi) to both be
considered candidate chain initiators.  For example, this allows chains
starting from an __int128 function argument to be processed by STV.

Compiled with -O2 on x86_64:

__int128 m0,m1,m2,m3;
void foo(__int128 m)
{
    m0 = m;
    m1 = m;
    m2 = m;
    m3 = m;
}

Previously generated:

foo:    xchgq   %rdi, %rsi
        movq    %rsi, m0(%rip)
        movq    %rdi, m0+8(%rip)
        movq    %rsi, m1(%rip)
        movq    %rdi, m1+8(%rip)
        movq    %rsi, m2(%rip)
        movq    %rdi, m2+8(%rip)
        movq    %rsi, m3(%rip)
        movq    %rdi, m3+8(%rip)
        ret

With the patch, we now generate:

foo: movq    %rdi, %xmm0
        movq    %rsi, %xmm1
        punpcklqdq      %xmm1, %xmm0
        movaps  %xmm0, m0(%rip)
        movaps  %xmm0, m1(%rip)
        movaps  %xmm0, m2(%rip)
        movaps  %xmm0, m3(%rip)
        ret

or with -mavx2:

foo: vmovq   %rdi, %xmm1
        vpinsrq $1, %rsi, %xmm1, %xmm0
        vmovdqa %xmm0, m0(%rip)
        vmovdqa %xmm0, m1(%rip)
        vmovdqa %xmm0, m2(%rip)
        vmovdqa %xmm0, m3(%rip)
        ret

Likewise, for zero-extension:

__int128 m0,m1,m2,m3;
void bar(unsigned long x)
{
    __int128 m = x;
    m0 = m;
    m1 = m;
    m2 = m;
    m3 = m;
}

Previously with -O2:

bar:    movq    %rdi, m0(%rip)
        movq    $0, m0+8(%rip)
        movq    %rdi, m1(%rip)
        movq    $0, m1+8(%rip)
        movq    %rdi, m2(%rip)
        movq    $0, m2+8(%rip)
        movq    %rdi, m3(%rip)
        movq    $0, m3+8(%rip)
        ret

with this patch:

bar: movq    %rdi, %xmm0
        movaps  %xmm0, m0(%rip)
        movaps  %xmm0, m1(%rip)
        movaps  %xmm0, m2(%rip)
        movaps  %xmm0, m3(%rip)
        ret

As shown in the examples above, the scalar-to-vector (STV) conversion of
*concatditi has an overhead [treating two DImode registers as a TImode
value is free on x86_64], but specifying this penalty allows the STV
pass to make an informed decision if the total cost/gain of the chain
is a net win.

2025-10-21  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-features.cc (timode_concatdi_p): New
function to recognize the various variants of *concatditi3_[1-7].
(scalar_chain::add_insn): Like VEC_SELECT, ZERO_EXTEND and
timode_concatdi_p instructions don't require their input
operands to be converted (to TImode).
(timode_scalar_chain::compute_convert_gain): Split/clone XOR and
IOR cases from AND case, to handle timode_concatdi_p costs.
<case PLUS>: Handle timode_concatdi_p conversion costs.
<case ZERO_EXTEND>: Provide costs of DImode to TImode extension.
(timode_convert_concatdi): Helper function to transform
a *concatditi3 instruction into a vec_concatv2di instruction.
(timode_scalar_chain::convert_insn): Split/clone XOR and IOR
cases from ANS case, to handle timode_concatdi_p using the new
timode_convert_concatdi helper function.
<case ZERO_EXTEND>: Convert zero_extendditi2 to *vec_concatv2di_0.
<case PLUS>: Handle timode_concatdi_p using the new
timode_convert_concatdi helper function.
(timode_scalar_to_vector_candidate_p): Support timode_concatdi_p
instructions in IOR, XOR and PLUS cases.
<case ZERO_EXTEND>: Consider zero extension of a register from
DImode to TImode to be a candidate.

gcc/testsuite/ChangeLog
* gcc.target/i386/sse4_1-stv-10.c: New test case.
* gcc.target/i386/sse4_1-stv-11.c: Likewise.
* gcc.target/i386/sse4_1-stv-12.c: Likewise.

2 weeks agoOpenMP: Update directive arrays used for 'omp assume(s)' with contains/absent
Tobias Burnus [Tue, 21 Oct 2025 11:31:19 +0000 (13:31 +0200)] 
OpenMP: Update directive arrays used for 'omp assume(s)' with contains/absent

Both Fortran and C/C++ have an array with classifications of directives;
currently, this array is only used to handle the restrictions of the
contains/absent clauses to the assume/assumes directives.

For C/C++, uncommenting 'declare mapper' was missed. Additionally,
'end ...' is a directive but not a directive name; hence, those
are now rejected as 'unknown directive' instead of as 'invalid'
directive.

Additionally, both lists now list newer entries (commented out) for
OpenMP 6.x - and a note (comment) was added for C/C++'s
'begin metadirective' and for Fortran's 'allocate', respectively.

gcc/c-family/ChangeLog:

* c-omp.cc (c_omp_directives): Uncomment 'declare mapper',
add comment to 'begin metadirective', add 6.x unimplemented
directives as comment-out entries.

gcc/c/ChangeLog:

* c-parser.cc (c_parser_omp_assumption_clauses): Switch to
'unknown' not 'invalid' directive name for end directives.

gcc/cp/ChangeLog:

* parser.cc (cp_parser_omp_assumption_clauses): Switch to
'unknown' not 'invalid' directive name for end directives.

gcc/fortran/ChangeLog:

* openmp.cc (gfc_omp_directive): Add comment to 'allocate';
add 6.x unimplemented directives as comment-out entries.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/assumes-2.c: Change for 'invalid'
to 'unknown' change for end directives.
* c-c++-common/gomp/begin-assumes-2.c: Likewise.
* c-c++-common/gomp/assume-2.c: Likewise. Check 'declare
mapper'.

2 weeks agotree-optimization/120687 - reduction chain with UB on signed overflow
Richard Biener [Mon, 20 Oct 2025 12:44:27 +0000 (14:44 +0200)] 
tree-optimization/120687 - reduction chain with UB on signed overflow

The following adds the ability to discover a reduction chain on a
series of statements that invoke undefined behavior on integer overflow.
This inhibits the reassoc pass from associating stmts in the way
naturally leading to a reduction chain.  The common mistake on the
source side is to rely on the += operator to sum multiple inputs.

After the refactoring of how we handle reduction chains we can
easily use vect_slp_linearize_chain to do this our selves and
rely on the vectorizer punning operations to unsigned given reduction
vectorization always associates.

PR tree-optimization/120687
* tree-vect-slp.cc (vect_analyze_slp_reduc_chain): When
there's no natural reduction chain see if vect_slp_linearize_chain
can recover one and built the SLP instance manually in that
case.
(vect_schedule_slp): Deal with NULL lanes when looking for
stores to remove.
* tree-vect-loop.cc (vect_transform_cycle_phi): Dump when we
are successfully transforming a reduction chain.

* gcc.dg/vect/vect-reduc-chain-4.c: New testcase.

2 weeks agoFix partial epilog for bool vectors
Richard Biener [Tue, 21 Oct 2025 09:09:41 +0000 (11:09 +0200)] 
Fix partial epilog for bool vectors

When we do epilogue vectorization the partial reduction of a bool
vector via vect_create_partial_epilog ends up being done on an
integer vector but we fail to pun back to a bool vector at the end,
causing an ICE later.  I couldn't manage to create a testcase
running into the failure but a pending patch will expose this on
gcc.dg/vect/vect-switch-ifcvt-3.c

* tree-vect-loop.cc (vect_create_partial_epilog): Pun back
to the requested type if necessary.

2 weeks agovect: Fix regression for PR104116
Avinash Jayakar [Tue, 21 Oct 2025 09:33:41 +0000 (15:03 +0530)] 
vect: Fix regression for PR104116

The commit gcc-16-4464-g6883d51304f added 30 new tests for testing
vectorization of {FLOOR,MOD,ROUND}_{DIV,MOD}_EXPR. Few of them failed
for certain targets due to the vectorization of runtime-check loop which
was not intended.
This patch disables optimization for all of the run-time check loops so
that the count of vectorized loop is always 1.

2025-10-21  Avinash Jayakar  <avinashd@linux.ibm.com>

gcc/testsuite/ChangeLog:
PR target/104116
* gcc.dg/vect/pr104116-ceil-div-2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-div-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-div.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-mod-2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-mod-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-mod.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-udiv-2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-udiv-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-udiv.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-umod-2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-umod-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-ceil-umod.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-div-2.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-div-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-div.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-mod-2.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-mod-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-floor-mod.c: disable vectorization.
* gcc.dg/vect/pr104116-round-div-2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-div-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-div.c: disable vectorization.
* gcc.dg/vect/pr104116-round-mod-2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-mod-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-mod.c: disable vectorization.
* gcc.dg/vect/pr104116-round-udiv-2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-udiv-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-udiv.c: disable vectorization.
* gcc.dg/vect/pr104116-round-umod-2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-umod-pow2.c: disable vectorization.
* gcc.dg/vect/pr104116-round-umod.c: disable vectorization.
* gcc.dg/vect/pr104116.h (init_arr): use std idiom, correct
indentation.
(init_uarr): use std idiom.

