platform/upstream/gcc.git
2 years agoImprove handling of table overflows in modref_ref_node
Jan Hubicka [Sat, 28 Aug 2021 18:57:08 +0000 (20:57 +0200)]
Improve handling of table overflows in modref_ref_node

gcc/ChangeLog:

* ipa-modref-tree.h (modref_access_node::merge): Break out
logic combining offsets and logic merging ranges to ...
(modref_access_node::combined_offsets): ... here
(modref_access_node::update2): ... here
(modref_access_node::closer_pair_p): New member function.
(modref_access_node::forced_merge): New member function.
(modre_ref_node::insert): Do merging when table is full.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/modref-9.c: New test.

2 years agoFortran - reject function entries with mismatched characteristics
Harald Anlauf [Sat, 28 Aug 2021 18:09:44 +0000 (20:09 +0200)]
Fortran - reject function entries with mismatched characteristics

gcc/fortran/ChangeLog:

PR fortran/87737
* resolve.c (resolve_entries): For functions of type CHARACTER
tighten the checks for matching characteristics.

gcc/testsuite/ChangeLog:

PR fortran/87737
* gfortran.dg/entry_24.f90: New test.

2 years agolibstdc++: Fix std::allocator<void> for versioned namespace
Jonathan Wakely [Sat, 28 Aug 2021 10:05:58 +0000 (11:05 +0100)]
libstdc++: Fix std::allocator<void> for versioned namespace

Removing the allocator<void> specialization for the versioned namespace
breaks _Extptr_allocator<void> because the allocator<void>
specialization was still declared in <bits/memoryfwd.h>, making it an
incomplete type.  It wrong to remove that specialization anyway, because
it is still needed pre-C++20.

This removes the #if ! _GLIBCXX_INLINE_VERSION check, so that
allocator<void> is still explicitly specialized for the versioned
namespace, consistent with the normal unversioned namespace mode.

To make _Extptr_allocator<void> usable as a ProtoAllocator, this change
adds a default constructor and converting constructor. That is
consistent with std::allocator<void> since C++20 (and harmless to do for
earlier standards).

I'm also explicitly specializing allocator_traits<allocator<void>> so
that it doesn't need to use allocator<void>::construct and destroy.
Doing that allows those members to be removed, further simplifying
allocator<void>.  That new explicit specialization can delete the
allocate, deallocate and max_size members, which are always ill-formed
for allocator<void>.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/alloc_traits.h (allocator_traits): Add explicit
specialization for allocator<void>. Improve doxygen comments.
* include/bits/allocator.h (allocator<void>): Restore for the
versioned namespace.
(allocator<void>::construct, allocator<void>::destroy): Remove.
* include/ext/extptr_allocator.h (_Extptr_allocator<void>):
Add default constructor and converting constructor.

2 years agolibstdc++: Fix comment typo
Jonathan Wakely [Fri, 27 Aug 2021 21:10:43 +0000 (22:10 +0100)]
libstdc++: Fix comment typo

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/stl_uninitialized.h: Fix typo in comment.

2 years agolibstdc++: Fix inefficiency in filesystem::absolute [PR99876]
Jonathan Wakely [Fri, 27 Aug 2021 09:59:54 +0000 (10:59 +0100)]
libstdc++: Fix inefficiency in filesystem::absolute [PR99876]

When the path is already absolute, the call to current_path() is
wasteful, because operator/ will ignore the left operand anyway.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/99876
* src/c++17/fs_ops.cc (fs::absolute): Call non-throwing form,
to avoid unnecessary current_path() call.

2 years agolibstdc++: Name std::function template parameter
Jonathan Wakely [Thu, 26 Aug 2021 23:20:31 +0000 (00:20 +0100)]
libstdc++: Name std::function template parameter

This avoids "<template-parameter-2-2>" being shown in the diagnostics
for ill-formed uses of std::function constructor:

In instantiation of 'std::function<_Res(_ArgTypes ...)>::function(_Functor&&)
[with _Functor = f(f()::_Z1fv.frame*)::<lambda()>;
<template-parameter-2-2> = void; _Res = void; _ArgTypes = {}]'

Instead we get:

In instantiation of 'std::function<_Res(_ArgTypes ...)>::function(_Functor&&)
[with _Functor = f(f()::_Z1fv.frame*)::<lambda()>;
_Constraints = void; _Res = void; _ArgTypes = {}]'

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/std_function.h (function::function(F&&)): Give
name to defaulted template parameter, to improve diagnostics.
Use markdown for more doxygen comments.

