platform/upstream/gcc.git
4 years agoUpdate GCC zh_TW.po.
Joseph Myers [Mon, 20 Jan 2020 18:56:48 +0000 (18:56 +0000)]
Update GCC zh_TW.po.

* zh_TW.po: Update.

4 years agoUpdate OpenACC tests for amdgcn
Andrew Stubbs [Thu, 14 Nov 2019 16:16:04 +0000 (16:16 +0000)]
Update OpenACC tests for amdgcn

2020-01-20  Andrew Stubbs  <ams@codesourcery.com>

libgomp/
* testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c: Skip test on gcn.
* testsuite/libgomp.oacc-c-c++-common/loop-dim-default.c (main):
Adjust test dimensions for amdgcn.
* testsuite/libgomp.oacc-c-c++-common/loop-gwv-1.c (main): Adjust
gang/worker/vector expectations dynamically.
* testsuite/libgomp.oacc-c-c++-common/loop-red-gwv-1.c
(main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-v-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-v-2.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-w-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-w-2.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-wv-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-v-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-w-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-wv-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
(acc_gang): Recognise acc_device_radeon.
(acc_worker): Likewise.
(acc_vector): Likewise.
(main): Set expectations for amdgcn.
* testsuite/libgomp.oacc-c-c++-common/routine-gwv-1.c
(main): Adjust gang/worker/vector expectations dynamically.
* testsuite/libgomp.oacc-c-c++-common/routine-v-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-w-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-wv-1.c (main): Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-wv-2.c: Set expectations
for amdgcn.

4 years ago[AArch64] Set jump-align=4 for neoversen1
Wilco Dijkstra [Mon, 20 Jan 2020 14:29:40 +0000 (14:29 +0000)]
[AArch64] Set jump-align=4 for neoversen1

Testing shows the setting of 32:16 for jump alignment has a significant
codesize cost, however it doesn't make a difference in performance.
So set jump-align to 4 to get 1.6% codesize improvement.

gcc/
* config/aarch64/aarch64.c (neoversen1_tunings): Set jump_align to 4.

4 years agoFix PR 93242: patchable-function-entry broken on MIPS
Andrew Pinski [Sat, 18 Jan 2020 00:41:06 +0000 (00:41 +0000)]
Fix PR 93242: patchable-function-entry broken on MIPS

On MIPS, .set noreorder/reorder needs to emitted around
the nop.  The template for the nop instruction uses %(/%) to
do that.  But default_print_patchable_function_entry uses
fprintf rather than output_asm_insn to output the instruction.

This fixes the problem by using output_asm_insn to emit the nop
instruction.

ChangeLog:

PR middle-end/93242
* targhooks.c (default_print_patchable_function_entry): Use
output_asm_insn to emit the nop instruction.

4 years agoPR testsuite/92829 - several -Wstringop-overflow test case failures on powerpc64
Martin Sebor [Mon, 20 Jan 2020 13:53:33 +0000 (14:53 +0100)]
PR testsuite/92829 - several -Wstringop-overflow test case failures on powerpc64

* g++.dg/warn/Wstringop-overflow-4.C: Adjust test to avoid failures
due to an aparrent VRP limtation.
* gcc.dg/Wstringop-overflow-25.c: Same.

4 years agoAlign __patchable_function_entries to POINTER_SIZE [PR93194]
Fangrui Song [Wed, 8 Jan 2020 04:46:26 +0000 (20:46 -0800)]
Align __patchable_function_entries to POINTER_SIZE [PR93194]

2020-01-20  Fangrui Song  <maskray@google.com>

gcc/
PR middle-end/93194
* targhooks.c (default_print_patchable_function_entry): Align to
POINTER_SIZE.

4 years ago[PR 80005] Fix __has_include
Nathan Sidwell [Mon, 20 Jan 2020 13:39:59 +0000 (05:39 -0800)]
[PR 80005]  Fix __has_include

__has_include is funky in that it is macro-like from the POV of #ifdef and
friends, but lexes its parenthesize argument #include-like.  We were
failing the second part of that, because we used a forwarding macro to an
internal name, and hence always lexed the argument in macro-parameter
context.  We componded that by not setting the right flag when lexing, so
it didn't even know.  Mostly users got lucky.

This reimplements the handline.
1) Remove the forwarding, but declare object-like macros that
expand to themselves.  This satisfies the #ifdef requirement

2) Correctly set angled_brackets when lexing the parameter.  This tells
the lexer (a) <...> is a header name and (b) "..." is too (not a string).

3) Remove the in__has_include lexer state, just tell find_file that that's
what's happenning, so it doesn't emit an error.

We lose the (undocumented) ability to #undef __has_include.  That may well
have been an accident of implementation.  There are no tests for it.

We gain __has_include behaviour for all users of the preprocessors -- not
just the C-family ones that defined a forwarding macro.

libcpp/
PR preprocessor/80005
* include/cpplib.h (BT_HAS_ATTRIBUTE): Fix comment.
* internal.h (struct lexer_state): Delete in__has_include field.
(struct spec_nodes): Rename n__has_include{,_next}__ fields.
(_cpp_defined_macro_p): New.
(_cpp_find_file): Add has_include parm.
* directives.c (lex_macro_node): Combine defined,
__has_inline{,_next} checking.
(do_ifdef, do_ifndef): Use _cpp_defined_macro_p.
(_cpp_init_directives): Refactor.
* expr.c (parse_defined): Use _cpp_defined_macro_p.
(eval_token): Adjust parse_has_include calls.
(parse_has_include): Add OP parameter.  Reimplement.
* files.c (_cpp_find_file): Add HAS_INCLUDE parm.  Use it to
inhibit error message.
(_cpp_stack_include): Adjust _cpp_find_file call.
(_cpp_fake_include, _cpp_compare_file_date): Likewise.
(open_file_failed): Remove in__has_include check.
(_cpp_has_header): Adjust _cpp_find_file call.
* identifiers.c (_cpp_init_hashtable): Don't init
__has_include{,_next} here ...
* init.c (cpp_init_builtins): ... init them here.  Define as
macros.
(cpp_read_main_file): Adjust _cpp_find_file call.
* pch.c (cpp_read_state): Adjust __has_include{,_next} access.
* traditional.c (_cpp_scan_out_locgical_line): Likewise.

gcc/c-family/
PR preprocessor/80005
* c-cppbuiltins.c (c_cpp_builtins): Don't define __has_include{,_next}.

gcc/testsuite/
PR preprocessor/80005
* g++.dg/cpp1y/feat-cxx14.C: Adjust.
* g++.dg/cpp1z/feat-cxx17.C: Adjust.
* g++.dg/cpp2a/feat-cxx2a.C: Adjust.
* g++.dg/cpp/pr80005.C: New.

4 years ago[PATCH] PR Fortran/93263 Correct test case
Mark Eggleston [Mon, 20 Jan 2020 13:23:07 +0000 (13:23 +0000)]
[PATCH] PR Fortran/93263 Correct test case

Should've have checked for the existance of a non static integer
using scan-tree-dump instead of scan-tree-dump-not. A cut and paste
error.

4 years agox32: Add x32 support to -mtls-dialect=gnu2
H.J. Lu [Mon, 20 Jan 2020 13:02:14 +0000 (05:02 -0800)]
x32: Add x32 support to -mtls-dialect=gnu2

To add x32 support to -mtls-dialect=gnu2, we need to replace DI with
P in GNU2 TLS patterns.  Since DEST set by tls_dynamic_gnu2_64 is in
ptr_mode, PLUS in GNU2 TLS address computation must be done in ptr_mode
to support -maddress-mode=long.  Also replace the "{q}" suffix on lea
with "%z0" to support both 32-bit and 64-bit destination register.

Tested on Linux/x86-64.

gcc/

PR target/93319
* config/i386/i386.c (legitimize_tls_address): Pass Pmode to
gen_tls_dynamic_gnu2_64.  Compute GNU2 TLS address in ptr_mode.
* config/i386/i386.md (tls_dynamic_gnu2_64): Renamed to ...
(@tls_dynamic_gnu2_64_<mode>): This.  Replace DI with P.
(*tls_dynamic_gnu2_lea_64): Renamed to ...
(*tls_dynamic_gnu2_lea_64_<mode>): This.  Replace DI with P.
Remove the {q} suffix from lea.
(*tls_dynamic_gnu2_call_64): Renamed to ...
(*tls_dynamic_gnu2_call_64_<mode>): This.  Replace DI with P.
(*tls_dynamic_gnu2_combine_64): Renamed to ...
(*tls_dynamic_gnu2_combine_64_<mode>): This.  Replace DI with P.
Pass Pmode to gen_tls_dynamic_gnu2_64.

gcc/testsuite/

PR target/93319
* gcc.target/i386/pr93319-1a.c: New test.
* gcc.target/i386/pr93319-1b.c: Likewise.
* gcc.target/i386/pr93319-1c.c: Likewise.
* gcc.target/i386/pr93319-1d.c: Likewise.

4 years ago[AArch64] Set SLOW_BYTE_ACCESS
Wilco Dijkstra [Mon, 20 Jan 2020 13:09:15 +0000 (13:09 +0000)]
[AArch64] Set SLOW_BYTE_ACCESS

Contrary to all documentation, SLOW_BYTE_ACCESS simply means accessing
bitfields by their declared type, which results in better codegeneration.

gcc/
* config/aarch64/aarch64.h (SLOW_BYTE_ACCESS): Set to 1.

4 years agoaarch64: Remove parameter name and ATTRIBUTE_UNUSED
Richard Sandiford [Sat, 18 Jan 2020 14:49:39 +0000 (14:49 +0000)]
aarch64: Remove parameter name and ATTRIBUTE_UNUSED

2020-01-20  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-sve-builtins-base.cc
(svld1ro_impl::memory_vector_mode): Remove parameter name.

4 years agodebug/92763 keep DIEs that might be used in DW_TAG_inlined_subroutine
Richard Biener [Mon, 20 Jan 2020 09:36:09 +0000 (10:36 +0100)]
debug/92763 keep DIEs that might be used in DW_TAG_inlined_subroutine

We were pruning type-local subroutine DIEs if their context is unused
despite us later needing those DIEs as abstract origins for inlines.
The patch makes code already present for -fvar-tracking-assignments
unconditional.

2020-01-20  Richard Biener  <rguenther@suse.de>

PR debug/92763
* dwarf2out.c (prune_unused_types): Unconditionally mark
called function DIEs.

* g++.dg/debug/pr92763.C: New testcase.

4 years agocontrib: New remotes structure for vendor and personal refs
Richard Earnshaw [Mon, 20 Jan 2020 10:37:29 +0000 (10:37 +0000)]
contrib: New remotes structure for vendor and personal refs

The initial structure for vendor and personal branches makes use of
the default remote (normally origin) for the upstream
repository).  Unfortunately, this causes some confusion, especially for
personal branches because a push will not push to the correct upstream
location.  This can be 'fixed' by adding a push refspec for the remote,
but that has the unfortunate consequence of breaking the push.default
behaviour for git push, and it becomes too easy to accidentally commit
something unintended to the main parts of the repository.

To work around this, this patch changes the configuration to use
separate 'remotes' for these additional refs, with one remote for the
personal space and another remote for each vendor's space.  The
personal space is called after the user's preferred branch-space
prefix (default 'me'), the vendor spaces are called
vendors/<vendor-name>.

As far as possible, I've made the script automatically restructure any
existing fetch or push lines that earlier versions of the scripts may
have created - the gcc-git-customization.sh script will convert all
vendor refs that it can find, so it is not necessary to re-add any
vendors you've already added.

You might, however, want to run
  git remote prune <origin>
after running to clean up any stale upstream-refs that might still be
in your local repo, and then
  git fetch vendors/<vendor>
or
  git fetch <me>
to re-populate the remotes/ structures.

Also, for any branch you already have that tracks a personal or vendor
branch upstream, you might need to run
  git config branch.<name>.remote <new-remote>

so that merges and pushes go to the right place (I haven't attempted
to automate this last part).

For vendors, the new structure means that

  git checkout -b <vendor>/<branch> remotes/vendors/<vendor>/<branch>

will correctly set up a remote tracking branch.

Please be aware that if you have multiple personal branches set up, then

  git push <me>

will still consider all of them for pushing.  If you only want to push
one branch, then either write
  git push <me> HEAD
or
  git push <me> <me>/branch
as appropriate.

And don't forget '-n' (--dry-run) to see what would be done if this
were not a dry run.

Finally, now that the vendors spaces are isolated from each other and
from the other spaces, I've added an option "--enable-push" to
git-fetch-vendor.sh.  If passed, then a "push" spec will be added for
that vendor to enable pushing to the upstream.  If you re-run the
script for the same vendor without the option, the push spec will be
removed.

* gcc-git-customization.sh: Check that user-supplied remote
name exists before continuting.  Use a separate remotes for the
personal commit area.  Convert existing personal and vendor
fetch rules to new layout.
* git-fetch-vendor.sh: New vendor layout.  Add --enable-push
option.

4 years agoAdd testcase of PR c++/92536, already fixed.
Paolo Carlini [Mon, 20 Jan 2020 10:17:19 +0000 (11:17 +0100)]
Add testcase of PR c++/92536, already fixed.

PR c++/92536
* g++.dg/cpp1z/pr92536.C: New.

4 years agoRecord outer non-cleanup region in TREE EH.
Martin Liska [Mon, 20 Jan 2020 10:10:30 +0000 (11:10 +0100)]
Record outer non-cleanup region in TREE EH.

PR tree-optimization/93199
* tree-eh.c (struct leh_state): Add
new field outer_non_cleanup.
(cleanup_is_dead_in): Pass leh_state instead
of eh_region.  Add a checking that state->outer_non_cleanup
points to outer non-clean up region.
(lower_try_finally): Record outer_non_cleanup
for this_state.
(lower_catch): Likewise.
(lower_eh_filter): Likewise.
(lower_eh_must_not_throw): Likewise.
(lower_cleanup): Likewise.

4 years agotree-optimization/93094 pass down VECTORIZED_CALL to versioning
Richard Biener [Wed, 15 Jan 2020 13:25:38 +0000 (14:25 +0100)]
tree-optimization/93094 pass down VECTORIZED_CALL to versioning

When versioning is run the IL is already mangled and finding
a VECTORIZED_CALL IFN can fail.

2020-01-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/93094
* tree-vectorizer.h (vect_loop_versioning): Adjust.
(vect_transform_loop): Likewise.
* tree-vectorizer.c (try_vectorize_loop_1): Pass down
loop_vectorized_call to vect_transform_loop.
* tree-vect-loop.c (vect_transform_loop): Pass down
loop_vectorized_call to vect_loop_versioning.
* tree-vect-loop-manip.c (vect_loop_versioning): Use
the earlier discovered loop_vectorized_call.

* gcc.dg/vect/pr93094.c: New testcase.

4 years agoClean up references to Subversion in documentation sources.
Eric S. Raymond [Mon, 20 Jan 2020 01:12:29 +0000 (17:12 -0800)]
Clean up references to Subversion in documentation sources.

Clean up references to SVN in in the GCC docs, redirecting to Git
documentation as appropriate.