2 weeks agomatch: Add support for convert `((signed)x) < 0` to `x >= (unsigned)SIGNED_TYPE_MIN...
Andrew Pinski [Mon, 20 Oct 2025 22:48:43 +0000 (15:48 -0700)] 
match: Add support for convert `((signed)x) < 0` to `x >= (unsigned)SIGNED_TYPE_MIN` while detecting min/max [PR110068]

This copies the optimization which was done to fix PR 95699 to match detection of MIN/MAX
from minmax_replacement to match.
This is another step in getting rid of minmax_replacement in phiopt.  There are still a few
more min/max detections that needs to be handled before the removal. pr101024-1.c adds one
example of that but since the testcase currently passes I didn't xfail it.

pr110068-1.c adds a testcase which was not detected beforehand either.

Changes since v1:
* v2: Fix comment about how it is transformed.
      Use SIGNED_TYPE_MIN everywhere instead of mxing in SIGNED_TYPE_MAX too.

Bootstrapped and tested on x86_64-linux-gnu.

PR tree-optimization/95699
PR tree-optimization/101024
PR tree-optimization/110068

gcc/ChangeLog:

* match.pd (`(type1)x CMP CST1 ? (type2)x : CST2`): Treat
`(signed)x </>= 0` as `x >=/< SIGNED_TYPE_MIN`

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/pr101024-1.c: New test.
* gcc.dg/tree-ssa/pr110068-1.c: New test.

Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
2 weeks agoRedefine ASM_PREFERRED_EH_DATA_FORMAT for ppc[64]-vxworks
Olivier Hainque [Sat, 20 Apr 2024 11:43:56 +0000 (08:43 -0300)] 
Redefine ASM_PREFERRED_EH_DATA_FORMAT for ppc[64]-vxworks

This patch redefines ASM_PREFERRED_EH_DATA_FORMAT from the
otherwise inherited linux variant, preventing DW_EH_PE_indirect
in 64bit DKMs, where they are not strictly
needed and where the runtime load could resolve the DW.refs to
symbols of the same name within a different DKM loaded previously.

gcc/
* config/rs6000/vxworks.h (ASM_PREFERRED_EH_DATA_FORMAT):
Redefine.

2 weeks agoReplace VSB_DIR by sysroot ref in VXWORKS_ADDITIONAL_CPP_SPEC
Olivier Hainque [Sat, 20 Apr 2024 15:37:51 +0000 (12:37 -0300)] 
Replace VSB_DIR by sysroot ref in VXWORKS_ADDITIONAL_CPP_SPEC

VXWORKS_ADDITIONAL_CPP_SPEC has an artificial guard on
-fself-test to prevent all-gcc build failures from self-tests
in environments where VSB_DIR is not defined.

The libraries are not built during such
checks; having a VxWorks installation at hand is not necessary, and
requiring VSB_DIR to be defined is inappropriate.

This patch replaces the use of %getenv(VSB_DIR) by $sysroot references
which allows removing the artifical guard of -fself-tests.

gcc/
* config/vxworks.h (VXWORKS_ADDITIONAL_CPP_SPEC):
Remove guard on -fself-tests and replace %:getenv(VSB_DIR) by
sysroot references.

2 weeks agoDaily bump.
GCC Administrator [Tue, 21 Oct 2025 00:20:03 +0000 (00:20 +0000)] 
Daily bump.

2 weeks agoFix minor RISC-V testsuite failure
Jeff Law [Mon, 20 Oct 2025 22:31:20 +0000 (16:31 -0600)] 
Fix minor RISC-V testsuite failure

This fixes reduc-8 yet again.  This time the required "a2" moved to the other source operand of the add.  So the regexp is further expanded to allow add anyreg,anyreg,a2 or add anyreg,a2,anyreg.

gcc/testsuite
* gcc.target/riscv/rvv/autovec/reduc/reduc-8.c: Adjust expected output.

2 weeks agoAda: Add missing qualifier for integer literal
Eric Botcazou [Mon, 20 Oct 2025 22:21:37 +0000 (00:21 +0200)] 
Ada: Add missing qualifier for integer literal

gcc/ada/
PR ada/102078
* affinity.c (__gnat_set_affinity_mask): Add U qualifier.

2 weeks agoipa: Delete callback edges when redirecting to unreachable.
Josef Melcr [Sat, 18 Oct 2025 10:47:17 +0000 (12:47 +0200)] 
ipa: Delete callback edges when redirecting to unreachable.

When a callback-carrying edge is redirected to __builtin_unreachable,
the associated callbacks will never get called, so the corresponding
callback edges must be deleted, as they no longer reflect the reality.

The line in analyze_function_body is an obvious typo I discovered during
debugging, so I decided to bundle it in.

gcc/ChangeLog:

* ipa-fnsummary.cc (redirect_to_unreachable): Purge callback
edges when redirecting the carrying edge.
(analyze_function_body): Fix typo.

Signed-off-by: Josef Melcr <jmelcr02@gmail.com>
2 weeks agolibgccjit: Add gcc_jit_context_new_array_type_u64
Antoni Boucher [Sat, 4 Mar 2023 05:44:49 +0000 (00:44 -0500)] 
libgccjit: Add gcc_jit_context_new_array_type_u64

gcc/jit/ChangeLog:

* docs/topics/compatibility.rst (LIBGCCJIT_ABI_37): New ABI tag.
* docs/topics/types.rst: Document
gcc_jit_context_new_array_type_u64.
* jit-playback.cc (new_array_type): Change num_elements type to
uint64_t.
* jit-playback.h (new_array_type): Change num_elements type to
uint64_t.
* jit-recording.cc (recording::context::new_array_type): Change
num_elements type to uint64_t.
(recording::array_type::make_debug_string): Use uint64_t
format.
(recording::array_type::write_reproducer): Switch to
gcc_jit_context_new_array_type_u64.
* jit-recording.h (class array_type): Change num_elements type
to uint64_t.
(new_array_type): Change num_elements type to uint64_t.
(num_elements): Change return type to uint64_t.
* libgccjit.cc (gcc_jit_context_new_array_type_u64):
New function.
* libgccjit.h (gcc_jit_context_new_array_type_u64):
New function.
* libgccjit.exports: New function.
* libgccjit.map: New function.

gcc/testsuite/ChangeLog:

* jit.dg/all-non-failing-tests.h: Add test-arrays-u64.c.
* jit.dg/test-arrays-u64.c: New test.

2 weeks agotestsuite: Move ipcp-cb* from ipa to libgomp
Josef Melcr [Mon, 20 Oct 2025 21:52:16 +0000 (23:52 +0200)] 
testsuite: Move ipcp-cb* from ipa to libgomp

This patch addresses the incorrectly placed tests, which fail if the
testsuite is ran and gcc has not been installed yet, as discussed
here:
https://gcc.gnu.org/pipermail/gcc-patches/2025-October/698095.html.

gcc/testsuite/ChangeLog:
* gcc.dg/ipa/ipcp-cb-spec1.c: Moved to libgomp/testsuite/libgomp.c/.
* gcc.dg/ipa/ipcp-cb-spec2.c: Likewise.
* gcc.dg/ipa/ipcp-cb1.c: Likewise.
libgomp/ChangeLog:
* testsuite/libgomp.c/ipcp-cb-spec1.c: Moved from
gcc/testsuite/gcc.dg/ipa/.
* testsuite/libgomp.c/ipcp-cb-spec2.c: Likewise.
* testsuite/libgomp.c/ipcp-cb1.c: Likewise.

Signed-off-by: Josef Melcr <jmelcr02@gmail.com>
2 weeks agoAda: Fix incorrect specification of GNAT.Calendar.Time_IO "%c"
Eric Botcazou [Mon, 20 Oct 2025 21:57:01 +0000 (23:57 +0200)] 
Ada: Fix incorrect specification of GNAT.Calendar.Time_IO "%c"

The timezone is not printed by the "%c" specifier.

gcc/ada/
PR ada/32318
* libgnat/g-catiio.adb (Image_Helper) <'c'>: Fix comment.

2 weeks agolibgccjit: Do not treat warnings as errors
Antoni Boucher [Mon, 12 Feb 2024 23:49:43 +0000 (19:49 -0400)] 
libgccjit: Do not treat warnings as errors

gcc/jit/ChangeLog:

* jit-playback.cc (add_error, add_error_va): Send DK_ERROR to
add_error_va.
(add_diagnostic): Call add_diagnostic instead of add_error.
* jit-recording.cc (DEFINE_DIAGNOSTIC_KIND): New define.
(recording::context::add_diagnostic): New function.
(recording::context::add_error): Send DK_ERROR to add_error_va.
(recording::context::add_error_va): New parameter diagnostic_kind.
* jit-recording.h (add_diagnostic): New function.
(add_error_va): New parameter diagnostic_kind.
* libgccjit.cc (jit_error): Send DK_ERROR to add_error_va.

gcc/testsuite/ChangeLog:

* jit.dg/test-error-array-bounds.c: Fix test.

