The changes in r12-4381 were intended to reduce memory usage, but
replacing the __contiguous constant in __string_from_range with the new
__is_contiguous variable template caused a regression. The old code
checked is_pointer_v<decltype(std::__niter_base(__first))> but he new
code just checks is_pointer_v<_InputIterator>. This means that we no
longer recognise basic_string::iterator and vector::iterator as
contiguous, and so return a temporary basic_string instead of a
basic_string_view. This only affects C++17 mode, because the
std::contiguous_iterator concept is used in C++20 which gives the right
answer for __normal_iterator (and more types as well).
The fix is to specialize the new __is_contiguous variable template so it
is true for __normal_iterator<T*, C> specializations. The new partial
specializations are defined for C++20 too, because it should be cheaper
to match the partial specialization than to check whether the
std::contiguous_iterator concept is satisfied.
libstdc++-v3/ChangeLog:
* include/bits/fs_path.h (__detail::__is_contiguous): Add
partial specializations for pointers and __normal_iterator.
Adjust the __detail::__effective_range overloads so they always return a
string or string view using std::char_traits, because we don't care
about the traits of an incoming string.
Use std::contiguous_iterator in the __effective_range(const Source&)
overload, to allow returning a basic_string_view in more cases. For the
non-contiguous casecl in both __effective_range and __string_from_range,
return a std::string instead of std::u8string when the value type of the
range is char8_t. These changes avoid unnecessary basic_string
temporaries.
Also simplify __string_from_range(Iter, Iter) to not need
std::__to_address for the contiguous case.
Combine the _S_convert(string_type) and _S_convert(const T&) overloads
into a single _S_convert(T) function which also avoids the dangling
view problem of PR 102592 (should that recur somehow).
libstdc++-v3/ChangeLog:
* include/bits/fs_path.h (__detail::__is_contiguous): New
variable template to identify contiguous iterators.
(__detail::__unified_char8_t): New alias template to decide when
to treat char8_t as char without encoding conversion.
(__detail::__effective_range(const basic_string<C,T>&)): Use
std::char_traits<C> for returned string view.
(__detail::__effective_range(const basic_string_view<C,T>&)):
Likewise.
(__detail::__effective_range(const Source&)): Use
__is_contiguous to detect mode cases of contiguous iterators.
Use __unified_char8_t to return a std::string instead of
std::u8string.
When creating a path from a pair of non-contiguous iterators we pass the
iterators to _S_convert(Iter, Iter). That function passes the iterators
to __string_from_range to get a contiguous sequence of characters, and
then calls _S_convert(const C*, const C*) to perform the encoding
conversions. If the value type, C, is char8_t, then no conversion is
needed and the _S_convert<char8_t>(const char8_t*, const char8_t*)
specialization casts the pointer to const char* and returns a
std::string_view that refs to the char8_t sequence. However, that
sequence is owned by the std::u8string rvalue returned by
__string_from_range, which goes out of scope when _S_convert(Iter, Iter)
returns. That means the std::string_view is dangling and we get
undefined behaviour when parsing it as a path.
The same problem does not exist for the path members taking a "Source"
argument, because those functions all convert a non-contiguous range
into a basic_string<C> immediately, using __effective_range(__source).
That means that the rvalue string returned by that function is still in
scope for the full expression, so the string_view does not dangle.
The solution for the buggy functions is to do the same thing, and call
__string_from_range immediately, so that the returned rvalue is still in
scope for the lifetime of the string_view returned by _S_convert. To
avoid reintroducing the same problem, remove the _S_convert(Iter, Iter)
overload that calls __string_from_range and returns a dangling view.
libstdc++-v3/ChangeLog:
PR libstdc++/102592
* include/bits/fs_path.h (path::path(Iter, Iter, format))
(path::append(Iter, Iter), path::concat(Iter, Iter)): Call
__string_from_range directly, instead of two-argument overload
of _S_convert.
(path::_S_convert(Iter, Iter)): Remove.
* testsuite/27_io/filesystem/path/construct/102592.C: New test.
This patch contains two SUBREG-related optimization enabling tweaks to
the x86 backend.
