amdgcn: Add instruction patterns for conditional min/max operations
gcc/ChangeLog:
* config/gcn/gcn-valu.md (<expander><mode>3_exec): Add patterns for
{s|u}{max|min} in QI, HI and DI modes.
(<expander><mode>3): Add pattern for {s|u}{max|min} in DI mode.
(cond_<fexpander><mode>): Add pattern for cond_f{max|min}.
(cond_<expander><mode>): Add pattern for cond_{s|u}{max|min}.
* config/gcn/gcn.cc (gcn_spill_class): Allow the exec register to be
saved in SGPRs.
gcc/testsuite/ChangeLog:
* gcc.target/gcn/cond_fmaxnm_1.c: New test.
* gcc.target/gcn/cond_fmaxnm_1_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_2.c: New test.
* gcc.target/gcn/cond_fmaxnm_2_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_3.c: New test.
* gcc.target/gcn/cond_fmaxnm_3_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_4.c: New test.
* gcc.target/gcn/cond_fmaxnm_4_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_5.c: New test.
* gcc.target/gcn/cond_fmaxnm_5_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_6.c: New test.
* gcc.target/gcn/cond_fmaxnm_6_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_7.c: New test.
* gcc.target/gcn/cond_fmaxnm_7_run.c: New test.
* gcc.target/gcn/cond_fmaxnm_8.c: New test.
* gcc.target/gcn/cond_fmaxnm_8_run.c: New test.
* gcc.target/gcn/cond_fminnm_1.c: New test.
* gcc.target/gcn/cond_fminnm_1_run.c: New test.
* gcc.target/gcn/cond_fminnm_2.c: New test.
* gcc.target/gcn/cond_fminnm_2_run.c: New test.
* gcc.target/gcn/cond_fminnm_3.c: New test.
* gcc.target/gcn/cond_fminnm_3_run.c: New test.
* gcc.target/gcn/cond_fminnm_4.c: New test.
* gcc.target/gcn/cond_fminnm_4_run.c: New test.
* gcc.target/gcn/cond_fminnm_5.c: New test.
* gcc.target/gcn/cond_fminnm_5_run.c: New test.
* gcc.target/gcn/cond_fminnm_6.c: New test.
* gcc.target/gcn/cond_fminnm_6_run.c: New test.
* gcc.target/gcn/cond_fminnm_7.c: New test.
* gcc.target/gcn/cond_fminnm_7_run.c: New test.
* gcc.target/gcn/cond_fminnm_8.c: New test.
* gcc.target/gcn/cond_fminnm_8_run.c: New test.
* gcc.target/gcn/cond_smax_1.c: New test.
* gcc.target/gcn/cond_smax_1_run.c: New test.
* gcc.target/gcn/cond_smin_1.c: New test.
* gcc.target/gcn/cond_smin_1_run.c: New test.
* gcc.target/gcn/cond_umax_1.c: New test.
* gcc.target/gcn/cond_umax_1_run.c: New test.
* gcc.target/gcn/cond_umin_1.c: New test.
* gcc.target/gcn/cond_umin_1_run.c: New test.
* gcc.target/gcn/smax_1.c: New test.
* gcc.target/gcn/smax_1_run.c: New test.
* gcc.target/gcn/smin_1.c: New test.
* gcc.target/gcn/smin_1_run.c: New test.
* gcc.target/gcn/umax_1.c: New test.
* gcc.target/gcn/umax_1_run.c: New test.
* gcc.target/gcn/umin_1.c: New test.
* gcc.target/gcn/umin_1_run.c: New test.
Tobias Burnus [Wed, 1 Mar 2023 14:18:40 +0000 (15:18 +0100)]
OpenMP/Fortran: Fix handling of optional is_device_ptr + bind(C) [PR108546]
For is_device_ptr, optional checks should only be done before calling
libgomp, afterwards they are NULL either because of absent or, by
chance, because it is unallocated or unassociated (for pointers/allocatables).
Additionally, it fixes an issue with explicit mapping for 'type(c_ptr)'.
PR middle-end/108546
gcc/fortran/ChangeLog:
* trans-openmp.cc (gfc_trans_omp_clauses): Fix mapping of
type(C_ptr) variables.
gcc/ChangeLog:
* omp-low.cc (lower_omp_target): Remove optional handling
on the receiver side, i.e. inside target (data), for
use_device_ptr.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/is_device_ptr-3.f90: New test.
* testsuite/libgomp.fortran/use_device_ptr-optional-4.f90: New test.
Tobias Burnus [Mon, 27 Feb 2023 11:47:54 +0000 (12:47 +0100)]
Update dg-dump-scan for "Fortran/OpenMP: Fix mapping of array descriptors and deferred-length strings"
Follow-up to commit 55a18d4744258e3909568e425f9f473c49f9d13f
"Fortran/OpenMP: Fix mapping of array descriptors and deferred-length strings"
updating the dumps.
* For the goacc testcase, 'to' changed to 'release' and due to 'finally' then
to 'delete', which can be regarded as bugfix.
* For pr78260-2.f90, the calculation moved inside the 'if(...->data == NULL)'
block to handle deferred-string length vars better, esp. when 'optional'.
Martin Liska [Fri, 17 Feb 2023 14:11:02 +0000 (15:11 +0100)]
asan: adjust module name for global variables
As mentioned in the PR, when we use LTO, we wrongly use ltrans output
file name as a module name of a global variable. That leads to a
non-reproducible output.
After the suggested change, we emit context name of normal global
variables. And for artificial variables (like .Lubsan_data3), we use
aux_base_name (e.g. "./a.ltrans0.ltrans").
PR sanitizer/108834
gcc/ChangeLog:
* asan.cc (asan_add_global): Use proper TU name for normal
global variables (and aux_base_name for the artificial one).
gcc/testsuite/ChangeLog:
* c-c++-common/asan/global-overflow-1.c: Test line and column
info for a global variable.
Kewen Lin [Tue, 14 Feb 2023 02:03:26 +0000 (20:03 -0600)]
rs6000/test: Adjust some test cases on partial vector [PR96373]
As Richard pointed out in [1] and the testing on Power10, the
proposed fix for PR96373 requires some updates on a few rs6000
test cases which adopt partial vector. This patch is to fix
all of them with one extra option "-fno-trapping-math" as
Richard suggested.
Besides, the original test case also failed on Power10 without
Richard's proposed fix, this patch adds it together for a bit
better testing coverage.
Richard Biener [Thu, 23 Feb 2023 10:03:03 +0000 (11:03 +0100)]
tree-optimization/108888 - call if-conversion
The following makes sure to only predicate calls necessary.
PR tree-optimization/108888
* tree-if-conv.cc (if_convertible_stmt_p): Set PLF_2 on
calls to predicate.
(predicate_statements): Only predicate calls with PLF_2.
Andrew Stubbs [Thu, 28 Jul 2022 15:07:22 +0000 (16:07 +0100)]
vect: inbranch SIMD clones
There has been support for generating "inbranch" SIMD clones for a long time,
but nothing actually uses them (as far as I can see).
This patch add supports for a sub-set of possible cases (those using
mask_mode == VOIDmode). The other cases fail to vectorize, just as before,
so there should be no regressions.
The sub-set of support should cover all cases needed by amdgcn, at present.
gcc/ChangeLog:
* internal-fn.cc (expand_MASK_CALL): New.
* internal-fn.def (MASK_CALL): New.
* internal-fn.h (expand_MASK_CALL): New prototype.
* omp-simd-clone.cc (simd_clone_adjust_argument_types): Set vector_type
for mask arguments also.