Where references to "the source code repository" rather than a
specific VCS make sense, I have used them. You might, after
all, change VCSes again someday.

I have not modified either generated HTML files nor maintainer scripts.
These changes should be complete with repect to the documentation tree.

2020-01-19  Eric S. Raymond <esr@thyrsus.com>

gcc/
* doc/contribute.texi: Update for SVN -> Git transition.
* doc/install.texi: Likewise.

libstdc++-v3
* doc/xml/faq.xml: Update for SVN -> Git transition.
* doc/xml/manual/appendix_contributing.xml: Likewise.
* doc/xml/manual/status_cxx1998.xml: Likewise.
* doc/xml/manual/status_cxx2011.xml: Likewise.
* doc/xml/manual/status_cxx2014.xml: Likewise.
* doc/xml/manual/status_cxx2017.xml: Likewise.
* doc/xml/manual/status_cxx2020.xml: Likewise.
* doc/xml/manual/status_cxxtr1.xml: Likewise.
* doc/xml/manual/status_cxxtr24733.xml: Likewise.

4 years agoanalyzer: ensure that all DejaGnu tests have unique names
David Malcolm [Fri, 17 Jan 2020 18:12:25 +0000 (13:12 -0500)]
analyzer: ensure that all DejaGnu tests have unique names

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/CVE-2005-1689-dedupe-issue.c: Ensure that all
test names are unique.
* gcc.dg/analyzer/attribute-nonnull.c: Likewise.
* gcc.dg/analyzer/conditionals-notrans.c: Likewise.
* gcc.dg/analyzer/data-model-1.c: Likewise.
* gcc.dg/analyzer/data-model-18.c: Likewise.
* gcc.dg/analyzer/data-model-8.c: Likewise.
* gcc.dg/analyzer/data-model-9.c: Likewise.
* gcc.dg/analyzer/file-1.c: Likewise.
* gcc.dg/analyzer/file-paths-1.c: Likewise.
* gcc.dg/analyzer/loop-2.c: Likewise.
* gcc.dg/analyzer/loop-2a.c: Likewise.
* gcc.dg/analyzer/loop-4.c: Likewise.
* gcc.dg/analyzer/loop.c: Likewise.
* gcc.dg/analyzer/malloc-1.c: Likewise.
* gcc.dg/analyzer/malloc-2.c: Likewise.
* gcc.dg/analyzer/malloc-ipa-10.c: Likewise.
* gcc.dg/analyzer/malloc-ipa-2.c: Likewise.
* gcc.dg/analyzer/malloc-macro-separate-events.c: Likewise.
* gcc.dg/analyzer/malloc-paths-1.c: Likewise.
* gcc.dg/analyzer/malloc-paths-2.c: Likewise.
* gcc.dg/analyzer/malloc-paths-3.c: Likewise.
* gcc.dg/analyzer/malloc-paths-4.c: Likewise.
* gcc.dg/analyzer/malloc-paths-5.c: Likewise.
* gcc.dg/analyzer/malloc-paths-7.c: Likewise.
* gcc.dg/analyzer/malloc-paths-9.c: Likewise.
* gcc.dg/analyzer/operations.c: Likewise.
* gcc.dg/analyzer/params.c: Likewise.
* gcc.dg/analyzer/pattern-test-1.c: Likewise.
* gcc.dg/analyzer/pattern-test-2.c: Likewise.
* gcc.dg/analyzer/sensitive-1.c: Likewise.
* gcc.dg/analyzer/switch.c: Likewise.
* gcc.dg/analyzer/taint-1.c: Likewise.
* gcc.dg/analyzer/unknown-fns.c: Likewise.

4 years agoDaily bump.
GCC Administrator [Mon, 20 Jan 2020 00:16:31 +0000 (00:16 +0000)]
Daily bump.

4 years agotestsuite: effective_target_march_option: support checking for -march=*
Hans-Peter Nilsson [Sun, 19 Jan 2020 19:38:06 +0000 (20:38 +0100)]
testsuite: effective_target_march_option: support checking for -march=*

* lib/target-supports.exp (effective_target_march_option): New.

I see no (other) way to, depending on the absence of an option,
add an option for a specific target.

For gcc.dg/torture/pr26515.c and cris-elf, you get an error for
supplying multiple (different) -march=... options (where that
error is desirable), like testing cris-elf with
RUNTESTFLAGS=--target_board=cris-sim/arch=v8, where otherwise
-march=v10 and -march=v8 will both be given, and the test would
fail.

For historians, this was accidentally misordered and committed after
the (first) patch using march_option.  Oops.

4 years agotestsuite: Make use of effective-target march_option for cris
Hans-Peter Nilsson [Sun, 19 Jan 2020 19:29:52 +0000 (20:29 +0100)]
testsuite: Make use of effective-target march_option for cris

* gcc.dg/torture/pr26515.c (cris*-*-*): Conditionalize
-march=v10 option on target ! march_option.
* gcc.target/cris/asm-v10.S, gcc.target/cris/inasm-v10.c,
gcc.target/cris/sync-1-v10.c: Similar.

4 years agoPR c++/33799 - destroy return value, take 2.
Jason Merrill [Sun, 19 Jan 2020 14:14:54 +0000 (09:14 -0500)]
PR c++/33799 - destroy return value, take 2.

This patch differs from the reverted patch for 33799 in that it adds the
CLEANUP_STMT for the return value at the end of the function, and only if
we've seen a cleanup that might throw, so it should not affect most C++11
code.

* cp-tree.h (current_retval_sentinel): New macro.
(struct language_function): Add throwing_cleanup bitfield.
* decl.c (cxx_maybe_build_cleanup): Set it.
* except.c (maybe_set_retval_sentinel)
(maybe_splice_retval_cleanup): New functions.
* parser.c (cp_parser_compound_statement): Call
maybe_splice_retval_cleanup.
* typeck.c (check_return_expr): Call maybe_set_retval_sentinel.

4 years agoSimplify lambda parsing.
Jason Merrill [Sun, 19 Jan 2020 14:15:26 +0000 (09:15 -0500)]
Simplify lambda parsing.

Since we removed the special parsing for C++11 lambdas, it's just been an
open-coded copy of cp_parser_function_body.  So let's call it instead.  This
avoids the need to change this code in my revised 33799 patch.

* parser.c (cp_parser_lambda_body): Use cp_parser_function_body.

4 years agoImplement speculative call verifier
Jan Hubicka [Sun, 19 Jan 2020 15:41:11 +0000 (16:41 +0100)]
Implement speculative call verifier

this patch implements verifier and fixes one bug where speculative calls
produced by ipa-devirt ended up having num_speculative_call_targets = 0
instead of 1.

* cgraph.c (cgraph_edge::make_speculative): Increase number of
speculative targets.
(verify_speculative_call): New function
(cgraph_node::verify_node): Use it.
* ipa-profile.c (ipa_profile): Fix formating; do not set number of
speculations.

4 years agoFix ICE in speculative_call_info
Jan Hubicka [Sun, 19 Jan 2020 12:49:38 +0000 (13:49 +0100)]
Fix ICE in speculative_call_info

this fixes two issues with the new multi-target speculation code which reproduce
on Firefox.  I can now build firefox with FDO locally but on Mozilla build bots
it still fails with ICE in speculative_call_info.

One problem is that speuclative code compares call_stmt and lto_stmt_uid in
a way that may get unwanted effect when these gets out of sync.  It does not
make sense to have both non-zero so I added code clearing it and sanity check
that it is kept this way.

Other problem is cgraph_edge::make_direct not working well with multiple
targets.  In this case it removed one speuclative target and the indirect call
leaving other targets in the tree.

This is fixed by iterating across all targets and removing all except the good
one (if it exists).

PR lto/93318
* cgraph.c (cgraph_edge::resolve_speculation): Fix foramting.
(cgraph_edge::make_direct): Remove all indirect targets.
(cgraph_edge::redirect_call_stmt_to_callee): Use make_direct..
(cgraph_node::verify_node): Verify that only one call_stmt or
lto_stmt_uid is set.
* cgraphclones.c (cgraph_edge::clone): Set only one call_stmt or
lto_stmt_uid.
* lto-cgraph.c (lto_output_edge): Simplify streaming of stmt.
(lto_output_ref): Simplify streaming of stmt.
* lto-streamer-in.c (fixup_call_stmt_edges_1): Clear lto_stmt_uid.

4 years agoCheck for illegal reference in function.
Thomas König [Thu, 16 Jan 2020 21:09:37 +0000 (22:09 +0100)]
Check for illegal reference in function.

4 years agoDaily bump.
GCC Administrator [Sun, 19 Jan 2020 00:16:30 +0000 (00:16 +0000)]
Daily bump.

4 years agoc++: Fix coroutines.cc build for nvptx-none target
Jakub Jelinek [Sat, 18 Jan 2020 18:22:21 +0000 (19:22 +0100)]
c++: Fix coroutines.cc build for nvptx-none target

When building offloading cross-compiler from x86_64-linux to nvptx-none, the
build fails with:
../../gcc/cp/coroutines.cc: In function 'tree_node* get_fn_local_identifier(tree, const char*)':
../../gcc/cp/coroutines.cc:2255:12: error: expected ';' before 'char'
 2255 |   sep = "$"
      |            ^
      |            ;
......
 2262 |   char *an;
      |   ~~~~

2020-01-18  Jakub Jelinek  <jakub@redhat.com>

* coroutines.cc (get_fn_local_identifier): Fix NO_DOT_IN_LABEL
but non-NO_DOLLAR_IN_LABEL case build.

4 years agolibgcc: cris: config/cris/arit.c (DS): Apply attribute __fallthrough__.
Hans-Peter Nilsson [Sat, 18 Jan 2020 15:31:38 +0000 (16:31 +0100)]
libgcc: cris: config/cris/arit.c (DS): Apply attribute __fallthrough__.

* config/cris/arit.c (DS): Apply attribute fallthrough.

Without this, there are, for each compilation of arit.c, 30ish
occurrences of "this statement may fall through
[-Wimplicit-fallthrough=]", for lines that look like
    case 32: DS; case 31: DS; case 30: DS; case 29: DS;

4 years agoFix compilation of crtstuff.c when DEFAULT_USE_CXA_ATEXIT is false
John David Anglin [Sat, 18 Jan 2020 15:30:29 +0000 (10:30 -0500)]
Fix compilation of crtstuff.c when DEFAULT_USE_CXA_ATEXIT is false

PR libgcc/92988
* crtstuff.c (__do_global_dtors_aux): Only call __cxa_finalize if
DEFAULT_USE_CXA_ATEXIT is true.

4 years agoAArch64: Fix unused variable warning breaking bootstrap.
Tamar Christina [Sat, 18 Jan 2020 14:03:31 +0000 (14:03 +0000)]
AArch64: Fix unused variable warning breaking bootstrap.

This marks the parameter &fi as unused so it doesn't
cause a boostrap failure.

committed under the obvious rule.

gcc/ChangeLog:

* config/aarch64/aarch64-sve-builtins-base.cc (memory_vector_mode):
Mark parameter unused.

4 years agotestsuite: Fix dg-message syntax.
Jakub Jelinek [Sat, 18 Jan 2020 12:54:08 +0000 (13:54 +0100)]
testsuite: Fix dg-message syntax.

2020-01-18  Jakub Jelinek  <jakub@redhat.com>

PR c/92833
* c-c++-common/pr92833-4.c: Fix dg-message syntax.

4 years agoconfig.gcc <obsolete targets>: Add crisv32-*-* and cris-*-linux*
Hans-Peter Nilsson [Sat, 18 Jan 2020 13:23:59 +0000 (14:23 +0100)]
config.gcc <obsolete targets>: Add crisv32-*-* and cris-*-linux*

I'm sorry to say that there's no incentive to maintain
crisv32-*-* and cris-*-linux* configurations beyond nostalgia,
(and I'm out of that for the moment).  Support in the Linux
kernel for either applicable CRIS variant (CRIS v10 and CRIS
v32) is gone since 2018.  Their related part of the cc0
transition workload would be noticable.  Note that cris-elf
remains, but crisv32-elf and the CRIS v32 multilib will be
removed, at least for now.

I'm not completely happy about the message (the next-next line
after the context) "*** unless a maintainer comes forward"
because it'd have to be at an infinitesimal maintenance cost to
the cris-elf support.  Still, I'm not bothered enough to add
another case construct or means for "planned obsolescence".

4 years agoReword a comment in varpool_node::ctor_useable_for_folding_p.
Gerald Pfeifer [Sat, 18 Jan 2020 12:49:05 +0000 (13:49 +0100)]
Reword a comment in varpool_node::ctor_useable_for_folding_p.

* varpool.c (ctor_useable_for_folding_p): Fix grammar.

4 years ago[libiberty] Update demangler to handle co_await operator.
Iain Sandoe [Wed, 4 Dec 2019 11:39:38 +0000 (11:39 +0000)]
[libiberty] Update demangler to handle co_await operator.

C++20 coroutines introduces a new operator with a mangling of 'aw'.
This patch adds that to libiberty's demangler.

libiberty/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* cp-demangle.c (cplus_demangle_operators): Add the co_await
operator.
* testsuite/demangle-expected: Test co_await operator mangling.

4 years ago[C++ coroutines] Initial implementation.
Iain Sandoe [Sat, 18 Jan 2020 11:54:46 +0000 (11:54 +0000)]
[C++ coroutines] Initial implementation.

This is the squashed version of the first 6 patches that were split to
facilitate review.

The changes to libiberty (7th patch) to support demangling the co_await
operator stand alone and are applied separately.

The patch series is an initial implementation of a coroutine feature,
expected to be standardised in C++20.

Standardisation status (and potential impact on this implementation)
--------------------------------------------------------------------

The facility was accepted into the working draft for C++20 by WG21 in
February 2019.  During following WG21 meetings, design and national body
comments have been reviewed, with no significant change resulting.

The current GCC implementation is against n4835 [1].

At this stage, the remaining potential for change comes from:

* Areas of national body comments that were not resolved in the version we
  have worked to:
  (a) handling of the situation where aligned allocation is available.
  (b) handling of the situation where a user wants coroutines, but does not
      want exceptions (e.g. a GPU).

* Agreed changes that have not yet been worded in a draft standard that we
  have worked to.

It is not expected that the resolution to these can produce any major
change at this phase of the standardisation process.  Such changes should be
limited to the coroutine-specific code.

ABI
---

The various compiler developers 'vendors' have discussed a minimal ABI to
allow one implementation to call coroutines compiled by another.

This amounts to:

1. The layout of a public portion of the coroutine frame.

 Coroutines need to preserve state across suspension points, the storage for
 this is called a "coroutine frame".

 The ABI mandates that pointers into the coroutine frame point to an area
 begining with two function pointers (to the resume and destroy functions
 described below); these are immediately followed by the "promise object"
 described in the standard.

 This is sufficient that the builtins can take a coroutine frame pointer and
 determine the address of the promise (or call the resume/destroy functions).

2. A number of compiler builtins that the standard library might use.

  These are implemented by this patch series.

3. This introduces a new operator 'co_await' the mangling for which is also
agreed between vendors (and has an issue filed for that against the upstream
c++abi).  Demangling for this is added to libiberty in a separate patch.