2 years agoMIPS: use N64 ABI by default if the triple end with -gnuabi64
YunQiang Su [Fri, 27 Aug 2021 03:56:58 +0000 (23:56 -0400)]
MIPS: use N64 ABI by default if the triple end with -gnuabi64

gcc/ChangeLog:

PR target/102089
* config.gcc: MIPS: use N64 ABI by default if the triple end
  with -gnuabi64, which is used by Debian since 2013.

2 years agofix latent bootstrap-debug issue
Alexandre Oliva [Sat, 28 Aug 2021 03:40:14 +0000 (00:40 -0300)]
fix latent bootstrap-debug issue

I've hit a bootstrap-debug error involving large subprograms in
gcc/ada/sem_ch12.adb.  I'm afraid I couldn't narrow it down to a
reasonable testcase.

thread1 made different decisions about a block containing a
builtin_eh_filter call because in one compilation, estimate_num_insns
found a cgraph_node for the builtin and could thus get to the
is_simple_builtin test, but in the other it didn't.  With different
insn counts, one stage jump-threaded and the other didn't, and the
resulting code diverged quite a bit.

The reason the builtin had a cgraph_node in one case but not the other
was that modref got a chance to analyze the builtin call when it was
the first stmt in the block, and that created the cgraph_node.
However, when it was preceded by debug stmts, the loop in
analyze_function was cut short after the first debug stmt, because the
summary so far was not useful.

This patch fixes both issues: skip debug stmts in the analyze_function
loop, so as to prevent them from affecting any decisions in the loop,
and enable the insn count estimator to get to the is_simple_builtin
test when a cgraph_node has not been created for the builtin.

for  gcc/ChangeLog

* ipa-modref.c (analyze_function): Skip debug stmts.
* tree-inline.c (estimate_num_insn): Consider builtins even
without a cgraph_node.

2 years agoDaily bump.
GCC Administrator [Sat, 28 Aug 2021 00:16:42 +0000 (00:16 +0000)]
Daily bump.

2 years agoc++: Set type on dependent ARROW_EXPR
Jason Merrill [Fri, 27 Aug 2021 14:00:49 +0000 (10:00 -0400)]
c++: Set type on dependent ARROW_EXPR

Even if the operand of -> has dependent type, if it's a pointer we know
that the result will be the target type of that pointer.  This should avoid
some unnecessary TYPEOF_EXPR when looking up a name after ->.

gcc/cp/ChangeLog:

* typeck2.c (build_x_arrow): Do set TREE_TYPE when operand is
a dependent pointer.

2 years agoSupport limited setcc for H8
Jeff Law [Fri, 27 Aug 2021 21:01:37 +0000 (17:01 -0400)]
Support limited setcc for H8

gcc/

* config/h8300/bitfield.md (cstore<mode>4): Remove expander.
* config/h8300/h8300.c (h8300_expand_branch): Remove function.
* config/h8300/h8300-protos.h (h8300_expadn_branch): Remove prototype.
* config/h8300/h8300.md (eqne): New code iterator.
(geultu, geultu_to_c): Similarly.
* config/h8300/testcompare.md (cstore<mode>4): Dummy expander.
(store_c_<mode>, store_c_i_<mode>): New define_insn_and_splits
(cmp<mode>_c): New pattern

2 years agoUpdate comments in float128-call.c test.
Michael Meissner [Fri, 27 Aug 2021 20:07:50 +0000 (16:07 -0400)]
Update comments in float128-call.c test.

Segher asked that I update the comments to include the d-form vector stores
(even though they wouldn't be generated by this test).

