This patch fixes an awkward corner case in which:
(a) we apply if-conversion to a loop;
(b) the original scalar loop doesn't have a vdef, and thus doesn't
need a virtual phi;
(c) the vectorised main loop does need a vdef and a virtual phi (see below);
(d) we also vectorise the epilogue; and
(e) the vectorised epilogue still needs a scalar epilogue
The specific case in which (c) applies is if a read-only loop is
vectorised using IFN_LOAD_LANES, which uses clobber statements to
mark the lifetime of the temporary array.
The vectoriser relies on the SSA renamer to update virtual operands.
All would probably be well if it postponed this update until after
it had vectorised both the main loop and the epilogue loop. However,
when vectorising the epilogue, vect_do_peeling does:
create_lcssa_for_virtual_phi (loop);
update_ssa (TODO_update_ssa_only_virtuals);
(with "loop" in this case being the to-be-vectorised epilogue loop).
So the vectoriser puts the virtual operand into SSA form for the
vectorised main loop as a separate step, during the early stages
of vectorising the epilogue.
I wasn't sure at first why that update_ssa was there. It looked
initially like it was related to create_lcssa_for_virtual_phi,
which seemed strange when create_lcssa_for_virtual_phi keeps the
SSA form up-to-date. But before r241099 it had the following comment,
which AFAICT is still the reason:
/* We might have a queued need to update virtual SSA form. As we
delete the update SSA machinery below after doing a regular
incremental SSA update during loop copying make sure we don't
lose that fact.
??? Needing to update virtual SSA form by renaming is unfortunate
but not all of the vectorizer code inserting new loads / stores
properly assigns virtual operands to those statements. */
The patch restores that comment since IMO it's helpful.
(a), (d) and (e) mean that we copy the original un-if-converted scalar
loop to act as the scalar epilogue. The update_ssa above means that this
copying needs to cope with any new virtual SSA names in the main loop.
The code to do that (reasonably) assumed that one of two things was true:
(1) the scalar loop and the vector loops don't have vdefs, and so no
virtual operand update is needed. The definition that applies
on entry to the loops is the same in all cases.
(2) the scalar loop and the vector loops have virtual phis, and so --
after applying create_lcssa_for_virtual_phi on the to-be-vectorised
epilogue loop -- the virtual operand update can be handled in the
same way as for normal SSA names.
But (b) and (c) together mean that the scalar loop and the
still-to-be-vectorised epilogue loop have no virtual phi that (2)
can use. We'd therefore keep the original vuses when duplicating,
rather than updating them to the definition that applies on exit
from the epilogue loop. (Since the epilogue is still unvectorised
and has no vdefs, the definition that applies on exit is the same
as the one that applies on entry.)
This patch therefore adds a third case: the scalar loop and
to-be-vectorised epilogue have no virtual defs, but the main loop does.
2019-12-31 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-loop-manip.c (create_lcssa_for_virtual_phi): Return
the incoming virtual operand definition.
(vect_do_peeling): When vectorizing an epilogue loop, handle the
case in which the main loop has a virtual phi and the epilogue
and scalar loops don't. Restore an earlier comment about the
update_ssa call.
gcc/testsuite/
* gcc.dg/vect/vect-epilogues-2.c: New test.
From-SVN: r279802
gcc/testsuite/gcc.dg/cpp/ucs.c #include <limits.h>
and then crafts a definition of WCHAR_MAX depending
on __WCHAR_TYPE__.
The test fails in VxWorks configurations because WCHAR_MAX
is already exposed by the system limits.h.
The patch simply guards the tentative definition
by a check verifying if the macro is defined already, so
we're using the value exposed by limits.h in this case.
2019-12-30 Olivier Hainque <hainque@adacore.com>
* testsuite/gcc.dg/cpp/ucs.c: Prevent redefinition
of WCHAR_MAX if already exposed by limits.h.
From-SVN: r279795
This change adjusts a few scan-tree-dump expressions
to allow '$' as well as '.' when matching symbol names,
This improves results on VxWorks targets configured with:
#undef NO_DOLLAR_IN_LABEL
#define NO_DOT_IN_LABEL
2019-12-20 Olivier Hainque <hainque@adacore.com>
Jerome Lambourg <lambourg@adacore.com>
* c-c++-common/pr56493.c: Allow '$' in addition to '.'
scan-tree-dump expressions matching symbol names.