The ABI has currently no target-specific content (a given psABI might elect
to mandate alignment, but the common ABI does not do this).

Standard Library impact
-----------------------

The current implementations require addition of only a single header to
the standard library (no change to the runtime).  This header is part of
the patch.

GCC Implementation outline
--------------------------

The standard's design for coroutines does not decorate the definition of
a coroutine in any way, so that a function is only known to be a coroutine
when one of the keywords (co_await, co_yield, co_return) is encountered.

This means that we cannot special-case such functions from the outset, but
must process them differently when they are finalised - which we do from
"finish_function ()".

At a high level, this design of coroutine produces four pieces from the
original user's function:

  1. A coroutine state frame (taking the logical place of the activation
     record for a regular function).  One item stored in that state is the
     index of the current suspend point.
  2. A "ramp" function
     This is what the user calls to construct the coroutine frame and start
     the coroutine execution.  This will return some object representing the
     coroutine's eventual return value (or means to continue it when it it
     suspended).
  3. A "resume" function.
     This is what gets called when a the coroutine is resumed when suspended.
  4. A "destroy" function.
     This is what gets called when the coroutine state should be destroyed
     and its memory released.

The standard's coroutines involve cooperation of the user's authored function
with a provided "promise" class, which includes mandatory methods for
handling the state transitions and providing output values.  Most realistic
coroutines will also have one or more 'awaiter' classes that implement the
user's actions for each suspend point.  As we parse (or during template
expansion) the types of the promise and awaiter classes become known, and can
then be verified against the signatures expected by the standard.

Once the function is parsed (and templates expanded) we are able to make the
transformation into the four pieces noted above.

The implementation here takes the approach of a series of AST transforms.
The state machine suspend points are encoded in three internal functions
(one of which represents an exit from scope without cleanups).  These three
IFNs are lowered early in the middle end, such that the majority of GCC's
optimisers can be run on the resulting output.

As a design choice, we have carried out the outlining of the user's function
in the front end, and taken advantage of the existing middle end's abilities
to inline and DCE where that is profitable.

Since the state machine is actually common to both resumer and destroyer
functions, we make only a single function "actor" that contains both the
resume and destroy paths.  The destroy function is represented by a small
stub that sets a value to signal the use of the destroy path and calls the
actor.  The idea is that optimisation of the state machine need only be done
once - and then the resume and destroy paths can be identified allowing the
middle end's inline and DCE machinery to optimise as profitable as noted
above.

The middle end components for this implementation are:

A pass that:
 1. Lowers the coroutine builtins that allow the standard library header to
    interact with the coroutine frame (these fairly simple logical or
    numerical substitution of values, given a coroutine frame pointer).
 2. Lowers the IFN that represents the exit from state without cleanup.
    Essentially, this becomes a gimple goto.
 3. Sets the final size of the coroutine frame at this stage.

A second pass (that requires the revised CFG that results from the lowering
of the scope exit IFNs in the first).

 1. Lower the IFNs that represent the state machine paths for the resume and
    destroy cases.

Patches squashed into this commit:

[C++ coroutines 1] Common code and base definitions.

This part of the patch series provides the gating flag, the keywords,
cpp defines etc.

[C++ coroutines 2] Define builtins and internal functions.

This part of the patch series provides the builtin functions
used by the standard library code and the internal functions
used to implement lowering of the coroutine state machine.

[C++ coroutines 3] Front end parsing and transforms.

There are two parts to this.

1. Parsing, template instantiation and diagnostics for the standard-
   mandated class entries.

  The user authors a function that becomes a coroutine (lazily) by
  making use of any of the co_await, co_yield or co_return keywords.

  Unlike a regular function, where the activation record is placed on the
  stack, and is destroyed on function exit, a coroutine has some state that
  persists between calls - the 'coroutine frame' (thus analogous to a stack
  frame).

  We transform the user's function into three pieces:
  1. A so-called ramp function, that establishes the coroutine frame and
     begins execution of the coroutine.
  2. An actor function that contains the state machine corresponding to the
     user's suspend/resume structure.
  3. A stub function that calls the actor function in 'destroy' mode.

  The actor function is executed:
   * from "resume point 0" by the ramp.
   * from resume point N ( > 0 ) for handle.resume() calls.
   * from the destroy stub for destroy point N for handle.destroy() calls.

  The C++ coroutine design described in the standard makes use of some helper
  methods that are authored in a so-called "promise" class provided by the
  user.

  At parse time (or post substitution) the type of the coroutine promise
  will be determined.  At that point, we can look up the required promise
  class methods and issue diagnostics if they are missing or incorrect.  To
  avoid repeating these actions at code-gen time, we make use of temporary
  'proxy' variables for the coroutine handle and the promise - which will
  eventually be instantiated in the coroutine frame.

  Each of the keywords will expand to a code sequence (although co_yield is
  just syntactic sugar for a co_await).

  We defer the analysis and transformatin until template expansion is
  complete so that we have complete types at that time.

2. AST analysis and transformation which performs the code-gen for the
   outlined state machine.

   The entry point here is morph_fn_to_coro () which is called from
   finish_function () when we have completed any template expansion.

   This is preceded by helper functions that implement the phases below.

   The process proceeds in four phases.

   A Initial framing.
     The user's function body is wrapped in the initial and final suspend
     points and we begin building the coroutine frame.
     We build empty decls for the actor and destroyer functions at this
     time too.
     When exceptions are enabled, the user's function body will also be
     wrapped in a try-catch block with the catch invoking the promise
     class 'unhandled_exception' method.

   B Analysis.
     The user's function body is analysed to determine the suspend points,
     if any, and to capture local variables that might persist across such
     suspensions.  In most cases, it is not necessary to capture compiler
     temporaries, since the tree-lowering nests the suspensions correctly.
     However, in the case of a captured reference, there is a lifetime
     extension to the end of the full expression - which can mean across a
     suspend point in which case it must be promoted to a frame variable.

     At the conclusion of analysis, we have a conservative frame layout and
     maps of the local variables to their frame entry points.

   C Build the ramp function.
     Carry out the allocation for the coroutine frame (NOTE; the actual size
     computation is deferred until late in the middle end to allow for future
     optimisations that will be allowed to elide unused frame entries).
     We build the return object.

   D Build and expand the actor and destroyer function bodies.
     The destroyer is a trivial shim that sets a bit to indicate that the
     destroy dispatcher should be used and then calls into the actor.

     The actor function is the implementation of the user's state machine.
     The current suspend point is noted in an index.
     Each suspend point is encoded as a pair of internal functions, one in
     the relevant dispatcher, and one representing the suspend point.

     During this process, the user's local variables and the proxies for the
     self-handle and the promise class instanceare re-written to their
     coroutine frame equivalents.

     The complete bodies for the ramp, actor and destroy function are passed
     back to finish_function for folding and gimplification.

[C++ coroutines 4] Middle end expanders and transforms.

The first part of this is a pass that provides:
 * expansion of the library support builtins, these are simple boolean
   or numerical substitutions.

 * The functionality of implementing an exit from scope without cleanup
   is performed here by lowering an IFN to a gimple goto.

This pass has to run for non-coroutine functions, since functions calling
the builtins are not necessarily coroutines (i.e. they are implementing the
library interfaces which may be called from anywhere).

The second part is the expansion of the coroutine IFNs that describe the
state machine connections to the dispatchers.  This only has to be run
for functions that are coroutine components.  The work done by this pass
is:

   In the front end we construct a single actor function that contains
   the coroutine state machine.

   The actor function has three entry conditions:
    1. from the ramp, resume point 0 - to initial-suspend.
    2. when resume () is executed (resume point N).
    3. from the destroy () shim when that is executed.

   The actor function begins with two dispatchers; one for resume and
   one for destroy (where the initial entry from the ramp is a special-
   case of resume point 0).

   Each suspend point and each dispatch entry is marked with an IFN such
   that we can connect the relevant dispatchers to their target labels.

   So, if we have:

   CO_YIELD (NUM, FINAL, RES_LAB, DEST_LAB, FRAME_PTR)

   This is await point NUM, and is the final await if FINAL is non-zero.
   The resume point is RES_LAB, and the destroy point is DEST_LAB.

   We expect to find a CO_ACTOR (NUM) in the resume dispatcher and a
   CO_ACTOR (NUM+1) in the destroy dispatcher.

   Initially, the intent of keeping the resume and destroy paths together
   is that the conditionals controlling them are identical, and thus there
   would be duplication of any optimisation of those paths if the split
   were earlier.

   Subsequent inlining of the actor (and DCE) is then able to extract the
   resume and destroy paths as separate functions if that is found
   profitable by the optimisers.

   Once we have remade the connections to their correct postions, we elide
   the labels that the front end inserted.

[C++ coroutines 5] Standard library header.

This provides the interfaces mandated by the standard and implements
the interaction with the coroutine frame by means of inline use of
builtins expanded at compile-time.  There should be a 1:1 correspondence
with the standard sections which are cross-referenced.

There is no runtime content.

At this stage, we have the content in an inline namespace "__n4835" for
the CD we worked to.

[C++ coroutines 6] Testsuite.

There are two categories of test:

1. Checks for correctly formed source code and the error reporting.
2. Checks for transformation and code-gen.

The second set are run as 'torture' tests for the standard options
set, including LTO.  These are also intentionally run with no options
provided (from the coroutines.exp script).

gcc/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* Makefile.in: Add coroutine-passes.o.
* builtin-types.def (BT_CONST_SIZE): New.
(BT_FN_BOOL_PTR): New.
(BT_FN_PTR_PTR_CONST_SIZE_BOOL): New.
* builtins.def (DEF_COROUTINE_BUILTIN): New.
* coroutine-builtins.def: New file.
* coroutine-passes.cc: New file.
* function.h (struct GTY function): Add a bit to indicate that the
function is a coroutine component.
* internal-fn.c (expand_CO_FRAME): New.
(expand_CO_YIELD): New.
(expand_CO_SUSPN): New.
(expand_CO_ACTOR): New.
* internal-fn.def (CO_ACTOR): New.
(CO_YIELD): New.
(CO_SUSPN): New.
(CO_FRAME): New.
* passes.def: Add pass_coroutine_lower_builtins,
pass_coroutine_early_expand_ifns.
* tree-pass.h (make_pass_coroutine_lower_builtins): New.
(make_pass_coroutine_early_expand_ifns): New.
* doc/invoke.texi: Document the fcoroutines command line
switch.

gcc/c-family/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* c-common.c (co_await, co_yield, co_return): New.
* c-common.h (RID_CO_AWAIT, RID_CO_YIELD,
RID_CO_RETURN): New enumeration values.
(D_CXX_COROUTINES): Bit to identify coroutines are active.
(D_CXX_COROUTINES_FLAGS): Guard for coroutine keywords.
* c-cppbuiltin.c (__cpp_coroutines): New cpp define.
* c.opt (fcoroutines): New command-line switch.

gcc/cp/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* Make-lang.in: Add coroutines.o.
* cp-tree.h (lang_decl-fn): coroutine_p, new bit.
(DECL_COROUTINE_P): New.
* lex.c (init_reswords): Enable keywords when the coroutine flag
is set,
* operators.def (co_await): New operator.
* call.c (add_builtin_candidates): Handle CO_AWAIT_EXPR.
(op_error): Likewise.
(build_new_op_1): Likewise.
(build_new_function_call): Validate coroutine builtin arguments.
* constexpr.c (potential_constant_expression_1): Handle
CO_AWAIT_EXPR, CO_YIELD_EXPR, CO_RETURN_EXPR.
* coroutines.cc: New file.
* cp-objcp-common.c (cp_common_init_ts): Add CO_AWAIT_EXPR,
CO_YIELD_EXPR, CO_RETRN_EXPR as TS expressions.
* cp-tree.def (CO_AWAIT_EXPR, CO_YIELD_EXPR, (CO_RETURN_EXPR): New.
* cp-tree.h (coro_validate_builtin_call): New.
* decl.c (emit_coro_helper): New.
(finish_function): Handle the case when a function is found to
be a coroutine, perform the outlining and emit the outlined
functions. Set a bit to signal that this is a coroutine component.
* parser.c (enum required_token): New enumeration RT_CO_YIELD.
(cp_parser_unary_expression): Handle co_await.
(cp_parser_assignment_expression): Handle co_yield.
(cp_parser_statement): Handle RID_CO_RETURN.
(cp_parser_jump_statement): Handle co_return.
(cp_parser_operator): Handle co_await operator.
(cp_parser_yield_expression): New.
(cp_parser_required_error): Handle RT_CO_YIELD.
* pt.c (tsubst_copy): Handle CO_AWAIT_EXPR.
(tsubst_expr): Handle CO_AWAIT_EXPR, CO_YIELD_EXPR and
CO_RETURN_EXPRs.
* tree.c (cp_walk_subtrees): Likewise.

libstdc++-v3/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* include/Makefile.am: Add coroutine to the std set.
* include/Makefile.in: Regenerated.
* include/std/coroutine: New file.

gcc/testsuite/ChangeLog:

2020-01-18  Iain Sandoe  <iain@sandoe.co.uk>

* g++.dg/coroutines/co-await-syntax-00-needs-expr.C: New test.
* g++.dg/coroutines/co-await-syntax-01-outside-fn.C: New test.
* g++.dg/coroutines/co-await-syntax-02-outside-fn.C: New test.
* g++.dg/coroutines/co-await-syntax-03-auto.C: New test.
* g++.dg/coroutines/co-await-syntax-04-ctor-dtor.C: New test.
* g++.dg/coroutines/co-await-syntax-05-constexpr.C: New test.
* g++.dg/coroutines/co-await-syntax-06-main.C: New test.
* g++.dg/coroutines/co-await-syntax-07-varargs.C: New test.
* g++.dg/coroutines/co-await-syntax-08-lambda-auto.C: New test.
* g++.dg/coroutines/co-return-syntax-01-outside-fn.C: New test.
* g++.dg/coroutines/co-return-syntax-02-outside-fn.C: New test.
* g++.dg/coroutines/co-return-syntax-03-auto.C: New test.
* g++.dg/coroutines/co-return-syntax-04-ctor-dtor.C: New test.
* g++.dg/coroutines/co-return-syntax-05-constexpr-fn.C: New test.
* g++.dg/coroutines/co-return-syntax-06-main.C: New test.
* g++.dg/coroutines/co-return-syntax-07-vararg.C: New test.
* g++.dg/coroutines/co-return-syntax-08-bad-return.C: New test.
* g++.dg/coroutines/co-return-syntax-09-lambda-auto.C: New test.
* g++.dg/coroutines/co-yield-syntax-00-needs-expr.C: New test.
* g++.dg/coroutines/co-yield-syntax-01-outside-fn.C: New test.
* g++.dg/coroutines/co-yield-syntax-02-outside-fn.C: New test.
* g++.dg/coroutines/co-yield-syntax-03-auto.C: New test.
* g++.dg/coroutines/co-yield-syntax-04-ctor-dtor.C: New test.
* g++.dg/coroutines/co-yield-syntax-05-constexpr.C: New test.
* g++.dg/coroutines/co-yield-syntax-06-main.C: New test.
* g++.dg/coroutines/co-yield-syntax-07-varargs.C: New test.
* g++.dg/coroutines/co-yield-syntax-08-needs-expr.C: New test.
* g++.dg/coroutines/co-yield-syntax-09-lambda-auto.C: New test.
* g++.dg/coroutines/coro-builtins.C: New test.
* g++.dg/coroutines/coro-missing-gro.C: New test.
* g++.dg/coroutines/coro-missing-promise-yield.C: New test.
* g++.dg/coroutines/coro-missing-ret-value.C: New test.
* g++.dg/coroutines/coro-missing-ret-void.C: New test.
* g++.dg/coroutines/coro-missing-ueh-1.C: New test.
* g++.dg/coroutines/coro-missing-ueh-2.C: New test.
* g++.dg/coroutines/coro-missing-ueh-3.C: New test.
* g++.dg/coroutines/coro-missing-ueh.h: New test.
* g++.dg/coroutines/coro-pre-proc.C: New test.
* g++.dg/coroutines/coro.h: New file.
* g++.dg/coroutines/coro1-ret-int-yield-int.h: New file.
* g++.dg/coroutines/coroutines.exp: New file.
* g++.dg/coroutines/torture/alloc-00-gro-on-alloc-fail.C: New test.
* g++.dg/coroutines/torture/alloc-01-overload-newdel.C: New test.
* g++.dg/coroutines/torture/call-00-co-aw-arg.C: New test.
* g++.dg/coroutines/torture/call-01-multiple-co-aw.C: New test.
* g++.dg/coroutines/torture/call-02-temp-co-aw.C: New test.
* g++.dg/coroutines/torture/call-03-temp-ref-co-aw.C: New test.
* g++.dg/coroutines/torture/class-00-co-ret.C: New test.
* g++.dg/coroutines/torture/class-01-co-ret-parm.C: New test.
* g++.dg/coroutines/torture/class-02-templ-parm.C: New test.
* g++.dg/coroutines/torture/class-03-operator-templ-parm.C: New test.
* g++.dg/coroutines/torture/class-04-lambda-1.C: New test.
* g++.dg/coroutines/torture/class-05-lambda-capture-copy-local.C: New test.
* g++.dg/coroutines/torture/class-06-lambda-capture-ref.C: New test.
* g++.dg/coroutines/torture/co-await-00-trivial.C: New test.
* g++.dg/coroutines/torture/co-await-01-with-value.C: New test.
* g++.dg/coroutines/torture/co-await-02-xform.C: New test.
* g++.dg/coroutines/torture/co-await-03-rhs-op.C: New test.
* g++.dg/coroutines/torture/co-await-04-control-flow.C: New test.
* g++.dg/coroutines/torture/co-await-05-loop.C: New test.
* g++.dg/coroutines/torture/co-await-06-ovl.C: New test.
* g++.dg/coroutines/torture/co-await-07-tmpl.C: New test.
* g++.dg/coroutines/torture/co-await-08-cascade.C: New test.
* g++.dg/coroutines/torture/co-await-09-pair.C: New test.
* g++.dg/coroutines/torture/co-await-10-template-fn-arg.C: New test.
* g++.dg/coroutines/torture/co-await-11-forwarding.C: New test.
* g++.dg/coroutines/torture/co-await-12-operator-2.C: New test.
* g++.dg/coroutines/torture/co-await-13-return-ref.C: New test.
* g++.dg/coroutines/torture/co-ret-00-void-return-is-ready.C: New test.
* g++.dg/coroutines/torture/co-ret-01-void-return-is-suspend.C: New test.
* g++.dg/coroutines/torture/co-ret-03-different-GRO-type.C: New test.
* g++.dg/coroutines/torture/co-ret-04-GRO-nontriv.C: New test.
* g++.dg/coroutines/torture/co-ret-05-return-value.C: New test.
* g++.dg/coroutines/torture/co-ret-06-template-promise-val-1.C: New test.
* g++.dg/coroutines/torture/co-ret-07-void-cast-expr.C: New test.
* g++.dg/coroutines/torture/co-ret-08-template-cast-ret.C: New test.
* g++.dg/coroutines/torture/co-ret-09-bool-await-susp.C: New test.
* g++.dg/coroutines/torture/co-ret-10-expression-evaluates-once.C: New test.
* g++.dg/coroutines/torture/co-ret-11-co-ret-co-await.C: New test.
* g++.dg/coroutines/torture/co-ret-12-co-ret-fun-co-await.C: New test.
* g++.dg/coroutines/torture/co-ret-13-template-2.C: New test.
* g++.dg/coroutines/torture/co-ret-14-template-3.C: New test.
* g++.dg/coroutines/torture/co-yield-00-triv.C: New test.
* g++.dg/coroutines/torture/co-yield-01-multi.C: New test.
* g++.dg/coroutines/torture/co-yield-02-loop.C: New test.
* g++.dg/coroutines/torture/co-yield-03-tmpl.C: New test.
* g++.dg/coroutines/torture/co-yield-04-complex-local-state.C: New test.
* g++.dg/coroutines/torture/co-yield-05-co-aw.C: New test.
* g++.dg/coroutines/torture/co-yield-06-fun-parm.C: New test.
* g++.dg/coroutines/torture/co-yield-07-template-fn-param.C: New test.
* g++.dg/coroutines/torture/co-yield-08-more-refs.C: New test.
* g++.dg/coroutines/torture/co-yield-09-more-templ-refs.C: New test.
* g++.dg/coroutines/torture/coro-torture.exp: New file.
* g++.dg/coroutines/torture/exceptions-test-0.C: New test.
* g++.dg/coroutines/torture/func-params-00.C: New test.
* g++.dg/coroutines/torture/func-params-01.C: New test.
* g++.dg/coroutines/torture/func-params-02.C: New test.
* g++.dg/coroutines/torture/func-params-03.C: New test.
* g++.dg/coroutines/torture/func-params-04.C: New test.
* g++.dg/coroutines/torture/func-params-05.C: New test.
* g++.dg/coroutines/torture/func-params-06.C: New test.
* g++.dg/coroutines/torture/lambda-00-co-ret.C: New test.
* g++.dg/coroutines/torture/lambda-01-co-ret-parm.C: New test.
* g++.dg/coroutines/torture/lambda-02-co-yield-values.C: New test.
* g++.dg/coroutines/torture/lambda-03-auto-parm-1.C: New test.
* g++.dg/coroutines/torture/lambda-04-templ-parm.C: New test.
* g++.dg/coroutines/torture/lambda-05-capture-copy-local.C: New test.
* g++.dg/coroutines/torture/lambda-06-multi-capture.C: New test.
* g++.dg/coroutines/torture/lambda-07-multi-yield.C: New test.
* g++.dg/coroutines/torture/lambda-08-co-ret-parm-ref.C: New test.
* g++.dg/coroutines/torture/local-var-0.C: New test.
* g++.dg/coroutines/torture/local-var-1.C: New test.
* g++.dg/coroutines/torture/local-var-2.C: New test.
* g++.dg/coroutines/torture/local-var-3.C: New test.
* g++.dg/coroutines/torture/local-var-4.C: New test.
* g++.dg/coroutines/torture/mid-suspend-destruction-0.C: New test.
* g++.dg/coroutines/torture/pr92933.C: New test.

4 years agoarm: Remove yet another unused variable.
Jakub Jelinek [Sat, 18 Jan 2020 08:48:48 +0000 (09:48 +0100)]
arm: Remove yet another unused variable.

Bootstrap found yet another unused variable:
../../gcc/config/arm/vfp.md:1651:17: warning: unused variable 'regname' [-Wunused-variable]

2020-01-18  Jakub Jelinek  <jakub@redhat.com>

* config/arm/vfp.md (*clear_vfp_multiple): Remove unused variable.

4 years agoarm: fix rtl checking bootstrap (PR target/93312)
Jakub Jelinek [Sat, 18 Jan 2020 08:41:59 +0000 (09:41 +0100)]
arm: fix rtl checking bootstrap (PR target/93312)

As reported in PR93312, the:
> > > > > >         * config/arm/arm.c (clear_operation_p): New function.
change broke RTL checking bootstrap.

On the testcase from the PR (which is distilled from libgcc2.c, so I think
we don't need to add it into testsuite) we ICE because SET_DEST (elt) is
not a REG, but SUBREG.  The code uses REGNO on it, which is invalid, but
only stores it into a variable, then performs REG_P (reg) check,
determines it is not a REG and bails early.

The following patch just moves the regno variable initialization after that
check, it isn't used in between.  And, as a small optimization, because
reg doesn't change, doesn't use REGNO (reg) a second time to set last_regno.

2020-01-18  Jakub Jelinek  <jakub@redhat.com>

PR target/93312
* config/arm/arm.c (clear_operation_p): Don't use REGNO until
after checking the argument is a REG.  Don't use REGNO (reg)
again to set last_regno, reuse regno variable instead.

4 years agoPR93234 INQUIRE on pre-assigned files of ROUND and SIGN properties
Jerry DeLisle [Sat, 18 Jan 2020 03:36:03 +0000 (19:36 -0800)]
PR93234 INQUIRE on pre-assigned files of ROUND and SIGN properties

PR libfortran/93234
* io/unit.c (set_internal_unit): Set round and sign flags
correctly.

* gfortran.dg/inquire_pre.f90: New test.

4 years agoDaily bump.
GCC Administrator [Sat, 18 Jan 2020 00:16:30 +0000 (00:16 +0000)]
Daily bump.

4 years agoanalyzer: prevent ICE on isnan (PR 93290)
David Malcolm [Thu, 16 Jan 2020 14:46:30 +0000 (09:46 -0500)]
analyzer: prevent ICE on isnan (PR 93290)

PR analyzer/93290 reports an ICE on calls to isnan().
The root cause is that an UNORDERED_EXPR is passed
to region_model::eval_condition_without_cm, and there's
a stray gcc_unreachable () in the case where we're comparing
an svalue against itself.

I attempted a more involved patch that properly handled NaN in general
but it seems I've baked the assumption of reflexivity too deeply into
the constraint_manager code.

For now, this patch avoids the ICE and documents the limitation.

gcc/analyzer/ChangeLog:
PR analyzer/93290
* region-model.cc (region_model::eval_condition_without_cm): Avoid
gcc_unreachable for unexpected operations for the case where
we're comparing an svalue against itself.

gcc/ChangeLog
* doc/analyzer.texi (Limitations): Add note about NaN.

gcc/testsuite/ChangeLog:
PR analyzer/93290
* gcc.dg/analyzer/pr93290.c: New test.

4 years agoPR90374 Zero width format specifiers.
Jerry DeLisle [Fri, 17 Jan 2020 19:26:10 +0000 (11:26 -0800)]
PR90374 Zero width format specifiers.

PR libfortran/90374
* io/format.c (parse_format_list): Zero width not allowed with
FMT_D.
* io/write_float.def (build_float_string): Include range of
higher exponent values that require wider width.

4 years agoAdd testcase of PR c++/92542, already fixed.
Paolo Carlini [Fri, 17 Jan 2020 19:03:58 +0000 (20:03 +0100)]
Add testcase of PR c++/92542, already fixed.

PR c++/92542
* g++.dg/pr92542.C: New.

4 years agoAdd testcase of PR c++/92542, already fixed.
Paolo Carlini [Fri, 17 Jan 2020 19:02:21 +0000 (20:02 +0100)]
Add testcase of PR c++/92542, already fixed.

PR c++/92542
* g++.dg/pr92542.C: New.

4 years ago[GCC/ARM, 2/2] Add support for ASRL(imm), LSLL(imm) and LSRL(imm) instructions for...
Mihail Ionescu [Fri, 17 Jan 2020 18:14:54 +0000 (18:14 +0000)]
[GCC/ARM, 2/2] Add support for ASRL(imm), LSLL(imm) and LSRL(imm) instructions for Armv8.1-M Mainline

This patch is adding the following instructions:

ASRL (imm)
LSLL (imm)
LSRL (imm)

*** gcc/ChangeLog ***

2020-01-17  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
    Sudakshina Das  <sudi.das@arm.com>

* config/arm/arm.md (ashldi3): Generate thumb2_lsll for both reg
and valid immediate.
(ashrdi3): Generate thumb2_asrl for both reg and valid immediate.
(lshrdi3): Generate thumb2_lsrl for valid immediates.
* config/arm/constraints.md (Pg): New.
* config/arm/predicates.md (long_shift_imm): New.
(arm_reg_or_long_shift_imm): Likewise.
* config/arm/thumb2.md (thumb2_asrl): New immediate alternative.
(thumb2_lsll): Likewise.
(thumb2_lsrl): New.

*** gcc/testsuite/ChangeLog ***

2020-01-17  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
    Sudakshina Das  <sudi.das@arm.com>

* gcc.target/arm/armv8_1m-shift-imm_1.c: New test.

4 years ago[GCC/ARM, 1/2] Add support for ASRL(reg) and LSLL(reg) instructions for Armv8.1-M...
Mihail Ionescu [Fri, 17 Jan 2020 17:56:41 +0000 (17:56 +0000)]
[GCC/ARM, 1/2] Add support for ASRL(reg) and LSLL(reg) instructions for Armv8.1-M Mainline

This patch is adding the following instructions:

ASRL (reg)
LSLL (reg)

*** gcc/ChangeLog ***

2020-01-17  Mihail-Calin Ionescu <mihail.ionescu@arm.com>
    Sudakshina Das  <sudi.das@arm.com>

* config/arm/arm.md (ashldi3): Generate thumb2_lsll for TARGET_HAVE_MVE.
(ashrdi3): Generate thumb2_asrl for TARGET_HAVE_MVE.
* config/arm/arm.c (arm_hard_regno_mode_ok): Allocate even odd
register pairs for doubleword quantities for ARMv8.1M-Mainline.
* config/arm/thumb2.md (thumb2_asrl): New.
(thumb2_lsll): Likewise.

2020-01-17  Mihail-Calin Ionescu <mihail.ionescu@arm.com>
    Sudakshina Das  <sudi.das@arm.com>

* gcc.target/arm/armv8_1m-shift-reg_1.c: New test.

4 years agoFix up ChangeLog.
Jakub Jelinek [Fri, 17 Jan 2020 18:45:35 +0000 (19:45 +0100)]
Fix up ChangeLog.

4 years agoarm: Unbreak bootstrap
Jakub Jelinek [Fri, 17 Jan 2020 18:41:42 +0000 (19:41 +0100)]
arm: Unbreak bootstrap

2020-01-17  Jakub Jelinek  <jakub@redhat.com>

* config/arm/arm.c (cmse_nonsecure_call_inline_register_clear): Remove
unused variable.

