gcc.git
4 years agoPR 93567, G edit descriptor uses E instead of F editing in rounding mode UP.
Dominique d'Humieres [Fri, 24 Jul 2020 18:27:53 +0000 (20:27 +0200)]
PR 93567, G edit descriptor uses E instead of F editing in rounding mode UP.

The switch between FMT_E and FMT_F is based on the absolute value.
Set r=0 for rounding toward zero and r = 1 otherwise.
If (exp_d - m) == 1 there is no rounding needed.

libgfortran/ChangeLog:

PR fortran/93567
* io/write_float.def (determine_en_precision): Fix switch between
FMT_E and FMT_F.

gcc/testsuite/ChangeLog:

PR fortran/93567
* gfortran.dg/round_3.f08: Add test cases.

4 years agoPR 93592 - Invalid UP/DOWN rounding with EN descriptor.
Dominique d'Humieres [Fri, 24 Jul 2020 18:07:12 +0000 (20:07 +0200)]
PR 93592 - Invalid UP/DOWN rounding with EN descriptor.

The fix is obvious (I have added a comment). The tests are probably
an overkill, but it does not hurt.

libgfortran/ChangeLog:

PR fortran/93592
* io/write_float.def (build_float_string): Do not reset
  nbefore for FMT_F and FMT_EN.

gcc/testsuite/ChangeLog:

PR fortran/93592
* gfortran.dg/fmt_en.f90: Adjust test.
* gfortran.dg/fmt_en_rd.f90: New test.
* gfortran.dg/fmt_en_rn.f90: New test.
* gfortran.dg/fmt_en_ru.f90: New test.
* gfortran.dg/fmt_en_rz.f90: New test.

4 years agoc++: Add testcase for [PR81339]
Patrick Palka [Fri, 24 Jul 2020 18:31:36 +0000 (14:31 -0400)]
c++: Add testcase for [PR81339]

We correctly reject this testcase since r11-434, i.e. since the fix for
PR c++/57943.

gcc/testsuite/ChangeLog:

PR c++/81339
* g++.dg/cpp0x/decltype78.C: New test.

4 years agoiDefine TARGET_OFFLOAD_OPTIONS for AArch64
Matthias Klose [Fri, 24 Jul 2020 14:17:44 +0000 (16:17 +0200)]
iDefine TARGET_OFFLOAD_OPTIONS for AArch64

2020-07-24  Matthias Klose  <doko@ubuntu.com>

* config/aarch64/aarch64.c (+aarch64_offload_options,
TARGET_OFFLOAD_OPTIONS): New.

4 years agoi386: Emit mfence_sse2 for -Os [PR95750]
Uros Bizjak [Fri, 24 Jul 2020 13:59:38 +0000 (15:59 +0200)]
i386: Emit mfence_sse2 for -Os [PR95750]

2020-07-24  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:
PR target/95750
* config/i386/sync.md (mmem_thread_fence): Emit mfence_sse2 for -Os.

4 years ago[testsuite] Unset 'offload_target' after use
Thomas Schwinge [Wed, 1 Jul 2020 10:33:59 +0000 (12:33 +0200)]
[testsuite] Unset 'offload_target' after use

..., so that we don't leak this into '*.exp' files running later.

This is relevant after commit efc16503ca10bc0e934e0bace5777500e4dc757a "handle
dumpbase in offloading, adjust testsuite" -- I was confused why in a
(simplified) testing sequence as follows:

  default 'libgomp.c/c.exp'
  default 'libgomp.oacc-c/c.exp'
  '-m32' 'libgomp.c/c.exp'
  '-m32' 'libgomp.oacc-c/c.exp'

..., the "'-m32' 'libgomp.c/c.exp'" variant would not execute any offloading
dump scanning.  The reason is that the "default 'libgomp.oacc-c/c.exp'" variant
ends with 'offload_target=disable' set, so that's what the "'-m32'
'libgomp.c/c.exp'" variant would then see, in particular
'gcc/testsuite/lib/scanoffload.exp:scoff'.

libgomp/
* testsuite/libgomp.oacc-c++/c++.exp: Unset 'offload_target' after
use.
* testsuite/libgomp.oacc-c/c.exp: Likewise.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.

4 years ago[testsuite] Fix TCL syntax error in 'scan-dump-dem-not'
Thomas Schwinge [Tue, 30 Jun 2020 05:18:19 +0000 (07:18 +0200)]
[testsuite] Fix TCL syntax error in 'scan-dump-dem-not'

    ERROR: [...]: error executing dg-final: missing close-bracket

Bug introduced in commit 2ef1bce664881cd59dac7db6c3f3b5709d858bfe (r188996),
long ago.

gcc/testsuite/
* lib/scandump.exp (scan-dump-dem-not): Fix TCL syntax error.

4 years agoaarch64: add PAC-RET protection to libitm sjlj.S
Szabolcs Nagy [Thu, 16 Jul 2020 15:39:23 +0000 (16:39 +0100)]
aarch64: add PAC-RET protection to libitm sjlj.S

_ITM_beginTransaction is a 'returns_twice' function that saves x30
on the stack as part of gtm_jmpbuf (that is passed down to
GTM_begin_transaction), but the saved x30 is also used for return.

The return path should be protected so we don't leave an
  ldp x29, x30, [sp]
  ret
gadget in the code, so x30 is signed on function entry. This
exposes the signed address in the gtm_jmpbuf too. The jmpbuf does
not need a signed address since GTM_longjmp uses
  ldp x29, x30, [x1]
  br x30
and with BTI there is a BTI j at the _ITM_beginTransaction call site
where this jump returns. Using PAC does not hurt: the gtm_jmpbuf is
internal to libitm and its layout is only used by sjlj.S so the
signed address does not escape. Saving signed x30 into gtm_jmpbuf
provides a bit of extra protection, but more importantly it allows
adding the PAC-RET support without changing the existing code much.

In theory bti and pac-ret protection can be added unconditionally
since the instructions are in the nop space, in practice they
can cause trouble if some tooling does not understand the gnu
property note (e.g. old binutils) or some unwinder or debugger
does not understand the new dwarf op code used for pac-ret (e.g
old gdb). So the code is written to only support branch-protection
according to the code generation options.

libitm/ChangeLog:

* config/aarch64/sjlj.S: Add conditional pac-ret protection.

4 years agoaarch64: add PAC GNU property note to libgcc lse.S
Szabolcs Nagy [Thu, 16 Jul 2020 15:26:10 +0000 (16:26 +0100)]
aarch64: add PAC GNU property note to libgcc lse.S

This note is not used anywhere currently but it is supposed to mark
objects if the return address is protected with PAC on the stack.
Since lse.S only has leaf functions the return address is never
saved on the stack so we can add the note.

The note is only added if pac-ret is enabled because it can cause
problems with old linkers and we don't have checks for that. This
can be changed later to be unconditional, for now it is consistent
with how gcc generates the notes.

libgcc/ChangeLog:

* config/aarch64/lse.S: Add PAC property note.

4 years agoRestrict PR middle-end/95886 x86 test to !ia32
H.J. Lu [Fri, 24 Jul 2020 02:14:06 +0000 (19:14 -0700)]
Restrict PR middle-end/95886 x86 test to !ia32

Since gcc.target/i386/memcpy-pr95886.c requires 64-bit register, restrict
it to !ia32.

PR middle-end/95886
* gcc.target/i386/memcpy-pr95886.c: Restrict test to !ia32.

4 years agoDaily bump.
GCC Administrator [Fri, 24 Jul 2020 00:16:20 +0000 (00:16 +0000)]
Daily bump.

4 years agolibgo: add AIX FAT libraries support
Clément Chigot [Thu, 16 Jul 2020 09:29:49 +0000 (11:29 +0200)]
libgo: add AIX FAT libraries support

AIX-style libraries contains both 32 and 64 bit shared objects.
This patch follows the adding of FAT libraries support in other gcc
libraries (libgcc, listdc++, etc).

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/242957

4 years agoRestrict test to LP64.
Martin Sebor [Thu, 23 Jul 2020 20:08:22 +0000 (14:08 -0600)]
Restrict test to LP64.

gcc/testsuite/ChangeLog:

PR testsuite/95886
* gcc.target/i386/memcpy-pr95886.c: Restrict test to LP64.

4 years agoopenacc: Remove unnecessary detach finalization
Julian Brown [Thu, 2 Jul 2020 21:18:20 +0000 (14:18 -0700)]
openacc: Remove unnecessary detach finalization

The call to gomp_detach_pointer in gomp_unmap_vars_internal does not
need to force finalization, and doing so may mask mismatched pointer
attachments/detachments. This patch removes the forcing.

2020-07-16  Julian Brown  <julian@codesourcery.com>
    Thomas Schwinge  <thomas@codesourcery.com>

libgomp/
* target.c (gomp_unmap_vars_internal): Remove unnecessary forcing of
finalization for detach operation.
* testsuite/libgomp.oacc-c-c++-common/structured-detach-underflow.c:
New test.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
4 years agoResolve regression rtl-optimization/96298. Sorry for the breakage.
Roger Sayle [Thu, 23 Jul 2020 17:48:54 +0000 (18:48 +0100)]
Resolve regression rtl-optimization/96298.  Sorry for the breakage.

2020-07-23  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
PR rtl-optimization/96298
* simplify-rtx.c (simplify_binary_operation_1) [XOR]: Xor doesn't
distribute over xor, so (a^b)^(c^b) is not the same as (a^c)^b.

4 years agocontrib/vimrc: detect more C-like files
Patrick Palka [Thu, 23 Jul 2020 14:38:58 +0000 (10:38 -0400)]
contrib/vimrc: detect more C-like files

Currently this script doesn't set the indentation style for the standard
library headers under libstdc++/ because they lack a file extension.
But they do have a modeline, so the file type is still set appropriately
by Vim.  So by inspecting &filetype, we can also detect these standard
library headers as C-like files.

contrib/ChangeLog:

* vimrc (SetStyle): Also inspect &filetype to determine whether
a file is C-like.