* tree-if-conv.cc: Include cgraph.h.
(if_convertible_stmt_p): Do if conversions for calls to SIMD calls.
(predicate_statements): Convert functions to IFN_MASK_CALL.
* tree-vect-loop.cc (vect_get_datarefs_in_loop): Recognise
IFN_MASK_CALL as a SIMD function call.
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Handle
IFN_MASK_CALL as an inbranch SIMD function call.
Generate the mask vector arguments.
gcc/testsuite/ChangeLog:
* gcc.dg/vect/vect-simd-clone-16.c: New test.
* gcc.dg/vect/vect-simd-clone-16b.c: New test.
* gcc.dg/vect/vect-simd-clone-16c.c: New test.
* gcc.dg/vect/vect-simd-clone-16d.c: New test.
* gcc.dg/vect/vect-simd-clone-16e.c: New test.
* gcc.dg/vect/vect-simd-clone-16f.c: New test.
* gcc.dg/vect/vect-simd-clone-17.c: New test.
* gcc.dg/vect/vect-simd-clone-17b.c: New test.
* gcc.dg/vect/vect-simd-clone-17c.c: New test.
* gcc.dg/vect/vect-simd-clone-17d.c: New test.
* gcc.dg/vect/vect-simd-clone-17e.c: New test.
* gcc.dg/vect/vect-simd-clone-17f.c: New test.
* gcc.dg/vect/vect-simd-clone-18.c: New test.
* gcc.dg/vect/vect-simd-clone-18b.c: New test.
* gcc.dg/vect/vect-simd-clone-18c.c: New test.
* gcc.dg/vect/vect-simd-clone-18d.c: New test.
* gcc.dg/vect/vect-simd-clone-18e.c: New test.
* gcc.dg/vect/vect-simd-clone-18f.c: New test.
Tobias Burnus [Wed, 22 Feb 2023 20:18:33 +0000 (21:18 +0100)]
Fortran/OpenMP: Fix mapping of array descriptors and deferred-length strings
Previously, array descriptors might have been mapped as 'alloc'
instead of 'to' for 'alloc', not updating the array bounds. The
'alloc' could also appear for 'data exit', failing with a libgomp
assert. In some cases, either array descriptors or deferred-length
string's length variable was not mapped. And, finally, some offset
calculations with array-sections mappings went wrong.
The testcases contain some comment-out tests which require follow-up
work and for which PR exist. Those mostly relate to deferred-length
strings which have several issues beyong OpenMP support.
This is the OG12 variant of the submitted but unreviewed GCC 13/mainline
patch at https://gcc.gnu.org/pipermail/gcc-patches/2023-February/612387.html
gcc/fortran/ChangeLog:
* trans-decl.cc (gfc_get_symbol_decl): Add attributes
such as 'declare target' also to hidden artificial
variable for deferred-length character variables.
* trans-openmp.cc (gfc_trans_omp_array_section,
gfc_trans_omp_clauses, gfc_trans_omp_target_exit_data):
Improve mapping of array descriptors and deferred-length
string variables.
gcc/ChangeLog:
* gimplify.cc (gimplify_scan_omp_clauses): Remove Fortran
special case.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/target-enter-data-3.f90: Uncomment
'target exit data'.
* testsuite/libgomp.fortran/target-enter-data-4.f90: New test.
* testsuite/libgomp.fortran/target-enter-data-5.f90: New test.
* testsuite/libgomp.fortran/target-enter-data-6.f90: New test.
* testsuite/libgomp.fortran/target-enter-data-7.f90: New test.
Tobias Burnus [Wed, 22 Feb 2023 11:35:29 +0000 (12:35 +0100)]
Fix: Fortran/OpenMP: align/allocator modifiers to the allocate clause
When merging r13-4584-gb2e1c49b4a4 to OG12 as commit 58e0579ed87,
the 'align' handling seemingly ended up in the wrong clause.
(Result: libgomp.fortran/allocate-2a.f90 FAILED; now fixed.)
gcc/fortran/
* trans-openmp.cc (gfc_trans_omp_clauses): Move align modifier
handling from OMP_LIST_ALLOCATOR to OMP_LIST_ALLOCATE.
Jonathan Wakely [Wed, 31 Aug 2022 12:57:34 +0000 (13:57 +0100)]
libstdc++: Add noexcept-specifier to std::reference_wrapper::operator()
This isn't required by the standard, but there's an LWG issue suggesting
to add it.
Also use __invoke_result instead of result_of, to match the spec in
recent standards.
libstdc++-v3/ChangeLog:
* include/bits/refwrap.h (reference_wrapper::operator()): Add
noexcept-specifier and use __invoke_result instead of result_of.
* testsuite/20_util/reference_wrapper/invoke-noexcept.cc: New test.
Jonathan Wakely [Thu, 2 Feb 2023 14:06:40 +0000 (14:06 +0000)]
libstdc++: Fix std::filesystem errors with -fkeep-inline-functions [PR108636]
With -fkeep-inline-functions there are linker errors when including
<filesystem>. This happens because there are some filesystem::path
constructors defined inline which call non-exported functions defined in
the library. That's usually not a problem, because those constructors
are only called by code that's also inside the library. But when the
header is compiled with -fkeep-inline-functions those inline functions
are emitted even though they aren't called. That then creates an
undefined reference to the other library internsl. The fix is to just
move the private constructors into the library where they are called.
That way they are never even seen by users, and so not compiled even if
-fkeep-inline-functions is used.
Marek Polacek [Thu, 16 Feb 2023 22:41:24 +0000 (17:41 -0500)]
c++: ICE with redundant capture [PR108829]
Here we crash in is_capture_proxy:
/* Location wrappers should be stripped or otherwise handled by the
caller before using this predicate. */
gcc_checking_assert (!location_wrapper_p (decl));
We only crash with the redundant capture:
int abyPage = [=, abyPage] { ... }
because prune_lambda_captures is only called when there was a default
capture, and with [=] only abyPage won't be in LAMBDA_EXPR_CAPTURE_LIST.
The problem is that LAMBDA_CAPTURE_EXPLICIT_P wasn't propagated
correctly and so var_to_maybe_prune proceeded where it shouldn't.
Co-Authored by: Patrick Palka <ppalka@redhat.com>
PR c++/108829
gcc/cp/ChangeLog:
* pt.cc (prepend_one_capture): Set LAMBDA_CAPTURE_EXPLICIT_P.
(tsubst_lambda_expr): Pass LAMBDA_CAPTURE_EXPLICIT_P to
prepend_one_capture.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/lambda/lambda-108829-2.C: New test.
* g++.dg/cpp0x/lambda/lambda-108829.C: New test.
Alex Coplan [Mon, 6 Feb 2023 14:32:21 +0000 (14:32 +0000)]
aarch64: Fix up bfmlal lane pattern [PR104921]
As the testcase shows, this pattern had an incorrect constraint leading
to GCC's output getting rejected by the assembler.
This patch fixes the constraint accordingly.
The test is split into two: one that can run without bf16 support from
the assembler and another that checks that the output actually assembles
when such support is available.
gcc/ChangeLog:
PR target/104921
* config/aarch64/aarch64-simd.md (aarch64_bfmlal<bt>_lane<q>v4sf):
Use correct constraint for operand 3.
gcc/testsuite/ChangeLog:
PR target/104921
* gcc.target/aarch64/pr104921-1.c: New test.
* gcc.target/aarch64/pr104921-2.c: New test.
* gcc.target/aarch64/pr104921.x: Include file for new tests.