The first change, to ix86_expand_vector_extract, cures the strange
-march=cascadelake related non-determinism that affected my new test
cases last week. Extracting a QImode or HImode element from an SSE
vector performs a zero-extension to SImode, which is currently
represented as:
(set (subreg:SI (reg:QI target)) (zero_extend:SI (...))
Unfortunately, the semantics of this RTL doesn't quite match what was
intended. A set of a paradoxical subreg allows the high-bits to take
an arbitrary value (hence the non-determinism). A more correct
representation should be:
(set (reg:SI temp) (zero_extend:SI (...))
(set (reg:QI target) (subreg:QI (reg:SI temp))
Optionally with the SUBREG rtx annotated as SUBREG_PROMOTED_VAR_P to
indicate that value is already zero-extended in the SUBREG_REG.
The second change is very similar, which is why I've included it in
this patch, where currently the early RTL optimizers can produce:
(set (reg:V?? hardreg) (subreg ...))
where this instruction may require a spill/reload from memory when
the modes aren't tieable. Alas the presence of the hard register
prevents combine/gcse etc. optimizing this away, or reusing the result
which would increase the lifetime of the hard register before reload.
The solution is to treat vector hard registers the same way as the
x86 backend handles scalar hard registers, and only allow sets from
pseudos before register allocation, which is achieved by checking
ix86_hardreg_mov_ok. Hence the above instruction is expanded and
maintained as:
(set (reg:V?? pseudo) (subreg ...))
(set (reg:V?? hardreg) (reg:V?? pseudo))
which allows the RTL optimizers freedom to optimize the SUBREG.
2021-10-13 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-expand.c (ix86_expand_vector_move): Use a
pseudo intermediate when moving a SUBREG into a hard register,
by checking ix86_hardreg_mov_ok.
(ix86_expand_vector_extract): Store zero-extended SImode
intermediate in a pseudo, then set target using a SUBREG_PROMOTED
annotated subreg.
* config/i386/sse.md (mov<VMOVE>_internal): Prevent CSE creating
complex (SUBREG) sets of (vector) hard registers before reload, by
checking ix86_hardreg_mov_ok.
This also helps get rid of warning
ctfc.h:215:18: warning: comma at end of enumerator list [-Wpedantic]
CTF_DTU_D_SLICE,
gcc/ChangeLog:
* ctfc.h (enum ctf_dtu_d_union_enum): Remove redundant comma.
The code that checks to see if objects have LTO content via
simple-object was not releasing resources, fixed thus.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* collect2.c (is_lto_object_file): Release simple-object
resources, close files.
The way we were previously dealing with addressing modes for MVE was preventing
the use of pre, post and offset addressing modes for the normal loads and
stores, including widening and narrowing. This patch fixes that and
adds tests to ensure we are capable of using all the available addressing
modes.
gcc/ChangeLog:
2021-10-12 Andre Vieira <andre.simoesdiasvieira@arm.com>
* config/arm/arm.c (thumb2_legitimate_address_p): Use VALID_MVE_MODE
when checking mve addressing modes.
(mve_vector_mem_operand): Fix the way we handle pre, post and offset
addressing modes.
(arm_print_operand): Fix printing of POST_ and PRE_MODIFY.
* config/arm/mve.md: Use mve_memory_operand predicate everywhere where
there is a single Ux constraint.
gcc/testsuite/ChangeLog:
2021-10-12 Andre Vieira <andre.simoesdiasvieira@arm.com>
* gcc.target/arm/mve/mve.exp: Make it test main directory.
* gcc.target/arm/mve/mve_load_memory_modes.c: New test.
* gcc.target/arm/mve/mve_store_memory_modes.c: New test.
2021-10-13 John David Anglin <danglin@gcc.gnu.org>
gcc/ChangeLog:
* config/pa/pa.md (muldi3): Add support for inlining 64-bit
multiplication on 32-bit PA 1.1 and 2.0 targets.
Support merging of profiles that are built from a different .o files
but belong to the same source file. Moreover, a checksum is verified
during profile merging and so we can safely combine such profile.
PR gcov-profile/90364
gcc/ChangeLog:
* coverage.c (build_info): Emit checksum to the global variable.
(build_info_type): Add new field for checksum.