4 years agoOpenMP: Support 'lastprivate (conditional:' in Fortran
Tobias Burnus [Thu, 23 Jul 2020 15:36:41 +0000 (17:36 +0200)]
OpenMP: Support 'lastprivate (conditional:' in Fortran

gcc/fortran/ChangeLog:

* gfortran.h (gfc_omp_namelist): Add lastprivate_conditional.
* openmp.c (gfc_match_omp_clauses): Handle 'conditional:'
modifier of 'lastprivate'.
* trans-openmp.c (gfc_omp_clause_default_ctor): Don't assert
on OMP_CLAUSE__CONDTEMP_ and other OMP_*TEMP_.
(gfc_trans_omp_variable_list): Handle lastprivate_conditional.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/lastprivate-conditional-1.f90: New test.
* gfortran.dg/gomp/lastprivate-conditional-2.f90: New test.
* gfortran.dg/gomp/lastprivate-conditional-3.f90: New test.
* gfortran.dg/gomp/lastprivate-conditional-4.f90: New test.
* gfortran.dg/gomp/lastprivate-conditional-5.f90: New test.

4 years ago[Ada] Stub CUDA_Execute and CUDA_Global pragmas
Arnaud Charlet [Thu, 23 Jul 2020 13:54:45 +0000 (09:54 -0400)]
[Ada] Stub CUDA_Execute and CUDA_Global pragmas

This commit adds CUDA_Execute and CUDA_Global to the list of allowed
pragmas. It also implements basic validation of said pragmas.

gcc/ada/

* aspects.ads: Declare CUDA_Global as aspect.
* einfo.ads: Use Flag118 for the Is_CUDA_Kernel flag.
(Set_Is_CUDA_Kernel): New function.
(Is_CUDA_Kernel): New function.
* einfo.adb (Set_Is_CUDA_Kernel): New function.
(Is_CUDA_Kernel): New function.
* par-prag.adb (Prag): Ignore Pragma_CUDA_Execute and
Pragma_CUDA_global.
* rtsfind.ads: Define CUDA.Driver_Types.Stream_T and
CUDA.Vector_Types.Dim3 entities
* rtsfind.adb: Define CUDA_Descendant subtype.
(Get_Unit_Name): Handle CUDA_Descendant packages.
* sem_prag.ads: Mark CUDA_Global as aspect-specifying pragma.
* sem_prag.adb (Analyze_Pragma): Validate Pragma_CUDA_Execute and
Pragma_CUDA_Global.
* snames.ads-tmpl: Define Name_CUDA_Execute and Name_CUDA_Global.

4 years ago[Ada] Ada2020: AI12-0027 Access values and unaliased component
Arnaud Charlet [Thu, 23 Jul 2020 13:11:56 +0000 (09:11 -0400)]
[Ada] Ada2020: AI12-0027 Access values and unaliased component

Access values should never designate unaliased components.
This new feature is documented in AI12-0027-1.

gcc/ada/

* sem_ch13.ads (Same_Representation): Renamed as
Has_Compatible_Representation because now the order of the arguments
are taken into account; its formals are also renamed as Target_Type
and Operand_Type.
* sem_ch13.adb (Same_Representation): Renamed and moved to place the
routine in alphabetic order.
* sem_attr.adb (Prefix_With_Safe_Accessibility_Level): New subprogram.
(Resolve_Attribute): Check that the prefix of attribute Access
does not have a value conversion of an array type.
* sem_res.adb (Resolve_Actuals): Remove restrictive check on view
conversions which required matching value of Has_Aliased_Components of
formals and actuals.
* exp_ch4.adb (Handle_Changed_Representation): Update call to
Same_Representation.
(Expand_N_Type_Conversion): Update call to Same_Representation.
* exp_ch5.adb (Change_Of_Representation): Update call to
Same_Representation.
* exp_ch6.adb (Add_Call_By_Copy_Code): Update call to
Same_Representation.
(Expand_Actuals): Update call to Same_Representation.
(Expand_Call_Helper): Update call to Same_Representation.

4 years agogcov-tool: Fix merging of different endian coverage data [PR96267]
Dong JianQiang [Tue, 21 Jul 2020 17:24:52 +0000 (01:24 +0800)]
gcov-tool: Fix merging of different endian coverage data [PR96267]

2020-07-21  Dong JianQiang  <dongjianqiang2@huawei.com>

gcc/ChangeLog:

PR gcov-profile/96267
* gcov-io.c (gcov_open): enable if IN_GCOV_TOOL.

4 years agolibomp: Add omp_depend_kind to omp_lib.{f90,h}
Tobias Burnus [Thu, 23 Jul 2020 13:02:15 +0000 (15:02 +0200)]
libomp: Add omp_depend_kind to omp_lib.{f90,h}

gcc/fortran/ChangeLog:

* intrinsic.texi (OMP_LIB_KINDS): Add omp_depend_kind.

libgomp/ChangeLog:

* configure.ac: Add OMP_DEPEND_KIND and OMP_INT128_SIZE.
* libgomp_f.h.in (omp_check_defines): Check whether
sizeof of determined Fortran kind and C typedef match.
* omp_lib.f90.in: Add omp_depened_kind.
* omp_lib.h.in: Likewise; fix omp_alloctrait_key_kind.
* configure: Regenerate.
* Makefile.in: Regenerate.
* testsuite/Makefile.in: Regenerate.

4 years ago[Ada] Add push/pop capability in Output
Arnaud Charlet [Thu, 23 Jul 2020 12:57:22 +0000 (08:57 -0400)]
[Ada] Add push/pop capability in Output

Add the capability to use the Write_* procedures in an environment where
you want to write debugging info but still use them to write to other
files (such a C source files).

gcc/ada/

* output.ads (Push_Output, Pop_Output): New procedures.
* output.adb (FD_Array, FD_Stack, FD_Stack_Idx): New type and vars.
(Push_Output, Pop_Output): New procedures.

4 years agors6000: Rename function adjust_vectorization_cost
Kewen Lin [Thu, 23 Jul 2020 10:01:36 +0000 (05:01 -0500)]
rs6000: Rename function adjust_vectorization_cost

This patch is to rename the existing function adjust_vectorization_cost
to rs6000_adjust_vect_cost_per_stmt, to avoid some confusion.

gcc/ChangeLog:

* config/rs6000/rs6000.c (adjust_vectorization_cost): Renamed to ...
(rs6000_adjust_vect_cost_per_stmt): ... here.
(rs6000_add_stmt_cost): Rename adjust_vectorization_cost to
rs6000_adjust_vect_cost_per_stmt.

4 years agoivopts: Handle vector with length IFNs
Kewen Lin [Thu, 23 Jul 2020 09:27:50 +0000 (04:27 -0500)]
ivopts: Handle vector with length IFNs

This patch is to handle vector with length internal functions
IFN_LEN_LOAD and IFN_LEN_STORE in IVOPTS.

gcc/ChangeLog:

* tree-ssa-loop-ivopts.c (get_mem_type_for_internal_fn): Handle
IFN_LEN_LOAD and IFN_LEN_STORE.
(get_alias_ptr_type_for_ptr_address): Likewise.

4 years agoOpenMP: Update gcc/fortran/*.texi
Tobias Burnus [Thu, 23 Jul 2020 10:55:53 +0000 (12:55 +0200)]
OpenMP: Update gcc/fortran/*.texi

gcc/fortran/ChangeLog:

* gfortran.texi (Standards): Update URL; state that OpenMP 4.5
is supported and 5.0 is partially.
* intrinsic.texi (OpenMP Modules): Refer also to OpenMP 5.0;
(OMP_LIB): Add missing derived type and new named constants.

4 years agogcov: create folders with 0777.
Martin Liska [Fri, 10 Jul 2020 08:56:34 +0000 (10:56 +0200)]
gcov: create folders with 0777.

libgcc/ChangeLog:

PR gcov-profile/96035
* libgcov-driver-system.c: Create gcov folders with 0777
so that it respects a system umask.

4 years agogcc-changelog: fix when somebody reverts a backport
Martin Liska [Thu, 23 Jul 2020 08:39:00 +0000 (10:39 +0200)]
gcc-changelog: fix when somebody reverts a backport

contrib/ChangeLog:

* gcc-changelog/git_commit.py: When reverting a backport,
we should print only Revert header.

4 years agoPR target/96260 - KASAN should work even back-end not porting anything.
Kito Cheng [Wed, 22 Jul 2020 06:50:40 +0000 (14:50 +0800)]
PR target/96260 - KASAN should work even back-end not porting anything.

 - Most KASAN function don't need any porting anything in back-end
   except asan stack protection.

 - However kernel will given shadow offset when enable asan stack
   protection, so eveything in KASAN can work if shadow offset is given.

 - Verified with x86 and risc-v.

 - Verified with RISC-V linux kernel.

gcc/ChangeLog:

PR target/96260
* asan.c (asan_shadow_offset_set_p): New.
* asan.h (asan_shadow_offset_set_p): Ditto.
* toplev.c (process_options): Allow -fsanitize=kernel-address
even TARGET_ASAN_SHADOW_OFFSET not implemented, only check when
asan stack protection is enabled.

gcc/testsuite/ChangeLog:

PR target/96260
* gcc.target/riscv/pr91441.c: Update warning message.
* gcc.target/riscv/pr96260.c: New.

4 years agoc-c++-common/builtin-has-attribute-4.c: Require visibility.
Hans-Peter Nilsson [Thu, 23 Jul 2020 02:27:25 +0000 (04:27 +0200)]
c-c++-common/builtin-has-attribute-4.c: Require visibility.

Another missed attribute-visibility-requirement, causing a failure for
e.g. mmix-knuth-mmixware.  Committed as obvious.

gcc/testsuite:
* c-c++-common/builtin-has-attribute-4.c: Require visibility.

4 years agoDaily bump.
GCC Administrator [Thu, 23 Jul 2020 00:16:28 +0000 (00:16 +0000)]
Daily bump.

4 years agolibstdc++: Add static assertions to futures and promises [LWG 3458]
Jonathan Wakely [Wed, 22 Jul 2020 19:10:38 +0000 (20:10 +0100)]
libstdc++: Add static assertions to futures and promises [LWG 3458]

LWG recently decided it should be ill-formed to instantiate std::future
and std::shared_future for types that can't be returned from a function.
This adds static assertions to enforce it (std::future already failed,
but this makes the error more understandable).