4 years agoRename acc_device_gcn to acc_device_radeon
Andrew Stubbs [Fri, 17 Jan 2020 14:46:59 +0000 (14:46 +0000)]
Rename acc_device_gcn to acc_device_radeon

2020-01-17  Andrew Stubbs  <ams@codesourcery.com>

libgomp/
* config/accel/openacc.f90 (openacc_kinds): Rename acc_device_gcn to
acc_device_radeon.
(openacc): Likewise.
* openacc.f90 (openacc_kinds): Likewise.
(openacc): Likewise.
* openacc.h (acc_device_t): Likewise.
* openacc_lib.h: Likewise.
* testsuite/lib/libgomp.exp
(check_effective_target_openacc_amdgcn_accel_present): Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-init-1.c
(cb_compute_construct_end): Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-kernels-1.c
(cb_enqueue_launch_start): Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-parallel-1.c
(cb_enter_data_end): Likewise.
(cb_exit_data_start): Likewise.
(cb_exit_data_end): Likewise.
(cb_compute_construct_end): Likewise.
(cb_enqueue_launch_start): Likewise.
(cb_enqueue_launch_end): Likewise.
* testsuite/libgomp.oacc-c-c++-common/asyncwait-nop-1.c
(main): Likewise.

4 years agolibstdc++: Fix freestanding build PR 92376)
Jonathan Wakely [Fri, 17 Jan 2020 15:49:02 +0000 (15:49 +0000)]
libstdc++: Fix freestanding build PR 92376)

In a freestanding library we don't install the <pstl/pstl_config.h>
header, so don't try to include it unless it exists.

Explicitly declare aligned alloc functions for freestanding, because
<cstdlib> doesn't declare them.

PR libstdc++/92376
* include/bits/c++config: Only do PSTL config when the header is
present, to fix freestanding.
* libsupc++/new_opa.cc [!_GLIBCXX_HOSTED]: Declare allocation
functions if they were detected by configure.

4 years agogdbinit.in: make shorthands accept an explicit argument
Alexander Monakov [Fri, 17 Jan 2020 15:15:44 +0000 (18:15 +0300)]
gdbinit.in: make shorthands accept an explicit argument

Make gdb shorthands such as 'pr' accept an argument, in addition to
implictly taking register '$' as the thing to examine.

The 'eval ...' one-liners are used to workaround GDB bug #22466.

* gdbinit.in (help-gcc-hooks): New command.
(pp, pr, prl, pt, pct, pgg, pgq, pgs, pge, pmz, ptc, pdn, ptn, pdd, prc,
pi, pbm, pel, trt): Take $arg0 instead of $ if supplied. Update
documentation.

4 years ago[AArch64] [Obvious] Correct pattern target requirement
Matthew Malcomson [Fri, 17 Jan 2020 15:08:21 +0000 (15:08 +0000)]
[AArch64] [Obvious] Correct pattern target requirement

Had mistakenly used a target macro that was not defined and not the
relevant one instead of the macro that should be used.

TARGET_ARMV8_6 is not defined, and also not the macro we want to check.
Instead check TARGET_F64MM.

gcc/ChangeLog:

2020-01-17  Matthew Malcomson  <matthew.malcomson@arm.com>

* config/aarch64/aarch64-sve.md (@aarch64_sve_ld1ro<mode>): Use
the correct target macro.

4 years agoFix g++ testsuite failure caused by std::is_pod deprecation
Jonathan Wakely [Fri, 10 Jan 2020 13:04:09 +0000 (13:04 +0000)]
Fix g++ testsuite failure caused by std::is_pod deprecation

PR testsuite/93227
* g++.dg/cpp0x/std-layout1.C: Use -Wno-deprecated-declarations for
C++20, due to std::is_pod being deprecated.

4 years ago[AArch64] [SVE] Implement svld1ro intrinsic.
Matthew Malcomson [Fri, 17 Jan 2020 11:50:54 +0000 (11:50 +0000)]
[AArch64] [SVE] Implement svld1ro intrinsic.

We take no action to ensure the SVE vector size is large enough.  It is
left to the user to check that before compiling this intrinsic or before
running such a program on a machine.

The main difference between ld1ro and ld1rq is in the allowed offsets,
the implementation difference is that ld1ro is implemented using integer
modes since there are no pre-existing vector modes of the relevant size.
Adding new vector modes simply for this intrinsic seems to make the code
less tidy.

Specifications can be found under the "Arm C Language Extensions for
Scalable Vector Extension" title at
https://developer.arm.com/architectures/system-architectures/software-standards/acle

gcc/ChangeLog:

2020-01-17  Matthew Malcomson  <matthew.malcomson@arm.com>

* config/aarch64/aarch64-protos.h
(aarch64_sve_ld1ro_operand_p): New.
* config/aarch64/aarch64-sve-builtins-base.cc
(class load_replicate): New.
(class svld1ro_impl): New.
(class svld1rq_impl): Change to inherit from load_replicate.
(svld1ro): New sve intrinsic function base.
* config/aarch64/aarch64-sve-builtins-base.def (svld1ro):
New DEF_SVE_FUNCTION.
* config/aarch64/aarch64-sve-builtins-base.h
(svld1ro): New decl.
* config/aarch64/aarch64-sve-builtins.cc
(function_expander::add_mem_operand): Modify assert to allow
OImode.
* config/aarch64/aarch64-sve.md (@aarch64_sve_ld1ro<mode>): New
pattern.
* config/aarch64/aarch64.c
(aarch64_sve_ld1rq_operand_p): Implement in terms of ...
(aarch64_sve_ld1rq_ld1ro_operand_p): This.
(aarch64_sve_ld1ro_operand_p): New.
* config/aarch64/aarch64.md (UNSPEC_LD1RO): New unspec.
* config/aarch64/constraints.md (UOb,UOh,UOw,UOd): New.
* config/aarch64/predicates.md
(aarch64_sve_ld1ro_operand_{b,h,w,d}): New.

gcc/testsuite/ChangeLog:

2020-01-17  Matthew Malcomson  <matthew.malcomson@arm.com>

* gcc.target/aarch64/sve/acle/asm/ld1ro_f16.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_f32.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_f64.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_s16.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_s32.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_s64.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_s8.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_u16.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_u32.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_u64.c: New test.
* gcc.target/aarch64/sve/acle/asm/ld1ro_u8.c: New test.

4 years ago[AArch64] Enable CLI for Armv8.6-A f64mm
Matthew Malcomson [Fri, 17 Jan 2020 11:22:28 +0000 (11:22 +0000)]
[AArch64] Enable CLI for Armv8.6-A f64mm

This patch is necessary for sve-ld1ro intrinsic I posted in
https://gcc.gnu.org/ml/gcc-patches/2020-01/msg00466.html .

I had mistakenly thought this option was already enabled upstream.

This provides the option +f64mm, that turns on the 64 bit floating point
matrix multiply extension.  This extension is only available for
AArch64.  Turning on this extension also turns on the SVE extension.

This extension is optional and only available at Armv8.2-A and onward.

We also add the ACLE defined macro for this extension.

gcc/ChangeLog:

2020-01-17  Matthew Malcomson  <matthew.malcomson@arm.com>

* config/aarch64/aarch64-c.c (_ARM_FEATURE_MATMUL_FLOAT64):
Introduce this ACLE specified predefined macro.
* config/aarch64/aarch64-option-extensions.def (f64mm): New.
(fp): Disabling this disables f64mm.
(simd): Disabling this disables f64mm.
(fp16): Disabling this disables f64mm.
(sve): Disabling this disables f64mm.
* config/aarch64/aarch64.h (AARCH64_FL_F64MM): New.
(AARCH64_ISA_F64MM): New.
(TARGET_F64MM): New.
* doc/invoke.texi (f64mm): Document new option.

gcc/testsuite/ChangeLog:

2020-01-17  Matthew Malcomson  <matthew.malcomson@arm.com>

* gcc.target/aarch64/pragma_cpp_predefs_2.c: Check for f64mm
predef.

4 years ago[AArch64] Enable compare branch fusion
Wilco Dijkstra [Fri, 17 Jan 2020 14:27:14 +0000 (14:27 +0000)]
[AArch64] Enable compare branch fusion

Enable the most basic form of compare-branch fusion since various CPUs
support it. This has no measurable effect on cores which don't support
branch fusion, but increases fusion opportunities on cores which do.

gcc/
* config/aarch64/aarch64.c (generic_tunings): Add branch fusion.
(neoversen1_tunings): Likewise.

4 years agoPR c++/92531 - ICE with noexcept(lambda).
Jason Merrill [Fri, 17 Jan 2020 13:37:49 +0000 (08:37 -0500)]
PR c++/92531 - ICE with noexcept(lambda).

This was failing because uses_template_parms didn't recognize LAMBDA_EXPR as
a kind of expression.  Instead of trying to enumerate all the different
varieties of expression and then aborting if what's left isn't
error_mark_node, let's handle error_mark_node and then assume anything else
is an expression.

* pt.c (uses_template_parms): Don't try to enumerate all the
expression cases.

4 years agoc++: Fix deprecated attribute handling on templates (PR c++/93228)
Jakub Jelinek [Fri, 17 Jan 2020 14:22:22 +0000 (15:22 +0100)]
c++: Fix deprecated attribute handling on templates (PR c++/93228)

As the following testcase shows, when deprecated attribute is on a template,
we'd never print the message if any, because the attribute is not
present on the TEMPLATE_DECL with which warn_deprecated_use is called,
but on its DECL_TEMPLATE_RESULT or its type.

2020-01-17  Jakub Jelinek  <jakub@redhat.com>

PR c++/93228
* parser.c (cp_parser_template_name): Look up deprecated attribute
in DECL_TEMPLATE_RESULT or its type's attributes.

* g++.dg/cpp1y/attr-deprecated-3.C: New test.

4 years ago[PR93306] Short-circuit has_include
Nathan Sidwell [Fri, 17 Jan 2020 13:39:47 +0000 (05:39 -0800)]
[PR93306] Short-circuit has_include

the preprocessor evaluator has a skip_eval counter, but we weren't
checking it after parsing has_include(foo), but before looking for
foo.  Resulting in unnecessary io for 'FALSE_COND && has_include <foo>'

PR preprocessor/93306
* expr.c (parse_has_include): Refactor.  Check skip_eval before
looking.

4 years agoanalyzer: fix handling of negative byte offsets (v2) (PR 93281)
David Malcolm [Wed, 15 Jan 2020 20:55:11 +0000 (15:55 -0500)]
analyzer: fix handling of negative byte offsets (v2) (PR 93281)

Various 32-bit targets show failures in gcc.dg/analyzer/data-model-1.c
with tests of the form:
  __analyzer_eval (q[-2].x == 107024); /* { dg-warning "TRUE" } */
  __analyzer_eval (q[-2].y == 107025); /* { dg-warning "TRUE" } */
where they emit UNKNOWN instead.

The root cause is that gimple has a byte-based twos-complement offset
of -16 expressed like this:
  _55 = q_92 + 4294967280;  (32-bit)
or:
  _55 = q_92 + 18446744073709551600; (64-bit)

Within region_model::convert_byte_offset_to_array_index that unsigned
offset was being divided by the element size to get an offset within
an array.

This happened to work on 64-bit target and host, but not elsewhere;
the offset needs to be converted to a signed type before the division
is meaningful.

This patch does so, fixing the failures.

gcc/analyzer/ChangeLog:
PR analyzer/93281
* region-model.cc
(region_model::convert_byte_offset_to_array_index): Convert to
ssizetype before dividing by byte_size.  Use fold_binary rather
than fold_build2 to avoid needlessly constructing a tree for the
non-const case.

4 years ago[AArch64] Fix shrinkwrapping interactions with atomics (PR92692)
Wilco Dijkstra [Fri, 17 Jan 2020 13:17:21 +0000 (13:17 +0000)]
[AArch64] Fix shrinkwrapping interactions with atomics (PR92692)

The separate shrinkwrapping pass may insert stores in the middle
of atomics loops which can cause issues on some implementations.
Avoid this by delaying splitting atomics patterns until after
prolog/epilog generation.

gcc/
PR target/92692
* config/aarch64/aarch64.c (aarch64_split_compare_and_swap)
Add assert to ensure prolog has been emitted.
(aarch64_split_atomic_op): Likewise.
* config/aarch64/atomics.md (aarch64_compare_and_swap<mode>)
Use epilogue_completed rather than reload_completed.
(aarch64_atomic_exchange<mode>): Likewise.
(aarch64_atomic_<atomic_optab><mode>): Likewise.
(atomic_nand<mode>): Likewise.
(aarch64_atomic_fetch_<atomic_optab><mode>): Likewise.
(atomic_fetch_nand<mode>): Likewise.
(aarch64_atomic_<atomic_optab>_fetch<mode>): Likewise.
(atomic_nand_fetch<mode>): Likewise.

4 years agoAdd PR number to change log
Richard Sandiford [Fri, 17 Jan 2020 12:22:58 +0000 (12:22 +0000)]
Add PR number to change log

4 years agoaarch64: Don't raise FE_INVALID for -__builtin_isgreater [PR93133]
Richard Sandiford [Thu, 16 Jan 2020 19:22:20 +0000 (19:22 +0000)]
aarch64: Don't raise FE_INVALID for -__builtin_isgreater [PR93133]

AIUI, the main purpose of REVERSE_CONDITION is to take advantage of
any integer vs. FP information encoded in the CC mode, particularly
when handling LT, LE, GE and GT.  For integer comparisons we can
safely map LT->GE, LE->GT, GE->LT and GT->LE, but for float comparisons
this would usually be invalid without -ffinite-math-only.

The aarch64 definition of REVERSE_CONDITION used
reverse_condition_maybe_unordered for FP comparisons, which had the
effect of converting an unordered-signalling LT, LE, GE or GT into a
quiet UNGE, UNGT, UNLT or UNLE.  And it would do the same in reverse:
convert a quiet UN* into an unordered-signalling comparison.

This would be safe in practice (although a little misleading) if we
always used a compare:CCFP or compare:CCFPE to do the comparison and
then used (gt (reg:CCFP/CCFPE CC_REGNUM) (const_int 0)) etc. to test
the result.  In that case any signal is raised by the compare and the
choice of quiet vs. signalling relations doesn't matter when testing
the result.  The problem is that we also want to use GT directly on
float registers, where any signal is raised by the comparison operation
itself and so must follow the normal rtl rules (GT signalling,
UNLE quiet).

I think the safest fix is to make REVERSIBLE_CC_MODE return false
for FP comparisons.  We can then use the default REVERSE_CONDITION
for integer comparisons and the usual conservatively-correct
reversed_comparison_code_parts behaviour for FP comparisons.
Unfortunately reversed_comparison_code_parts doesn't yet handle
-ffinite-math-only, but that's probably GCC 11 material.

A downside is that:

    int f (float x, float y) { return !(x < y); }

now generates:

        fcmpe   s0, s1
        cset    w0, mi
        eor     w0, w0, 1
        ret

without -ffinite-math-only.  Maybe for GCC 11 we should define rtx
codes for all IEEE comparisons, so that we don't have this kind of
representational gap.

Changing REVERSE_CONDITION itself is pretty easy.  However, the macro
was also used in the ccmp handling, which relied on being able to
reverse all comparisons.  The patch adds new reversed patterns for
cases in which the original condition needs to be kept.

The test is based on gcc.dg/torture/pr91323.c.  It might well fail
on other targets that have similar bugs; please XFAIL as appropriate
if you don't want to fix the target for GCC 10.