(coverage_obj_finish): Pass object_checksum.
(coverage_init): Use 0 as checksum for .gcno files.
* gcov-dump.c (dump_gcov_file): Dump also new checksum field.
* gcov.c (read_graph_file): Read also checksum.
* doc/invoke.texi: Document the behaviour change.
libgcc/ChangeLog:
* libgcov-driver.c (merge_one_data): Skip timestamp and verify
checksums.
(write_one_data): Write also checksum.
* libgcov-util.c (read_gcda_file): Read also checksum field.
* libgcov.h (struct gcov_info): Add new field.
Currently when adding a sequence before there's no way to get the
iterator placed at the last added stmt which results in convoluted
code in the if-conversion usecase. The following adds
GSI_LAST_NEW_STMT and corrects one obvious mistake in
execute_update_addresses_taken as well as tries to avoid the
just filed PR102726 by biasing the enum values to be outside of
the boolean 0/1 range.
2021-10-13 Richard Biener <rguenther@suse.de>
* gimple-iterator.h (gsi_iterator_update): Add GSI_LAST_NEW_STMT,
start at integer value 2.
* gimple-iterator.c (gsi_insert_seq_nodes_before): Update
the iterator for GSI_LAST_NEW_STMT.
(gsi_insert_seq_nodes_after): Likewise.
* tree-if-conv.c (predicate_statements): Use GSI_LAST_NEW_STMT.
* tree-ssa.c (execute_update_addresses_taken): Correct bogus
arguments to gsi_replace.
PR target/102688
gcc/ChangeLog:
* common.opt: Use EnabledBy instead of detection in
finish_options and process_options.
* opts.c (finish_options): Remove handling of
x_flag_unroll_all_loops.
* toplev.c (process_options): Likewise for flag_web and
flag_rename_registers.
The following makes sure to rewrite arithmetic with undefined behavior
on overflow to a well-defined variant when moving them to be always
executed as part of doing if-conversion for loop vectorization.
2021-10-11 Richard Biener <rguenther@suse.de>
PR tree-optimization/102659
* tree-if-conv.c (need_to_rewrite_undefined): New flag.
(if_convertible_gimple_assign_stmt_p): Mark the loop for
rewrite when stmts with undefined behavior on integer
overflow appear.
(combine_blocks): Predicate also when we need to rewrite stmts.
(predicate_statements): Rewrite affected stmts to something
with well-defined behavior on overflow.
(tree_if_conversion): Initialize need_to_rewrite_undefined.
* gcc.dg/torture/pr69760.c: Adjust the testcase.
* gcc.target/i386/avx2-vect-mask-store-move1.c: Expect to move
the conversions to unsigned as well.
This is needed because people still find it necessary to do:
extern "C" {
#include <stdlib.h>
}
libstdc++-v3/ChangeLog:
* include/bits/c++config (__terminate): Add extern "C++".
The following fixes the volatileness check of IPA SRA which was
looking at the innermost reference when checking TREE_THIS_VOLATILE
but the reference to check is the outermost one.
2021-10-13 Richard Biener <rguenther@suse.de>
PR ipa/102714
* ipa-sra.c (ptr_parm_has_nonarg_uses): Fix volatileness
check.
* gcc.dg/ipa/pr102714.c: New testcase.
Commit r12-4240 enables vectorization at O2, this patch is to
adjust some test cases for rs6000 port accordingly.
It simply adds -fno-tree-vectorize to retain original test points.
gcc/testsuite/ChangeLog:
PR testsuite/102658
* gcc.target/powerpc/dform-1.c: Adjust as vectorization enabled at O2.
* gcc.target/powerpc/dform-2.c: Likewise.
* gcc.target/powerpc/pr80510-2.c: Likewise.
C2X adds a %b printf format to print integers in binary (analogous to
%x, including %#b printing a leading 0b on nonzero integers), with
recommended practice for a corresponding %B (where %#B uses 0B instead
of 0b) where that doesn't conflict with existing implementation
extensions. See N2630 for details (accepted for C2X, not yet in the
latest working draft). There is also a scanf %b format.