2 weeks agolibgccjit: Fix infinite recursion in gt_ggc_mx_lang_tree_node
Antoni Boucher [Fri, 3 Jun 2022 01:14:06 +0000 (21:14 -0400)] 
libgccjit: Fix infinite recursion in gt_ggc_mx_lang_tree_node

2022-06-02  Antoni Boucher  <bouanto@zoho.com>

gcc/jit/
PR jit/105827
* dummy-frontend.cc: Fix lang_tree_node.
* jit-common.h: New function (jit_tree_chain_next) used by
lang_tree_node.

2 weeks agolibgccjit: Support more target builtin types
Antoni Boucher [Tue, 18 Mar 2025 16:51:23 +0000 (12:51 -0400)] 
libgccjit: Support more target builtin types

This also adds option to abort on unsupported type in order to be able
to detect new unsupported types more easily.

gcc/jit/ChangeLog:
PR jit/117886
* dummy-frontend.cc: Support some missing types.
* jit-playback.h (get_abort_on_unsupported_target_builtin): New
function.
* jit-recording.cc (get_abort_on_unsupported_target_builtin,
set_abort_on_unsupported_target_builtin): New functions.
* jit-recording.h (get_abort_on_unsupported_target_builtin,
set_abort_on_unsupported_target_builtin): New functions.
(m_abort_on_unsupported_target_builtin): New field.
* libgccjit.cc
(gcc_jit_context_set_abort_on_unsupported_target_builtin): New
function.
* libgccjit.h
(gcc_jit_context_set_abort_on_unsupported_target_builtin): New
function.
* libgccjit.exports (LIBGCCJIT_ABI_36): New ABI tag.
* libgccjit.map (LIBGCCJIT_ABI_36): New ABI tag.
* docs/topics/compatibility.rst (LIBGCCJIT_ABI_36): New ABI tag.
* docs/topics/contexts.rst: Document new function.

2 weeks agohurd: Add OPTION_GLIBC_P and OPTION_GLIBC
Svante Signell [Sun, 6 Feb 2022 11:43:23 +0000 (11:43 +0000)] 
hurd: Add OPTION_GLIBC_P and OPTION_GLIBC

GNU/Hurd uses glibc just like GNU/Linux.

This is needed for gcc to notice that glibc supports split stack in
finish_options.

PR go/104290
gcc/ChangeLog:
* config/gnu.h (OPTION_GLIBC_P, OPTION_GLIBC): Define.

2 weeks agoc++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads...
Thomas Schwinge [Mon, 20 Oct 2025 15:36:49 +0000 (17:36 +0200)] 
c++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads: Adjust 'libgomp.c++/{target-flex-101.C,target-std__flat_map-concurrent.C,target-std__flat_multimap-concurrent.C}' [PR114457, PR122268, PR120450]

With commit r16-4212-gf256a13f8aed833fe964a2ba541b7b30ad9b4a76
"c++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads [PR114457]",
we acquired:

    {+FAIL: libgomp.c++/target-flex-101.C (internal compiler error: in assign_temp, at function.cc:990)+}
    [-PASS:-]{+FAIL:+} libgomp.c++/target-flex-101.C (test for excess errors)
    [-PASS:-]{+UNRESOLVED:+} libgomp.c++/target-flex-101.C [-execution test-]{+compilation failed to produce executable+}

... for GCN, nvptx offloading compilation, and on the other hand:

    [-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_map-concurrent.C (internal compiler error[-: in assign_temp, at function.cc:990)-]
    [-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_map-concurrent.C (test for excess errors)
    [-UNRESOLVED:-]{+PASS:+} libgomp.c++/target-std__flat_map-concurrent.C [-compilation failed to produce executable-]{+execution test+}

    [-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C (internal compiler error[-: in assign_temp, at function.cc:990)-]
    [-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C (test for excess errors)
    [-UNRESOLVED:-]{+PASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C [-compilation failed to produce executable-]{+execution test+}

... for GCN offloading compilation (already PASSed for nvptx).

Note that these test cases explicitly use '-std=c++23', so don't undergo the
new C++26 P2795R5 functionality.  Yet, comparing before vs. after that commit,
in the 'gimple' dumps (that is, early host compilation), there are a lot of
changes where 'gimple_assign <constructor, [...], {CLOBBER(bob)}, NULL, NULL>'s
and relatedly 'gimple_bind's newly appear/no longer appear elsewhere.  This
leads to correspondingly different code at the beginning of offloading
compilation.  Why/how that now ('libgomp.c++/target-flex-101.C') vs. before
('libgomp.c++/{target-std__flat_map-concurrent.C,target-std__flat_multimap-concurrent.C}')
translates into 'expand' ICEs, I can't tell.

PR c++/114457
PR c++/122268
PR c++/120450
libgomp/
* testsuite/libgomp.c++/target-flex-101.C: XFAIL GCN, nvptx
offloading compilation.
* testsuite/libgomp.c++/target-std__flat_map-concurrent.C:
Un-XFAIL GCN offloading compilation.
* testsuite/libgomp.c++/target-std__flat_multimap-concurrent.C:
Likewise.

2 weeks agoc++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads...
Thomas Schwinge [Mon, 20 Oct 2025 15:36:49 +0000 (17:36 +0200)] 
c++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads: Adjust 'c-c++-common/goacc/kernels-decompose-pr100280-1.c' [PR114457]

With commit r16-4212-gf256a13f8aed833fe964a2ba541b7b30ad9b4a76
"c++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads [PR114457]",
we acquired:

    @@ -181180,8 +184423,8 @@ PASS: c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26  at line 14
    PASS: c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26  at line 15 (test for warnings, line 12)
    PASS: c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26  at line 16 (test for warnings, line 12)
    PASS: c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26 (test for excess errors)
    [-XFAIL:-]{+XPASS:+} c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26 TODO at line 18 (test for warnings, line 19)
    [-XFAIL:-]{+XPASS:+} c-c++-common/goacc/kernels-decompose-pr100280-1.c  -std=c++26 TODO location at line 17 (test for bogus messages, line 10)

As in other OpenACC 'kernels' test cases, the underlying issue again is
PR121975 "Various goacc failures with -ftrivial-auto-var-init=zero" (to be
resolved later on).

PR c++/114457
gcc/testsuite/
* c-c++-common/goacc/kernels-decompose-pr100280-1.c: Skip for
c++26 until PR121975 is fixed.

2 weeks agoAda: Fix Default_Component_Value aspect wrongly ignored on derived type
Eric Botcazou [Mon, 20 Oct 2025 19:01:06 +0000 (21:01 +0200)] 
Ada: Fix Default_Component_Value aspect wrongly ignored on derived type

This is again an old issue, which was mostly fixed a few releases ago except
for the specific case of an array type derived from String.

gcc/ada/
PR ada/68179
* exp_ch3.adb (Expand_Freeze_Array_Type): Build an initialization
procedure for a type derived from String declared with the aspect
Default_Aspect_Component_Value.

gcc/testsuite/
* gnat.dg/component_value1.adb: New test.

2 weeks agoAda: Fix use type clause invalidated by use clause in nested package
Eric Botcazou [Mon, 20 Oct 2025 18:48:39 +0000 (20:48 +0200)] 
Ada: Fix use type clause invalidated by use clause in nested package

This is an old issue, whereby a use type clause is partially invalidated by
a use clause in a nested package, a variant of PR ada/64869 recently fixed.
The problem occurs only for unusual primitive operators because of a small
oversight in the implementation.  The fix simply aligns this implementation
with the one exercised by PR ada/64869, which is more robust.

gcc/ada/
PR ada/52319
* sem_ch7.adb (Uninstall_Declarations): Use direct test on Nkind
to spot operators.
* sem_ch8.adb (End_Use_Package): Also test the Etype of operators
to spot those which are primitive operators of use-visible types.

gcc/testsuite/
* gnat.dg/use_type3.adb: New test.

2 weeks agoEnsure use of gcc's version of stdatomic.h in gthr-vxworks
Ashley Gay [Fri, 8 Mar 2024 11:30:35 +0000 (11:30 +0000)] 
Ensure use of gcc's version of stdatomic.h in gthr-vxworks

VxWorks provides its own version of the standard stdatomic.h, possibly
relying on non-gcc builtins, and our implementation of the gthr API resorts
to VxWorks specific functions for atomicity features.

When compiling libgcc (with gcc), make sure gcc's version of stdatomic.h
is used: #include it here, first, then define the macro used to guard the
system version so it doesn't get expanded when included indirectly by
other system headers.

2025-10-20  Olivier Hainque  <hainque@adacore.com>
    Ashley Gay  <gay@adacore.com>

libgcc/
* config/gthr-vxworks.h: Include stdatomic.h and prevent indirect
inclusion of contents from the system version of that header.