* gcc.dg/tree-ssa/sra-17.c: Likewise.
* gcc.dg/tree-ssa/sra-18.c: Likewise.
Co-Authored-By: Jerome Lambourg <lambourg@adacore.com>
From-SVN: r279794
This change reworks the VxWorks specific os_defines.h internal
lisbstdc++ header to help fix build and runtime failures of various
kinds in environments from 6.4/6.9 to 7 SR640, based on experiments
and observations conducted against real installs of these OSes for
different CPU architectures.
2019-12-30 Jerome Lambourg <lambourg@adacore.com>
Olivier Hainque <hainque@adacore.com>
libstdc++
* config/os/vxworks/os_defines.h
(NOMINMAX): Always redefine to 1.
(_NO_CPP_INLINES): Likewise.
(_GLIBCXX_USE_WEAK_REF): Define to 1 for RTP on
VxWorks >= 7, to 0 otherwise.
(_GLIBCXX_HAVE_TLS): Define to 1.
For VxWorks >= 7:
(_GLIBCXX_USE_C99_MATH): Define to 1.
(_GLIBCXX_USE_C99_MATH_FP_MACROS_DYNAMIC): Define to 0.
(_HAS_TR1_DECLARATIONS): Redefine to 0.
For VxWorks < 7, RTP:
(_GLIBCXX_INCLUDE_NEXT_C_HEADERS): Define to 1.
(_GLIBCXX_USE_C99_FP_MACROS_DYNAMIC): Redefine to 1.
(__CORRECT_ISO_CPP11_MATH_H_PROTO_FP): Define.
For VxWorks < 7, kernel: #include <vxWorks.h>
Co-Authored-By: Olivier Hainque <hainque@adacore.com>
From-SVN: r279792
On Vxworks 7, includers of stdint.h (which we currently "provide")
need yvals.h to have been included ahead.
Instead of altering the common stdint-gcc.h with unpleasant
vxworks specific bits to do that, we arrange to provide stdint-gcc.h
on its own along with a stdint.h wrapper which preincludes yvals.h
on vx7 then includes stdint-gcc.h.
2019-12-30 Olivier Hainque <hainque@adacore.com>
* config/vxworks/stdint.h: New file. Include _yvals.h
then stdint-gcc.h.
* config/t-vxworks: Arrange to install the stdint.h wrapper.
* config.gcc (*-*-vxworks*): Add stdint-gcc.h to $extra_headers
so it gets copied. Set use_gcc_stdint to request _not_ crafting
stdint.h through the common Makefile rules.
From-SVN: r279791
* config/vxworks/_yvals.h: New file.
* config/vxworks/_yvals-wrapper.h: New file.
* config/vxworks/math.h: Use it to wrap the VxWorks
math.h header.
* config/vxworks/complex.h: Likewise.
* config/vxworks/setjmp.h: Likewise.
* config/vxworks/inttypes.h: Likewise.
From-SVN: r279787
Starting from VxWorks 7, the system comes with a Dinkumware
environment which requires the inclusion of "yvals.h" before other
system headers. We provide wrapped versions of a few headers to
accommodate such constraints.
2019-12-30 Jerome Lambourg <lambourg@adaccore.com>
Olivier Hainque <hainque@adacore.com>
gcc/
* config/vxworks/_yvals.h: New file.
* config/vxworks/_yvals-wrapper.h: New file.
* gcc/config/vxworks/math.h: Use it to wrap the VxWorks
math.h header.
* gcc/config/vxworks/complex.h: Likewise.
* gcc/config/vxworks/setjmp.h: Likewise.
* gcc/config/vxworks/inttypes.h: Likewise.
* config.gcc (*-*-vxworks*): Add system header wrappers
to extra_headers.
(powerpc-*-vxworks*): Reuse the common extra_headers.
From-SVN: r279784
This changes introduces an internal API for VxWorks version checks
within runtime files, a prerequisite to a few fixes coming up for libstdc++
builds on more recent versions of the OS.
2019-12-30 Olivier Hainque <hainque@adacore.com>
gcc/
* config/vxworks/_vxworks-versions.h: New file.
* config.gcc (*-*-vxworks*): Add it to extra_headers.
libgcc/
* config/gthr-vxworks.h: Use _vxworks-versions.h.