Add corresponding format checking support (%b accepted by -std=c2x
-Wformat -pedantic, %B considered an extension to be diagnosed with
-Wformat -pedantic). glibc support for the printf formats has been
proposed at
<https://sourceware.org/pipermail/libc-alpha/2021-October/131764.html>
(scanf support to be done in a separate patch).
Note that this does not add any support for these formats to the code
for bounding the amount of output produces by a printf function,
although that would also be useful.
Bootstrapped with no regressions for x86_64-pc-linux-gnu.
gcc/c-family/
* c-format.c (print_char_table): Add %b and %B formats.
(scan_char_table): Add %b format.
* c-format.h (T2X_UI, T2X_UL, T2X_ULL, T2X_US, T2X_UC, T2X_ST)
(T2X_UPD, T2X_UIM): New macros.
gcc/testsuite/
* gcc.dg/format/c11-printf-1.c, gcc.dg/format/c11-scanf-1.c,
gcc.dg/format/c2x-printf-1.c, gcc.dg/format/c2x-scanf-1.c,
gcc.dg/format/ext-9.c, gcc.dg/format/ext-10.c: New tests.
The vec_cpsgn built-in function API differs in argument order from the
copysign<mode>3 convention. Currently that pattern is incorrctly used to
implement vec_cpsgn. Fix that by reversing the operand order of the
builtin while leaving the existing pattern in place to implement copysignf
for vector modes.
Part of the fix when using the new built-in support requires an adjustment
to a pending patch that replaces much of altivec.h with an automatically
generated file.
Also fix a bug in the new built-in overload infrastructure where we were
using the VSX form of the VEC_COPYSIGN built-in when we should default to
the VMX form.
2021-10-12 Bill Schmidt <wschmidt@linux.ibm.com>
gcc/
PR target/101985
* config/rs6000/altivec.h (vec_cpsgn): Swap operand order.
* config/rs6000/rs6000-overload.def (VEC_COPYSIGN): Use SKIP to
avoid generating an automatic #define of vec_cpsgn. Use the
correct built-in for V4SFmode that doesn't depend on VSX.
gcc/testsuite/
PR target/101985
* gcc.target/powerpc/pr101985-1.c: New.
* gcc.target/powerpc/pr101985-2.c: New.
The subdirectories in question are libcody,
libdecnumber, c++tools, libgcc, and libobjc.
This is progress towards allowing "make tags" to
work from the top-level directory; a few additional
changes may also be necessary, though.
ChangeLog:
* Makefile.def: Mark libcody, libdecnumber,
c++tools, libgcc, and libobjc as missing TAGS
targets.
* Makefile.in: Regenerate.
As explained in PR82524, LRA is not able to reload strict_low_part inout
operand with matched input operand. The patch introduces a workaround,
where we allow LRA to generate an instruction with non-matched input operand
which is split post reload to an instruction that inserts non-matched input
operand to an inout operand and the instruction that uses matched operand.
The generated code improves from:
movsbl %dil, %edx
movl %edi, %eax
sall $3, %edx
movb %dl, %al
to:
movl %edi, %eax
movb %dil, %al
salb $3, %al
which is still not optimal, but the code is one instruction shorter and
does not use a temporary register.
2021-10-12 Uroš Bizjak <ubizjak@gmail.com>
gcc/
PR target/85730
PR target/82524
* config/i386/i386.md (*add<mode>_1_slp): Rewrite as
define_insn_and_split pattern. Add alternative 1 and split it
post reload to insert operand 1 into the low part of operand 0.
(*sub<mode>_1_slp): Ditto.
(*and<mode>_1_slp): Ditto.
(*<any_or:code><mode>_1_slp): Ditto.
(*ashl<mode>3_1_slp): Ditto.
(*<any_shiftrt:insn><mode>3_1_slp): Ditto.
(*<any_rotate:insn><mode>3_1_slp): Ditto.
(*neg<mode>_1_slp): New insn_and_split pattern.
(*one_cmpl<mode>_1_slp): Ditto.
gcc/testsuite/
PR target/85730
PR target/82524
* gcc.target/i386/pr85730.c: New test.
Also restore the test for 'a < a' that was removed by r12-2537 because
it is ill-formed. We still want to test operator< for tuple, we just
need to not use std::nullptr_t in that tuple type.
libstdc++-v3/ChangeLog:
* testsuite/20_util/tuple/comparison_operators/overloaded.cc:
Restore test for operator<.