LWG 3466 extends that to std::promise. The actual constraint is that
t.~T() is well-formed for the primary template, but rejecting arrays and
functions as done for futures matches that condition.

libstdc++-v3/ChangeLog:

* include/std/future (future, shared_future, promise): Add
static assertions to the primary template to reject array and
function types.
* testsuite/30_threads/future/requirements/lwg3458.cc: New test.
* testsuite/30_threads/promise/requirements/lwg3466.cc: New test.
* testsuite/30_threads/shared_future/requirements/lwg3458.cc: New test.

4 years agoOpenMP: Support 'if (simd:/cancel:' in Fortran
Tobias Burnus [Wed, 22 Jul 2020 19:02:01 +0000 (21:02 +0200)]
OpenMP: Support 'if (simd:/cancel:' in Fortran

gcc/fortran/ChangeLog:

* gfortran.h (enum gfc_omp_if_kind): Add OMP_IF_CANCEL and OMP_IF_SIMD.
* openmp.c (OMP_SIMD_CLAUSES): Add OMP_CLAUSE_IF.
(gfc_match_omp_clauses, resolve_omp_clauses): Handle 'if (simd/cancel:'.
* dump-parse-tree.c (show_omp_clauses): Likewise.
* trans-openmp.c (gfc_trans_omp_clauses, gfc_trans_omp_cancel,
(gfc_split_omp_clauses): Likewise.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/cancel-2.f90: New test.
* gfortran.dg/gomp/cancel-3.f90: New test.
* gfortran.dg/gomp/if-1.f90: New test.

4 years agors6000: __builtin_mma_disassemble_acc() doesn't store elements correctly in LE mode
Peter Bergner [Wed, 22 Jul 2020 16:44:35 +0000 (11:44 -0500)]
rs6000: __builtin_mma_disassemble_acc() doesn't store elements correctly in LE mode

PR96236 shows a problem where we don't correctly store our 512-bit accumulators
correctly in little-endian mode.  The patch below detects when we're doing a
little-endian memory access and stores to the correct memory locations.

2020-07-22  Peter Bergner  <bergner@linux.ibm.com>

gcc/
PR target/96236
* config/rs6000/rs6000-call.c (rs6000_gimple_fold_mma_builtin): Handle
little-endian memory ordering.

gcc/testsuite/
PR target/96236
* gcc.target/powerpc/mma-double-test.c: Update storing results for
correct little-endian ordering.
* gcc.target/powerpc/mma-single-test.c: Likewise.

4 years agoc++: Don't add enums to class's decl_list
Nathan Sidwell [Wed, 22 Jul 2020 15:38:58 +0000 (08:38 -0700)]
c++: Don't add enums to class's decl_list

We don't need to add CONST_DECLs to a template decl's decl list.  Also made the
code flow a bit clearer.

gcc/cp/
* class.c (maybe_add_class_template_decl_list): Don't add CONST_DECLs.

4 years agodump: filenames containing -
Nathan Sidwell [Wed, 22 Jul 2020 15:34:22 +0000 (08:34 -0700)]
dump: filenames containing -

I discovered the dump machinery would get confused by filenames containing '-'.
Fixed thusly.

gcc/
* dumpfile.c (parse_dump_option): Deal with filenames
containing '-'

4 years agoc++: structural_comptypes addition
Nathan Sidwell [Wed, 22 Jul 2020 15:27:30 +0000 (08:27 -0700)]
c++: structural_comptypes addition

I had to debug structural_comptypes, and its complex if conditions and
tail calling of same_type_p made that hard.  I'd hope we can turn the
eqivalent of return boolean_fn () ? true : false; into a tail call of
the boolean.  We also were not dealing with TYPEOF_TYPE.

gcc/cp/
* typeck.c (structural_comptypes): [DECLTYPE_TYPE] break
apart complex if.
[UNDERLYING_TYPE]: Use an if.
[TYPEOF_TYPE]: New.

4 years agoc++: More cleanups for modern C++
Nathan Sidwell [Wed, 22 Jul 2020 15:20:31 +0000 (08:20 -0700)]
c++: More cleanups for modern C++

Here are some more places where we can declare variables at the
assignment point, rather than use C89.  Also, let's name our variables
by what they contain -- the register allocator is perfectly able to
track liveness for us.

gcc/cp/
* decl.c (decls_match): Move variables into scopes
they're needed in.
(duplicate_decls): Use STRIP_TEMPLATE.
(build_typename_type): Move var decls to their assignments.
(begin_function_body): Likewise.
* decl2.c (get_guard): Likewise.
(mark_used): Use true for truthiness.
* error.c (dump_aggr_type): Hold the decl in a var called
'decl', not 'name'.

4 years agoc++: Shrink lambda-expr
Nathan Sidwell [Wed, 22 Jul 2020 15:07:25 +0000 (08:07 -0700)]
c++: Shrink lambda-expr

I noticed the default capture mode and the discriminator both used
ints.  That seems excessive.  This shrinks them to 8 bits and 16 bits
respectively.  I suppose the discriminator could use the remaining 24
bits of an int allocation unit, if we're worried about more that 64K
lambdas per function.  I know, users are strange :)  On a 64 bit system
this saves 64 bits, because we also had 32 bits of padding added.

gcc/cp/
* cp-tree.h (struct tree_lambda_expr): Shrink
default_capture_mode & discriminator.

4 years agopreprocessor: add_path simplifications
Nathan Sidwell [Wed, 22 Jul 2020 15:01:45 +0000 (08:01 -0700)]
preprocessor: add_path simplifications

I noticed add_path was calling strlen more than once on the same
string.  Let's not do that.

gcc/
* incpath.c (add_path): Avoid multiple strlen calls.

4 years agoc++: mangling cleanups
Nathan Sidwell [Wed, 22 Jul 2020 14:52:31 +0000 (07:52 -0700)]
c++: mangling cleanups

I noticed the mangler's handling of templates could be simplified.
We know template_info is non-null, which is sufficiently boolean --
no need for an explicit bool return.  also some of the internals of
template_args_equal had crept into find_substitution.  Let's not do
that.

gcc/cp/
* mangle.c (decl_is_template_id): Rename to ...
(maybe_template_info): ... here.  Return the template info,
rather than use a pointer.  Adjust all callers.
(find_substitution): Use template_args_equal, rather than
local check.

4 years agoLimit pr95237-6.c to x86 targets
H.J. Lu [Wed, 22 Jul 2020 14:33:54 +0000 (07:33 -0700)]
Limit pr95237-6.c to x86 targets

Since c-c++-common/pr95237-6.c is x86 specific, limit it to x86 targets.

PR target/95237
* c-c++-common/pr95237-6.c: Only run for x86 targets.

4 years agolibstdc++: Constrain reverse_iterator and move_iterator conversions [LWG 3435]
Jonathan Wakely [Wed, 22 Jul 2020 12:25:11 +0000 (13:25 +0100)]
libstdc++: Constrain reverse_iterator and move_iterator conversions [LWG 3435]

libstdc++-v3/ChangeLog:

* include/bits/stl_iterator.h (reverse_iterator): Constrain
converting constructor and converting assignment operator.
Access source iterator's data member directly instead of
calling base().
(move_iterator): Likewise.
* testsuite/24_iterators/move_iterator/dr3435.cc: New test.
* testsuite/24_iterators/reverse_iterator/dr3435.cc: New test.

4 years agoexpmed: Fix possible use of NULL_RTX return value from emit_store_flag
Jozef Lawrynowicz [Wed, 22 Jul 2020 11:37:29 +0000 (12:37 +0100)]
expmed: Fix possible use of NULL_RTX return value from emit_store_flag

MSP430 does not support have any store-flag instructions, so
emit_store_flag can return NULL_RTX.  Catch the NULL_RTX in
expmed.c:expand_sdiv_pow2.

gcc/ChangeLog:

* expmed.c (expand_sdiv_pow2): Check return value from emit_store_flag
is not NULL_RTX before use.

4 years agoexpr: Allow scalar_int_mode target mode when converting a constant
Jozef Lawrynowicz [Wed, 22 Jul 2020 11:34:30 +0000 (12:34 +0100)]
expr: Allow scalar_int_mode target mode when converting a constant

is_int_mode does not allow MODE_PARTIAL_INT modes, so convert_modes was
not allowing a constant value to be converted to a MODE_PARTIAL_INT for
use as operand 2 in patterns such as ashlpsi3. The constant had
to be copied into a register before it could be used, but now can be
used directly as an operand without any copying.

gcc/ChangeLog:

* expr.c (convert_modes): Allow a constant integer to be converted to
any scalar int mode.

4 years agocritical-hint-*.{c,f90}: Move from gcc/testsuite to libgomp/testsuite
Tobias Burnus [Wed, 22 Jul 2020 10:12:48 +0000 (12:12 +0200)]
critical-hint-*.{c,f90}: Move from gcc/testsuite to libgomp/testsuite

libgomp/ChangeLog:

* testsuite/libgomp.c-c++-common/critical-hint-1.c: New; moved from
gcc/testsuite/c-c++-common/gomp/.
* testsuite/libgomp.c-c++-common/critical-hint-2.c: Likewise.
* testsuite/libgomp.fortran/critical-hint-1.f90: New; moved
from gcc/testsuite/gfortran.dg/gomp/.
* testsuite/libgomp.fortran/critical-hint-2.f90: Likewise.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/critical-hint-1.c: Moved to libgomp/.
* c-c++-common/gomp/critical-hint-2.c: Moved to libgomp/.
* gfortran.dg/gomp/critical-hint-1.f90: Moved to libgomp/.
* gfortran.dg/gomp/critical-hint-2.f90: Moved to libgomp/.

4 years agoGeneration of adjusted ldp/stp for vector types
Przemyslaw Wirkus [Wed, 22 Jul 2020 09:14:39 +0000 (10:14 +0100)]
Generation of adjusted ldp/stp for vector types

Introduce simple peephole2 optimization which substitutes a sequence of
four consecutive load or store (LDR, STR) instructions with two load or
store pair (LDP, STP) instructions for 2 element supported vector modes
(V2SI, V2SF, V2DI, and V2DF).
Generated load / store pair instruction offset is adjusted accordingly.