2 weeks agoTidy bits of libgcc/config/gthr-vxworks
Olivier Hainque [Wed, 9 Jul 2025 10:48:34 +0000 (10:48 +0000)] 
Tidy bits of libgcc/config/gthr-vxworks

This addresses a variety of warnings about missing prototypes
or suspicious ptr-to-function conversions.

libgcc/
* config/gthr-vxworks-thread.c (__init_gthread_tcb): Make static.
(__delete_gthread_tcb): Likewise.
(__task_wrapper): Likewise.
(__gthread_create): Convert __task_wrapper to (void *) before going
to (FUNCPTR).
* config/gthr-vxworks-tls.c (tls_delete_hook): Accommodate prototype
variations between kernel and rtp. Return STATUS.

2 weeks agoxtensa: Make all memory constraints special
Takayuki 'January June' Suwa [Thu, 9 Oct 2025 21:33:36 +0000 (06:33 +0900)] 
xtensa: Make all memory constraints special

In a previous commit (fb7b82964f54192d0723a45c0657d2eb7c5ac97c), we fixed an issue
where loads from literal pool to a hardware floating-point register were double-
indirected; that is, the address of the literal pool entry was temporarily loaded
from another entry into the address (GP) register, and then loaded from that
address into the FP register.  However, we discovered that the same issue could
occur in rare cases when loading FP constants into address registers.

Similarly, this problem can be avoided by prefixing the corresponding alternative
constraint with '^' to increase the cost of Reload/LRA, but as a more fundamental
and comprehensive solution, this patch defines all memory constraint definitions
using define_special_memory_constraint, so that reloads cannot occur for addresses
(based on a good suggestion from Jeff Law).

gcc/ChangeLog:

* config/xtensa/constraints.md (R, U):
Change define_memory_constraint to define_special_memory_constraint.
* config/xtensa/xtensa.md
(movsi_internal, movhi_internal, movqi_internal):
Rearrange their alternatives in the order of constant assignment, register-
register move, load, store and special.  And also consolidate overlapping
alternatives.
(movsf_internal): Rearrange the alternatives as above, and remove the '^'
alternative character which is no longer needed.

2 weeks agoxtensa: Make individual use of CONST16 instruction
Takayuki 'January June' Suwa [Thu, 9 Oct 2025 21:29:16 +0000 (06:29 +0900)] 
xtensa: Make individual use of CONST16 instruction

Until now, in Xtensa ISA, the CONST16 machine instruction (which shifts a
specified register left by half a word and stores a 16-bit constant value
in the low halfword of the register) has always been used in pairs and
only for full-word constant value assignments.

This patch provides a new insn definition for using CONST16 alone, and
also adds a constantsynth method that saves one byte for constant assign-
ments within a certain range when TARGET_DENSITY is also enabled.

gcc/ChangeLog:

* config/xtensa/xtensa.cc
(constantsynth_method_const16): New.
(constantsynth_methods): Append constantsynth_method_const16().
(constantsynth_info): Add cost calculation for full-word constant
assignment when TARGET_CONST16 is enabled.
(constantsynth_pass1): Change it so that it works regardless of
TARGET_CONST16.
* config/xtensa/xtensa.md (*xtensa_const16): New.

2 weeks agoxtensa: Apply split_DI_SF_DF_const() even if TARGET_CONST16 or TARGET_AUTOLITPOOLS
Takayuki 'January June' Suwa [Thu, 9 Oct 2025 21:26:54 +0000 (06:26 +0900)] 
xtensa: Apply split_DI_SF_DF_const() even if TARGET_CONST16 or TARGET_AUTOLITPOOLS

Otherwise, if TARGET_CONST16 or TARGET_AUTOLITPOOLS is enabled, DI/SF/DFmode
constant assignments will not benefit from their splitting or constantsynth.

gcc/ChangeLog:

* config/xtensa/xtensa.cc (do_largeconst):
Change split_DI_SF_DF_const() to be called unconditionally.

2 weeks agolibstdc++: Implement P3060R3: Add std::views::indices(n)
Yuao Ma [Tue, 14 Oct 2025 16:28:48 +0000 (00:28 +0800)] 
libstdc++: Implement P3060R3: Add std::views::indices(n)

This patch adds the views::indices function using iota.

libstdc++-v3/ChangeLog:

* include/bits/version.def: Add ranges_indices FTM.
* include/bits/version.h: Regenerate.
* include/std/ranges: Implement views::indices.
* testsuite/std/ranges/indices/1.cc: New test.

2 weeks agoInclude linux-protos.h for ppc*vxworks7r2
Olivier Hainque [Thu, 10 Jul 2025 09:38:30 +0000 (09:38 +0000)] 
Include linux-protos.h for ppc*vxworks7r2

This provides prototypes for target hooks dragged in through linux.h,
in a similar fashion as the ppc*-linux ports do.

gcc/
* config.gcc (powerpc*-wrs-vxworks7r*): Add linux-protos.h
to tm_p_file.

2 weeks agolibstdc++: Deduce function_ref<M&() noexcept> from member object pointers.
Tomasz Kamiński [Mon, 20 Oct 2025 13:19:43 +0000 (15:19 +0200)] 
libstdc++: Deduce function_ref<M&() noexcept> from member object pointers.

Implement resolution of LWG4425.

libstdc++-v3/ChangeLog:

* include/bits/funcwrap.h (__polyfunc::__deduce_funcref):
Adjust signature produced for member object pointers.
* testsuite/20_util/function_ref/deduction.cc: Update tests.

2 weeks agoInfer TOOL/TOOL_FAMILY from vxworks-predef.h on VxWorks7
Olivier Hainque [Sat, 20 Apr 2024 15:35:41 +0000 (12:35 -0300)] 
Infer TOOL/TOOL_FAMILY from vxworks-predef.h on VxWorks7

This change moves, for VxWorks 7, the setting of the TOOL
and TOOL_FAMILY macros from a builtin_define to a run-time
computation from vxworks-predefs.h.

This is useful on Vx7 to allow a single toolchain to be used
for instances of VxWorks based on either a gnu or an llvm system
toolchain for a given cpu (typically, powerpc).

This is achieved by leveraging the existence of a very basic
autoconf.h file in all VxWorks 7 VSBs, #included directly from
vxworks-predef.h.

gcc/
* config/vxworks.h (VXWORKS_OS_CPP_BUILTINS): Only
builtin_define TOOL and TOOL_FAMILY for !TARGET_VXWORKS7.
Augment comment on VXWORKS_PERSONALITY.
* config/vxworks/vxworks-predef.h: Infer TOOL and TOOL_FAMILY
from the VSB autoconf.h when we have one, determined by the presence
of a _VSB_CONFIG_FILE definition.

libgcc/
* config/t-vxworks: -include vxworks-predef.h explicitly, as the
automatic inclusion is disabled by -nostdinc.

2 weeks agoaarch64: Add support for menable-sysreg-checking flag.
Srinath Parvathaneni [Mon, 20 Oct 2025 13:07:55 +0000 (14:07 +0100)] 
aarch64: Add support for menable-sysreg-checking flag.

Hi All,

In the current Binutils we have disabled the feature gating for sysreg
by default and we have introduced a new flag "-menable-sysreg-checking"
to renable some of this checking.

However in GCC, we have disabled the feature gating of sysreg to read/write
intrinsics __arm_[wr]sr* and we have not added any mechanism to check the
feature gating if needed similar to Binutils.

This patch adds the support for the flag "-menable-sysreg-checking" which
renables some of the feature checking of sysreg to read/write intrinsics
__arm_[wr]sr* similar to Binutils.

For inline assembly, sysreg checks are not performed by CC1 and are
instead delegated to the assembler. By default, the assembler does not
perform these checks either. With this patch, the -menable-sysreg-checking
flag passed to the compiler will also be propagated to the assembler,
enabling sysreg checking for inline assembly.

gcc/ChangeLog:

* config/aarch64/aarch64-elf.h (ASM_SPEC): Update the macro.
* config/aarch64/aarch64.cc (aarch64_valid_sysreg_name_p):
Add feature check condition.
(aarch64_retrieve_sysreg): Likewise.
* config/aarch64/aarch64.opt (menable-sysreg-checking):
Define new flag.
* doc/invoke.texi (menable-sysreg-checking): Document new flag.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/acle/asm-inlined-sysreg-1.c: New test.
* gcc.target/aarch64/acle/asm-inlined-sysreg-2.c: Likewise.
* gcc.target/aarch64/acle/rwsr-gated-1.c: Likewise.
* gcc.target/aarch64/acle/rwsr-gated-2.c: Likewise.
* lib/target-supports.exp
(check_effective_target_aarch64_sysreg_guarding_ok): Check
assembler support of -menable-sysreg-checking flag.

2 weeks agoMAINTAINERS: Add myself to vectorizer maintainer list
Tamar Christina [Mon, 20 Oct 2025 13:12:31 +0000 (14:12 +0100)] 
MAINTAINERS: Add myself to vectorizer maintainer list

Following the announcement on
https://gcc.gnu.org/pipermail/gcc/2025-October/246833.html
adding myself to vectorizer maintainer list.

ChangeLog:

* MAINTAINERS (Various Maintainers): Add myself for the vectorizer.