2020-01-17  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.h (REVERSIBLE_CC_MODE): Return false
for FP modes.
(REVERSE_CONDITION): Delete.
* config/aarch64/iterators.md (CC_ONLY): New mode iterator.
(CCFP_CCFPE): Likewise.
(e): New mode attribute.
* config/aarch64/aarch64.md (ccmp<GPI:mode>): Rename to...
(@ccmp<CC_ONLY:mode><GPI:mode>): ...this, using CC_ONLY instead of CC.
(fccmp<GPF:mode>, fccmpe<GPF:mode>): Merge into...
(@ccmp<CCFP_CCFPE:mode><GPF:mode>): ...this combined pattern.
(@ccmp<CC_ONLY:mode><GPI:mode>_rev): New pattern.
(@ccmp<CCFP_CCFPE:mode><GPF:mode>_rev): Likewise.
* config/aarch64/aarch64.c (aarch64_gen_compare_reg): Update
name of generator from gen_ccmpdi to gen_ccmpccdi.
(aarch64_gen_ccmp_next): Use code_for_ccmp.  If we want to reverse
the previous comparison but aren't able to, use the new ccmp_rev
patterns instead.

4 years agogimplifier: handle POLY_INT_CST-sized TARGET_EXPRs
Richard Sandiford [Wed, 15 Jan 2020 16:52:18 +0000 (16:52 +0000)]
gimplifier: handle POLY_INT_CST-sized TARGET_EXPRs

If a TARGET_EXPR has poly-int size, the gimplifier would treat it
like a VLA and use gimplify_vla_decl.  gimplify_vla_decl in turn
would use an alloca and expect all references to be gimplified
via the DECL_VALUE_EXPR.  This caused confusion later in
gimplify_var_or_parm_decl_1 when we (correctly) had direct rather
than indirect references.

For completeness, the patch also fixes similar tests in the RETURN_EXPR
handling and OpenMP depend clauses.

2020-01-17  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* gimplify.c (gimplify_return_expr): Use poly_int_tree_p rather
than testing directly for INTEGER_CST.
(gimplify_target_expr, gimplify_omp_depend): Likewise.

gcc/testsuite/
* g++.target/aarch64/sve/acle/general-c++/gimplify_1.C: New test.

4 years agoPATCH] Fortran: PR93263 -fno-automatic and RECURSIVE
Mark Eggleston [Fri, 17 Jan 2020 08:49:25 +0000 (08:49 +0000)]
PATCH] Fortran: PR93263 -fno-automatic and RECURSIVE

The use of -fno-automatic should not affect the save attribute of a
recursive procedure. The first test case checks unsaved variables
and the second checks saved variables.

4 years agovect: Fix ICE in vectorizable_comparison PR93292
Jakub Jelinek [Fri, 17 Jan 2020 08:39:45 +0000 (09:39 +0100)]
vect: Fix ICE in vectorizable_comparison PR93292

The following testcase ICEs on powerpc64le-linux.  The problem is that
get_vectype_for_scalar_type returns NULL, and while most places in
tree-vect-stmts.c handle that case, this spot doesn't and punts only
if it is non-NULL, but with different number of elts than expected.

2020-01-17  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/93292
* tree-vect-stmts.c (vectorizable_comparison): Punt also if
get_vectype_for_scalar_type returns NULL.

* g++.dg/opt/pr93292.C: New test.

4 years agotestsuite: Unbreak compat.exp testing with alt compiler PR93294
Jakub Jelinek [Fri, 17 Jan 2020 08:30:17 +0000 (09:30 +0100)]
testsuite: Unbreak compat.exp testing with alt compiler PR93294

2020-01-17  Jakub Jelinek  <jakub@redhat.com>

PR testsuite/93294
* lib/c-compat.exp (compat-use-alt-compiler): Handle
-fdiagnostics-urls=never similarly to -fdiagnostics-color=never.
(compat_setup_dfp): Likewise.

4 years agoChangeLog fixes.
Jakub Jelinek [Fri, 17 Jan 2020 08:29:28 +0000 (09:29 +0100)]
ChangeLog fixes.

4 years agocontrib/gcc_update: Insert "tformat:" for git log --pretty=tformat:%p:%t:%H
Hans-Peter Nilsson [Fri, 17 Jan 2020 07:42:32 +0000 (08:42 +0100)]
contrib/gcc_update: Insert "tformat:" for git log --pretty=tformat:%p:%t:%H

Really old git versions (like 1.6.0) require
"git log --pretty=tformat:%p:%t:%H"
or else we see:

Updating GIT tree
Current branch master is up to date.
fatal: invalid --pretty format: %p:%t:%H
Adjusting file timestamps
Touching gcc/config.in...
Touching gcc/config/arm/arm-tune.md...

...and an empty revision in LAST_UPDATED and gcc/REVISION.
In its absence, for newer git versions, "tformat" is the default
qualifier, documented as such default for at least git-2.11.0.

4 years agoPR c++/93286 - ICE with __is_constructible and variadic template.
Jason Merrill [Thu, 16 Jan 2020 21:55:39 +0000 (16:55 -0500)]
PR c++/93286 - ICE with __is_constructible and variadic template.

Here we had been recursing in tsubst_copy_and_build if type2 was a TREE_LIST
because that function knew how to deal with pack expansions, and tsubst
didn't.  But tsubst_copy_and_build expects to be dealing with expressions,
so we crash when trying to convert_from_reference a type.

* pt.c (tsubst) [TREE_LIST]: Handle pack expansion.
(tsubst_copy_and_build) [TRAIT_EXPR]: Always use tsubst for type2.

4 years agoDaily bump.
GCC Administrator [Fri, 17 Jan 2020 00:16:37 +0000 (00:16 +0000)]
Daily bump.

4 years agoExtern -param=max-predicted-iterations range.
Jan Hubicka [Thu, 16 Jan 2020 23:50:22 +0000 (00:50 +0100)]
Extern -param=max-predicted-iterations range.

* params.opt (-param=max-predicted-iterations): Increase range from 0.
* predict.c (estimate_loops): Add 1 to param_max_predicted_iterations.

4 years agoFix ICE caused by swallowing a token in c_parser_consume_token
Kerem Kat [Thu, 16 Jan 2020 23:42:11 +0000 (23:42 +0000)]
Fix ICE caused by swallowing a token in c_parser_consume_token

This patch fixes ICE on invalid code, specifically files that have
conflict-marker-like signs before EOF.

PR c/92833
gcc/c/
* c-parser.c (c_parser_consume_token): Fix peeked token stack pop
to support 4 available tokens.

gcc/testsuite/
* c-c++-common/pr92833-1.c, c-c++-common/pr92833-2.c,
c-c++-common/pr92833-3.c, c-c++-common/pr92833-4.c: New tests.

4 years agoMake profile estimation more precise
Jan Hubicka [Thu, 16 Jan 2020 22:55:44 +0000 (23:55 +0100)]
Make profile estimation more precise

While analyzing code size regression in SPEC2k GCC binary I noticed that we
perform some inline decisions because we think that number of executions are
very high.
In particular there was inline decision inlining gen_rtx_fmt_ee to find_reloads
believing that it is called 4 billion times.  This turned out to be cummulation
of roundoff errors in propagate_freq which was bit mechanically updated from
original sreals to C++ sreals and later to new probabilities.

This led us to estimate that a loopback edge is reached with probability 2.3
which was capped to 1-1/10000 and since this happened in nested loop it quickly
escalated to large values.

Originally capping to REG_BR_PROB_BASE avoided such problems but now we have
much higher range.

This patch avoids going from probabilites to REG_BR_PROB_BASE so precision is
kept.  In addition it makes the propagation to not estimate more than
param-max-predicted-loop-iterations.  The first change makes the cap to not
be triggered on the gcc build, but it is still better to be safe than sorry.

* ipa-fnsummary.c (estimate_calls_size_and_time): Fix formating of
dump.
* params.opt: (max-predicted-iterations): Set bounds.
* predict.c (real_almost_one, real_br_prob_base,
real_inv_br_prob_base, real_one_half, real_bb_freq_max): Remove.
(propagate_freq): Add max_cyclic_prob parameter; cap cyclic
probabilities; do not truncate to reg_br_prob_bases.
(estimate_loops_at_level): Pass max_cyclic_prob.
(estimate_loops): Compute max_cyclic_prob.
(estimate_bb_frequencies): Do not initialize real_*; update calculation
of back edge prob.
* profile-count.c (profile_probability::to_sreal): New.
* profile-count.h (class sreal): Move up in file.
(profile_probability::to_sreal): Declare.

4 years agoPR c++/93280 - ICE with aggregate assignment and DMI.
Jason Merrill [Thu, 16 Jan 2020 15:46:40 +0000 (10:46 -0500)]
PR c++/93280 - ICE with aggregate assignment and DMI.

I recently added an assert to cp-gimplify to catch any
TARGET_EXPR_DIRECT_INIT_P being expanded without a target object, and this
testcase found one.  We started out with a TARGET_EXPR around the
CONSTRUCTOR, which would normally mean that the member initializer would be
used to directly initialize the appropriate member of whatever object the
TARGET_EXPR ends up initializing.  But then gimplify_modify_expr_rhs
stripped the TARGET_EXPR in order to assign directly from the elements of
the CONSTRUCTOR, leaving no object for the TARGET_EXPR_DIRECT_INIT_P to
initialize.  I considered setting CONSTRUCTOR_PLACEHOLDER_BOUNDARY in that
case, which implies TARGET_EXPR_NO_ELIDE, but decided that there's no
particular reason the A initializer needs to initialize a member of a B
rather than a distinct A object, so let's only set TARGET_EXPR_DIRECT_INIT_P
when we're using the DMI in a constructor.

* init.c (get_nsdmi): Set TARGET_EXPR_DIRECT_INIT_P here.
* typeck2.c (digest_nsdmi_init): Not here.

4 years agoFix noreorder symbol partitioning reversion.
Martin Liska [Thu, 16 Jan 2020 19:44:27 +0000 (20:44 +0100)]
Fix noreorder symbol partitioning reversion.

* lto-partition.c (lto_balanced_map): Remember
best_noreorder_pos and then restore to it
when we revert.

4 years agolibstdc++: std::ctype fixes for recent versions of NetBSD
Jonathan Wakely [Fri, 10 Jan 2020 16:01:19 +0000 (16:01 +0000)]
libstdc++: std::ctype fixes for recent versions of NetBSD

This removes support for EOL versions of NetBSD and syncs the
definitions with patches from NetBSD upstream.

The only change here that isn't from upstream is to use _CTYPE_BL for
the isblank class, which is correct but wasn't previously done either in
FSF GCC or the NetBSD packages.

2020-01-16  Kai-Uwe Eckhardt  <kuehro@gmx.de>
    Matthew Bauer  <mjbauer95@gmail.com>
    Jonathan Wakely  <jwakely@redhat.com>

PR bootstrap/64271 (partial)
* config/os/bsd/netbsd/ctype_base.h (ctype_base::mask): Change type
to unsigned short.
(ctype_base::alpha, ctype_base::digit, ctype_base::xdigit)
(ctype_base::print, ctype_base::graph, ctype_base::alnum): Sync
definitions with NetBSD upstream.
(ctype_base::blank): Use _CTYPE_BL.
* config/os/bsd/netbsd/ctype_configure_char.cc (_C_ctype_): Remove
Declaration.
(ctype<char>::classic_table): Use _C_ctype_tab_ instead of _C_ctype_.
(ctype<char>::do_toupper, ctype<char>::do_tolower): Cast char
parameters to unsigned char.
* config/os/bsd/netbsd/ctype_inline.h (ctype<char>::is): Likewise.

4 years ago[GCC][PATCH][ARM] Add Bfloat16_t scalar type, vector types and machine modes to ARM...
Stam Markianos-Wright [Thu, 16 Jan 2020 15:54:53 +0000 (15:54 +0000)]
[GCC][PATCH][ARM] Add Bfloat16_t scalar type, vector types and machine modes to ARM back-end [2/2]

gcc/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* config/arm/arm.c
(arm_invalid_conversion): New function for target hook.
(arm_invalid_unary_op): New function for target hook.
(arm_invalid_binary_op): New function for target hook.

gcc/testsuite/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* g++.target/arm/bfloat_cpp_typecheck.C: New test.
* gcc.target/arm/bfloat16_scalar_typecheck.c: New test.
* gcc.target/arm/bfloat16_vector_typecheck_1.c: New test.
* gcc.target/arm/bfloat16_vector_typecheck_2.c: New test.

4 years ago[GCC][PATCH][ARM] Add Bfloat16_t scalar type, vector types and machine modes to ARM...
Stam Markianos-Wright [Thu, 16 Jan 2020 15:50:08 +0000 (15:50 +0000)]
[GCC][PATCH][ARM] Add Bfloat16_t scalar type, vector types and machine modes to ARM back-end [1/2]

gcc/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* config.gcc: Add arm_bf16.h.
* config/arm/arm-builtins.c (arm_mangle_builtin_type): Fix comment.
(arm_simd_builtin_std_type): Add BFmode.
(arm_init_simd_builtin_types): Define element types for vector types.
(arm_init_bf16_types): New function.
(arm_init_builtins): Add arm_init_bf16_types function call.
* config/arm/arm-modes.def: Add BFmode and V4BF, V8BF vector modes.
* config/arm/arm-simd-builtin-types.def: Add V4BF, V8BF.
* config/arm/arm.c (aapcs_vfp_sub_candidate):  Add BFmode.
(arm_hard_regno_mode_ok): Add BFmode and tidy up statements.
(arm_vector_mode_supported_p): Add V4BF, V8BF.
(arm_mangle_type):  Add __bf16.
* config/arm/arm.h: Add V4BF, V8BF to VALID_NEON_DREG_MODE,
VALID_NEON_QREG_MODE respectively. Add export arm_bf16_type_node,
arm_bf16_ptr_type_node.
* config/arm/arm.md: Add BFmode to movhf expand, mov pattern and
define_split between ARM registers.
* config/arm/arm_bf16.h: New file.
* config/arm/arm_neon.h: Add arm_bf16.h and Bfloat vector types.
* config/arm/iterators.md: (ANY64_BF, VDXMOV, VHFBF, HFBF, fporbf): New.
(VQXMOV): Add V8BF.
* config/arm/neon.md: Add BF vector types to movhf NEON move patterns.
* config/arm/vfp.md: Add BFmode to movhf patterns.

gcc/testsuite/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* g++.dg/abi/mangle-neon.C: Add BF16 SIMD types.
* g++.dg/ext/arm-bf16/bf16-mangle-1.C: New test.
* gcc.target/arm/bfloat16_scalar_1_1.c: New test.
* gcc.target/arm/bfloat16_scalar_1_2.c: New test.
* gcc.target/arm/bfloat16_scalar_2_1.c: New test.
* gcc.target/arm/bfloat16_scalar_2_2.c: New test.
* gcc.target/arm/bfloat16_scalar_3_1.c: New test.
* gcc.target/arm/bfloat16_scalar_3_2.c: New test.
* gcc.target/arm/bfloat16_scalar_4.c: New test.
* gcc.target/arm/bfloat16_simd_1_1.c: New test.
* gcc.target/arm/bfloat16_simd_1_2.c: New test.
* gcc.target/arm/bfloat16_simd_2_1.c: New test.
* gcc.target/arm/bfloat16_simd_2_2.c: New test.
* gcc.target/arm/bfloat16_simd_3_1.c: New test.
* gcc.target/arm/bfloat16_simd_3_2.c: New test.