Xi Ruoyao [Mon, 13 Feb 2023 10:38:53 +0000 (18:38 +0800)]
LoongArch: Fix multiarch tuple canonization
Multiarch tuple will be coded in file or directory names in
multiarch-aware distros, so one ABI should have only one multiarch
tuple. For example, "--target=loongarch64-linux-gnu --with-abi=lp64s"
and "--target=loongarch64-linux-gnusf" should both set multiarch tuple
to "loongarch64-linux-gnusf". Before this commit,
"--target=loongarch64-linux-gnu --with-abi=lp64s --disable-multilib"
will produce wrong result (loongarch64-linux-gnu).
A recent LoongArch psABI revision mandates "loongarch64-linux-gnu" to be
used for -mabi=lp64d (instead of "loongarch64-linux-gnuf64") for some
non-technical reason [1]. Note that we cannot make
"loongarch64-linux-gnuf64" an alias for "loongarch64-linux-gnu" because
to implement such an alias, we must create thousands of symlinks in the
distro and doing so would be completely unpractical. This commit also
aligns GCC with the revision.
Tested by building cross compilers with --enable-multiarch and multiple
combinations of --target=loongarch64-linux-gnu*, --with-abi=lp64{s,f,d},
and --{enable,disable}-multilib; and run "xgcc --print-multiarch" then
manually verify the result with eyesight.
* config.gcc (triplet_abi): Set its value based on $with_abi,
instead of $target.
(la_canonical_triplet): Set it after $triplet_abi is set
correctly.
* config/loongarch/t-linux (MULTILIB_OSDIRNAMES): Make the
multiarch tuple for lp64d "loongarch64-linux-gnu" (without
"f64" suffix).
Thomas Schwinge [Thu, 16 Feb 2023 14:57:37 +0000 (15:57 +0100)]
Attempt to register OpenMP pinned memory using a device instead of 'mlock'
Implemented for nvptx offloading via 'cuMemHostRegister'. This means: (a) not
running into 'mlock' limitations, and (b) the device is aware of this and may
optimize host <-> device memory transfers.
Thomas Schwinge [Thu, 16 Feb 2023 20:59:55 +0000 (21:59 +0100)]
Un-break nvptx libgomp build
In file included from [...]/libgomp/config/nvptx/allocator.c:49:
[...]/libgomp/config/nvptx/../../basic-allocator.c:52:2: error: invalid preprocessing directive #deine; did you mean #define?
52 | #deine BASIC_ALLOC_YIELD
| ^~~~~
| define
* include/experimental/bits/simd_x86.h
(_SimdImplX86::_S_not_equal_to, _SimdImplX86::_S_less)
(_SimdImplX86::_S_less_equal): Do not call
__builtin_is_constant_evaluated in constexpr-if.
* include/experimental/bits/simd.h
(_SimdWrapper::_M_is_constprop_none_of)
(_SimdWrapper::_M_is_constprop_all_of): Return false unless the
computed result still satisfies __builtin_constant_p.
Andrew Stubbs [Mon, 30 Jan 2023 14:43:00 +0000 (14:43 +0000)]
amdgcn, libgomp: low-latency allocator
This implements the OpenMP low-latency memory allocator for AMD GCN using the
small per-team LDS memory (Local Data Store).
Since addresses can now refer to LDS space, the "Global" address space is
no-longer compatible. This patch therefore switches the backend to use
entirely "Flat" addressing (which supports both memories). A future patch
will re-enable "global" instructions for cases where it is known to be safe
to do so.
gcc/ChangeLog:
* config/gcn/gcn-builtins.def (DISPATCH_PTR): New built-in.
* config/gcn/gcn.cc (gcn_init_machine_status): Disable global
addressing.
(gcn_expand_builtin_1): Implement GCN_BUILTIN_DISPATCH_PTR.
libgomp/ChangeLog:
* config/gcn/libgomp-gcn.h (TEAM_ARENA_START): Move to here.
(TEAM_ARENA_FREE): Likewise.
(TEAM_ARENA_END): Likewise.
(GCN_LOWLAT_HEAP): New.
* config/gcn/team.c (LITTLEENDIAN_CPU): New, and import hsa.h.
(__gcn_lowlat_init): New prototype.
(gomp_gcn_enter_kernel): Initialize the low-latency heap.
* libgomp.h (TEAM_ARENA_START): Move to libgomp.h.
(TEAM_ARENA_FREE): Likewise.
(TEAM_ARENA_END): Likewise.
* plugin/plugin-gcn.c (lowlat_size): New variable.
(print_kernel_dispatch): Label the group_segment_size purpose.
(init_environment_variables): Read GOMP_GCN_LOWLAT_POOL.
(create_kernel_dispatch): Pass low-latency head allocation to kernel.
(run_kernel): Use shadow; don't assume values.
* testsuite/libgomp.c/allocators-7.c: Enable for amdgcn.
* config/gcn/allocator.c: New file.
Jakub Jelinek [Thu, 16 Feb 2023 11:32:00 +0000 (12:32 +0100)]
libgomp: Fix comment typo
I saw
FAIL: libgomp.fortran/target-nowait-array-section.f90 -O execution test
in my last x86_64-linux bootstrap. From quick skimming, it might be just
unreliable test, which assumes that asynchronous execution wouldn't produce
ordered sequence, but can't it happen even with asynchronous execution?
That said, while skimming the test, I've noticed a comment typo and
this patch fixes that up.
2023-02-16 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.fortran/target-nowait-array-section.f90: Fix
comment typo and improve its wording.
Eric Botcazou [Wed, 15 Feb 2023 22:32:12 +0000 (23:32 +0100)]
Fix PR target/90458
This is the incompatibility of -fstack-clash-protection with Windows SEH.
Now the Windows ports always enable TARGET_STACK_PROBE, which means that
the stack is always probed (out of line) so -fstack-clash-protection does
nothing more.
gcc/
PR target/90458
* config/i386/i386.cc (ix86_compute_frame_layout): Disable the
effects of -fstack-clash-protection for TARGET_STACK_PROBE.
(ix86_expand_prologue): Likewise.
Thomas Schwinge [Wed, 15 Feb 2023 11:39:19 +0000 (12:39 +0100)]
Fix 'libgomp.{c-c++-common,fortran}/target-present-*' test cases
Their execution isn't expected to error out if we've been *compiling for any
offload target*, but rather if they're *executing on a non-shared memory
offload device*. For example, if (any) offloading is configured but not
effective (no device available, for example), you'd get:
PASS: libgomp.c/../libgomp.c-c++-common/target-present-1.c (test for excess errors)
FAIL: libgomp.c/../libgomp.c-c++-common/target-present-1.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/target-present-2.c (test for excess errors)
FAIL: libgomp.c/../libgomp.c-c++-common/target-present-2.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/target-present-3.c (test for excess errors)
FAIL: libgomp.c/../libgomp.c-c++-common/target-present-3.c execution test
PASS: libgomp.c++/../libgomp.c-c++-common/target-present-1.c (test for excess errors)
FAIL: libgomp.c++/../libgomp.c-c++-common/target-present-1.c execution test
PASS: libgomp.c++/../libgomp.c-c++-common/target-present-2.c (test for excess errors)
FAIL: libgomp.c++/../libgomp.c-c++-common/target-present-2.c execution test
PASS: libgomp.c++/../libgomp.c-c++-common/target-present-3.c (test for excess errors)
FAIL: libgomp.c++/../libgomp.c-c++-common/target-present-3.c execution test
PASS: libgomp.fortran/target-present-1.f90 -O0 (test for excess errors)
FAIL: libgomp.fortran/target-present-1.f90 -O0 execution test
[...]