* testsuite/20_util/tuple/comparison_operators/overloaded2.cc:
Adjust expected errors for C++20.
The r12-3022 commit only fixed the case where an array is the last
element of the tuple. This fixes the other cases too. We can just define
the move constructor as defaulted, which does the right thing. Changing
the move constructor to be trivial would be an ABI break, but since the
last base class still has a non-trivial move constructor, defining the
derived ones as defaulted doesn't change anything.
libstdc++-v3/ChangeLog:
PR libstdc++/101960
* include/std/tuple (_Tuple_impl(_Tuple_impl&&)): Define as
defauled.
* testsuite/20_util/tuple/cons/101960.cc: Check tuples with
array elements before the last element.
This adds deleted overloads so that the errors for invalid uses of
std::advance and std::distance are easier to understand (see for example
PR 102181).
libstdc++-v3/ChangeLog:
* include/bits/stl_iterator_base_funcs.h (__advance): Add
deleted overload to improve diagnostics.
(__distance): Likewise.
This patch releases the device lock on a sanity-checking error path in
transfer combining (cbuf) handling in libgomp:target.c. This shouldn't
happen when handling well-formed mapping clauses, but erroneous clauses
can currently cause a hang if the condition triggers.
2021-12-10 Julian Brown <julian@codesourcery.com>
libgomp/
* target.c (gomp_copy_host2dev): Release device lock on cbuf
error path.
This fixes a forgotten adjustment of matches[] when we fail SLP
discovery.
2021-10-12 Richard Biener <rguenther@suse.de>
PR tree-optimization/102696
* tree-vect-slp.c (vect_build_slp_tree_2): Properly mark
the tree fatally failed when we reject a BIT_FIELD_REF.
* g++.dg/vect/pr102696.cc: New testcase.
This fixes the vector def gathering for invariant masks which
failed to pass in the desired vector type resulting in a non-mask
type to be generate.
2021-10-12 Richard Biener <rguenther@suse.de>
PR tree-optimization/102572
* tree-vect-stmts.c (vect_build_gather_load_calls): When
gathering the vectorized defs for the mask pass in the
desired mask vector type so invariants will be handled
correctly.
* g++.dg/vect/pr102572.cc: New testcase.
The following example
void f10(double * restrict z, double * restrict w, double * restrict x,
double * restrict y, int n)
{
for (int i = 0; i < n; i++) {
z[i] = (w[i] > 0) ? x[i] + w[i] : y[i] - w[i];
}
}
generates currently:
ld1d z1.d, p1/z, [x1, x5, lsl 3]
fcmgt p2.d, p1/z, z1.d, #0.0
fcmgt p0.d, p3/z, z1.d, #0.0
ld1d z2.d, p2/z, [x2, x5, lsl 3]
bic p0.b, p3/z, p1.b, p0.b
ld1d z0.d, p0/z, [x3, x5, lsl 3]
where a BIC is generated between p1 and p0 where a NOT would be better here
since we won't require the use of p3 and opens the pattern up to being CSEd.
After this patch using a 2 -> 2 split we generate:
ld1d z1.d, p0/z, [x1, x5, lsl 3]
fcmgt p2.d, p0/z, z1.d, #0.0
not p1.b, p0/z, p2.b
The additional scratch is needed such that we can CSE the two operations. If
both statements wrote to the same register then CSE won't be able to CSE the
values if there are other statements in between that use the register.
A second pattern is needed to capture the nor case as combine will match the
longest sequence first. So without this pattern we end up de-optimizing nor
and instead emit two nots. I did not find a better way to do this.
gcc/ChangeLog:
* config/aarch64/aarch64-sve.md (*fcm<cmp_op><mode>_bic_combine,
*fcm<cmp_op><mode>_nor_combine, *fcmuo<mode>_bic_combine,
*fcmuo<mode>_nor_combine): New.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/sve/pred-not-gen-1.c: New test.
* gcc.target/aarch64/sve/pred-not-gen-2.c: New test.
* gcc.target/aarch64/sve/pred-not-gen-3.c: New test.