4 years agoAdd CLI and multilib support for Armv8.1-M Mainline MVE extensions
Mihail Ionescu [Wed, 15 Jan 2020 13:25:30 +0000 (13:25 +0000)]
Add CLI and multilib support for Armv8.1-M Mainline MVE extensions

gcc/ChangeLog:

2020-01-16  Mihail Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Andre Vieira  <andre.simoesdiasvieira@arm.com>

* config/arm/arm-cpus.in (mve, mve_float): New features.
(dsp, mve, mve.fp): New options.
* config/arm/arm.h (TARGET_HAVE_MVE, TARGET_HAVE_MVE_FLOAT): Define.
* config/arm/t-rmprofile: Map v8.1-M multilibs to v8-M.
* doc/invoke.texi: Document the armv8.1-m mve and dps options.

gcc/testsuite/ChangeLog:

2020-01-16  Mihail Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Andre Vieira  <andre.simoesdiasvieira@arm.com>

* testsuite/gcc.target/arm/multilib.exp: Add v8.1-M entries.

4 years ago[PATCH, GCC/ARM, 10/10] Enable -mcmse
Mihail Ionescu [Wed, 15 Jan 2020 11:49:32 +0000 (11:49 +0000)]
[PATCH, GCC/ARM, 10/10] Enable -mcmse

The patch is straightforward: it redefines ARMv8_1m_main as having the
same features as ARMv8m_main (and thus as having the cmse feature) with
the extra features represented by armv8_1m_main.  It also removes the
error for using -mcmse on Armv8.1-M Mainline.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm-cpus.in (ARMv8_1m_main): Redefine as an extension to
Armv8-M Mainline.
* config/arm/arm.c (arm_options_perform_arch_sanity_checks): Remove
error for using -mcmse when targeting Armv8.1-M Mainline.

4 years ago[PATCH, GCC/ARM, 9/10] Call nscall function with blxns
Mihail Ionescu [Wed, 15 Jan 2020 11:45:53 +0000 (11:45 +0000)]
[PATCH, GCC/ARM, 9/10] Call nscall function with blxns

This change to use BLXNS to call a nonsecure function from secure
directly (not using a libcall) is made in 2 steps:
- change nonsecure_call patterns to use blxns instead of calling
  __gnu_cmse_nonsecure_call
- loosen requirement for function address to allow any register when
  doing BLXNS.

The former is a straightforward check over whether instructions added in
Armv8.1-M Mainline are available while the latter consist in making the
nonsecure call pattern accept any register by using match_operand and
changing the nonsecure_call_internal expander to no force r4 when
targeting Armv8.1-M Mainline.

The tricky bit is actually in the test update, specifically how to check
that register lists for CLRM have all registers except for the one
holding parameters (already done) and the one holding the address used
by BLXNS. This is achieved with 3 scan-assembler directives.

1) The first one lists all registers that can appear in CLRM but make
   each of them optional.
   Property guaranteed: no wrong register is cleared and none appears
   twice in the register list.
2) The second directive check that the CLRM is made of a fixed number
   of the right registers to be cleared. The number used is the number
   of registers that could contain a secret minus one (used to hold the
   address of the function to call.
   Property guaranteed: register list has the right number of registers
   Cumulated property guaranteed: only registers with a potential secret
   are cleared and they are all listed but ont
3) The last directive checks that we cannot find a CLRM with a register
   in it that also appears in BLXNS. This is check via the use of a
   back-reference on any of the allowed register in CLRM, the
   back-reference enforcing that whatever register match in CLRM must be
   the same in the BLXNS.
   Property guaranteed: register used for BLXNS is different from
   registers cleared in CLRM.

Some more care needs to happen for the gcc.target/arm/cmse/cmse-1.c
testcase due to there being two CLRM generated. To ensure the third
directive match the right CLRM to the BLXNS, a negative lookahead is
used between the CLRM register list and the BLXNS. The way negative
lookahead work is by matching the *position* where a given regular
expression does not match. In this case, since it comes after the CLRM
register list it is requesting that what comes after the register list
does not have a CLRM again followed by BLXNS. This guarantees that the
.*blxns after only matches a blxns without another CLRM before.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm.md (nonsecure_call_internal): Do not force memory
address in r4 when targeting Armv8.1-M Mainline.
(nonsecure_call_value_internal): Likewise.
* config/arm/thumb2.md (nonsecure_call_reg_thumb2): Make memory address
a register match_operand again.  Emit BLXNS when targeting
Armv8.1-M Mainline.
(nonsecure_call_value_reg_thumb2): Likewise.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/cmse-1.c: Add check for BLXNS when instructions
introduced in Armv8.1-M Mainline Security Extensions are available and
restrict checks for libcall to __gnu_cmse_nonsecure_call to Armv8-M
targets only.  Adapt CLRM check to verify register used for BLXNS is
not in the CLRM register list.
* gcc.target/arm/cmse/cmse-14.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-4.c: Likewise and adapt
check for LSB clearing bit to be using the same register as BLXNS when
targeting Armv8.1-M Mainline.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-6.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-9.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-and-union.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/union-1.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/union-2.c: Likewise.
* gcc.target/arm/cmse/cmse-15.c: Count BLXNS when targeting Armv8.1-M
Mainline and restrict libcall count to Armv8-M.

4 years ago[PATCH, GCC/ARM, 8/10] Do lazy store & load inline when calling nscall function
Mihail Ionescu [Wed, 15 Jan 2020 11:35:21 +0000 (11:35 +0000)]
[PATCH, GCC/ARM, 8/10] Do lazy store & load inline when calling nscall function

This patch adds two new patterns for the VLSTM and VLLDM instructions.
cmse_nonsecure_call_inline_register_clear is then modified to
generate VLSTM and VLLDM respectively before and after calls to
functions with the cmse_nonsecure_call attribute in order to have lazy
saving, clearing and restoring of VFP registers. Since these
instructions do not do writeback of the base register, the stack is adjusted
prior the lazy store and after the lazy load with appropriate frame
debug notes to describe the effect on the CFA register.

As with CLRM, VSCCLRM and VSTR/VLDR, the instruction is modeled as an
unspecified operation to the memory pointed to by the base register.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm.c (arm_add_cfa_adjust_cfa_note): Declare early.
(cmse_nonsecure_call_inline_register_clear): Define new lazy_fpclear
variable as true when floating-point ABI is not hard.  Replace
check against TARGET_HARD_FLOAT_ABI by checks against lazy_fpclear.
Generate VLSTM and VLLDM instruction respectively before and
after a function call to cmse_nonsecure_call function.
* config/arm/unspecs.md (VUNSPEC_VLSTM): Define unspec.
(VUNSPEC_VLLDM): Likewise.
* config/arm/vfp.md (lazy_store_multiple_insn): New define_insn.
(lazy_load_multiple_insn): Likewise.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13.c: Add check for VLSTM and
VLLDM.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-8.c: Likewise.

4 years ago[PATCH, GCC/ARM, 7/10] Clear all VFP regs inline in hardfloat nscall functions
Mihail Ionescu [Wed, 15 Jan 2020 11:33:30 +0000 (11:33 +0000)]
[PATCH, GCC/ARM, 7/10] Clear all VFP regs inline in hardfloat nscall functions

The patch is fairly straightforward in its approach and consist of the
following 3 logical changes:
- abstract the number of floating-point register to clear in
  max_fp_regno
- use max_fp_regno to decide how many registers to clear so that the
  same code works for Armv8-M and Armv8.1-M Mainline
- emit vpush and vpop instruction respectively before and after a
  nonsecure call

Note that as in the patch to clear GPRs inline, debug information has to
be disabled for VPUSH and VPOP due to VPOP adding CFA adjustment note
for SP when R7 is sometimes used as CFA.

ChangeLog entries are as follows:

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm.c (vfp_emit_fstmd): Declare early.
(arm_emit_vfp_multi_reg_pop): Likewise.
(cmse_nonsecure_call_inline_register_clear): Abstract number of VFP
registers to clear in max_fp_regno.  Emit VPUSH and VPOP to save and
restore callee-saved VFP registers.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: Add check for
VPUSH and VPOP and update expectation for VSCCLRM.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: Likewise.

4 years ago[PATCH, GCC/ARM, 6/10] Clear GPRs inline when calling nscall function
Mihail Ionescu [Wed, 15 Jan 2020 11:31:35 +0000 (11:31 +0000)]
[PATCH, GCC/ARM, 6/10] Clear GPRs inline when calling nscall function

Besides changing the set of registers that needs to be cleared inline,
this patch also generates the push and pop to save and restore
callee-saved registers without trusting the callee inline. To make the
code more future-proof, this (currently) Armv8.1-M specific behavior is
expressed in terms of clearing of callee-saved registers rather than
directly based on the targets.

The patch contains 1 subtlety:

Debug information is disabled for push and pop because the
REG_CFA_RESTORE notes used to describe popping of registers do not stack.
Instead, they just reset the debug state for the register to the one at
the beginning of the function, which is incorrect for a register that is
pushed twice (in prologue and before nonsecure call) and then popped for
the first time. In particular, this occasionally trips CFI note creation
code when there are two codepaths to the epilogue, one of which does not
go through the nonsecure call. Obviously this mean that debugging
between the push and pop is not reliable.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm.c (arm_emit_multi_reg_pop): Declare early.
(cmse_nonsecure_call_clear_caller_saved): Rename into ...
(cmse_nonsecure_call_inline_register_clear): This.  Save and clear
callee-saved GPRs as well as clear ip register before doing a nonsecure
call then restore callee-saved GPRs after it when targeting
Armv8.1-M Mainline.
(arm_reorg): Adapt to function rename.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/cmse-1.c: Add check for PUSH and POP and update
CLRM check.
* gcc.target/arm/cmse/cmse-14.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-4.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-6.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-9.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-and-union.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/union-1.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/union-2.c: Likewise.

4 years ago[PATCH, GCC/ARM, 5/10] Clear VFP registers with VSCCLRM
Mihail Ionescu [Wed, 15 Jan 2020 11:29:42 +0000 (11:29 +0000)]
[PATCH, GCC/ARM, 5/10] Clear VFP registers with VSCCLRM

This patch adds a new pattern for the VSCCLRM instruction.
cmse_clear_registers () is then modified to use the new VSCCLRM
instruction when targeting Armv8.1-M Mainline, thus, making the Armv8-M
register clearing code specific to Armv8-M.

Since the VSCCLRM instruction mandates VPR in the register list, the
pattern is encoded with a parallel which only requires an unspecified
VUNSPEC_CLRM_VPR constant modelling the APSR clearing. Other expression
in the parallel are expected to be set expression for clearing the VFP
registers.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm-protos.h (clear_operation_p): Adapt prototype.
* config/arm/arm.c (clear_operation_p): Extend to be able to check a
clear_vfp_multiple pattern based on a new vfp parameter.
(cmse_clear_registers): Generate VSCCLRM to clear VFP registers when
targeting Armv8.1-M Mainline.
(cmse_nonsecure_entry_clear_before_return): Clear VFP registers
unconditionally when targeting Armv8.1-M Mainline architecture.  Check
whether VFP registers are available before looking call_used_regs for a
VFP register.
* config/arm/predicates.md (clear_multiple_operation): Adapt to change
of prototype of clear_operation_p.
(clear_vfp_multiple_operation): New predicate.
* config/arm/unspecs.md (VUNSPEC_VSCCLRM_VPR): New volatile unspec.
* config/arm/vfp.md (clear_vfp_multiple): New define_insn.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/bitfield-1.c: Add check for VSCCLRM.
* gcc.target/arm/cmse/bitfield-2.c: Likewise.
* gcc.target/arm/cmse/bitfield-3.c: Likewise.
* gcc.target/arm/cmse/cmse-1.c: Likewise.
* gcc.target/arm/cmse/struct-1.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-5.c: Likewise.

4 years ago[PATCH, GCC/ARM, 4/10] Clear GPR with CLRM
Mihail Ionescu [Wed, 15 Jan 2020 10:38:44 +0000 (10:38 +0000)]
[PATCH, GCC/ARM, 4/10] Clear GPR with CLRM

This patch adds a new pattern for the CLRM instruction and guards the
current clearing code in output_return_instruction() and thumb_exit()
on Armv8.1-M Mainline instructions not being present.
cmse_clear_registers () is then modified to use the new CLRM instruction
when targeting Armv8.1-M Mainline while keeping Armv8-M register
clearing code for VFP registers.

For the CLRM instruction, which does not mandated APSR in the register
list, checking whether it is the right volatile unspec or a clearing
register is done in clear_operation_p.

Note that load/store multiple were deemed sufficiently different in
terms of RTX structure compared to the CLRM pattern for a different
function to be used to validate the match_parallel.

ChangeLog entries are as follows:

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm-protos.h (clear_operation_p): Declare.
* config/arm/arm.c (clear_operation_p): New function.
(cmse_clear_registers): Generate clear_multiple instruction pattern if
targeting Armv8.1-M Mainline or successor.
(output_return_instruction): Only output APSR register clearing if
Armv8.1-M Mainline instructions not available.
(thumb_exit): Likewise.
* config/arm/predicates.md (clear_multiple_operation): New predicate.
* config/arm/thumb2.md (clear_apsr): New define_insn.
(clear_multiple): Likewise.
* config/arm/unspecs.md (VUNSPEC_CLRM_APSR): New volatile unspec.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/bitfield-1.c: Add check for CLRM.
* gcc.target/arm/cmse/bitfield-2.c: Likewise.
* gcc.target/arm/cmse/bitfield-3.c: Likewise.
* gcc.target/arm/cmse/struct-1.c: Likewise.
* gcc.target/arm/cmse/cmse-14.c: Likewise.
* gcc.target/arm/cmse/cmse-1.c: Likewise.  Restrict checks for Armv8-M
GPR clearing when CLRM is not available.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-4.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-6.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-9.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-5.c: likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: likewise.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-5.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/union-1.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/union-2.c: Likewise.

4 years ago[PATCH, GCC/ARM, 3/10] Save/restore FPCXTNS in nsentry functions
Mihail Ionescu [Wed, 15 Jan 2020 10:33:52 +0000 (10:33 +0000)]
[PATCH, GCC/ARM, 3/10] Save/restore FPCXTNS in nsentry functions

This patch consists mainly of creating 2 new instruction patterns to
push and pop special FP registers via vldm and vstr and using them in
prologue and epilogue. The patterns are defined as push/pop with an
unspecified operation on the memory accessed, with an unspecified
constant indicating what special FP register is being saved/restored.

Other aspects of the patch include:
  * defining the set of special registers that can be saved/restored and
    their name
  * reserving space in the stack frames for these push/pop
  * preventing return via pop
  * guarding the clearing of FPSCR to target architecture not having
    Armv8.1-M Mainline instructions.