PASS: libgomp.fortran/target-present-2.f90 -O0 (test for excess errors)
FAIL: libgomp.fortran/target-present-2.f90 -O0 execution test
[...]
PASS: libgomp.fortran/target-present-3.f90 -O0 (test for excess errors)
FAIL: libgomp.fortran/target-present-3.f90 -O0 execution test
[...]
Also, verify reaching a checkpoint before the expected error condition -- and
fix up one case where that didn't happen; missing OpenMP 'map' clauses
('libgomp.fortran/target-present-2.f90').
Marek Polacek [Tue, 14 Feb 2023 22:05:44 +0000 (17:05 -0500)]
warn-access: wrong -Wdangling-pointer with labels [PR106080]
-Wdangling-pointer warns when the address of a label escapes. This
causes grief in OCaml (<https://github.com/ocaml/ocaml/issues/11358>) as
well as in the kernel:
<https://bugzilla.kernel.org/show_bug.cgi?id=215851> because it uses
to get the PC. -Wdangling-pointer is documented to warn about pointers
to objects. However, it uses is_auto_decl which checks DECL_P, but DECL_P
is also true for a label/enumerator/function declaration, none of which is
an object. Rather, it should use auto_var_p which correctly checks VAR_P
and PARM_DECL.
PR middle-end/106080
gcc/ChangeLog:
* gimple-ssa-warn-access.cc (is_auto_decl): Remove. Use auto_var_p
instead.
gcc/testsuite/ChangeLog:
* c-c++-common/Wdangling-pointer-10.c: New test.
* c-c++-common/Wdangling-pointer-9.c: New test.
Tobias Burnus [Wed, 15 Feb 2023 10:43:21 +0000 (11:43 +0100)]
OpenMP/Fortran: Fix loop-iter var privatization with !$OMP LOOP [PR108512]
For 'parallel', loop-iteration variables are marked are marked as 'private',
unless they either appear in an omp do/simd loop or an data-sharing clause
already exists for those on 'parallel'. 'omp loop' wasn't handled, leading
to (potentially) multiple data-sharing clauses in gfc_resolve_do_iterator
as omp_current_ctx pointed to the 'parallel' directive, ignoring the
in-betwen 'loop' directive.
The latter lead to a bogus diagnostic - or rather an ICE as the source
location var contained only '\0'.
Additionally, several 'case EXEC_OMP...LOOP' have been added to call the
right resolution function and likewise for '{masked,master} taskloop'.
gcc/fortran/ChangeLog:
PR fortran/108512
* openmp.cc (gfc_resolve_omp_parallel_blocks): Handle combined 'loop'
directives.
(gfc_resolve_do_iterator): Set a source location for added
'private'-clause arguments.
* resolve.cc (gfc_resolve_code): Call gfc_resolve_omp_do_blocks
also for EXEC_OMP_LOOP and gfc_resolve_omp_parallel_blocks for
combined directives with loop + '{masked,master} taskloop (simd)'.
Tobias Burnus [Wed, 15 Feb 2023 10:42:00 +0000 (11:42 +0100)]
libgomp: Fix reverse-offload for GOMP_MAP_TO_PSET
libgomp/
* target.c (gomp_target_rev): Dereference ptr
to get device address.
* testsuite/libgomp.fortran/reverse-offload-5.f90: Add test
for unallocated allocatable.
Tobias Burnus [Wed, 15 Feb 2023 10:40:21 +0000 (11:40 +0100)]
libgomp: Fix 'target enter data' with always pointer
As GOMP_MAP_ALWAYS_POINTER operates on the previous map item, ensure that
with 'target enter data' both are passed together to gomp_map_vars_internal.
libgomp/ChangeLog:
* target.c (gomp_map_vars_internal): Add 'i > 0' before doing a
kind check.
(GOMP_target_enter_exit_data): If the next map item is
GOMP_MAP_ALWAYS_POINTER map it together with the current item.
* testsuite/libgomp.fortran/target-enter-data-4.f90: New test.
Marek Polacek [Fri, 10 Feb 2023 22:26:57 +0000 (17:26 -0500)]
c++: fix ICE in joust_maybe_elide_copy [PR106675]
joust_maybe_elide_copy checks that the last conversion in the ICS for
the first argument is ck_ref_bind, which is reasonable, because we've
checked that we're dealing with a copy/move constructor. But it can
also happen that we couldn't figure out which conversion function is
better to convert the argument, as in this testcase: joust couldn't
decide if we should go with
operator foo &()
or
operator foo const &()
so we get a ck_ambig, which then upsets joust_maybe_elide_copy. Since
a ck_ambig can validly occur, I think we should just return early, as
in the patch below.
PR c++/106675
gcc/cp/ChangeLog:
* call.cc (joust_maybe_elide_copy): Return false for ck_ambig.
Thomas Schwinge [Tue, 14 Feb 2023 22:34:45 +0000 (23:34 +0100)]
Address cast to pointer from integer of different size in 'libgomp/target.c:gomp_target_rev'
For example, for '-m32' multilib of x86_64-pc-linux-gnu:
[...]/libgomp/target.c: In function ‘gomp_target_rev’:
[...]/libgomp/target.c:3699:33: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast]
3699 | (void *) devaddrs[i],
| ^
While mainline (GCC 13) converts assumptions to the ASSUME internal
function, OG12 has only parsing-only support. Thus, remove the
failing dump scan from the testcase.
"vec_vsubcuqP" should be "vec_vsubcuq", this typo caused
us to define vec_vsubcuqP in rs6000-vecdefines.h instead
of vec_vsubcuq, so that compiler is not able to realize
the built-in function name vec_vsubcuq any more.
Co-authored-By: Andrew Pinski <apinski@marvell.com>
PR target/108396
gcc/ChangeLog:
* config/rs6000/rs6000-overload.def (VEC_VSUBCUQ): Fix typo
vec_vsubcuqP with vec_vsubcuq.
Kewen Lin [Wed, 18 Jan 2023 08:34:19 +0000 (02:34 -0600)]
rs6000: Teach rs6000_opaque_type_invalid_use_p about gcall [PR108348]
PR108348 shows one special case that MMA opaque types are
used in function arguments and treated as pass by reference,
it results in one copying from argument to a temp variable,
since this copying happens before rs6000_function_arg check,
it can cause ICE without MMA support then. This patch is to
teach function rs6000_opaque_type_invalid_use_p to check if
any function argument in a gcall stmt has the invalid use of
MMA opaque types.
btw, I checked the handling on return value, it doesn't have
this kind of issue as its checking and error emission is quite
early, so this doesn't handle function return value.
PR target/108348
gcc/ChangeLog:
* config/rs6000/rs6000.cc (rs6000_opaque_type_invalid_use_p): Add the
support for invalid uses of MMA opaque type in function arguments.
gcc/testsuite/ChangeLog:
* gcc.target/powerpc/pr108348-1.c: New test.
* gcc.target/powerpc/pr108348-2.c: New test.
Kewen Lin [Mon, 16 Jan 2023 08:15:39 +0000 (02:15 -0600)]
rs6000: Teach rs6000_opaque_type_invalid_use_p about inline asm [PR108272]
As PR108272 shows, there are some invalid uses of MMA opaque
types in inline asm statements. This patch is to teach the
function rs6000_opaque_type_invalid_use_p for inline asm,
check and error any invalid use of MMA opaque types in input
and output operands.
PR target/108272
gcc/ChangeLog:
* config/rs6000/rs6000.cc (rs6000_opaque_type_invalid_use_p): Add the
support for invalid uses in inline asm, factor out the checking and
erroring to lambda function check_and_error_invalid_use.
gcc/testsuite/ChangeLog:
* gcc.target/powerpc/pr108272-1.c: New test.