* gcc.target/aarch64/sve/pred-not-gen-4.c: New test.
We need a 32-byte wide integer mode (OImode) in order to handle structure
returns in the 64-bit ABI.
gcc/
PR target/102588
* config/sparc/sparc-modes.def (OI): New integer mode.
This adds the Fortran testsuite coverage of
omp_{get_max,set_num}_threads and omp_{s,g}et_teams_thread_limit
libgomp/
* testsuite/libgomp.fortran/icv-3.f90: New.
* testsuite/libgomp.fortran/icv-4.f90: New.
Starting point was PR102541, were a previous patch caused an invalid
e->ref access for class. When testing, it turned out that for
CLASS to CLASS the code was never executed - additionally, issues
appeared for optional and a bogus error for -fcheck=all. In particular:
There were a bunch of issues related to optional CLASS, can have the
'attr.dummy' set in CLASS_DATA (sym) - but sometimes also in 'sym'!?!
Additionally, gfc_variable_attr could return pointer = 1 for nonpointers
when the expr is no longer "var" but "var%_data".
PR fortran/102541
gcc/fortran/ChangeLog:
* check.c (gfc_check_present): Handle optional CLASS.
* interface.c (gfc_compare_actual_formal): Likewise.
* trans-array.c (gfc_trans_g77_array): Likewise.
* trans-decl.c (gfc_build_dummy_array_decl): Likewise.
* trans-types.c (gfc_sym_type): Likewise.
* primary.c (gfc_variable_attr): Fixes for dummy and
pointer when 'class%_data' is passed.
* trans-expr.c (set_dtype_for_unallocated, gfc_conv_procedure_call):
For assumed-rank dummy, fix setting rank for dealloc/notassoc actual
and setting ubound to -1 for assumed-size actuals.
gcc/testsuite/ChangeLog:
* gfortran.dg/assumed_rank_24.f90: New test.
We can use the clear_padding_type_may_have_padding_p function, which
is conservative for e.g. RECORD_TYPE/UNION_TYPE, but for the floating and
complex floating types is accurate. clear_type_padding_in_mask is
more expensive because we need to allocate memory, fill it, call the function
which itself is more expensive and then analyze the memory, so for the
common case of float/double atomics or even long double on most targets
we can avoid that.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* gimple-fold.h (clear_padding_type_may_have_padding_p): Declare.
* gimple-fold.c (clear_padding_type_may_have_padding_p): No longer
static.
gcc/c-family/
* c-omp.c (c_finish_omp_atomic): Use
clear_padding_type_may_have_padding_p.
This patch adds documentation for these new OpenMP 5.1 APIs as well as
two new environment variables - OMP_NUM_TEAMS and OMP_TEAMS_THREAD_LIMIT.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (omp_get_max_teams, omp_get_teams_thread_limit,
omp_set_num_teams, omp_set_teams_thread_limit, OMP_NUM_TEAMS,
OMP_TEAMS_THREAD_LIMIT): Document.
When building libgomp documentation, I see
makeinfo --split-size=5000000 -I ../../../libgomp/../gcc/doc/include -I ../../../libgomp -o libgomp.info ../../../libgomp/libgomp.texi
../../../libgomp/libgomp.texi:503: warning: node next `omp_get_default_device' in menu `omp_get_device_num' and in sectioning `omp_get_dynamic' differ
../../../libgomp/libgomp.texi:528: warning: node prev `omp_get_dynamic' in menu `omp_get_device_num' and in sectioning `omp_get_default_device' differ
../../../libgomp/libgomp.texi:560: warning: node next `omp_get_initial_device' in menu `omp_get_level' and in sectioning `omp_get_device_num' differ
../../../libgomp/libgomp.texi:587: warning: node next `omp_get_device_num' in menu `omp_get_dynamic' and in sectioning `omp_get_level' differ
../../../libgomp/libgomp.texi:587: warning: node prev `omp_get_device_num' in menu `omp_get_default_device' and in sectioning `omp_get_initial_device' differ
../../../libgomp/libgomp.texi:615: warning: node prev `omp_get_level' in menu `omp_get_initial_device' and in sectioning `omp_get_device_num' differ
warnings. This patch fixes those.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (omp_get_device_num): Move @node before omp_get_dynamic
to avoid makeinfo warnings.