Bootstrapped and tested on aarch64-none-linux-gnu.

Example:
$ cat stp_vec_v2sf.c
typedef float __attribute__((vector_size(8))) vec;

void
store_adjusted(vec *out, vec x, vec y)
{
  out[400] = x;
  out[401] = y;
  out[402] = y;
  out[403] = x;
}

Example compiled with:
$ ./aarch64-none-linux-gnu-gcc -S -O2 stp_vec_v2sf.c -dp

Before the patch:

store_adjusted:
    str     d0, [x0, 3200]    // 9    [c=4 l=4]  *aarch64_simd_movv2si/2
    str     d1, [x0, 3208]    // 11   [c=4 l=4]  *aarch64_simd_movv2si/2
    str     d1, [x0, 3216]    // 13   [c=4 l=4]  *aarch64_simd_movv2si/2
    str     d0, [x0, 3224]    // 15   [c=4 l=4]  *aarch64_simd_movv2si/2
    ret       // 26       [c=0 l=4]  *do_return

After the patch:

store_adjusted:
    add     x1, x0, 3200    // 27   [c=4 l=4]  *adddi3_aarch64/0
    stp     d0, d1, [x1]    // 28   [c=0 l=4]  vec_store_pairv2siv2si
    stp     d1, d0, [x1, 16]        // 29   [c=0 l=4]  vec_store_pairv2siv2si
    ret             // 22   [c=0 l=4]  *do_return

gcc/ChangeLog:

* config/aarch64/aarch64-ldpstp.md: Add two peepholes for adjusted vector
V2SI, V2SF, V2DI, V2DF load pair and store pair modes.
* config/aarch64/aarch64-protos.h (aarch64_gen_adjusted_ldpstp):
Change mode parameter to machine_mode.
(aarch64_operands_adjust_ok_for_ldpstp): Change mode parameter to
machine_mode.
* config/aarch64/aarch64.c (aarch64_operands_adjust_ok_for_ldpstp):
Change mode parameter to machine_mode.
(aarch64_gen_adjusted_ldpstp): Change mode parameter to machine_mode.
* config/aarch64/iterators.md (VP_2E): New iterator for 2 element vectors.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/ldp_vec_v2sf.c: New test.
* gcc.target/aarch64/ldp_vec_v2si.c: New test.
* gcc.target/aarch64/stp_vec_v2df.c: New test.
* gcc.target/aarch64/stp_vec_v2di.c: New test.
* gcc.target/aarch64/stp_vec_v2sf.c: New test.
* gcc.target/aarch64/stp_vec_v2si.c: New test.

4 years agodoc: fix a typo in languages.texi
Wei Wentao [Wed, 22 Jul 2020 09:16:34 +0000 (10:16 +0100)]
doc: fix a typo in languages.texi

gcc/
* doc/languages.texi: Fix “then”/“than” typo.

4 years agoMAINTAINERS: Add myself for write after approval
Przemyslaw Wirkus [Wed, 22 Jul 2020 09:13:00 +0000 (10:13 +0100)]
MAINTAINERS: Add myself for write after approval

2020-07-22  Przemyslaw Wirkus  <przemyslaw.wirkus@arm.com>

* MAINTAINERS (Write After Approval): Add myself.

4 years agoOpenMP: Fixes for omp critical + hint
Tobias Burnus [Tue, 21 Jul 2020 16:29:53 +0000 (18:29 +0200)]
OpenMP: Fixes for omp critical + hint

gcc/c-family/ChangeLog:

* c-omp.c (c_finish_omp_critical): Check for no name but
nonzero hint provided.

gcc/c/ChangeLog:

* c-parser.c (c_parser_omp_clause_hint): Require nonnegative hint clause.
(c_parser_omp_critical): Permit hint(0) clause without named critical.
(c_parser_omp_construct): Don't assert if error_mark_node is returned.

gcc/cp/ChangeLog:

* parser.c (cp_parser_omp_clause_hint): Require nonnegative hint.
(cp_parser_omp_critical): Permit hint(0) clause without named critical.
* pt.c (tsubst_expr): Re-check the latter for templates.

gcc/fortran/ChangeLog:

* openmp.c (gfc_match_omp_critical): Fix handling hints; permit
hint clause without named critical.
(resolve_omp_clauses): Require nonnegative constant integer
for the hint clause.
(gfc_resolve_omp_directive): Check for no name but
nonzero value for hint clause.
* parse.c (parse_omp_structured_block): Fix same-name check
for critical.
* trans-openmp.c (gfc_trans_omp_critical): Handle hint clause properly.

libgomp/ChangeLog:

* omp_lib.f90.in: Add omp_sync_hint_* and omp_sync_hint_kind.
* omp_lib.h.in: Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/gomp/critical-3.C: Add nameless critical with hint testcase.
* c-c++-common/gomp/critical-hint-1.c: New test.
* c-c++-common/gomp/critical-hint-2.c: New test.
* gfortran.dg/gomp/critical-hint-1.f90: New test.
* gfortran.dg/gomp/critical-hint-2.f90: New test.

4 years agogcc.dg/no_profile_instrument_function-attr-1.c: Adjust for NO_DOT_IN_LABEL
Hans-Peter Nilsson [Wed, 22 Jul 2020 01:46:46 +0000 (03:46 +0200)]
gcc.dg/no_profile_instrument_function-attr-1.c: Adjust for NO_DOT_IN_LABEL

mmix-knuth-mmixware is a NO_DOT_IN_LABEL target, so it gets a "_"
instead of the "." in the identifier of interest.  Also tested and
compared to the output for cris-elf which is "regular" regarding
labels: there are no "false positive" identifiers there.  The "." in a
TCL bracket expression matches only a literal ".".

Committed as obvious.

gcc/testsuite:
* gcc.dg/no_profile_instrument_function-attr-1.c: Adjust scanned
regex for NO_DOT_IN_LABEL.

4 years agoDaily bump.
GCC Administrator [Wed, 22 Jul 2020 00:16:27 +0000 (00:16 +0000)]
Daily bump.

4 years agoAdd TARGET_LOWER_LOCAL_DECL_ALIGNMENT [PR95237]
Sunil K Pandey [Sat, 18 Jul 2020 02:42:09 +0000 (19:42 -0700)]
Add TARGET_LOWER_LOCAL_DECL_ALIGNMENT [PR95237]

Default for this hook is NOP. For x86, in 32 bit mode, this hook
sets alignment of long long on stack to 32 bits if preferred stack
boundary is 32 bits.

  - This patch prevents lowering of alignment from following macros.
LOCAL_ALIGNMENT
STACK_SLOT_ALIGNMENT
LOCAL_DECL_ALIGNMENT
  - This patch fixes
gcc.target/i386/pr69454-2.c
gcc.target/i386/stackalign/longlong-1.c
  - Regression test on x86-64, no new fail introduced.

Tested on x86-64.

gcc/c/ChangeLog:

PR target/95237
* c-decl.c (finish_decl): Call target hook
lower_local_decl_alignment to lower local decl alignment.

gcc/ChangeLog:

PR target/95237
* config/i386/i386-protos.h (ix86_local_alignment): Add
another function parameter may_lower alignment. Default is
false.
* config/i386/i386.c (ix86_lower_local_decl_alignment): New
function.
(ix86_local_alignment): Amend ix86_local_alignment to accept
another parameter may_lower. If may_lower is true, new align
may be lower than incoming alignment. If may_lower is false,
new align will be greater or equal to incoming alignment.
(TARGET_LOWER_LOCAL_DECL_ALIGNMENT): Define.
* doc/tm.texi: Regenerate.
* doc/tm.texi.in (TARGET_LOWER_LOCAL_DECL_ALIGNMENT): New
hook.
* target.def (lower_local_decl_alignment): New hook.

gcc/cp/ChangeLog:

PR target/95237
* decl.c (cp_finish_decl): Call target hook
lower_local_decl_alignment to lower local decl alignment.

gcc/testsuite/ChangeLog:

PR target/95237
* c-c++-common/pr95237-1.c: New test.
* c-c++-common/pr95237-2.c: New test.
* c-c++-common/pr95237-3.c: New test.
* c-c++-common/pr95237-4.c: New test.
* c-c++-common/pr95237-5.c: New test.
* c-c++-common/pr95237-6.c: New test.
* c-c++-common/pr95237-7.c: New test.
* c-c++-common/pr95237-8.c: New test.
* c-c++-common/pr95237-9.c: New test.

4 years agors6000: Update test case count when compiling for power9
Peter Bergner [Tue, 21 Jul 2020 21:20:33 +0000 (16:20 -0500)]
rs6000: Update test case count when compiling for power9

2020-07-21  Peter Bergner  <bergner@linux.ibm.com>

gcc/testsuite/
PR target/92488
* gcc.target/powerpc/convert-fp-128.c (bl): Update POWER9 count.

4 years agoc++: Fix scan forward over pragma [PR96257]
Nathan Sidwell [Tue, 21 Jul 2020 19:37:58 +0000 (12:37 -0700)]
c++: Fix scan forward over pragma [PR96257]

It turns out that the paren scanning code is used for speculatively searching
to see if we're looking at a compound_literal.  So we shouldn't always purge
pragma tokens.

gcc/cp/
* parser.c (cp_lexer_consume_token): Drop PRAGMA_EOL assert.
(cp_parser_skip_to_closing_parenthesis_1): Only pass start token
to pragma skipper if recovering.
(cp_parser_skip_to_pragma_eol): Only purge and change pragma
state when recovering.
gcc/testsuite/
* g++.dg/parse/pr96257.C: New.

4 years agoPR fortran/89574 - ICE in conv_function_val, at fortran/trans-expr.c:3792
Harald Anlauf [Tue, 21 Jul 2020 19:37:30 +0000 (21:37 +0200)]
PR fortran/89574 - ICE in conv_function_val, at fortran/trans-expr.c:3792

When checking for an external procedure from the same file, do not
consider symbols from different modules.

gcc/fortran/
PR fortran/89574
* trans-decl.c (gfc_get_extern_function_decl): Check whether a
symbol belongs to a different module.