* gcc.target/powerpc/pr108272-2.c: New test.
* gcc.target/powerpc/pr108272-3.c: New test.
* gcc.target/powerpc/pr108272-4.c: New test.
Some package builds enable -fstack-protector and -Werror. Since
-fstack-protector is not supported on hppa because the stack grows
up, these packages must check for the warning generated by
-fstack-protector and suppress it on hppa. This is problematic
since hppa is the only significant architecture where the stack
grows up.
2022-12-16 John David Anglin <danglin@gcc.gnu.org>
Jakub Jelinek [Tue, 31 Jan 2023 08:20:34 +0000 (09:20 +0100)]
i386: Fix up -Wuninitialized warnings in avx512erintrin.h [PR105593]
As reported in the PR, there are some -Wuninitialized warnings in
avx512erintrin.h. One can see that by compiling sse-23.c testcase with
-Wuninitialized (or when actually using those intrinsics).
Those 6 spots use an uninitialized variable and pass it as one of the
argument to a builtin with constant mask -1, because there is no unmasked
builtin. It is true that expansion of those builtins into RTL will see
mask is all ones and ignore the unneeded argument, but -Wuninitialized
is diagnosed on GIMPLE and on GIMPLE these builtins are just builtin calls.
avx512fintrin.h and other headers use in these cases the _mm*_undefined_* ()
intrinsics, like:
return (__m512i) __builtin_ia32_psrav8di_mask ((__v8di) __X,
(__v8di) __Y,
(__v8di)
_mm512_undefined_epi32 (),
(__mmask8) -1);
etc. and the following patch does the same for avx512erintrin.h.
With the recent changes in C++ FE and the _mm*_undefined_* intrinsics,
we don't emit -Wuninitialized warnings for those (previously we didn't
just in C due to self-initialization). Of course we could also
just self-initialize these uninitialized vars and add the #pragma GCC
diagnostic dances around it, but using the intrinsics is consistent with
the rest and IMHO cleaner.
2023-01-31 Jakub Jelinek <jakub@redhat.com>
PR c++/105593
* config/i386/avx512erintrin.h (_mm512_exp2a23_round_pd,
_mm512_exp2a23_round_ps, _mm512_rcp28_round_pd, _mm512_rcp28_round_ps,
_mm512_rsqrt28_round_pd, _mm512_rsqrt28_round_ps): Use
_mm512_undefined_pd () or _mm512_undefined_ps () instead of using
uninitialized automatic variable __W.
* gcc.target/i386/sse-23.c: Add -Wuninitialized to dg-options.
Jakub Jelinek [Mon, 16 Jan 2023 08:41:38 +0000 (09:41 +0100)]
x86: Avoid -Wuninitialized warnings on _mm*_undefined_* in C++ [PR105593]
In https://gcc.gnu.org/pipermail/gcc-patches/2023-January/609844.html
I've posted a patch to allow ignoring -Winit-self using GCC diagnostic
pragmas, such that one can mark self-initialization as intentional
disabling of -Wuninitialized warnings.
The following incremental patch uses that in the x86 intrinsic
headers.
Jakub Jelinek [Mon, 16 Jan 2023 08:40:14 +0000 (09:40 +0100)]
c, c++: Allow ignoring -Winit-self through pragmas [PR105593]
As mentioned in the PR, various x86 intrinsics need to return
an uninitialized vector. Currently they use self initialization
to avoid -Wuninitialized warnings, which works fine in C, but
doesn't work in C++ where -Winit-self is enabled in -Wall.
We don't have an attribute to mark a variable as knowingly
uninitialized (the uninitialized attribute exists but means
something else, only in the -ftrivial-auto-var-init context),
and trying to suppress either -Wuninitialized or -Winit-self
inside of the _mm_undefined_ps etc. intrinsic definitions
doesn't work, one needs to currently disable through pragmas
-Wuninitialized warning at the point where _mm_undefined_ps etc.
result is actually used, but that goes against the intent of
those intrinsics.
The -Winit-self warning option actually doesn't do any warning,
all we do is record a suppression for -Winit-self if !warn_init_self
on the decl definition and later look that up in uninit pass.
The following patch changes those !warn_init_self tests which
are true only based on the command line option setting, not based
on GCC diagnostic pragma overrides to
!warning_enabled_at (DECL_SOURCE_LOCATION (decl), OPT_Winit_self)
such that it takes them into account.
2023-01-16 Jakub Jelinek <jakub@redhat.com>
PR c++/105593
gcc/c/
* c-parser.cc (c_parser_initializer): Check warning_enabled_at
at the DECL_SOURCE_LOCATION (decl) for OPT_Winit_self instead
of warn_init_self.
gcc/cp/
* decl.cc (cp_finish_decl): Check warning_enabled_at
at the DECL_SOURCE_LOCATION (decl) for OPT_Winit_self instead
of warn_init_self.
gcc/testsuite/
* c-c++-common/Winit-self3.c: New test.
* c-c++-common/Winit-self4.c: New test.
* c-c++-common/Winit-self5.c: New test.
because of the check on line 718 -- (int) i is not i. So -Wno-init-self
doesn't disable the warning as it's supposed to.
The following patch fixes it by moving the suppress_warning call from
c_gimplify_expr to the front ends, at points where we haven't created
the NOP_EXPR yet.
* c-parser.cc (c_parser_initializer): Add new tree parameter. Use it.
Call suppress_warning.
(c_parser_declaration_or_fndef): Pass d down to c_parser_initializer.
(c_parser_omp_declare_reduction): Pass omp_priv down to
c_parser_initializer.
Jakub Jelinek [Tue, 24 Jan 2023 10:28:00 +0000 (11:28 +0100)]
c++: Handle structured bindings like anon unions in initializers [PR108474]
As reported by Andrew Pinski, structured bindings (with the exception
of the ones using std::tuple_{size,element} and get which are really
standalone variables in addition to the binding one) also use
DECL_VALUE_EXPR and needs the same treatment in static initializers.
On Sun, Jan 22, 2023 at 07:19:07PM -0500, Jason Merrill wrote:
> Though, actually, why not instead fix expand_expr_real_1 (and staticp) to
> look through DECL_VALUE_EXPR?
Doing it when emitting the initializers seems to be too late to me,
we in various spots try to put parts of the static var DECL_INITIAL expressions
into the IL, or e.g. for varpool purposes remember which vars are referenced
there.
This patch moves it to record_reference, which is called from varpool_node::analyze
and so about the same time as gimplification of the bodies which also
replaces DECL_VALUE_EXPRs.
2023-01-24 Jakub Jelinek <jakub@redhat.com>
PR c++/108474
* cp-gimplify.cc (cp_fold_r): Handle structured bindings
vars like anon union artificial vars.
* g++.dg/cpp1z/decomp57.C: New test.
* g++.dg/cpp1z/decomp58.C: New test.
Jakub Jelinek [Thu, 19 Jan 2023 09:00:51 +0000 (10:00 +0100)]
forwprop: Further fixes for simplify_rotate [PR108440]
As mentioned in the simplify_rotate comment, for e.g.
((T) ((T2) X << (Y & (B - 1)))) | ((T) ((T2) X >> ((-Y) & (B - 1))))
we already emit
X r<< (Y & (B - 1))
as replacement. This PR is about the
((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y)))
((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y)))
forms if T2 is wider than T. Unlike e.g.