This adds (C/C++ only) testsuite coverage for these new OpenMP 5.1 APIs.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.c-c++-common/icv-3.c: New test.
* testsuite/libgomp.c-c++-common/icv-4.c: New test.
As reported, the alloc-9.c test and alloc-{1,2,3}.F* and alloc-11.f90
tests fail on powerpc64-linux with -m32.
The reason why it fails just there is that malloc doesn't guarantee there
128-bit alignment (historically glibc guaranteed 2 * sizeof (void *)
alignment from malloc).
There are two separate issues.
One is a thinko on my side.
In this part of alloc-9.c test (copied to alloc-11.f90), we have
2 allocators, a with pool size 1024B and alignment 16B and default fallback
and a2 with pool size 512B and alignment 32B and a as fallback allocator.
We start at no allocations in both at line 194 and do:
p = (int *) omp_alloc (sizeof (int), a2);
// This succeeds in a2 and needs 4+overhead bytes (which includes the 32B alignment)
p = (int *) omp_realloc (p, 420, a, a2);
// This allocates 420 bytes+overhead in a, with 16B alignment and deallocates the above
q = (int *) omp_alloc (sizeof (int), a);
// This allocates 4+overhead bytes in a, with 16B alignment
q = (int *) omp_realloc (q, 420, a2, a);
// This allocates 420+overhead in a2 with 32B alignment
q = (int *) omp_realloc (q, 768, a2, a2);
// This attempts to reallocate, but as there are elevated alignment
// requirements doesn't try to just realloc (even if it wanted to try that
// a2 is almost full, with 512-420-overhead bytes left in it), so it
// tries to alloc in a2, but there is no space left in the pool, falls
// back to a, which already has 420+overhead bytes allocated in it and
// 1024-420-overhead bytes left and so fails too and fails to default
// non-pool allocator that allocates it, but doesn't guarantee alignment
// higher than malloc guarantees.
// But, the test expected 16B alignment.
So, I've slightly lowered the allocation sizes in that part of the test
420->320 and 768 -> 568, so that the last test still fails to allocate
in a2 (568 > 512-320-overhead) but succeeds in a as fallback, which was
the intent of the test.
Another thing is that alloc-1.F90 seems to be transcription of
libgomp.c-c++-common/alloc-1.c into Fortran, but alloc-1.c had:
q = (int *) omp_alloc (768, a2);
if ((((uintptr_t) q) % 16) != 0)
abort ();
q[0] = 7;
q[767 / sizeof (int)] = 8;
r = (int *) omp_alloc (512, a2);
if ((((uintptr_t) r) % __alignof (int)) != 0)
abort ();
there but Fortran has:
cq = omp_alloc (768_c_size_t, a2)
if (mod (transfer (cq, intptr), 16_c_intptr_t) /= 0) stop 12
call c_f_pointer (cq, q, [768 / c_sizeof (i)])
q(1) = 7
q(768 / c_sizeof (i)) = 8
cr = omp_alloc (512_c_size_t, a2)
if (mod (transfer (cr, intptr), 16_c_intptr_t) /= 0) stop 13
I'm changing the latter to 4_c_intptr_t because other spots in the
testcase do that, Fortran sadly doesn't have c_alignof, but strictly
speaking it isn't correct, __alignof (int) could be on some architectures
smaller than 4.
So probably alloc-1.F90 etc. should also have
! { dg-additional-sources alloc-7.c }
! { dg-prune-output "command-line option '-fintrinsic-modules-path=.*' is valid for Fortran but not for C" }
and use get__alignof_int.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
PR libgomp/102628
PR libgomp/102668
* testsuite/libgomp.c-c++-common/alloc-9.c (main): Decrease
allocation sizes from 420 to 320 and from 768 to 568.
* testsuite/libgomp.fortran/alloc-11.f90: Likewise.
* testsuite/libgomp.fortran/alloc-1.F90: Change expected alignment
for cr from 16 to 4.
> * testsuite/libgomp.c++/scan-10.C: Add option -fvect-cost-model=cheap.
I don't think this is the right thing to do.