4 years agoi386: Fix insn conditions of mfence patterns [PR95750]
Uros Bizjak [Tue, 21 Jul 2020 18:22:05 +0000 (20:22 +0200)]
i386: Fix insn conditions of mfence patterns [PR95750]

2020-07-21  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:
PR target/95750
* config/i386/sync.md (mfence_sse2): Enable for
TARGET_64BIT and TARGET_SSE2.
(mfence_nosse): Always enable.

4 years agoMSP430: Remove do_no_relax_short_jumps
Jozef Lawrynowicz [Tue, 21 Jul 2020 16:24:04 +0000 (17:24 +0100)]
MSP430: Remove do_no_relax_short_jumps

This is an old cludge from from when the Binutils linker could not relax
BR to JMP and vice-versa when shuffling "either" sections between upper
and lower memory. This has been fixed since at least Binutils 2.30.

gcc/ChangeLog:

* config/msp430/msp430-protos.h (msp430_do_not_relax_short_jumps):
Remove.
* config/msp430/msp430.c (msp430_do_not_relax_short_jumps): Likewise.
* config/msp430/msp430.md (cbranchhi4_real): Remove special case for
msp430_do_not_relax_short_jumps.

4 years agoMSP430: Define extendqipsi2
Jozef Lawrynowicz [Tue, 21 Jul 2020 16:24:03 +0000 (17:24 +0100)]
MSP430: Define extendqipsi2

The SXT instruction extends the sign of the low byte of the operand
through the entire PSImode register.
SXTX.A can be used to sign extend the low byte of a memory operand
through to the 19th bit. Bits 31:20 are cleared.

gcc/ChangeLog:

* config/msp430/msp430.md: New "extendqipsi2" define_insn.

4 years agoMSP430: Define NO_FUNCTION_CSE
Jozef Lawrynowicz [Tue, 21 Jul 2020 16:24:03 +0000 (17:24 +0100)]
MSP430: Define NO_FUNCTION_CSE

Calling a constant function address costs the same number of clock
cycles as calling an address stored in a register. However, in terms of
instruction length, calling a constant address is more expensive.

Set NO_FUNCTION_CSE to true, only when optimizing for speed.

gcc/ChangeLog:

* config/msp430/msp430.h (NO_FUNCTION_CSE): Set to true at -O2 and
above.

4 years agoScale down long-running tree-prof.exp tests for non-FDO testing.
Sandra Loosemore [Tue, 21 Jul 2020 16:20:36 +0000 (09:20 -0700)]
Scale down long-running tree-prof.exp tests for non-FDO testing.

2020-07-21  Sandra Loosemore  <sandra@codesourcery.com>

gcc/testsuite/
* lib/profopt.exp (auto-profopt-execute): Pass -DFOR_AUTOFDO_TESTING
on command line for both compiles.
* gcc.dg/tree-prof/cold_partition_label.c: Scale down for
non-FDO testing.
* gcc.dg/tree-prof/crossmodule-indir-call-topn-1.c: Likewise.
* gcc.dg/tree-prof/crossmodule-indir-call-topn-2.c: Likewise.
* gcc.dg/tree-prof/indir-call-prof-topn.c: Likewise.
* gcc.dg/tree-prof/section-attr-1.c: Likewise.
* gcc.dg/tree-prof/section-attr-2.c: Likewise.
* gcc.dg/tree-prof/section-attr-3.c: Likewise.

4 years agotestsuite: Add default_packed filters
Dimitar Dimitrov [Thu, 9 Jul 2020 19:29:13 +0000 (22:29 +0300)]
testsuite: Add default_packed filters

Fix test cases assumptions that target has alignment constraints.

gcc/testsuite/ChangeLog:

* gcc.dg/attr-copy-4.c: Unpacked may still have alignment of 1
on targets with default_packed.
* gcc.dg/c11-align-9.c: Remove AVR target filter and replace
with default_packed filter.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
4 years agotestsuite: Relax pattern to include "packed" targets
Dimitar Dimitrov [Thu, 9 Jul 2020 19:29:37 +0000 (22:29 +0300)]
testsuite: Relax pattern to include "packed" targets

The actual warning message depends on the default alignment of the
target. With this update the test correctly passes on AVR and PRU
targets.

gcc/testsuite/ChangeLog:

* gcc.dg/pr53037-1.c: Relax warning pattern.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
4 years agotestsute: Add expected warning for packed attribute
Dimitar Dimitrov [Thu, 9 Jul 2020 19:04:38 +0000 (22:04 +0300)]
testsute: Add expected warning for packed attribute

Targets which pack structures by default get warnings for packed structure
attributes. This is expected, so add markers in the test cases.

gcc/testsuite/ChangeLog:

* c-c++-common/Waddress-of-packed-member-2.c: Add dg-warning for
ignored attribute if target is default_packed.
* c-c++-common/Wattributes.c: Ditto.
* c-c++-common/attr-copy.c: Ditto.
* c-c++-common/builtin-has-attribute-4.c: Ditto.
* c-c++-common/pr51628-29.c: Ditto.
* c-c++-common/pr51628-30.c: Ditto.
* c-c++-common/pr51628-32.c: Ditto.
* gcc.dg/Wattributes-6.c: Ditto.
* gcc.dg/attr-copy-4.c: Ditto.
* gcc.dg/attr-copy-8.c: Ditto.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
4 years agotestsuite: Filter unaligned pointer value warning
Dimitar Dimitrov [Thu, 9 Jul 2020 19:00:46 +0000 (22:00 +0300)]
testsuite: Filter unaligned pointer value warning

Targets which pack structures by default will not get warnings about
unaligned access to structure members.

gcc/testsuite/ChangeLog:

* c-c++-common/Waddress-of-packed-member-1.c: Filter dg-warning
for targets who pack by default.
* c-c++-common/Waddress-of-packed-member-2.c: Ditto.
* c-c++-common/pr51628-13.c: Ditto.
* c-c++-common/pr51628-15.c: Ditto.
* c-c++-common/pr51628-16.c: Ditto.
* c-c++-common/pr51628-26.c: Ditto.
* c-c++-common/pr51628-27.c: Ditto.
* c-c++-common/pr51628-28.c: Ditto.
* c-c++-common/pr51628-29.c: Ditto.
* c-c++-common/pr51628-3.c: Ditto.
* c-c++-common/pr51628-30.c: Ditto.
* c-c++-common/pr51628-31.c: Ditto.
* c-c++-common/pr51628-32.c: Ditto.
* c-c++-common/pr51628-33.c: Ditto.
* c-c++-common/pr51628-35.c: Ditto.
* c-c++-common/pr51628-4.c: Ditto.
* c-c++-common/pr51628-5.c: Ditto.
* c-c++-common/pr51628-6.c: Ditto.
* c-c++-common/pr51628-8.c: Ditto.
* c-c++-common/pr51628-9.c: Ditto.
* c-c++-common/pr88664-2.c: Ditto.
* gcc.dg/pr51628-17.c: Ditto.
* gcc.dg/pr51628-19.c: Ditto.
* gcc.dg/pr51628-20.c: Ditto.
* gcc.dg/pr51628-21.c: Ditto.
* gcc.dg/pr51628-22.c: Ditto.
* gcc.dg/pr51628-24.c: Ditto.
* gcc.dg/pr51628-25.c: Ditto.
* gcc.dg/pr51628-34.c: Ditto.
* gcc.dg/pr88928.c: Ditto.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
4 years agotestsuite: Add signal checking for signal related testcase in analyzer.
Kito Cheng [Mon, 20 Jul 2020 06:41:04 +0000 (14:41 +0800)]
testsuite: Add signal checking for signal related testcase in analyzer.

 - Verifed on RISC-V and x86.

gcc/testsuite/ChangeLog:

* gcc.dg/analyzer/signal-1.c: Add dg-require-effective-target
signal.
* gcc.dg/analyzer/signal-2.c: Ditto.
* gcc.dg/analyzer/signal-3.c: Ditto.
* gcc.dg/analyzer/signal-4a.c: Ditto.
* gcc.dg/analyzer/signal-4b.c: Ditto.
* gcc.dg/analyzer/signal-5.c: Ditto.
* gcc.dg/analyzer/signal-6.c: Ditto.
* gcc.dg/analyzer/signal-exit.c: Ditto.

4 years agors6000: Define movsf_from_si2 to extract high part SF element from DImode[PR89310]
Xionghu Luo [Tue, 21 Jul 2020 03:37:30 +0000 (22:37 -0500)]
rs6000: Define movsf_from_si2 to extract high part SF element from DImode[PR89310]

For extracting high part element from DImode register like:

{%1:SF=unspec[r122:DI>>0x20#0] 86;clobber scratch;}

split it before reload with "and mask" to avoid generating shift right
32 bit then shift left 32 bit.  This pattern also exists in PR42475 and
PR67741, etc.

srdi 3,3,32
sldi 9,3,32
mtvsrd 1,9
xscvspdpn 1,1

=>

rldicr 3,3,0,31
mtvsrd 1,3
xscvspdpn 1,1

Bootstrap and regression tested pass on Power8-LE.

gcc/ChangeLog:

2020-07-21  Xionghu Luo  <luoxhu@linux.ibm.com>

PR rtl-optimization/89310
* config/rs6000/rs6000.md (movsf_from_si2): New define_insn_and_split.

gcc/testsuite/ChangeLog:

2020-07-21  Xionghu Luo  <luoxhu@linux.ibm.com>

PR rtl-optimization/89310
* gcc.target/powerpc/pr89310.c: New test.