* config/gthr-vxworks-tls.c: Likewise.
From-SVN: r279783
This patch modifies the C & C++ VxWorks compiler to predefine
the __STDC_ macros verified by gcc.dg/c99-predef-1.c in the testsuite.
2019-12-13 Joel Brobecker <brobecker@adacore.com>
* config.gcc <*-*-vxworks*>: Add vxworks-c.o to c_target_objs
and cxx_target_objs. Set target_has_targetcm to "yes". Add
vxworks-predef.h to extra_headers.
* config/t-vxworks (vxworks-c.o): New target.
* config/vxworks-c.c: New file.
* config/vxworks/vxworks-predef.h: New file.
From-SVN: r279781
This change refines the VxWorks macro definitions configuring
wchar_t to accommodate the VxWorks7 environment, where wchar_t
is now typically a 32bit type.
We also ensure that the definitions for wint_t are always based
on those for wchar_t, so the two remain in sync in environments
where WCHAR_TYPE is redefined for a specific CPU architecture.
2019-12-30 Alexandre Oliva <oliva@adacore.com>
Olivier Hainque <hainque@adacore.com>
* config/vx-common.h (WCHAR_TYPE_SIZE): 32 on VxWorks 7.
(WCHAR_TYPE): Pick accordingly.
(WINT_TYPE_SIZE): Define in terms of WCHAR_TYPE_SIZE.
(WINT_TYPE): Define in terms of WCHAR_TYPE.
Co-Authored-By: Olivier Hainque <hainque@adacore.com>
From-SVN: r279780
2019-12-30 Olivier Hainque <hainque@adacore.com>
Jerome Lambourg <labourg@adacore.com>
* config/t-vxworks: Arrange to alter/restore glimits.h
before/after stmp-int-hdrs, so it uses a different macro
name to protect itself against recursive inclusions.
Co-Authored-By: Jerome Lambourg <lambourg@adacore.com>
From-SVN: r279777
Asserting !BRACE_ENCLOSED_INITIALIZER_P seems pretty pointless, since
that checks for init_list_type_node, and a compound literal won't have
that type, nor will we see that type if we just checked that it's
something else.
* decl.c (reshape_init_r): Remove assert.
From-SVN: r279771
2019-12-30 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/91541
* intrinsic.c (add_sym_4ind): New function.
(add_functions): Use it for INDEX.
(resolve_intrinsic): Also call f1m for INDEX.
* intrinsic.h (gfc_resolve_index_func): Adjust prototype to
take a gfc_arglist instead of individual arguments.
* iresolve.c (gfc_resolve_index_func): Adjust arguments.
Remove KIND argument if present, and make sure this is
not done twice.
* trans-decl.c: Include "intrinsic.h".
(gfc_get_extern_function_decl): Special case for resolving INDEX.
2019-12-30 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/91541
* gfortran.dg/index_3.f90: New test.
From-SVN: r279763
2019-12-30 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/92961
* gfortran.h (gfc_seen_div0): Add declaration.
* arith.h (gfc_seen_div0): Add definition.
(eval_intrinsic): For integer division by zero, set gfc_seen_div0.
* decl.c (variable_decl): If resolution resp. simplification
fails for array spec and a division of zero error has been
seen, return MATCH_ERROR.
2019-12-30 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/92961
* gfortran.dg/arith_divide_2.f90: New test.
From-SVN: r279762
This patch implements
<http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1327r1.html>.
When build_dynamic_cast realizes that a dynamic_cast needs a run-time check, it
generates a call to __dynamic_cast -- see dyncast.cc in libsupc++ for its
definition. The gist of my approach is to evaluate such a call at compile time.
* constexpr.c (cxx_dynamic_cast_fn_p): New function.
(extract_obj_from_addr_offset): New function.
(get_component_with_type): New function.
(cxx_eval_dynamic_cast_fn): New function.
(cxx_eval_call_expression): Call cxx_eval_dynamic_cast_fn for a call
to __dynamic_cast.
(potential_constant_expression_1): Don't give up on
cxx_dynamic_cast_fn_p.
* rtti.c (build_dynamic_cast_1): When creating a call to
__dynamic_cast, use the location of the original expression.
* g++.dg/cpp2a/constexpr-dynamic1.C: New test.