2 weeks agoFix minor testsuite scan failures for RISC-V
Jeff Law [Mon, 20 Oct 2025 11:48:54 +0000 (05:48 -0600)] 
Fix minor testsuite scan failures for RISC-V

This fixes minor testsuite fallout after some of Jan's recent changes, nothing
of real significance, just minor changes in codegen causing scan tests to fail.
It's mostly an -O1/-Og problem and we can just skip the tests for those.

gcc/testsuite
* gcc.target/riscv/rvv/vsetvl/imm_switch-6.c: Skip scan-asm test for -O1 too.
* gcc.target/riscv/rvv/vsetvl/imm_switch-7.c: Likewise.
* gcc.target/riscv/shrink-wrap-1.c: Likewise.  Skip for -Og as well.
* gcc.target/riscv/xandes/xandesperf-1.c: Adjust expected output.

2 weeks agoAda: Use Osint.Program_Name in gnatchop
Nicolas Boulenguez [Mon, 20 Oct 2025 11:08:22 +0000 (13:08 +0200)] 
Ada: Use Osint.Program_Name in gnatchop

This aligns gnatchop with the other GNAT tools when it comes to locating
GCC's driver executable.

gcc/ada/
PR ada/87777
* gnatchop.adb: Add with clause for Osint.
(Locate_Executable): Delete.
(Gnatchop): Use Osint.Program_Name and Locate_Exec_On_Path instead
of Locate_Executable to locate GCC's driver executable.

2 weeks agotop-level: Add forgejo sanity checks
Christophe Lyon [Mon, 1 Sep 2025 15:51:33 +0000 (15:51 +0000)] 
top-level: Add forgejo sanity checks

Add a sample workflow for Forgejo, as an example of integrated CI.

To keep it lightweight, we run only two small checks on each patch of
the series:
- contrib/check_GNU_style.py
  which catches common mistakes (spaces vs tab, missing spaces, ...)
  but has some false positive warnings.

- contrib/gcc-changelog/git_check_commit.py
  which checks the commit message and ChangeLog entry

In order to run both checks even if the other fails, we use two steps
with 'continue-on-error: true', and we need a 'final-result'
consolidation step to generate the global status.

ChangeLog:
* .forgejo/workflows/sanity-checks.yaml: New file.

2 weeks agolibstdc++: Remove undeclared macros from configure.ac [PR122322]
Jonathan Wakely [Sat, 18 Oct 2025 21:05:43 +0000 (22:05 +0100)] 
libstdc++: Remove undeclared macros from configure.ac [PR122322]

The additions inr16-4443-g651bf5126da124 cause errors when running
autoreconf.

libstdc++-v3/ChangeLog:

PR libstdc++/122322
* configure.ac (with_newlib) <*-rtems*>: Remove
HAVE_SYS_IOCT4YL_H, _GLIBCXX_USE_LINK, _GLIBCXX_USE_READLINK,
_GLIBCXX_USE_SYMLINK, _GLIBCXX_USE_TRUNCATE, and
_GLIBCXX_USE_FDOPENDIR. Remove duplicates.
* configure: Regenerate.

2 weeks agoAda: Fix spurious warning for renaming of component of VFA record
Eric Botcazou [Mon, 20 Oct 2025 09:21:21 +0000 (11:21 +0200)] 
Ada: Fix spurious warning for renaming of component of VFA record

This is a regression present on the mainline and all active branches: the
compiler gives a spurious "is not referenced" warning for the renaming of
a component of a Volatile_Full_Access record.

gcc/ada/
PR ada/107536
* exp_ch2.adb (Expand_Renaming): Mark the entity as referenced.

gcc/testsuite/
* gnat.dg/renaming18.adb: New test.

2 weeks agotree-optimization/121631 - UB in vector epilogue
Richard Biener [Fri, 17 Oct 2025 12:03:21 +0000 (14:03 +0200)] 
tree-optimization/121631 - UB in vector epilogue

The vectorizer fails to take UB due to signed overflow into account
when generating code for the epilogue of a signed reduction.  The
following tries to make sure to perform the actual reduction
computations in an unsigned type.  I did not bother to adjust
inputs to internal functions like .REDUC_PLUS.

PR tree-optimization/121631
* tree-vect-loop.cc (vect_create_epilog_for_reduction):
When the reduction operation invokes UB on signed overflow
make sure to perform operations with it on an unsigned type.

2 weeks agoImplement bool reduction vectorization
Richard Biener [Thu, 9 Oct 2025 12:03:29 +0000 (14:03 +0200)] 
Implement bool reduction vectorization

Currently we mess up here in two places.  One is pattern recognition
which computes a mask-precision for a bool reduction PHI that's
inconsistent with that of the latch definition.  This is solved by
iterating the mask-precision computation.  The second is that the
reduction epilogue generation and the code querying support for it
isn't ready for mask inputs.  The following fixes this by falling
back to doing all the epilogue processing on a data type again, if
the target does not support a direct mask reduction.  For that we
utilize the newly added reduc_sbool_{and,ior,xor}_scal optabs
so we can go the direct IFN path on masks if the target supports
that.  In the future we can also implement an additional fallback
for IOR and AND reductions using a scalar cond-expr like
mask != 0 ? true : false, but the new optabs provide more information
to the target.

PR tree-optimization/101639
PR tree-optimization/103495
* tree-vectorizer.h (vect_reduc_info_s): Add reduc_type_for_mask.
(VECT_REDUC_INFO_VECTYPE_FOR_MASK): New.
* tree-vect-patterns.cc (vect_determine_mask_precision):
Return whether the mask precision changed.
(vect_determine_precisions): Iterate mask precision computation
for loop vectorization.
* tree-vect-loop.cc (get_initial_defs_for_reduction): Properly
convert non-mask initial values to a mask initial def for
the reduction.
(sbool_reduction_fn_for_fn): New function.
(vect_create_epilog_for_reduction): For a mask input convert
it to the vector type analysis decided to use.  Use a regular
conversion for the final convert to the scalar code type.
(vectorizable_reduction): Support mask reductions.  Verify
we can compute a data vector from the mask result or a direct
maks reduction is provided by the target.

* gcc.dg/vect/vect-reduc-bool-1.c: New testcase.
* gcc.dg/vect/vect-reduc-bool-2.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-3.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-4.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-5.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-6.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-7.c: Likewise.
* gcc.dg/vect/vect-reduc-bool-8.c: Likewise.

2 weeks agoAdd reduc_sbool_{and,ior,xor}_scal optabs
Richard Biener [Tue, 14 Oct 2025 09:48:43 +0000 (11:48 +0200)] 
Add reduc_sbool_{and,ior,xor}_scal optabs

The following adds named patterns for reducing of vector masks with
AND, IOR and XOR to be used by the vectorizer.  A slight complication
are targets using scalar integer modes as mask modes, as for those
the mode for low-precision masks is ambiguous.  For this reason the
optab follows what vec_pack_sbool_trunc does and passes an additional
CONST_INT operand indicating the number of lanes in the input mask.
Note this is done always when the vector mask mode is an integer mode
and never otherwise.

* doc/md.texi (reduc_sbool_{and,ior,xor}_scal_<mode>): Document.
* optabs.def (reduc_sbool_and_scal_optab,
reduc_sbool_ior_scal_optab, reduc_sbool_xor_scal_optab): New.
* internal-fn.def (REDUC_SBOOL_AND, REDUC_SBOOL_IOR,
REDUC_SBOO_XOR): Likewise.
* internal-fn.cc (reduc_sbool_direct): New initializer.
(expand_reduc_sbool_optab_fn): New expander.
(direct_reduc_sbool_optab_supported_p): New.

2 weeks agoUpdate auto-vectorizer maintainance area
Richard Biener [Mon, 20 Oct 2025 07:31:15 +0000 (09:31 +0200)] 
Update auto-vectorizer maintainance area

The following adjusts the attribution of the auto-vectorizer area
to say 'vectorizer (+ tree-if-conv)' as approved by the SC.

* MAINTAINERS (auto-vectorizer): Change attribution to
vectorizer (+ tree-if-conv).

2 weeks agox86: Optimize copysign (x, const_double)
H.J. Lu [Sun, 19 Oct 2025 01:13:52 +0000 (09:13 +0800)] 
x86: Optimize copysign (x, const_double)

After

commit 3f176e1adc6bc9cc2c21222d776b51d9f43cb66b
Author: Tamar Christina <tamar.christina@arm.com>
Date:   Thu Nov 9 13:59:39 2023 +0000

    middle-end: optimize fneg (fabs (x)) to copysign (x, -1) [PR109154]

fneg (fabs (x)) is expanded to copysign (x, -1).  Swap constraints for
operands[1] and operands[2] in copysign<mode>3 pattern to optimize

  y = copysign (x, const_double)

instead of

  y = copysign (const_double, x)

Simplify

  y = copysign (x, positive_const_double)

to

  y = ~signbit_mask & x

and

  y = copysign (x, negative_const_double)

to

  y = signbit_mask | x

gcc/

PR target/99930
PR target/122323
* config/i386/i386-expand.cc (ix86_expand_copysign): Swap
operands[1] with operands[2].  Optimize copysign (x, const_double)
instead of copysign (const_double, x).
* config/i386/i386.md (copysign<mode>3): Swap constraints for
operands[1] and operands[2].

gcc/testsuite/

PR target/99930
PR target/122323
* gcc.target/i386/builtin-copysign-2.c: New test.
* gcc.target/i386/builtin-copysign-3.c: Likewise.
* gcc.target/i386/builtin-copysign-4.c: Likewise.
* gcc.target/i386/builtin-copysign-5.c: Likewise.
* gcc.target/i386/builtin-copysign-6.c: Likewise.
* gcc.target/i386/builtin-copysign-7.c: Likewise.
* gcc.target/i386/builtin-copysign-8a.c: Likewise.
* gcc.target/i386/builtin-copysign-8b.c: Likewise.
* gcc.target/i386/builtin-fabs-1.c: Likewise.
* gcc.target/i386/builtin-fabs-2.c: Likewise.

Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
2 weeks agoDaily bump.
GCC Administrator [Mon, 20 Oct 2025 00:18:13 +0000 (00:18 +0000)] 
Daily bump.

2 weeks agoPR modula2/122333: m2spellcheck.cc remove memset and tidyup
Gaius Mulley [Sun, 19 Oct 2025 17:48:18 +0000 (18:48 +0100)] 
PR modula2/122333: m2spellcheck.cc remove memset and tidyup

This patch removes memset from m2spellcheck_InitCandidates.
It corrects a comment boiler plate and removes an unused local
variable.  Finally it frees up memory used by the candidates_array
in KillCandidates.

gcc/m2/ChangeLog:

PR modula2/122333
* gm2-compiler/M2MetaError.mod (JoinSentances): Remove
unused variable.
* gm2-gcc/m2spellcheck.cc (m2spellcheck_InitCandidates): Rewrite.
(KillCandidates): Deallocate auto_vec candidates_array.
(candidates_array_vec_t): New declaration.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
2 weeks agoAVR: The nzb=1 patterns with IOR, XOR, AND work the same way with PLUS.
Georg-Johann Lay [Sun, 19 Oct 2025 14:42:07 +0000 (16:42 +0200)] 
AVR: The nzb=1 patterns with IOR, XOR, AND work the same way with PLUS.

gcc/
* config/avr/avr.cc (avr_nonzero_bits_lsr_operands_p): Also
handle PLUS.
* config/avr/avr.md (pixaop): New code iterator for PLUS,
IOR, XOR, AND.
(nzb=1 insns): Use pixaop instead of bitop code iterator.
Handle PLUS in outputs.

2 weeks agoad PR122212: Fix test case for 16-bit int targets.
Georg-Johann Lay [Sun, 19 Oct 2025 14:39:32 +0000 (16:39 +0200)] 
ad PR122212: Fix test case for 16-bit int targets.

PR testsuite/122212
PR testsuite/52641
gcc/testsuite/
* gcc.dg/torture/pr122212.c: Pass 0xffffffff instead of -1u
for all bits set in uint32_t.

2 weeks agoad PR122016: Fix test case for 16-bit size targets.
Georg-Johann Lay [Sun, 19 Oct 2025 14:32:43 +0000 (16:32 +0200)] 
ad PR122016: Fix test case for 16-bit size targets.

PR testsuite/122016
PR testsuite/52641
gcc/testsuite/
* gcc.dg/torture/pr122016.c (strncmp): Use __SIZE_TYPE__ instead
of long as type of the size argument.

2 weeks agoRISC-V: Add testcase for unsigned scalar SAT_MUL form 6
Pan Li [Fri, 17 Oct 2025 07:12:10 +0000 (15:12 +0800)] 
RISC-V: Add testcase for unsigned scalar SAT_MUL form 6

The form 6 of unsigned scalar SAT_MUL has supported from the
previous change.  Thus, add the test cases to make sure it
works well.

The below test suites are passed for this patch series.
 * The rv64gcv fully regression test.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/sat/sat_arith.h: Add test helper macros.
* gcc.target/riscv/sat/sat_u_mul-7-u16-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u16-from-u32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u16-from-u64.rv32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u16-from-u64.rv64.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u32-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u32-from-u64.rv32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u32-from-u64.rv64.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u64-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u8-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u8-from-u16.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u8-from-u32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u8-from-u64.rv32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-7-u8-from-u64.rv64.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u16-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u16-from-u32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u16-from-u64.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u32-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u32-from-u64.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u64-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u8-from-u128.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u8-from-u16.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u8-from-u32.c: New test.
* gcc.target/riscv/sat/sat_u_mul-run-7-u8-from-u64.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
2 weeks agocobol: Implement ENTRY statement; finish removing ascii/ebcdic dichotomy.
Robert Dubner [Sat, 18 Oct 2025 14:35:52 +0000 (10:35 -0400)] 
cobol: Implement ENTRY statement; finish removing ascii/ebcdic dichotomy.

The prior set of changes largely eliminated the assumption that the
internal codeset was either ascii or ebcdic.  These changes remove the
last vestiges of that assumption.

These changes also implement the COBOL ENTRY statement, which allows a
program-id to have more than one externally callable entry point. Since
GCC assumes the existence of an ABI that is not, repeat *not* capable of
that, it is implemented here by creating a separate function with the
name specified by the ENTRY statement.  That function sets up global
variables which cause control to be transferred to the ENTRY point when
the parent function is called re-entrantly, and then executes that call.

gcc/cobol/ChangeLog:

* genapi.cc (move_tree): Formatting.
(parser_enter_file): Incorporate global __gg__entry_label.
(enter_program_common): Remove calls to alphabet overrides.
(parser_alphabet): Change cbl_alphabet_e handling.
(parser_alphabet_use): Likewise.
(initialize_the_data): Likewise.
(establish_using): Process passed parameters in a subroutine.
(parser_division): Remove in-line parameter processing;
call establish_using() instead. Check for __gg__entry_label.
(parser_file_add): Temporary workaround for charset encoding.
(parser_file_open): Likewise.
(create_and_call): Push/pop program state around call to external.
(parser_entry): Implement new ENTRY statement feature.
(mh_source_is_literalN): Formatting.
* genapi.h (parser_entry): New ENTRY statement.
* gengen.cc (gg_create_goto_pair): Formatting.
(gg_goto_label_decl): Remove.
* gengen.h (gg_goto_label_decl): Remove.
* genutil.cc (internal_codeset_is_ebcdic): Remove.
* genutil.h (internal_codeset_is_ebcdic): Remove.
* symbols.cc (symbols_alphabet_set): Restrict alphabet scan to
program.
* symbols.h (is_elementary): Use defined constants instead of
explicit 'A'and 'N'

libgcobol/ChangeLog:

* charmaps.cc (__gg__set_internal_codeset): Eliminate ascii/ebcdic.
(__gg__text_conversion_override): Remove.
* charmaps.h (enum text_device_t):  Eliminate ascii/ebcdic.
(enum text_codeset_t): Remove.
(__gg__set_internal_codeset): Remove.
(__gg__text_conversion_override): Remove.
* gfileio.cc: Anticipate cbl_encoding_t fixes.
* libgcobol.cc (struct program_state): Incorporate
__gg__entry_label.
(__gg__pop_program_state): Eliminate unused defines.
(__gg__alphabet_use): Eliminate ascii/ebcdic dichotomy.
* valconv.cc (__gg__alphabet_create): Likewise.

2 weeks agoDaily bump.
GCC Administrator [Sun, 19 Oct 2025 00:17:28 +0000 (00:17 +0000)] 
Daily bump.

2 weeks agoRegenerate common.opt.urls
Mark Wielaard [Sat, 18 Oct 2025 23:48:07 +0000 (01:48 +0200)] 
Regenerate common.opt.urls

An alias for -ftree-parallelize-loops was added to common.opt, but
common.opt.urls wasn't regenerated.

Fixes: f708b83d197b ("tree-parloops: Enable runtime thread detection with -ftree-parallelize-loops")
gcc/ChangeLog:

* common.opt.urls: Regenerate.

2 weeks agolibstdc++: Implement P1494 and P3641 Partial program correctness [PR119060]
Iain Sandoe [Thu, 4 Sep 2025 15:21:16 +0000 (16:21 +0100)] 
libstdc++: Implement P1494 and P3641 Partial program correctness [PR119060]

This implements the library parts of P1494 as amended by P3641.  For GCC the
compiler itself treats stdio operations as equivalent to the observable
checkpoint and thus it does not appear to be necessary to add calls to those
functions (it will not alter the outcome).

This adds the facility for C++26, although there is no reason, in principle,
that it would not work back to C++11 at least.

PR c++/119060

libstdc++-v3/ChangeLog:

* include/bits/version.def: Add observable_checkpoint at present
allowed from C++26.
* include/bits/version.h: Regenerate.
* include/std/utility: Add std::observable_checkpoint().
* src/c++23/std.cc.in: Add obervable_checkpoint () to utility.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
2 weeks agoc++: Implement P1494 and P3641 Partial program correctness [PR119060].
Iain Sandoe [Sat, 6 Sep 2025 16:11:21 +0000 (17:11 +0100)] 
c++: Implement P1494 and P3641 Partial program correctness [PR119060].

P1494 provides a mechanism that serves to demarc epochs within the code
preventing UB-based optimisations from 'time traveling' across such
boundaries.  The additional paper, P3641, alters the name of the function
to 'observable_checkpoint' which is the name used here.