4 years agors6000/test: Add vector with length test cases
Kewen Lin [Tue, 21 Jul 2020 03:14:19 +0000 (22:14 -0500)]
rs6000/test: Add vector with length test cases

This patch is to add the test coverage for vector with
length feature on rs6000.  Tested on P9 LE, P7 BE and
P9 BE (aix), the results looked fine.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/p9-vec-length-1.h: New test.
* gcc.target/powerpc/p9-vec-length-2.h: New test.
* gcc.target/powerpc/p9-vec-length-3.h: New test.
* gcc.target/powerpc/p9-vec-length-4.h: New test.
* gcc.target/powerpc/p9-vec-length-5.h: New test.
* gcc.target/powerpc/p9-vec-length-6.h: New test.
* gcc.target/powerpc/p9-vec-length-7.h: New test.
* gcc.target/powerpc/p9-vec-length-8.h: New test.
* gcc.target/powerpc/p9-vec-length-epil-1.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-2.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-3.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-4.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-5.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-6.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-7.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-8.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-1.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-2.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-3.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-4.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-5.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-6.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-7.c: New test.
* gcc.target/powerpc/p9-vec-length-epil-run-8.c: New test.
* gcc.target/powerpc/p9-vec-length-full-1.c: New test.
* gcc.target/powerpc/p9-vec-length-full-2.c: New test.
* gcc.target/powerpc/p9-vec-length-full-3.c: New test.
* gcc.target/powerpc/p9-vec-length-full-4.c: New test.
* gcc.target/powerpc/p9-vec-length-full-5.c: New test.
* gcc.target/powerpc/p9-vec-length-full-6.c: New test.
* gcc.target/powerpc/p9-vec-length-full-7.c: New test.
* gcc.target/powerpc/p9-vec-length-full-8.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-1.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-2.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-3.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-4.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-5.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-6.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-7.c: New test.
* gcc.target/powerpc/p9-vec-length-full-run-8.c: New test.
* gcc.target/powerpc/p9-vec-length-run-1.h: New test.
* gcc.target/powerpc/p9-vec-length-run-2.h: New test.
* gcc.target/powerpc/p9-vec-length-run-3.h: New test.
* gcc.target/powerpc/p9-vec-length-run-4.h: New test.
* gcc.target/powerpc/p9-vec-length-run-5.h: New test.
* gcc.target/powerpc/p9-vec-length-run-6.h: New test.
* gcc.target/powerpc/p9-vec-length-run-7.h: New test.
* gcc.target/powerpc/p9-vec-length-run-8.h: New test.
* gcc.target/powerpc/p9-vec-length.h: New test.

4 years agogcc.dg/independent-cloneids-1.c: Skip for mmix.
Hans-Peter Nilsson [Tue, 21 Jul 2020 00:43:11 +0000 (02:43 +0200)]
gcc.dg/independent-cloneids-1.c: Skip for mmix.

Regular ELF label definitions for this test-case, matched by the
regexps, e.g.:
 /* { dg-final { scan-assembler-times {(?n)^_*bar[.$_]constprop[.$_]0:} 1 } } */
typically look like this:
bar_constprop.0:

For MMIX, they look like this:
bar_constprop::0 IS @

I think it's better to just skip the test for MMIX than further
uglifying the matching regexps, since the test is IIUC general
enough that nothing in the target port can reasonably make a
difference: it passes for all targets or fail for all targets.

gcc/testsuite:
* gcc.dg/independent-cloneids-1.c: Skip for mmix.
flag_stack_usage_info.

4 years agogcc.dg/cdce3.c: Update matched line-number.
Hans-Peter Nilsson [Tue, 21 Jul 2020 00:35:26 +0000 (02:35 +0200)]
gcc.dg/cdce3.c: Update matched line-number.

I missed updating the line-number when adding that dg-skip-if.
Committed as obvious.

* gcc.dg/cdce3.c: Update matched line-number.

4 years agoDaily bump.
GCC Administrator [Tue, 21 Jul 2020 00:16:28 +0000 (00:16 +0000)]
Daily bump.

4 years agommix: support -fstack-usage
Hans-Peter Nilsson [Mon, 20 Jul 2020 23:15:36 +0000 (01:15 +0200)]
mmix: support -fstack-usage

MMIX has two stacks; the regular one using register $254 as a
convention and the register-stack, pushed and popped by call
instructions (usually).  The decision to only report the stack usage
of the regular stack (and not of the register stack) may be updated,
perhaps the sum is better.  This initial decision is helped a little
bit by the order of passes: the size of the register-stack is
calculated only later (in the machine-dependent reorg pass), long
after finalization of the stack-usage info (in the prologue/epilogue
pass).  No regressions for mmix-knuth-mmixware (but a whole lot more
PASSes), committed.

gcc:
* config/mmix/mmix.c (mmix_expand_prologue): Calculate the total
allocated size and set current_function_static_stack_size, if
flag_stack_usage_info.

4 years agolibstdc++: Add std::from_chars for floating-point types
Jonathan Wakely [Mon, 20 Jul 2020 22:49:27 +0000 (23:49 +0100)]
libstdc++: Add std::from_chars for floating-point types

This adds the missing std::from_chars overloads for floating-point
types, as required for C++17 conformance.

The implementation is a hack and not intended to be used in the long
term. Rather than parsing the string directly, this determines the
initial portion of the string that matches the pattern determined by the
chars_format parameter, then creates a NTBS to be parsed by strtod (or
strtold or strtof).

Because creating a NTBS requires allocating memory, but std::from_chars
is noexcept, we need to be careful to minimise allocation. Even after
being careful, allocation failure is still possible, and so a
non-conforming std::no_more_memory error code might be returned.

Because strtod et al depend on the current locale, but std::from_chars
does not, we change the current thread's locale to "C" using newlocale
and uselocale before calling strtod, and restore it afterwards.

Because strtod doesn't have the equivalent of a std::chars_format
parameter, it has to examine the input to determine the format in use,
even though the std::from_chars code has already parsed it once (or
twice for large input strings!)

By replacing the use of strtod we could avoid allocation, avoid changing
locale, and use optimised code paths specific to each std::chars_format
case. We would also get more portable behaviour, rather than depending
on the presence of uselocale, and on any bugs or quirks of the target
libc's strtod. Replacing strtod is a project for a later date.

libstdc++-v3/ChangeLog:

* acinclude.m4 (libtool_VERSION): Bump version.
* config.h.in: Regenerate.
* config/abi/pre/gnu.ver: Add GLIBCXX_3.4.29 version and new
exports.
* config/os/gnu-linux/ldbl-extra.ver: Add _GLIBCXX_LDBL_3.4.29
version and new export.
* configure: Regenerate.
* configure.ac: Check for <xlocale.h> and uselocale.
* crossconfig.m4: Add macro or checks for uselocale.
* include/std/charconv (from_chars): Declare overloads for
float, double, and long double.
* src/c++17/Makefile.am: Add new file.
* src/c++17/Makefile.in: Regenerate.
* src/c++17/floating_from_chars.cc: New file.
(from_chars): Define for float, double, and long double.
* testsuite/20_util/from_chars/1_c++20_neg.cc: Prune extra
diagnostics caused by new overloads.
* testsuite/20_util/from_chars/1_neg.cc: Likewise.
* testsuite/20_util/from_chars/2.cc: Check leading '+'.
* testsuite/20_util/from_chars/4.cc: New test.
* testsuite/20_util/from_chars/5.cc: New test.
* testsuite/util/testsuite_abi.cc: Add new symbol versions.

4 years agoc++: Pseudo-destructor ends object lifetime.
Jason Merrill [Sun, 12 Jul 2020 21:31:24 +0000 (17:31 -0400)]
c++: Pseudo-destructor ends object lifetime.

P0593R6 is mostly about a new object model whereby malloc and the like are
treated as implicitly starting the lifetime of whatever trivial types are
necessary to give the program well-defined semantics; that seems only
relevant to TBAA, and is not implemented here.

The paper also specifies that a pseudo-destructor call (a destructor call
for a non-class type) ends the lifetime of the object like a destructor call
for an object of class type, even though it doesn't call a destructor; this
patch implements that change.

The paper was voted as a DR, so I'm applying this change to all standard
levels.  Like class end-of-life clobbers, it is controlled by
-flifetime-dse.

gcc/cp/ChangeLog:

* pt.c (type_dependent_expression_p): A pseudo-dtor can be
dependent.
* semantics.c (finish_call_expr): Use build_trivial_dtor_call for
pseudo-destructor.
(finish_pseudo_destructor_expr): Leave type NULL for dependent arg.

gcc/testsuite/ChangeLog:

* g++.dg/opt/flifetime-dse7.C: New test.

4 years agoc++: Allow subobject references in C++20.
Jason Merrill [Tue, 7 Jul 2020 05:41:35 +0000 (01:41 -0400)]
c++: Allow subobject references in C++20.

The last new thing allowed by P1907R1: subobject addresses as template
arguments.  The ABI group has discussed mangling for this; there has been
some talk of a compressed subobject mangling, but it hasn't been finalized,
so for now I'm using normal expression mangling.  In order for two array
subobject references to compare as equal template arguments, the offsets
need to have the same type, so I convert them to always be the same type,
currently ptrdiff_t.  Base class conversions are represented as a cast to
reference type, only if necessary to resolve an ambiguity.

This patch also updates the value of __cpp_nontype_template_args, since
the paper is fully implemented.

gcc/cp/ChangeLog:

* mangle.c (write_base_ref): New.
(write_expression): Use it for base field COMPONENT_REFs.
* pt.c (invalid_tparm_referent_p): Canonicalize the type
of array offsets.  Allow subobjects.

gcc/c-family/ChangeLog:

* c-cppbuiltin.c (c_cpp_builtins): Update
__cpp_nontype_template_args for C++20.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/nontype2.C: No error in C++20.
* g++.dg/template/nontype25.C: No error in C++20.
* g++.dg/template/nontype8.C: No error in C++20.
* g++.dg/cpp2a/nontype-subob1.C: New test.
* g++.dg/cpp2a/nontype-subob2.C: New test.
* g++.dg/cpp1z/nontype3.C: Now C++17-only.
* g++.dg/cpp2a/feat-cxx2a.C: Adjust expected value.

4 years agoc++: Aggregate CTAD and string constants.
Jason Merrill [Wed, 8 Jul 2020 05:11:44 +0000 (01:11 -0400)]
c++: Aggregate CTAD and string constants.

In CWG discussion, it was suggested that deduction from a string literal
should be to reference-to-const, so that we deduce 'char' rather than 'const
char' for T.

gcc/cp/ChangeLog:

* pt.c (collect_ctor_idx_types): Add 'const' when deducing from
a string constant.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/class-deduction-aggr7.C: New test.

4 years agosparc/sparc64: use crtendS.o for default-pie executables [PR96190]
Sergei Trofimovich [Tue, 14 Jul 2020 06:48:42 +0000 (07:48 +0100)]
sparc/sparc64: use crtendS.o for default-pie executables [PR96190]

In --enable-default-pie mode compiler should switch from
using crtend.o to crtendS.o. On sparc it is especially important
because crtend.o contains PIC-unfriendly code.