(X << Y) OP (X >> (B - Y))
which is valid just for Y in [1, B - 1], the above 2 forms are actually
valid and do the rotates for Y in [0, B] - for Y 0 the X value is preserved
by the left shift and right logical shift by B adds just zeros (but because
the shift is in wider precision B is still valid shift count), while for
Y equal to B X is preserved through the latter shift and the former adds
just zeros.
Now, it is unclear if we in the middle-end treat rotates with rotate count
equal or larger than precision as UB or not, unlike shifts there are less
reasons to do so, but e.g. expansion of X r<< Y if there is no rotate optab
for the mode is emitted as (X << Y) | (((unsigned) X) >> ((-Y) & (B - 1)))
and so with UB on Y == B.
The following patch does multiple things:
1) for the above 2, asks the ranger if Y could be equal to B and if so,
instead of using X r<< Y uses X r<< (Y & (B - 1))
2) for the
((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1))))
((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
forms that were fixed 2 days ago it only punts if Y might be in the
[B,B2-1] range but isn't known to be in the
[0,B][2*B,2*B][3*B,3*B]... range. Because for Y which is a multiple
of B but smaller than B2 it acts as a rotate too, left shift provides
0 and (-Y) & (B - 1) is 0 and so preserves X. Though, for the cases
where Y is not known to be in [0,B-1] the patch also uses
X r<< (Y & (B - 1)) rather than X r<< Y
3) as discussed with Aldy, instead of using global ranger it uses a pass
specific copy but lazily created on first simplify_rotate that needs it;
this e.g. handles rotate inside of if body where the guarding condition
limits the shift count to some range which will not work with the
global ranger (unless there is some SSA_NAME to attach the range to).
Note, e.g. on x86 X r<< (Y & (B - 1)) and X r<< Y actually emit the
same assembly because rotates work the same even for larger rotate counts,
but that is handled only during combine.
2023-01-19 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/108440
* tree-ssa-forwprop.cc: Include gimple-range.h.
(simplify_rotate): For the forms with T2 wider than T and shift counts of
Y and B - Y add & (B - 1) masking for the rotate count if Y could be equal
to B. For the forms with T2 wider than T and shift counts of
Y and (-Y) & (B - 1), don't punt if range could be [B, B2], but only if
range doesn't guarantee Y < B or Y = N * B. If range doesn't guarantee
Y < B, also add & (B - 1) masking for the rotate count. Use lazily created
pass specific ranger instead of get_global_range_query.
(pass_forwprop::execute): Disable that ranger at the end of pass if it has
been created.
* c-c++-common/rotate-10.c: New test.
* c-c++-common/rotate-11.c: New test.
Jakub Jelinek [Tue, 17 Jan 2023 11:14:25 +0000 (12:14 +0100)]
forwprop: Fix up rotate pattern matching [PR106523]
The comment above simplify_rotate roughly describes what patterns
are matched into what:
We are looking for X with unsigned type T with bitsize B, OP being
+, | or ^, some type T2 wider than T. For:
(X << CNT1) OP (X >> CNT2) iff CNT1 + CNT2 == B
((T) ((T2) X << CNT1)) OP ((T) ((T2) X >> CNT2)) iff CNT1 + CNT2 == B
transform these into:
X r<< CNT1
Or for:
(X << Y) OP (X >> (B - Y))
(X << (int) Y) OP (X >> (int) (B - Y))
((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y)))
((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y)))
(X << Y) | (X >> ((-Y) & (B - 1)))
(X << (int) Y) | (X >> (int) ((-Y) & (B - 1)))
((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1))))
((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
transform these into (last 2 only if ranger can prove Y < B):
X r<< Y
The following testcase shows that 2 of these are problematic.
If T2 is wider than T, then the 2 which yse (-Y) & (B - 1) on one
of the shift counts but Y on the can do something different from
rotate. E.g.:
__attribute__((noipa)) unsigned char
f7 (unsigned char x, unsigned int y)
{
unsigned int t = x;
return (t << y) | (t >> ((-y) & 7));
}
if y is [0, 7], then it is a normal rotate, and if y is in [32, ~0U]
then it is UB, but for y in [9, 31] the left shift in this case
will never leave any bits in the result, while in a rotate they are
left there. Say for y 5 and x 0xaa the expression gives
0x55 which is the same thing as rotate, while for y 19 and x 0xaa
0x5, which is different.
Now, I believe the
((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y)))
((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y)))
forms are ok, because B - Y still needs to be a valid shift count,
and if Y > B then B - Y should be either negative or very large
positive (for unsigned types).
And similarly the last 2 cases above which use & (B - 1) on both
shift operands are definitely ok.
The following patch disables the
((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1))))
((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
unless ranger says Y is not in [B, B2 - 1] range.
And, looking at it again this morning, actually the Y equal to B
case is still fine, if Y is equal to 0, then it is
(T) (((T2) X << 0) | ((T2) X >> 0))
and so X, for Y == B it is
(T) (((T2) X << B) | ((T2) X >> 0))
which is the same as
(T) (0 | ((T2) X >> 0))
which is also X. So instead of the [B, B2 - 1] range we could use
[B + 1, B2 - 1]. And, if we wanted to go further, even multiplies
of B are ok if they are smaller than B2, so we could construct a detailed
int_range_max if we wanted.
2023-01-17 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/106523
* tree-ssa-forwprop.cc (simplify_rotate): For the
patterns with (-Y) & (B - 1) in one operand's shift
count and Y in another, if T2 has wider precision than T,
punt if Y could have a value in [B, B2 - 1] range.
* c-c++-common/rotate-2.c (f5, f6, f7, f8, f13, f14, f15, f16,
f37, f38, f39, f40, f45, f46, f47, f48): Add assertions using
__builtin_unreachable about shift count.
* c-c++-common/rotate-2b.c: New test.
* c-c++-common/rotate-4.c (f5, f6, f7, f8, f13, f14, f15, f16,
f37, f38, f39, f40, f45, f46, f47, f48): Add assertions using
__builtin_unreachable about shift count.
* c-c++-common/rotate-4b.c: New test.
* gcc.c-torture/execute/pr106523.c: New test.
Jakub Jelinek [Sat, 14 Jan 2023 09:17:14 +0000 (10:17 +0100)]
c++: Avoid incorrect shortening of divisions [PR108365]
The following testcase is miscompiled, because we shorten the division
in a case where it should not be shortened.
Divisions (and modulos) can be shortened if it is unsigned division/modulo,
or if it is signed division/modulo where we can prove the dividend will
not be the minimum signed value or divisor will not be -1, because e.g.
on sizeof(long long)==sizeof(int)*2 && __INT_MAX__ == 0x7fffffff targets
(-2147483647 - 1) / -1 is UB
but
(int) (-2147483648LL / -1LL) is not, it is -2147483648.
The primary aim of both the C and C++ FE division/modulo shortening I assume
was for the implicit integral promotions of {,signed,unsigned} {char,short}
and because at this point we have no VRP information etc., the shortening
is done if the integral promotion is from unsigned type for the divisor
or if the dividend is an integer constant other than -1.
This works fine for char/short -> int promotions when char/short have
smaller precision than int - unsigned char -> int or unsigned short -> int
will always be a positive int, so never the most negative.
Now, the C FE checks whether orig_op0 is TYPE_UNSIGNED where op0 is either
the same as orig_op0 or that promoted to int, I think that works fine,
if it isn't promoted, either the division/modulo common type will have the
same precision as op0 but then the division/modulo is unsigned and so
without UB, or it will be done in wider precision (e.g. because op1 has
wider precision), but then op0 can't be minimum signed value. Or it has
been promoted to int, but in that case it was again from narrower type and
so never minimum signed int.