This just means that at some point between 2013 when -fsimd-cost-model has
been introduced and now -fsimd-cost-model= option at least partially stopped
working properly.
As documented, -fsimd-cost-model= overrides the -fvect-cost-model= setting
for OpenMP simd loops (loop->force_vectorize is true) if specified differently
from default.
In tree-vectorizer.h we have:
static inline bool
unlimited_cost_model (loop_p loop)
{
if (loop != NULL && loop->force_vectorize
&& flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
return flag_simd_cost_model == VECT_COST_MODEL_UNLIMITED;
return (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED);
}
and use it in various places, but we also just use flag_vect_cost_model
in lots of places (and in one spot use flag_simd_cost_model, not sure if
we are sure it is a force_vectorize loop or what).
So, IMHO we should change the above inline function to
loop_cost_model and let it return the cost model and then just
reimplement unlimited_cost_model as
return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
and then adjust the direct uses of the flag and revert these changes.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree-vectorizer.h (loop_cost_model): New function.
(unlimited_cost_model): Use it.
* tree-vect-loop.c (vect_analyze_loop_costing): Use loop_cost_model
call instead of flag_vect_cost_model.
* tree-vect-data-refs.c (vect_enhance_data_refs_alignment): Likewise.
(vect_prune_runtime_alias_test_list): Likewise. Also use it instead
of flag_simd_cost_model.
gcc/testsuite/
* gcc.dg/gomp/simd-2.c: Remove option -fvect-cost-model=cheap.
* gcc.dg/gomp/simd-3.c: Likewise.
libgomp/
* testsuite/libgomp.c/scan-11.c: Remove option -fvect-cost-model=cheap.
* testsuite/libgomp.c/scan-12.c: Likewise.
* testsuite/libgomp.c/scan-13.c: Likewise.
* testsuite/libgomp.c/scan-14.c: Likewise.
* testsuite/libgomp.c/scan-15.c: Likewise.
* testsuite/libgomp.c/scan-16.c: Likewise.
* testsuite/libgomp.c/scan-17.c: Likewise.
* testsuite/libgomp.c/scan-18.c: Likewise.
* testsuite/libgomp.c/scan-19.c: Likewise.
* testsuite/libgomp.c/scan-20.c: Likewise.
* testsuite/libgomp.c/scan-21.c: Likewise.
* testsuite/libgomp.c/scan-22.c: Likewise.
* testsuite/libgomp.c++/scan-9.C: Likewise.
* testsuite/libgomp.c++/scan-10.C: Likewise.
* testsuite/libgomp.c++/scan-11.C: Likewise.
* testsuite/libgomp.c++/scan-12.C: Likewise.
* testsuite/libgomp.c++/scan-13.C: Likewise.
* testsuite/libgomp.c++/scan-14.C: Likewise.
* testsuite/libgomp.c++/scan-15.C: Likewise.
* testsuite/libgomp.c++/scan-16.C: Likewise.
This issue was observed in rs6000 specific PR102658 as well.
I've looked into it a bit, it's caused by the "conditional store replacement" which
is originally disabled without vectorization as below code.
/* If either vectorization or if-conversion is disabled then do
not sink any stores. */
if (param_max_stores_to_sink == 0
|| (!flag_tree_loop_vectorize && !flag_tree_slp_vectorize)
|| !flag_tree_loop_if_convert)
return false;
The new change makes the innermost loop look like
for (int c1 = 0; c1 <= 1499; c1 += 1) {
if (c1 <= 500) {
S_10(c0, c1);
} else {
S_9(c0, c1);
}
S_11(c0, c1);
}
and can not be splitted as:
for (int c1 = 0; c1 <= 500; c1 += 1)
S_10(c0, c1);
for (int c1 = 501; c1 <= 1499; c1 += 1)
S_9(c0, c1);
So instead of disabling vectorization, could we just disable this cs replacement
with parameter "--param max-stores-to-sink=0"?
I tested this proposal on ppc64le, it should work as well.
2021-10-11 Kewen Lin <linkw@linux.ibm.com>
libgomp/ChangeLog:
* testsuite/libgomp.graphite/force-parallel-8.c: Add --param max-stores-to-sink=0.