We use GNU_USER_TARGET_ENDFILE_SPEC as a baseline spec to get
crtendS.o instead of crtend.o in !no-pie mode.

gcc:

2020-07-14  Sergei Trofimovich  <siarheit@google.com>

PR target/96190
* config/sparc/linux.h (ENDFILE_SPEC): Use GNU_USER_TARGET_ENDFILE_SPEC
to get crtendS.o for !no-pie mode.
* config/sparc/linux64.h (ENDFILE_SPEC): Ditto.

4 years agolibstdc++: Avoid overflow in istream::get(streambuf&) [LWG 3464]
Jonathan Wakely [Mon, 20 Jul 2020 19:06:46 +0000 (20:06 +0100)]
libstdc++: Avoid overflow in istream::get(streambuf&) [LWG 3464]

Similar to the recent changes to basic_istream::ignore, this change
ensures that _M_gcount doesn't overflow when extracting characters and
inserting them into another streambuf.

The solution used here is to use unsigned long long for the count. We
assume that the number of characters extracted won't exceed the maximum
value for that type, but even if it does we avoid any undefined
behaviour.

libstdc++-v3/ChangeLog:

* include/bits/istream.tcc
(basic_istream::get(__streambuf_type&, char_type): Use unsigned
long long for counter and check if it would overflow _M_gcount.
* testsuite/27_io/basic_istream/get/char/lwg3464.cc: New test.
* testsuite/27_io/basic_istream/get/wchar_t/lwg3464.cc: New test.

4 years agovect: Fix an ICE in vectorizable_simd_clone_call
y00520163 [Mon, 20 Jul 2020 18:47:05 +0000 (19:47 +0100)]
vect: Fix an ICE in vectorizable_simd_clone_call

In vectorizable_simd_clone_call, type compatibility is handled based on
the number of elements and the type compatibility of elements, which is
not enough. This patch add VIEW_CONVERT_EXPRs if the arguments types
and return type of simd clone function are distinct with the vectype of
stmt.

2020-07-20  Yang Yang  <yangyang305@huawei.com>

gcc/ChangeLog:

* tree-vect-stmts.c (vectorizable_simd_clone_call): Add
VIEW_CONVERT_EXPRs if the arguments types and return type
of simd clone function are distinct with the vectype of stmt.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/sve/pr96195.c: New test.

4 years agoi386: Use lock prefixed insn instead of MFENCE [PR95750]
Uros Bizjak [Mon, 20 Jul 2020 18:34:46 +0000 (20:34 +0200)]
i386: Use lock prefixed insn instead of MFENCE [PR95750]

Currently, __atomic_thread_fence(seq_cst) on x86 and x86-64 generates
mfence instruction. A dummy atomic instruction (a lock-prefixed instruction
or xchg with a memory operand) would provide the same sequential consistency
guarantees while being more efficient on most current CPUs. The mfence
instruction additionally orders non-temporal stores, which is not relevant
for atomic operations and are not ordered by seq_cst atomic operations anyway.

2020-07-20  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:
PR target/95750
* config/i386/i386.h (TARGET_AVOID_MFENCE):
Rename from TARGET_USE_XCHG_FOR_ATOMIC_STORE.
* config/i386/sync.md (mfence_sse2): Disable for TARGET_AVOID_MFENCE.
(mfence_nosse): Enable also for TARGET_AVOID_MFENCE. Emit stack
referred memory in word_mode.
(mem_thread_fence): Do not generate mfence_sse2 pattern when
TARGET_AVOID_MFENCE is true.
(atomic_store<mode>): Update for rename.
* config/i386/x86-tune.def (X86_TUNE_AVOID_MFENCE):
Rename from X86_TUNE_USE_XCHG_FOR_ATOMIC_STORE.

gcc/testsuite/ChangeLog:
PR target/95750
* gcc.target/i386/pr95750.c: New test.

4 years agoCorrect handling of constant representations containing embedded nuls.
Martin Sebor [Mon, 20 Jul 2020 18:06:18 +0000 (12:06 -0600)]
Correct handling of constant representations containing embedded nuls.

Resolves:
PR middle-end/95189 - memcmp being wrongly stripped like strcm
PR middle-end/95886 - suboptimal memcpy with embedded zero bytes

gcc/ChangeLog:

PR middle-end/95189
PR middle-end/95886
* builtins.c (inline_expand_builtin_string_cmp): Rename...
(inline_expand_builtin_bytecmp): ...to this.
(builtin_memcpy_read_str): Don't expect data to be nul-terminated.
(expand_builtin_memory_copy_args): Handle object representations
with embedded nul bytes.
(expand_builtin_memcmp): Same.
(expand_builtin_strcmp): Adjust call to naming change.
(expand_builtin_strncmp): Same.
* expr.c (string_constant): Create empty strings with nonzero size.
* fold-const.c (c_getstr): Rename locals and update comments.
* tree.c (build_string): Accept null pointer argument.
(build_string_literal): Same.
* tree.h (build_string): Provide a default.
(build_string_literal): Same.

gcc/testsuite/ChangeLog:

PR middle-end/95189
PR middle-end/95886
* gcc.dg/memcmp-pr95189.c: New test.
* gcc.dg/strncmp-3.c: New test.
* gcc.target/i386/memcpy-pr95886.c: New test.

4 years agoRegenerate .pot files.
Joseph Myers [Mon, 20 Jul 2020 18:05:30 +0000 (18:05 +0000)]
Regenerate .pot files.

gcc/po/
* gcc.pot: Regenerate.

libcpp/po/
* cpplib.pot: Regenerate.

4 years agoRemove stray text from option description (PR c/96249).
Martin Sebor [Mon, 20 Jul 2020 17:56:54 +0000 (11:56 -0600)]
Remove stray text from option description (PR c/96249).

gcc/c-family/ChangeLog:
PR c/96249
* c.opt: Remove stray text.

4 years agojit: Fix truncation of testsuite output [PR69435]
Alex Coplan [Mon, 20 Jul 2020 15:15:29 +0000 (16:15 +0100)]
jit: Fix truncation of testsuite output [PR69435]

This fixes a bug in jit.exp which causes the DejaGnu output of the
libgccjit testsuite to be nondeterministically truncated. This bug was
copied from DejaGnu's own implementation of the host_execute function.
See the upstream bug report [0] where the maintainers point out that the
regex patterns in host_execute should (but don't currently) explicitly
match newlines to avoid relying on DejaGnu not reading more than one
line of the output (which is not guaranteed).

This should make it easier to test jit patches in the future, since it
makes it possible to reliably compare the output of two jit.sum files
(as with the other tests in GCC).

[0] : https://debbugs.gnu.org/cgi/bugreport.cgi?bug=42399

gcc/testsuite/
PR jit/69435
* jit.dg/jit.exp (fixed_host_execute): Fix regex patterns to
always explicitly match newlines.

4 years agoremove write-only array in rev_post_order_and_mark_dfs_back_seme
Richard Biener [Mon, 20 Jul 2020 13:58:19 +0000 (15:58 +0200)]
remove write-only array in rev_post_order_and_mark_dfs_back_seme

This removes a write-only array in
rev_post_order_and_mark_dfs_back_seme.

2020-07-20  Richard Biener  <rguenther@suse.de>

* cfganal.c (rev_post_order_and_mark_dfs_back_seme): Remove
write-only post array.

4 years agopreprocessor: line-map cleanups
Nathan Sidwell [Mon, 20 Jul 2020 12:07:08 +0000 (05:07 -0700)]
preprocessor: line-map cleanups

I found the linemap logic dealing with running out of column numbers
confusing.  There's no need for completely separate code blocks there,
as we can rely on the masking operations working all the way down to
zero bits.  The two binary searches for linemap lookups could do with
modernization of placing the var decls at their initialization point.
(These two searches work in opposite directions, and while lower_bound
would work there, the caching got in the way and I decided to be
conservative.)

libcpp/
* line-map.c (linemap_add): Simplify column overflow calculation.
Add comment about range and column bit init.
(linemap_ordinary_map_lookup): Refactor for RAII
(linemap_macro_map_lookup): Likewise.

4 years agogimple-fold: Handle bitfields in fold_const_aggregate_ref_1 [PR93121]
Jakub Jelinek [Mon, 20 Jul 2020 08:24:19 +0000 (10:24 +0200)]
gimple-fold: Handle bitfields in fold_const_aggregate_ref_1 [PR93121]

When working on __builtin_bit_cast that needs to handle bitfields too,
I've made the following change to handle at least some bitfields in
fold_const_aggregate_ref_1 (those that have integral representative).
It already handles some, but only those that start and end at byte
boundaries.

2020-07-20  Jakub Jelinek  <jakub@redhat.com>

PR libstdc++/93121
* gimple-fold.c (fold_const_aggregate_ref_1): For COMPONENT_REF
of a bitfield not aligned on byte boundaries try to
fold_ctor_reference DECL_BIT_FIELD_REPRESENTATIVE if any and
adjust it depending on endianity.

* gcc.dg/tree-ssa/pr93121-2.c: New test.

4 years agofold-const: Handle bitfields in native_encode_initializer [PR93121]
Jakub Jelinek [Mon, 20 Jul 2020 08:08:28 +0000 (10:08 +0200)]
fold-const: Handle bitfields in native_encode_initializer [PR93121]

When working on __builtin_bit_cast that needs to handle bitfields too,
I've made the following change to handle at least some bitfields in
native_encode_initializer (those that have integral representative).

2020-07-20  Jakub Jelinek  <jakub@redhat.com>

PR libstdc++/93121
* fold-const.c (native_encode_initializer): Handle bit-fields.

* gcc.dg/tree-ssa/pr93121-1.c: New test.

4 years agogcc.dg/const-uniq-1.c: Adjust scanned pattern for mmix.
Hans-Peter Nilsson [Mon, 20 Jul 2020 03:06:49 +0000 (05:06 +0200)]
gcc.dg/const-uniq-1.c: Adjust scanned pattern for mmix.