But the C++ FE was checking if op0 is a NOP_EXPR from TYPE_UNSIGNED.
First of all, not sure if the operand of NOP_EXPR couldn't be non-integral
type where TYPE_UNSIGNED wouldn't be meaningful, but more importantly,
even if it is a cast from unsigned integral type, we only know it can't be
minimum signed value if it is a widening cast, if it is same precision or
narrowing cast, we know nothing.
So, the following patch for the NOP_EXPR cases checks just in case that
it is from integral type and more importantly checks it is a widening
conversion.
2023-01-14 Jakub Jelinek <jakub@redhat.com>
PR c++/108365
* typeck.cc (cp_build_binary_op): For integral division or modulo,
shorten if type0 is unsigned, or op0 is cast from narrower unsigned
integral type or stripped_op1 is INTEGER_CST other than -1.
* g++.dg/opt/pr108365.C: New test.
* g++.dg/warn/pr108365.C: New test.
Jakub Jelinek [Thu, 24 Nov 2022 09:38:42 +0000 (10:38 +0100)]
libstdc++: Another merge from fast_float upstream [PR107468]
Upstream fast_float came up with a cheaper test for
fegetround () == FE_TONEAREST using one float addition, one subtraction
and one comparison. If we know we are rounding to nearest, we can use
fast path in more cases as before.
The following patch merges those changes into libstdc++.
Jakub Jelinek [Mon, 7 Nov 2022 14:17:21 +0000 (15:17 +0100)]
libstdc++: Update from latest fast_float [PR107468]
The following patch partially updates from fast_float trunk.
That way it gets fix for the incorrect rounding case,
PR107468 aka https://github.com/fastfloat/fast_float/issues/149
Using std::fegetround showed in benchmarks too slow, so instead of
doing that the patch limits the fast path where it uses floating
point multiplication rather than integral to cases where we can
prove there will be no rounding (the multiplication will be exact, not
just that the two multiplication or division operation arguments are
exactly representable).
2022-11-07 Jakub Jelinek <jakub@redhat.com>
PR libstdc++/107468
* src/c++17/fast_float/fast_float.h: Partial merge from fast_float 662497742fea7055f0e0ee27e5a7ddc382c2c38e commit.
* testsuite/20_util/from_chars/pr107468.cc: New test.
Andrew Pinski [Thu, 9 Feb 2023 15:03:54 +0000 (16:03 +0100)]
match.pd: When simplifying BFR of an insert, require a mode precision integral type [PR108688]
The same problem as PR 88739 has crept in but
this time in match.pd when simplifying bit_field_ref of
an bit_insert. That is we are generating a BIT_FIELD_REF
of a non-mode-precision integral type.
PR tree-optimization/108688
* match.pd (bit_field_ref [bit_insert]): Avoid generating
BIT_FIELD_REFs of non-mode-precision integral operands.
Jakub Jelinek [Wed, 8 Feb 2023 17:41:21 +0000 (18:41 +0100)]
vect-patterns: Fix up vect_widened_op_tree [PR108692]
The following testcase is miscompiled on aarch64-linux since r11-5160.
Given
<bb 3> [local count: 955630225]:
# i_22 = PHI <i_20(6), 0(5)>
# r_23 = PHI <r_19(6), 0(5)>
...
a.0_5 = (unsigned char) a_15;
_6 = (int) a.0_5;
b.1_7 = (unsigned char) b_17;
_8 = (int) b.1_7;
c_18 = _6 - _8;
_9 = ABS_EXPR <c_18>;
r_19 = _9 + r_23;
...
where SSA_NAMEs 15/17 have signed char, 5/7 unsigned char and rest is int
we first pattern recognize c_18 as
patt_34 = (a.0_5) w- (b.1_7);
which is still correct, 5/7 are unsigned char subtracted in wider type,
but then vect_recog_sad_pattern turns it into
SAD_EXPR <a_15, b_17, r_23>
which is incorrect, because 15/17 are signed char and so it is
sum of absolute signed differences rather than unsigned sum of
absolute unsigned differences.
The reason why this happens is that vect_recog_sad_pattern calls
vect_widened_op_tree with MINUS_EXPR, WIDEN_MINUS_EXPR on the
patt_34 = (a.0_5) w- (b.1_7); statement's vinfo and vect_widened_op_tree
calls vect_look_through_possible_promotion on the operands of the
WIDEN_MINUS_EXPR, which looks through the further casts.
vect_look_through_possible_promotion has careful code to stop when there
would be nested casts that need to be preserved, but the problem here
is that the WIDEN_*_EXPR operation itself has an implicit cast on the
operands already - in this case of WIDEN_MINUS_EXPR the unsigned char
5/7 SSA_NAMEs are widened to unsigned short before the subtraction,
and vect_look_through_possible_promotion obviously isn't told about that.
Now, I think when we see those WIDEN_{MULT,MINUS,PLUS}_EXPR codes, we had
to look through possible promotions already when creating those and so
vect_look_through_possible_promotion again isn't really needed, all we need
to do is arrange what that function will do if the operand isn't result
of any cast. Other option would be let vect_look_through_possible_promotion
know about the implicit promotion from the WIDEN_*_EXPR, but I'm afraid
that would be much harder.
2023-02-08 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/108692
* tree-vect-patterns.cc (vect_widened_op_tree): If rhs_code is
widened_code which is different from code, don't call
vect_look_through_possible_promotion but instead just check op is
SSA_NAME with integral type for which vect_is_simple_use is true
and call set_op on this_unprom.
Jakub Jelinek [Fri, 3 Feb 2023 20:37:27 +0000 (21:37 +0100)]
fortran: Fix up hash table usage in gfc_trans_use_stmts [PR108451]
The first testcase in the PR (which I haven't included in the patch because
it is unclear to me if it is supposed to be valid or not) ICEs since extra
hash table checking has been added recently. The problem is that
gfc_trans_use_stmts does
tree *slot = entry->decls->find_slot_with_hash (rent->use_name, hash,
INSERT);
if (*slot == NULL)
and later on doesn't store anything into *slot and continues. Another spot
a few lines later correctly clears the slot if it decides not to use the
slot, so the following patch does the same.
2023-02-03 Jakub Jelinek <jakub@redhat.com>
PR fortran/108451
* trans-decl.cc (gfc_trans_use_stmts): Call clear_slot before
doing continue.
Jakub Jelinek [Thu, 2 Feb 2023 08:54:18 +0000 (09:54 +0100)]
nested, openmp: Wrap OMP_CLAUSE_*_GIMPLE_SEQ into GIMPLE_BIND for declare_vars [PR108435]
When gimplifying OMP_CLAUSE_{LASTPRIVATE,LINEAR}_STMT, we wrap it always
into a GIMPLE_BIND, but when putting statements directly into
OMP_CLAUSE_{LASTPRIVATE,LINEAR}_GIMPLE_SEQ, we do it only if needed (there
are any temporaries that need to be declared in the sequence).
convert_nonlocal_omp_clauses was relying on the GIMPLE_BIND to be there always
because it called declare_vars on it.
The following patch wraps it into GIMPLE_BIND in tree-nested if we need to
declare_vars on it on demand.
2023-02-02 Jakub Jelinek <jakub@redhat.com>
PR middle-end/108435
* tree-nested.cc (convert_nonlocal_omp_clauses)
<case OMP_CLAUSE_LASTPRIVATE>: If info->new_local_var_chain and *seq
is not a GIMPLE_BIND, wrap the sequence into a new GIMPLE_BIND
before calling declare_vars.