This implementation  maintains the observable function call through to
expand, where it produces no code.

PR c++/119060

gcc/ChangeLog:

* builtins.cc (expand_builtin): Handle BUILT_IN_OBSERVABLE_CHKPT.
* builtins.def (BUILT_IN_OBSERVABLE_CHKPT): New.
* tree.cc (build_common_builtin_nodes): Build observable
checkpoint builtin.

gcc/cp/ChangeLog:

* cxxapi-data.csv: Add observable_checkpoint to <utility>.
* std-name-hint.gperf: Add observable_checkpoint to <utility>.
* std-name-hint.h: Regenerate.

gcc/testsuite/ChangeLog:

* g++.dg/cpp26/observable-checkpoint.C: New test.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
2 weeks agoc++/modules: Import purview using-directives in the same module [PR122279]
Nathaniel Shead [Wed, 15 Oct 2025 07:13:38 +0000 (18:13 +1100)] 
c++/modules: Import purview using-directives in the same module [PR122279]

[namespace.qual] p1 says that a namespace nominated by a using-directive
is searched if the using-directive precedes that point.

[basic.lookup.general] p2 says that a declaration in a different TU
within a module purview is visible if either the declaration is
exported, or the other TU is part of the same module as the point of
lookup.  This patch implements the second half of that.

PR c++/122279

gcc/cp/ChangeLog:

* module.cc (depset::hash::add_namespace_entities): Seed any
purview using-decls.
(module_state::write_using_directives): Stream if the udir was
exported or not.
(module_state::read_using_directives): Add the using-directive
if it's either exported or part of this module.

gcc/testsuite/ChangeLog:

* g++.dg/modules/namespace-13_b.C: Adjust expected results.
* g++.dg/modules/namespace-13_c.C: Test non-exported
using-directive is not used.
* g++.dg/modules/namespace-14_a.C: New test.
* g++.dg/modules/namespace-14_b.C: New test.
* g++.dg/modules/namespace-14_c.C: New test.
* g++.dg/modules/namespace-14_d.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Patrick Palka <ppalka@redhat.com>
Reviewed-by: Jason Merrill <jason@redhat.com>
2 weeks agoAArch64: Implement widen_[us]sum using 2-way [US]UDOT for SVE2p1 [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:22:50 +0000 (08:22 +0100)] 
AArch64: Implement widen_[us]sum using 2-way [US]UDOT for SVE2p1 [PR122069]

SVE2p1 adds 2-way dotproduct which we can use when we have to do a single step
widening addition.  This is useful for instance when the value to be widened
does not come from a load.  For example for

int foo2_int(unsigned short *x, unsigned short * restrict y) {
  int sum = 0;
  for (int i = 0; i < 8000; i++)
    {
      x[i] = x[i] + y[i];
      sum += x[i];
    }
  return sum;
}

we used to generate

.L12:
        ld1h    z30.h, p7/z, [x0, x2, lsl 1]
        ld1h    z29.h, p7/z, [x1, x2, lsl 1]
        add     z30.h, z30.h, z29.h
        uaddwb  z31.s, z31.s, z30.h
        uaddwt  z31.s, z31.s, z30.h
        st1h    z30.h, p7, [x0, x2, lsl 1]
        mov     x3, x2
        inch    x2
        cmp     w2, w4
        bls     .L12
        inch    x3
        uaddv   d31, p7, z31.s

but with +sve2p1

.L12:
        ld1h    z31.h, p7/z, [x0, x2, lsl 1]
        ld1h    z29.h, p7/z, [x1, x2, lsl 1]
        add     z31.h, z31.h, z29.h
        udot    z30.s, z31.h, z28.h
        st1h    z31.h, p7, [x0, x2, lsl 1]
        mov     x3, x2
        inch    x2
        cmp     w2, w4
        bls     .L12
        inch    x3
        uaddv   d30, p7, z30.s

gcc/ChangeLog:

PR middle-end/122069
* config/aarch64/aarch64-sve2.md
(widen_ssum<mode><Vnarrow>3): Update.
(widen_usum<mode><Vnarrow>3): Update.

gcc/testsuite/ChangeLog:

PR middle-end/122069
* gcc.target/aarch64/sve2/pr122069_3.c: New test.
* gcc.target/aarch64/sve2/pr122069_4.c: New test.

2 weeks agoAArch64: Implement widen_[us]sum using [US]ADDW[TB] for SVE2 [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:22:18 +0000 (08:22 +0100)] 
AArch64: Implement widen_[us]sum using [US]ADDW[TB] for SVE2 [PR122069]

SVE2 adds [US]ADDW[TB] which we can use when we have to do a single step
widening addition.  This is useful for instance when the value to be widened
does not come from a load.  For example for

int foo2_int(unsigned short *x, unsigned short * restrict y) {
  int sum = 0;
  for (int i = 0; i < 8000; i++)
    {
      x[i] = x[i] + y[i];
      sum += x[i];
    }
  return sum;
}

we used to generate

.L6:
        ld1h    z1.h, p7/z, [x0, x2, lsl 1]
        ld1h    z29.h, p7/z, [x1, x2, lsl 1]
        add     z29.h, z29.h, z1.h
        punpklo p6.h, p7.b
        uunpklo z0.s, z29.h
        add     z31.s, p6/m, z31.s, z0.s
        punpkhi p6.h, p7.b
        uunpkhi z30.s, z29.h
        add     z31.s, p6/m, z31.s, z30.s
        st1h    z29.h, p7, [x0, x2, lsl 1]
        add     x2, x2, x4
        whilelo p7.h, w2, w3
        b.any   .L6
        ptrue   p7.b, all
        uaddv   d31, p7, z31.s

but with +sve2

.L12:
        ld1h    z30.h, p7/z, [x0, x2, lsl 1]
        ld1h    z29.h, p7/z, [x1, x2, lsl 1]
        add     z30.h, z30.h, z29.h
        uaddwb  z31.s, z31.s, z30.h
        uaddwt  z31.s, z31.s, z30.h
        st1h    z30.h, p7, [x0, x2, lsl 1]
        mov     x3, x2
        inch    x2
        cmp     w2, w4
        bls     .L12
        inch    x3
        uaddv   d31, p7, z31.s

gcc/ChangeLog:

PR middle-end/122069
* config/aarch64/aarch64-sve2.md: (widen_ssum<mode><Vnarrow>3): New.
(widen_usum<mode><Vnarrow>3): New.
* config/aarch64/iterators.md (Vnarrow): New, to match VNARROW.

gcc/testsuite/ChangeLog:

PR middle-end/122069
* gcc.target/aarch64/sve2/pr122069_1.c: New test.
* gcc.target/aarch64/sve2/pr122069_2.c: New test.

2 weeks agoAArch64: Implement widen_[us]sum using dotproduct for SVE [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:21:56 +0000 (08:21 +0100)] 
AArch64: Implement widen_[us]sum using dotproduct for SVE [PR122069]

This patch implements support for using dotproduct to do sum reductions by
changing += a into += (a * 1).  i.e. we seed the multiplication with 1.

Given the example

int foo_int(unsigned char *x, unsigned char * restrict y) {
  int sum = 0;
  for (int i = 0; i < 8000; i++)
     sum += char_abs(x[i] - y[i]);
  return sum;
}

we used to generate

.L2:
        ld1b    z1.b, p7/z, [x0, x2]
        ld1b    z29.b, p7/z, [x1, x2]
        sub     z29.b, z1.b, z29.b
        uunpklo z0.h, z29.b
        uunpkhi z29.h, z29.b
        uunpklo z30.s, z0.h
        add     z31.s, p6/m, z31.s, z30.s
        uunpkhi z0.s, z0.h
        add     z31.s, p5/m, z31.s, z0.s
        uunpklo z28.s, z29.h
        add     z31.s, p4/m, z31.s, z28.s
        uunpkhi z29.s, z29.h
        add     z31.s, p3/m, z31.s, z29.s
        add     x2, x2, x7
        whilelo p7.b, w2, w3
        whilelo p3.s, w2, w6
        whilelo p4.s, w2, w5
        whilelo p5.s, w2, w4
        whilelo p6.s, w2, w3
        b.any   .L2
        ptrue   p7.b, all
        uaddv   d31, p7, z31.s

but now generates with +dotprod

.L3:
        ld1b    z30.b, p7/z, [x5, x2]
        ld1b    z29.b, p7/z, [x1, x2]
        sub     z30.b, z30.b, z29.b
        udot    z31.s, z30.b, z28.b
        mov     x3, x2
        add     x2, x2, x6
        cmp     w2, w0
        bls     .L3
        incb    x3
        uaddv   d31, p7, z31.s

gcc/ChangeLog:

PR middle-end/122069
* config/aarch64/aarch64-sve.md (widen_<sur>sum<mode><vsi2qi>3): New.

gcc/testsuite/ChangeLog:

PR middle-end/122069
* gcc.target/aarch64/sve/pr122069_1.c: New test.
* gcc.target/aarch64/sve/pr122069_2.c: New test.