Apparently local labels end up in the gimple dumps.  For mmix,
local labels that for other targets look like ".LC0" or "LC.0"
instead look like "LC:0".  Committed as obvious.

gcc/testsuite:
* gcc.dg/const-uniq-1.c: Adjust scanned pattern for mmix.

4 years agogcc.dg/cdce3.c: Skip for mmix.
Hans-Peter Nilsson [Mon, 20 Jul 2020 02:34:39 +0000 (04:34 +0200)]
gcc.dg/cdce3.c: Skip for mmix.

The test is gated on effective-target hard_float but what it
really requires is a sqrtf insn (SFmode, not DFmode).  (It
indeed passes for mmix-knuth-mmixware if the sqrtf is changed to
sqrt and float to double; there is a DFmode sqrt insn.)

Committed.

gcc/testsuite:
* gcc.dg/cdce3.c: Skip for mmix.

4 years agovect: Support length-based partial vectors approach
Kewen Lin [Mon, 20 Jul 2020 01:40:10 +0000 (20:40 -0500)]
vect: Support length-based partial vectors approach

Power9 supports vector load/store instruction lxvl/stxvl which allow
us to operate partial vectors with one specific length.  This patch
extends some of current mask-based partial vectors support code for
length-based approach, also adds some length specific support code.
So far it assumes that we can only have one partial vectors approach
at the same time, it will disable to use partial vectors if both
approaches co-exist.

Like the description of optab len_load/len_store, the length-based
approach can have two flavors, one is length in bytes, the other is
length in lanes.  This patch is mainly implemented and tested for
length in bytes, but as Richard S. suggested, most of code has
considered both flavors.

This also introduces one parameter vect-partial-vector-usage allow
users to control when the loop vectorizer considers using partial
vectors as an alternative to falling back to scalar code.

gcc/ChangeLog:

* config/rs6000/rs6000.c (rs6000_option_override_internal):
Set param_vect_partial_vector_usage to 0 explicitly.
* doc/invoke.texi (vect-partial-vector-usage): Document new option.
* optabs-query.c (get_len_load_store_mode): New function.
* optabs-query.h (get_len_load_store_mode): New declare.
* params.opt (vect-partial-vector-usage): New.
* tree-vect-loop-manip.c (vect_set_loop_controls_directly): Add the
handlings for vectorization using length-based partial vectors, call
vect_gen_len for length generation, and rename some variables with
items instead of scalars.
(vect_set_loop_condition_partial_vectors): Add the handlings for
vectorization using length-based partial vectors.
(vect_do_peeling): Allow remaining eiters less than epilogue vf for
LOOP_VINFO_USING_PARTIAL_VECTORS_P.
* tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Init
epil_using_partial_vectors_p.
(_loop_vec_info::~_loop_vec_info): Call release_vec_loop_controls
for lengths destruction.
(vect_verify_loop_lens): New function.
(vect_analyze_loop): Add handlings for epilogue of loop when it's
marked to use vectorization using partial vectors.
(vect_analyze_loop_2): Add the check to allow only one vectorization
approach using partial vectorization at the same time.  Check param
vect-partial-vector-usage for partial vectors decision.  Mark
LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P if the epilogue is
considerable to use partial vectors.  Call release_vec_loop_controls
for lengths destruction.
(vect_estimate_min_profitable_iters): Adjust for loop vectorization
using length-based partial vectors.
(vect_record_loop_mask): Init factor to 1 for vectorization using
mask-based partial vectors.
(vect_record_loop_len): New function.
(vect_get_loop_len): Likewise.
* tree-vect-stmts.c (check_load_store_for_partial_vectors): Add
checks for vectorization using length-based partial vectors.  Factor
some code to lambda function get_valid_nvectors.
(vectorizable_store): Add handlings when using length-based partial
vectors.
(vectorizable_load): Likewise.
(vect_gen_len): New function.
* tree-vectorizer.h (struct rgroup_controls): Add field factor
mainly for length-based partial vectors.
(vec_loop_lens): New typedef.
(_loop_vec_info): Add lens and epil_using_partial_vectors_p.
(LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P): New macro.
(LOOP_VINFO_LENS): Likewise.
(LOOP_VINFO_FULLY_WITH_LENGTH_P): Likewise.
(vect_record_loop_len): New declare.
(vect_get_loop_len): Likewise.
(vect_gen_len): Likewise.

4 years agogcc.dg/pr87485.c: Require scheduling
Hans-Peter Nilsson [Mon, 20 Jul 2020 02:01:43 +0000 (04:01 +0200)]
gcc.dg/pr87485.c: Require scheduling

Committed as obvious, fixing one failure for mmix-knuth-mmixware.

gcc/testsuite:
* gcc.dg/pr87485.c: Require scheduling.

4 years agommix: When debug-dump, revert to "standard" pseudos for emitting integers
Hans-Peter Nilsson [Mon, 20 Jul 2020 01:17:04 +0000 (03:17 +0200)]
mmix: When debug-dump, revert to "standard" pseudos for emitting integers

The sole purpose of not providing pseudos and forcing use of
TARGET_ASM_INTEGER is to arrange for assembly output that people can,
instead of using gas, usefullt feed to mmixal (Knuth's assembler).  It
uses pseudos with slightly different semantics (BYTE, WYDE, TETRA,
OCTA).  Nice when it works, but that only happens for limited
use-cases and debug-information is excluded.

The dwarf2out.c shortcuts, looking up the pseudo-strings and doing the
output on its own, has the result that in absence of such
integer-emitting pseudos, the target TARGET_ASM_INTEGER hook is
called, which by definition outputs a newline, and then the
dwarf2out.c code redundantly adds another.  That uglifies output and
breaks the expectations of the dwarf2 parts of the test-suite.

The wart by which an extra newline is emitted is ripe for a cleanup,
and I started out thinking I'd add a default argument for emitting a
newline to the assemble_integer family of functions, but then
reconsidered.  I also can't bring myself to use different pseudos
with/without -g so instead I check for options for the more detailed
dumps, as used by the test-suite to get those DW_* strings in detailed
assembly output.

This eliminates all FAILs in the dwarf2-specific parts of the gcc
test-suite for mmix-knuth-mmixware, with no regressions.  Committed.

gcc:
* config/mmix/mmix.c (mmix_option_override): Reinstate default
integer-emitting targetm.asm_out pseudos when dumping detailed
assembly-code.
(mmix_assemble_integer): Update comment.

4 years agoDaily bump.
GCC Administrator [Mon, 20 Jul 2020 00:16:19 +0000 (00:16 +0000)]
Daily bump.

4 years agoAlways use name from c_interop_kinds_table for -fc-prototypes.
Thomas Koenig [Sun, 19 Jul 2020 15:27:45 +0000 (17:27 +0200)]
Always use name from c_interop_kinds_table for -fc-prototypes.

When a user specified a KIND that was a parameter taking the value
of an iso_c_binding KIND, the code used the name of that parameter
to look up the type name.  Corrected by always looking it up in
the table of C interop kinds (which was previously done for
non-C-interop types, anyway).

gcc/fortran/ChangeLog:

PR fortran/96220
* dump-parse-tree.c (get_c_type_name): Always use the entries from
c_interop_kinds_table to find the correct C type.

4 years agox86: Add __cpuidex and include guard to <cpuid.h>
H.J. Lu [Sat, 18 Jul 2020 11:43:10 +0000 (04:43 -0700)]
x86: Add __cpuidex and include guard to <cpuid.h>

Add

void __cpuidex (int __cpuid_info[4], int __leaf, int __subleaf);

as well as include guard to <cpuid.h>.

gcc/

PR target/95973
PR target/96238
* config/i386/cpuid.h: Add include guard.
(__cpuidex): New.

gcc/testsuite/

PR target/95973
PR target/96238
* gcc.target/i386/pr95973.c: New test.

4 years agoFix handling of implicit_pure by checking if non-pure procedures are called.
Thomas Koenig [Sun, 19 Jul 2020 10:23:43 +0000 (12:23 +0200)]
Fix handling of implicit_pure by checking if non-pure procedures are called.

Procedures are marked as implicit_pure if they fulfill the criteria of
pure procedures.  In this case, a procedure was not marked as not being
implicit_pure which called another procedure, which had not yet been
marked as not being implicit_impure.

Fixed by iterating over all procedures, setting callers of procedures
which are non-pure and non-implicit_pure as non-implicit_pure and
doing this until no more procedure has been changed.

gcc/fortran/ChangeLog:

2020-07-19  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/96018
* frontend-passes.c (gfc_check_externals): Adjust formatting.
(implicit_pure_call): New function.
(implicit_pure_expr): New function.
(gfc_fix_implicit_pure): New function.
* gfortran.h (gfc_fix_implicit_pure): New prototype.
* parse.c (translate_all_program_units): Call gfc_fix_implicit_pure.

4 years agogcc.dg/attr-copy-6.c: Require visibility.
Hans-Peter Nilsson [Sun, 19 Jul 2020 03:55:18 +0000 (05:55 +0200)]
gcc.dg/attr-copy-6.c: Require visibility.

Another trivial one.

gcc/testsuite:
* gcc.dg/attr-copy-6.c: Require visibility.

4 years agogcc.dg/Wno-frame-address.c: Skip for cris and mmix.
Hans-Peter Nilsson [Sun, 19 Jul 2020 03:18:57 +0000 (05:18 +0200)]
gcc.dg/Wno-frame-address.c: Skip for cris and mmix.

Long-standing FAIL remedied; committed.  Maybe better to list
the targets that *do* support arbitrary frame access?

gcc/testsuite:
* gcc.dg/Wno-frame-address.c: Skip for cris and mmix.

4 years agoDaily bump.
GCC Administrator [Sun, 19 Jul 2020 00:16:22 +0000 (00:16 +0000)]
Daily bump.

4 years agotestsuite: Fix pragma-eof.c failure due to added line to the test.
Jakub Jelinek [Sat, 18 Jul 2020 21:09:56 +0000 (23:09 +0200)]
testsuite: Fix pragma-eof.c failure due to added line to the test.

2020-07-18  Jakub Jelinek  <jakub@redhat.com>

* c-c++-common/cpp/pragma-eof.c: Use .+3 instead of 6 in
dg-error.