(convert_nonlocal_omp_clauses) <case OMP_CLAUSE_LINEAR>: Merge
with the OMP_CLAUSE_LASTPRIVATE handling except for whether
seq is initialized to &OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (clause)
or &OMP_CLAUSE_LINEAR_GIMPLE_SEQ (clause).
Jakub Jelinek [Wed, 1 Feb 2023 11:52:52 +0000 (12:52 +0100)]
ree: Fix -fcompare-debug issues in combine_reaching_defs [PR108573]
The PR78437 r7-4871 changes made combine_reaching_defs punt on
WORD_REGISTER_OPERATIONS targets if a setter of smaller than word
register has wider uses. This unfortunately breaks -fcompare-debug,
because if such a use appears only in DEBUG_INSN(s), while all other
uses aren't wider than the setter, we can REE optimize it without -g
and not with -g.
Such decisions shouldn't be based on debug instructions. We could try
to reset them or adjust in some other way after we decide to perform the
change, but at least on the testcase which used to fail on riscv64-linux
the
(debug_insn 8 7 9 2 (var_location:HI s (minus:HI (subreg:HI (and:DI (reg:DI 10 a0 [160])
(const_int 1 [0x1])) 0)
(subreg:HI (ashiftrt:DI (reg/v:DI 9 s1 [orig:151 l ] [151])
(debug_expr:SI D#1)) 0))) "pr108573.c":12:5 -1
(nil))
clearly doesn't care about the upper bits and I have hard time imaging how
could one end up with DEBUG_INSN which actually cares about those upper
bits.
So, the following patch just ignores uses on DEBUG_INSNs in this case,
if we run into something where we'd need to do something further later on,
let's deal with it when we have a testcase for it.
2023-02-01 Jakub Jelinek <jakub@redhat.com>
PR debug/108573
* ree.cc (combine_reaching_defs): Don't return false for paradoxical
subregs in DEBUG_INSNs.
Jakub Jelinek [Wed, 1 Feb 2023 09:38:46 +0000 (10:38 +0100)]
c++, openmp: Handle some OMP_*/OACC_* constructs during constant expression evaluation [PR108607]
While potential_constant_expression_1 handled most of OMP_* codes (by saying that
they aren't potential constant expressions), OMP_SCOPE was missing in that list.
I've also added OMP_SCAN, though that is less important (similarly to OMP_SECTION
it ought to appear solely inside of OMP_{FOR,SIMD} resp. OMP_SECTIONS).
As the testcase shows, it isn't enough, potential_constant_expression_1
can catch only some cases, as soon as one uses switch or ifs where at least
one of the possible paths could be constant expression, we can run into the
same codes during cxx_eval_constant_expression, so this patch handles those
there as well.
2023-02-01 Jakub Jelinek <jakub@redhat.com>
PR c++/108607
* constexpr.cc (cxx_eval_constant_expression): Handle OMP_*
and OACC_* constructs as non-constant.
(potential_constant_expression_1): Handle OMP_SCAN and OMP_SCOPE.
Jakub Jelinek [Tue, 31 Jan 2023 09:12:19 +0000 (10:12 +0100)]
i386: Fix up ix86_convert_const_wide_int_to_broadcast [PR108599]
The following testcase is miscompiled. The problem is that during
RTL DSE we see a V4DI register is being loaded { 16, 16, 0, 0 }
value and DSE mostly works in terms of scalar modes, so it calls
movoi to set an OImode REG to (const_wide_int 0x100000000000000010)
and ix86_convert_const_wide_int_to_broadcast thinks it can compute
that value by broadcasting DImode 0x10. While it is true that
for TImode result the broadcast could be used, for OImode/XImode
it can't be, because all but the lowest 2 HOST_WIDE_INTs aren't
present (so are 0 or -1 depending on sign), not 0x10 in this case.
The function checks if the least significant HOST_WIDE_INT elt
of the CONST_WIDE_INT is broadcastable from QI/HI/SI/DImode and then
/* Check if OP can be broadcasted from VAL. */
for (int i = 1; i < CONST_WIDE_INT_NUNITS (op); i++)
if (val != CONST_WIDE_INT_ELT (op, i))
return nullptr;
That is needed of course, but nothing checks that
CONST_WIDE_INT_NUNITS (op) isn't too small for the mode in question.
I think if op would be 0 or -1, it ought to be never CONST_WIDE_INT,
but CONST_INT and so we can just punt whenever the number of
CONST_WIDE_INT elts is not the expected one.
2023-01-31 Jakub Jelinek <jakub@redhat.com>
PR target/108599
* config/i386/i386-expand.cc
(ix86_convert_const_wide_int_to_broadcast): Return nullptr if
CONST_WIDE_INT_NUNITS (op) times HOST_BITS_PER_WIDE_INT isn't
equal to bitsize of mode.
Jakub Jelinek [Tue, 31 Jan 2023 08:46:35 +0000 (09:46 +0100)]
bbpart: Fix up ICE on asm goto [PR108596]
On the following testcase we have asm goto in hot block with 2 successors,
one cold to which it both falls through and has one of the label
pointing to it and another hot successor with another label.
Now, during bbpart we want to ensure that no blocks from one partition fall
through into a block in a different partition. fix_up_fall_thru_edges
does that by temporarily clearing the EDGE_CROSSING on the fallthrough edge,
calling force_nonfallthru and then depending on whether it created a new
bb either set EDGE_CROSSING on the single successor edge from the new bb
(the new bb is kept in the same partition as the predecessor block), or
if no new bb has been created setting EDGE_CROSSING back on the fallthru
edge which has been forced non-EDGE_FALLTHRU.
For asm goto this doesn't always work, force_nonfallthru can create a new bb
and change the fallthrough edge to point to that, but if the original
fallthru destination block has its label referenced among the asm goto
labels, it will create a new non-fallthru edge for the label(s).
But because we've temporarily cheated and cleared EDGE_CROSSING on the edge,
it is cleared on the new edge as well, then the caller sees we've created
a new bb and just sets EDGE_CROSSING on the single fallthru edge from the
new bb. But the direct edge from cur_bb to fallthru edge's destination
isn't handled and fails afterwards consistency checks, because it crosses
partitions.
The following patch notes the case and sets EDGE_CROSSING on that edge too.
2023-01-31 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/108596
* bb-reorder.cc (fix_up_fall_thru_edges): Handle the case where cur_bb
ends with asm goto and has a crossing fallthrough edge to the same bb
that contains at least one of its labels by restoring EDGE_CROSSING
flag even on possible edge from cur_bb to new_bb successor.
Jakub Jelinek [Fri, 27 Jan 2023 10:17:35 +0000 (11:17 +0100)]
doc: Fix up return type of __builtin_va_arg_pack_len [PR108560]
__builtin_va_arg_pack_len as implemented returned int since its introduction
in 2007. The initial documentation didn't mention any return type,
which changed in 2010 in r0-103077-gab940b73bfabe2cec4 during some
documentation formatting cleanups
https://gcc.gnu.org/legacy-ml/gcc-patches/2010-09/msg01632.html
I can understand that for formatting some type was needed there
but what exactly hasn't been really discussed.
So, I think we should change documentation to match the implementation,
rather than change implementation to match the documentation.
Most people don't use more than 2147483647 arguments to inline functions,
and on poor targets with 16-bit ints I bet even having more than 65535
arguments to inline functions would be highly unexpected.
2023-01-27 Jakub Jelinek <jakub@redhat.com>
PR other/108560
* doc/extend.texi: Fix up return type of __builtin_va_arg_pack_len
from size_t to int.
projects / thirdparty / gcc.git / log