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/arm.c (fp_sysreg_names): Declare and define.
(use_return_insn): Also return false for Armv8.1-M Mainline.
(output_return_instruction): Skip FPSCR clearing if Armv8.1-M
Mainline instructions are available.
(arm_compute_frame_layout): Allocate space in frame for FPCXTNS
when targeting Armv8.1-M Mainline Security Extensions.
(arm_expand_prologue): Save FPCXTNS if this is an Armv8.1-M
Mainline entry function.
(cmse_nonsecure_entry_clear_before_return): Clear IP and r4 if
targeting Armv8.1-M Mainline or successor.
(arm_expand_epilogue): Fix indentation of caller-saved register
clearing.  Restore FPCXTNS if this is an Armv8.1-M Mainline
entry function.
* config/arm/arm.h (TARGET_HAVE_FP_CMSE): New macro.
(FP_SYSREGS): Likewise.
(enum vfp_sysregs_encoding): Define enum.
(fp_sysreg_names): Declare.
* config/arm/unspecs.md (VUNSPEC_VSTR_VLDR): New volatile unspec.
* config/arm/vfp.md (push_fpsysreg_insn): New define_insn.
(pop_fpsysreg_insn): Likewise.

*** gcc/testsuite/Changelog ***

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* gcc.target/arm/cmse/bitfield-1.c: add checks for VSTR and VLDR.
* gcc.target/arm/cmse/bitfield-2.c: Likewise.
* gcc.target/arm/cmse/bitfield-3.c: Likewise.
* gcc.target/arm/cmse/cmse-1.c: Likewise.
* gcc.target/arm/cmse/struct-1.c: Likewise.
* gcc.target/arm/cmse/cmse.exp: Run existing Armv8-M Mainline tests
from mainline/8m subdirectory and new Armv8.1-M Mainline tests from
mainline/8_1m subdirectory.
* gcc.target/arm/cmse/mainline/bitfield-4.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-4.c: This.
* gcc.target/arm/cmse/mainline/bitfield-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-5.c: This.
* gcc.target/arm/cmse/mainline/bitfield-6.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-6.c: This.
* gcc.target/arm/cmse/mainline/bitfield-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-7.c: This.
* gcc.target/arm/cmse/mainline/bitfield-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-8.c: This.
* gcc.target/arm/cmse/mainline/bitfield-9.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-9.c: This.
* gcc.target/arm/cmse/mainline/bitfield-and-union-1.c: Move and rename
into ...
* gcc.target/arm/cmse/mainline/8m/bitfield-and-union.c: This.
* gcc.target/arm/cmse/mainline/hard-sp/cmse-13.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard-sp/cmse-13.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard-sp/cmse-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard-sp/cmse-5.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard-sp/cmse-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard-sp/cmse-7.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard-sp/cmse-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard-sp/cmse-8.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard/cmse-13.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard/cmse-13.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard/cmse-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard/cmse-5.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard/cmse-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard/cmse-7.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/hard/cmse-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/hard/cmse-8.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/soft/cmse-13.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/soft/cmse-13.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/soft/cmse-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/soft/cmse-5.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/soft/cmse-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/soft/cmse-7.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/soft/cmse-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/soft/cmse-8.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp-sp/cmse-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp-sp/cmse-5.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp-sp/cmse-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp-sp/cmse-7.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp-sp/cmse-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp-sp/cmse-8.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp/cmse-13.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp/cmse-13.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp/cmse-5.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp/cmse-5.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp/cmse-7.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp/cmse-7.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/softfp/cmse-8.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/softfp/cmse-8.c: This.  Clean up
dg-skip-if directive for float ABI.
* gcc.target/arm/cmse/mainline/union-1.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/union-1.c: This.
* gcc.target/arm/cmse/mainline/union-2.c: Move into ...
* gcc.target/arm/cmse/mainline/8m/union-2.c: This.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-4.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-6.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-9.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/bitfield-and-union.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-13.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard-sp/cmse-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-13.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/hard/cmse-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-5.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/union-1.c: New file.
* gcc.target/arm/cmse/mainline/8_1m/union-2.c: New file.
* lib/target-supports.exp (check_effective_target_arm_cmse_clear_ok):
New procedure.

4 years ago[PATCH, GCC/ARM, 2/10] Add command line support for Armv8.1-M Mainline
Mihail Ionescu [Tue, 14 Jan 2020 18:00:35 +0000 (18:00 +0000)]
[PATCH, GCC/ARM, 2/10] Add command line support for Armv8.1-M Mainline

Besides the expected enabling of the new value for the -march
command-line option (-march=armv8.1-m.main) and its extensions (see
below), this patch disables support of the Security Extensions for this
newly added architecture. This is done both by not including the cmse
bit in the architecture description and by throwing an error message
when user request Armv8.1-M Mainline Security Extensions. Note that
Armv8-M Baseline and Mainline Security Extensions are still enabled.

Only extensions for already supported instructions are implemented in
this patch. Other extensions (MVE integer and float) will be added in
separate patches. The following configurations are allowed for Armv8.1-M
Mainline with regards to FPU and implemented in this patch:
+ no FPU (+nofp)
+ single precision VFPv5 with FP16 (+fp)
+ double precision VFPv5 with FP16 (+fp.dp)

ChangeLog entry are as follow:

*** gcc/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme <thomas.preudhomme@arm.com>

* config/arm/arm-cpus.in (armv8_1m_main): New feature.
(ARMv4, ARMv4t, ARMv5t, ARMv5te, ARMv5tej, ARMv6, ARMv6j, ARMv6k,
ARMv6z, ARMv6kz, ARMv6zk, ARMv6t2, ARMv6m, ARMv7, ARMv7a, ARMv7ve,
ARMv7r, ARMv7m, ARMv7em, ARMv8a, ARMv8_1a, ARMv8_2a, ARMv8_3a,
ARMv8_4a, ARMv8_5a, ARMv8m_base, ARMv8m_main, ARMv8r): Reindent.
(ARMv8_1m_main): New feature group.
(armv8.1-m.main): New architecture.
* config/arm/arm-tables.opt: Regenerate.
* config/arm/arm.c (arm_arch8_1m_main): Define and default initialize.
(arm_option_reconfigure_globals): Initialize arm_arch8_1m_main.
(arm_options_perform_arch_sanity_checks): Error out when targeting
Armv8.1-M Mainline Security Extensions.
* config/arm/arm.h (arm_arch8_1m_main): Declare.

*** gcc/testsuite/ChangeLog ***

2020-01-16  Mihail-Calin Ionescu <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme <thomas.preudhomme@arm.com>

* lib/target-supports.exp
(check_effective_target_arm_arch_v8_1m_main_ok): Define.
(add_options_for_arm_arch_v8_1m_main): Likewise.
(check_effective_target_arm_arch_v8_1m_main_multilib): Likewise.

4 years ago[PATCH, GCC/ARM, 1/10] Fix -mcmse check in libgcc
Mihail Ionescu [Tue, 14 Jan 2020 17:47:39 +0000 (17:47 +0000)]
[PATCH, GCC/ARM, 1/10] Fix -mcmse check in libgcc

This patch is part of a patch series to add support for Armv8.1-M
Mainline Security Extensions architecture.

Code to detect whether cmse.c can be buit with -mcmse checks the output
of host GCC when invoked with -mcmse. However, an error from the
compiler does not prevent some minimal output so this always holds true.

2020-01-16  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
2020-01-16  Thomas Preud'homme  <thomas.preudhomme@arm.com>

* config/arm/t-arm: Check return value of gcc rather than lack of
output.

4 years agogcc-git-customization.sh: Avoid double expansion
Andreas Schwab [Thu, 16 Jan 2020 15:00:33 +0000 (16:00 +0100)]
gcc-git-customization.sh: Avoid double expansion

4 years agogcc-git-customization.sh: avoid double expansion
Andreas Schwab [Thu, 16 Jan 2020 14:26:31 +0000 (15:26 +0100)]
gcc-git-customization.sh: avoid double expansion

4 years ago[GCC][PATCH][AArch64]Add ACLE intrinsics for bfdot for ARMv8.6 Extension
Stam Markianos-Wright [Thu, 16 Jan 2020 14:47:30 +0000 (14:47 +0000)]
[GCC][PATCH][AArch64]Add ACLE intrinsics for bfdot for ARMv8.6 Extension

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* config/aarch64/aarch64-simd-builtins.def (aarch64_bfdot,
aarch64_bfdot_lane, aarch64_bfdot_laneq): New.
* config/aarch64/aarch64-simd.md (aarch64_bfdot, aarch64_bfdot_lane,
aarch64_bfdot_laneq): New.
* config/aarch64/arm_bf16.h (vbfdot_f32, vbfdotq_f32,
vbfdot_lane_f32, vbfdotq_lane_f32, vbfdot_laneq_f32,
vbfdotq_laneq_f32): New.
* config/aarch64/iterators.md (UNSPEC_BFDOT, Vbfdottype,
VBFMLA_W, VBF): New.
(isquadop): Add V4BF, V8BF.

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* gcc.target/aarch64/advsimd-intrinsics/bfdot-1.c: New.
* gcc.target/aarch64/advsimd-intrinsics/bfdot-2.c: New.
* gcc.target/aarch64/advsimd-intrinsics/bfdot-3.c: New.

4 years agolibstdc++: Improve unordered containers == operator (PR 91263)
François Dumont [Thu, 16 Jan 2020 08:34:21 +0000 (08:34 +0000)]
libstdc++: Improve unordered containers == operator (PR 91263)

Avoid comparing elements with operator== multiple times by replacing
uses of find and equal_range with equivalent inlined code that uses
operator== instead of the container's equality comparison predicate.
This is valid because the standard requires that operator== is a
refinement of the equality predicate.

Also replace the _S_is_permutation function with std::is_permutation,
which wasn't yet implemented when this code was first written.

PR libstdc++/91263
* include/bits/hashtable.h (_Hashtable<>): Make _Equality<> friend.
* include/bits/hashtable_policy.h: Include <bits/stl_algo.h>.
(_Equality_base): Remove.
(_Equality<>::_M_equal): Review implementation. Use
std::is_permutation.
* testsuite/23_containers/unordered_multiset/operators/1.cc
(Hash, Equal, test02, test03): New.
* testsuite/23_containers/unordered_set/operators/1.cc
(Hash, Equal, test02, test03): New.

4 years ago[GCC][PATCH][AArch64]Add ACLE intrinsics for dot product (usdot - vector, <us/su...
Stam Markianos-Wright [Thu, 16 Jan 2020 14:20:48 +0000 (14:20 +0000)]
[GCC][PATCH][AArch64]Add ACLE intrinsics for dot product (usdot - vector, <us/su>dot - by element) for AArch64 AdvSIMD ARMv8.6 Extension

gcc/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* config/aarch64/aarch64-builtins.c: (enum aarch64_type_qualifiers):
New qualifier_lane_quadtup_index, TYPES_TERNOP_SSUS,
TYPES_QUADOPSSUS_LANE_QUADTUP, TYPES_QUADOPSSSU_LANE_QUADTUP.
(aarch64_simd_expand_args): Add case SIMD_ARG_LANE_QUADTUP_INDEX.
(aarch64_simd_expand_builtin): Add qualifier_lane_quadtup_index.
* config/aarch64/aarch64-simd-builtins.def (usdot, usdot_lane,
usdot_laneq, sudot_lane,sudot_laneq): New.
* config/aarch64/aarch64-simd.md (aarch64_usdot): New.
(aarch64_<sur>dot_lane): New.
* config/aarch64/arm_neon.h (vusdot_s32): New.
(vusdotq_s32): New.
(vusdot_lane_s32): New.
(vsudot_lane_s32): New.
* config/aarch64/iterators.md (DOTPROD_I8MM): New iterator.
(UNSPEC_USDOT, UNSPEC_SUDOT): New unspecs.

gcc/testsuite/ChangeLog:

2020-01-16  Stam Markianos-Wright  <stam.markianos-wright@arm.com>

* gcc.target/aarch64/advsimd-intrinsics/vdot-compile-3-1.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/vdot-compile-3-2.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/vdot-compile-3-3.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/vdot-compile-3-4.c: New test.

4 years agocontrib: Check and if needed set user.name and user.email in gcc-git-customization.sh
Richard Earnshaw [Thu, 16 Jan 2020 13:48:37 +0000 (13:48 +0000)]
contrib: Check and if needed set user.name and user.email in gcc-git-customization.sh

As discussed on IRC, this adds a couple more checks in the
customization setup for git.  If the variables user.name and
user.email are not set anywhere in the git config hierarchy, we set
some local values.  We always ask about the values we detect and if
the user gives an answer that is new, we save that in the local
config: this gives the opportunity to use different values to those
configured for the global space.

Also cleaned up a couple of minor niggles, such as using $(cmd) rather
than `cmd` for subshells and some quoting issues when using eval.

* gcc-git-customization.sh: Check that user.name and user.email
are set.  Use $(cmd) instead of `cmd`.  Fix variable quoting when
using eval.

4 years agoFix spacing in a dump in value-prof.c.
Martin Liska [Thu, 16 Jan 2020 13:33:08 +0000 (14:33 +0100)]
Fix spacing in a dump in value-prof.c.

* value-prof.c (dump_histogram_value): Fix
obvious spacing issue.

4 years agoFix value numbering dealing with reverse byte order
Andrew Pinski [Thu, 16 Jan 2020 07:54:51 +0000 (07:54 +0000)]
Fix value numbering dealing with reverse byte order

Hi,
  While working on bit-field lowering pass, I came across this bug.
The IR looks like:
  VIEW_CONVERT_EXPR<unsigned long>(var1) = _12;
  _1 = BIT_FIELD_REF <var1, 64, 0>;

Where the BIT_FIELD_REF has REF_REVERSE_STORAGE_ORDER set on it
and var1's type has TYPE_REVERSE_STORAGE_ORDER set on it.
PRE/FRE would decided to prop _12 into the BFR statement
which would produce wrong code.
And yes _12 has the correct byte order already; bit-field lowering
removes the implicit byte swaps in the IR and adds the explicity
to make it easier optimize later on.

This patch adds a check for storage_order_barrier_p on the lhs tree
which returns true in the case where we had a reverse order with a VCE.

ChangeLog:
* tree-ssa-sccvn.c(vn_reference_lookup_3): Check lhs for
!storage_order_barrier_p.

4 years agoUninitialized padding in struct _dep.
Andrew Pinski [Thu, 16 Jan 2020 02:07:28 +0000 (02:07 +0000)]
Uninitialized padding in struct _dep.

In struct _dep, there is an implicit padding of 4bits.  This
bit-field padding is uninitialized when init_dep_1 is being called.
This means we access uninitialized memory but never use it for
anything.  Adding an unused bit-field field and initializing it
in init_dep_1 will improve code generation also as we initialize
the whole 32bits now rather than just part of it.

ChangeLog:
* sched-int.h (_dep): Add unused bit-field field for the padding.
* sched-deps.c (init_dep_1): Init unused field.

4 years agoFix uninitialized field in expand_operand.
Andrew Pinski [Thu, 16 Jan 2020 01:52:12 +0000 (01:52 +0000)]
Fix uninitialized field in expand_operand.

Commit g:f96bf49a0 added the target field to expand_operand.
But it leaves it uninitialized when doing a full initialization
inside create_expand_operand.  This fixes the problem and improves
the code generation inside create_expand_operand too.

ChangeLog:
* optabs.h (create_expand_operand): Initialize target field also.