2 weeks agors6000: convert widen_[us]sum into convert optab [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:21:30 +0000 (08:21 +0100)] 
rs6000: convert widen_[us]sum into convert optab [PR122069]

This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab.  The result of which requires the output mode to be added to
the existing patterns.

No change in functionality is expected.

gcc/ChangeLog:

PR middle-end/122069
* config/rs6000/altivec.md (widen_usum<mode>3): Rename ...
(widen_usumv4si<mode>3): ... to this.
(widen_ssumv16qi3): Rename ...
(widen_ssumv4siv16qi3): ... to this.
(widen_ssumv8hi3): Rename ...
(widen_ssumv4siv8hi3): ... to this.

2 weeks agoia64: convert widen_[us]sum into convert optab [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:21:09 +0000 (08:21 +0100)] 
ia64: convert widen_[us]sum into convert optab [PR122069]

The target does not seem to have a maintainer listed, I've CC'ed a group of
global maintainers instead hoping one of you could approve it.

This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab.  The result of which requires the output mode to be added to
the existing patterns.

No change in functionality is expected.

gcc/ChangeLog:

PR middle-end/122069
* config/ia64/vect.md (widen_usumv8qi3): Renamed ...
(widen_usumv4hiv8qi3): ... into this.
(widen_usumv4hi3): Renamed ...
(widen_usumv2siv4hi3): ... into this.
(widen_ssumv8qi3): Renamed ...
(widen_ssumv4hiv8qi3): ... into this.
(widen_ssumv4hi3): Renamed ...
(widen_ssumv2siv4hi3): ... into this.

2 weeks agoarm: convert widen_[us]sum into convert optab [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:20:47 +0000 (08:20 +0100)] 
arm: convert widen_[us]sum into convert optab [PR122069]

This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab.  The result of which requires the output mode to be added to
the existing patterns.

No change in functionality is expected.

gcc/ChangeLog:

PR middle-end/122069
* config/arm/iterators.md (v_double_width): New, matching
V_double_width.
* config/arm/neon.md (widen_ssum<mode>3): Renamed ...
(widen_ssum<v_double_width><mode>3, widen_ssum<V_widen_l><mode>3): ...
into these.
(widen_usum<mode>3): Renamed ...
(widen_usum<v_double_width><mode>3, widen_usum<V_widen_l><mode>3): ...
into these.

2 weeks agoAArch64: add double widen_sum optab using dotprod for Adv.SIMD [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:20:07 +0000 (08:20 +0100)] 
AArch64: add double widen_sum optab using dotprod for Adv.SIMD [PR122069]

This patch implements support for using dotproduct to do sum reductions by
changing += a into += (a * 1).  i.e. we seed the multiplication with 1.

Given the example

int foo_int(unsigned char *x, unsigned char * restrict y) {
  int sum = 0;
  for (int i = 0; i < 8000; i++)
     sum += char_abs(x[i] - y[i]);
  return sum;
}

we used to generate

.L2:
        ldr     q0, [x0, x2]
        ldr     q28, [x1, x2]
        sub     v28.16b, v0.16b, v28.16b
        zip1    v29.16b, v28.16b, v31.16b
        zip2    v28.16b, v28.16b, v31.16b
        uaddw   v30.4s, v30.4s, v29.4h
        uaddw2  v30.4s, v30.4s, v29.8h
        uaddw   v30.4s, v30.4s, v28.4h
        uaddw2  v30.4s, v30.4s, v28.8h
        add     x2, x2, 16
        cmp     x2, x3
        bne     .L2
        addv    s31, v30.4s

but now generates with +dotprod

.L2:
        ldr     q29, [x0, x2]
        ldr     q28, [x1, x2]
        sub     v28.16b, v29.16b, v28.16b
        udot    v31.4s, v28.16b, v30.16b
        add     x2, x2, 16
        cmp     x2, x3
        bne     .L2
        addv    s31, v31.4s

gcc/ChangeLog:

PR middle-end/122069
* config/aarch64/aarch64-simd.md (widen_ssum<mode><vsi2qi>3): New.
(widen_usum<mode><vsi2qi>3): New.

gcc/testsuite/ChangeLog:

PR middle-end/122069
* gcc.target/aarch64/pr122069_3.c: New test.
* gcc.target/aarch64/pr122069_4.c: New test.

2 weeks agoAArch64: convert widen_sum optabs to convert [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:19:28 +0000 (08:19 +0100)] 
AArch64: convert widen_sum optabs to convert [PR122069]

This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab.  The result of which requires the output mode to be added to
the existing patterns.

No change in functionality is expected.

gcc/ChangeLog:

PR middle-end/122069
* config/aarch64/aarch64-simd.md (widen_ssum<mode>3): Change into..
(widen_ssum<Vdblw><mode>3, widen_ssum<Vwide><mode>3): ... these.
(widen_usum<mode>3): Change into ...
(widen_usum<Vdblw><mode>3, widen_usum<Vwide><mode>3): ... these.
* config/aarch64/iterators.md (Vdblw): New.
(Vwide): Extend to match VWIDE.

gcc/testsuite/ChangeLog:

PR middle-end/122069
* gcc.target/aarch64/pr122069_1.c: New test.
* gcc.target/aarch64/pr122069_2.c: New test.

2 weeks agomiddle-end: refactor WIDEN_SUM_EXPR into convert optab [PR122069]
Tamar Christina [Sat, 18 Oct 2025 07:18:14 +0000 (08:18 +0100)] 
middle-end: refactor WIDEN_SUM_EXPR into convert optab [PR122069]

This patch changes the widen_[us]sum optabs into a convert optabs such that
targets and specify more than one conversion.

Following this patch are patches rewriting all targets using this change.

While working on this I noticed that the pattern does miss some cases it
could handle if it tried multiple attempts. e.g. if the promotion is from
qi to si, and the target doesn't have this, it should try hi -> si.

But I'm leaving that for now.

gcc/ChangeLog:

PR middle-end/122069
* doc/md.texi (widen_ssum@var{n}@var{m}3, widen_usum@var{n}@var{m}3):
Update docs.
* optabs.cc (expand_widen_pattern_expr): Add WIDEN_SUM_EXPR as widening.
* optabs.def (ssum_widen_optab, usum_widen_optab): Convert from direct
to a conversion optab.
* tree-vect-patterns.cc (vect_recog_widen_sum_pattern): Change
vect_supportable_direct_optab_p into vect_supportable_conv_optab_p.

2 weeks agofortran: allow character in conditional expression
Yuao Ma [Thu, 16 Oct 2025 14:32:52 +0000 (22:32 +0800)] 
fortran: allow character in conditional expression

This patch allows the use of character types in conditional expressions.

gcc/fortran/ChangeLog:

* resolve.cc (resolve_conditional): Allow character in cond-expr.
* trans-const.cc (gfc_conv_constant): Handle want_pointer.
* trans-expr.cc (gfc_conv_conditional_expr): Fill se->string_length.
(gfc_conv_string_parameter): Handle COND_EXPR tree code.

gcc/testsuite/ChangeLog:

* gfortran.dg/conditional_1.f90: Test character type.
* gfortran.dg/conditional_2.f90: Test print constants.
* gfortran.dg/conditional_4.f90: Test diagnostic message.
* gfortran.dg/conditional_6.f90: Test character cond-arg.

2 weeks agotree-object-size.cc: Fix assert constant offset in check_for_plus_in_loops [PR122012]
Linsen Zhou [Fri, 17 Oct 2025 03:05:04 +0000 (11:05 +0800)] 
tree-object-size.cc: Fix assert constant offset in check_for_plus_in_loops [PR122012]

After commit 51b85dfeb19652bf3e0aaec08828ba7cee1e641c, when the
pointer offset is a variable in the loop, the object size of the
pointer may also need to be reexamined.
Which make gcc_assert in the check_for_plus_in_loops failed.

gcc/ChangeLog:

PR tree-optimization/122012
* tree-object-size.cc (check_for_plus_in_loops): Skip check
for the variable offset

gcc/testsuite/ChangeLog:

PR tree-optimization/122012
* gcc.dg/torture/pr122012.c: New test.

Signed-off-by: Linsen Zhou <i@lin.moe>
2 weeks agoDaily bump.
GCC Administrator [Sat, 18 Oct 2025 00:18:06 +0000 (00:18 +0000)] 
Daily bump.

3 weeks agobpf: fix memset miscompilation with larger stores [PR122139]
David Faust [Wed, 15 Oct 2025 20:36:38 +0000 (13:36 -0700)] 
bpf: fix memset miscompilation with larger stores [PR122139]

The BPF backend expansion of setmem was broken, because it could elect
to use stores of HI, SI or DI modes based on the destination alignment
when the value was QI, but fail to duplicate the byte value across to
those larger sizes.  This resulted in not all bytes of the destination
actually being set to the desired value.

Fix bpf_expand_setmem to ensure the desired byte value is really
duplicated as necessary, whether it is constant or a (sub)reg:QI.

PR target/122139

gcc/

* config/bpf/bpf.cc (bpf_expand_setmem): Duplicate byte value
across to new mode when using larger modes for store.

gcc/testsuite/

* gcc.target/bpf/memset-3.c: New.
* gcc.target/bpf/memset-4.c: New.