* g++.dg/cpp2a/constexpr-dynamic10.C: New test.
* g++.dg/cpp2a/constexpr-dynamic11.C: New test.
* g++.dg/cpp2a/constexpr-dynamic12.C: New test.
* g++.dg/cpp2a/constexpr-dynamic13.C: New test.
* g++.dg/cpp2a/constexpr-dynamic14.C: New test.
* g++.dg/cpp2a/constexpr-dynamic15.C: New test.
* g++.dg/cpp2a/constexpr-dynamic16.C: New test.
* g++.dg/cpp2a/constexpr-dynamic17.C: New test.
* g++.dg/cpp2a/constexpr-dynamic2.C: New test.
* g++.dg/cpp2a/constexpr-dynamic3.C: New test.
* g++.dg/cpp2a/constexpr-dynamic4.C: New test.
* g++.dg/cpp2a/constexpr-dynamic5.C: New test.
* g++.dg/cpp2a/constexpr-dynamic6.C: New test.
* g++.dg/cpp2a/constexpr-dynamic7.C: New test.
* g++.dg/cpp2a/constexpr-dynamic8.C: New test.
* g++.dg/cpp2a/constexpr-dynamic9.C: New test.
From-SVN: r279755
In this testcase we use an unmasked SVE loop with an Advanced SIMD
epilogue (because we don't yet support fully-masked downward loops).
The main loop uses a gather load for the strided access while the
epilogue loop builds the access from scalars instead. In both cases
we gimplify expressions based on the DR_STEP and insert them in the
loop preheader.
The problem was that the gather load code didn't copy the DR_STEP before
gimplifying it, meaning that the epilogue loop tried to reuse a result
from the (non-dominating) main loop preheader.
It looks at first glance like there could be other instances of this too,
but this patch just deals with the gather/scatter case.
2019-12-29 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-stmts.c (vect_get_strided_load_store_ops): Copy
DR_STEP before gimplifying it.
gcc/testsuite/
* gcc.dg/vect/vect-strided-epilogue-1.c: New test.
From-SVN: r279753
The EXTRACT_LAST_REDUCTION handling needs to generate a separate
comparison instruction that feeds the vector mask argument of the
IFN_EXTRACT_LAST call. We weren't checking whether that comparison
was supported, leading to an ICE on the testcase.
2019-12-29 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-stmts.c (vectorizable_condition): For extract-last
reductions, check that the target supports the required comparison
operation.
gcc/testsuite/
* gcc.dg/vect/vect-cond-12.c: New test.
From-SVN: r279752
All VNx2 V_INT_CONTAINER entries should map to VNx2DI. The lower-case
version was already correct.
2019-12-27 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* config/aarch64/iterators.md (V_INT_CONTAINER): Fix VNx2SF entry.
gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_11.c: New test.
From-SVN: r279743
The fold-left reduction code has a (rarely-used) fallback that handles
cases in which the loop is fully-masked and the target has no native
support for the reduction. The fallback includea a VEC_COND_EXPR
between the reduction vector and a safe value, so we should check
whether that VEC_COND_EXPR is supported.
2019-12-27 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-loop.c (vectorizable_reduction): Check whether the
target supports the required VEC_COND_EXPR operation before
allowing the fallback handling of masked fold-left reductions.
gcc/testsuite/
* gcc.target/aarch64/sve/mixed_size_10.c: New test.
From-SVN: r279742
PR c++/92438
* parser.c (cp_parser_constructor_declarator_p): If open paren
is followed by RID_ATTRIBUTE, skip over the attribute tokens and
try to parse type specifier.
* g++.dg/ext/attrib61.C: New test.
From-SVN: r279736
When cross-building for vxworks, test for declarations of long double
functions in math.h. We don't normally test for these functions when
cross compiling, because link tests don't work, or ever really, but
not defining them as available causes replacements to be defined in
ways that may cause duplicate definition linker errors if the units
defining both the replacement and the actual implementation are
brought in because of other symbols.
for libstdc++-v3/ChangeLog
* crossconfig.m4 (GLIBCXX_CROSSCONFIG) [*-vxworks*]: Define
long double functions as available if declared by math.h.
(GLIBCXX_CHECK_MATH_DECL, GLIBCXX_CHECK_MATH_DECLS): New.
* configure: Rebuild.
From-SVN: r279731