2021-08-25  Michael Meissner  <meissner@linux.ibm.com>

gcc/testsuite/
* gcc.target/powerpc/float128-call.c: Update comments.

2 years agoReduce vector comparison of uniform vectors to a scalar comparison
Jeff Law [Fri, 27 Aug 2021 19:27:38 +0000 (15:27 -0400)]
Reduce vector comparison of uniform vectors to a scalar comparison

gcc/
* tree-ssa-dom.c (reduce_vector_comparison_to_scalar_comparison): New
function.
(dom_opt_dom_walker::optimize_stmt): Use it.

2 years agoFix float128-call.c test for power8 IEEE 128 and power10.
Michael Meissner [Fri, 27 Aug 2021 17:19:54 +0000 (13:19 -0400)]
Fix float128-call.c test for power8 IEEE 128 and power10.

I built a compiler on a little endian power8 system where the default long
double was IEEE 128-bit instead of IBM 128-bit.  I discovered that on
power8, we would generate a lxvd2x and xxpermdi to deal with the endianess
instead of the Altivec lxv.

In addition, I noticed the constant that was being loaded (1.0q) could be
loaded by the lxvkq instruction.

I rewrote the test to handle all forms of vector load and store that can
be generated.

2021-08-27  Michael Meissner  <meissner@linux.ibm.com>

gcc/testsuite/
* gcc.target/powerpc/float128-call.c: Fix test for IEEE 128-bit
long double and power10.

2 years agoDarwin : Mark the mod init/term section starts with a linker-visible sym.
Iain Sandoe [Sat, 29 Aug 2020 21:05:30 +0000 (22:05 +0100)]
Darwin : Mark the mod init/term section starts with a linker-visible sym.

Some newer assemblers emit section start temp symbols for mod init and term
sections if there is no suitable symbol present already.
The temp symbols are linker visible and therefore appear in the symbol tables.
Since the temp symbol number can vary when debug is enabled, that causes
compare-debug fails.  The solution is to provide a stable linker-visible
symbol.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* config/darwin.c (finalize_ctors): Add a section-start linker-
visible symbol.
(finalize_dtors): Likewise.
* config/darwin.h (MIN_LD64_INIT_TERM_START_LABELS): New.

2 years agors6000: Execute the automatic built-in initialization code
Bill Schmidt [Fri, 27 Aug 2021 14:55:29 +0000 (09:55 -0500)]
rs6000: Execute the automatic built-in initialization code

2021-08-27  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-call.c (rs6000-builtins.h): New #include.
(rs6000_init_builtins): Call rs6000_init_generated_builtins.  Skip the
old initialization logic when new builtins are enabled.
* config/rs6000/rs6000-gen-builtins.c (write_decls): Rename
rs6000_autoinit_builtins to rs6000_init_generated_builtins.
(write_init_file): Likewise.

2 years agolibgfortran: Use the libtool macro to determine libm availability.
Iain Sandoe [Sun, 28 Mar 2021 14:39:33 +0000 (15:39 +0100)]
libgfortran: Use the libtool macro to determine libm availability.

We recently had a report of build failure against a Darwin branch on
the latest OS release.  This was because (temporarily) the symlink
from libm.dylib => libSystem.dylib had been removed/omitted.

libm is not needed on Darwin, and should not be added unconditionally
even if that is (mostly) harmless since it is a symlink to libc.

There could be cases where the addition was not completely harmless
because the presentation of the symlink would cause the symbols exposed
in libSystem to be considered ahead of ones presented in convenience
libraries.

libgfortran/ChangeLog:

* Makefile.am: Use configured libm availability.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Use libtool macro to find libm availability.
* libgfortran.spec.in: Use configured libm availability.

2 years agotestsuite, Darwin : Do not claim 'GAS' for cctools assembler.
Iain Sandoe [Mon, 15 Mar 2021 21:38:31 +0000 (21:38 +0000)]
testsuite, Darwin : Do not claim 'GAS' for cctools assembler.

Although the cctools assembler is based of GNU GAS, it is from a
very old version (1.38) which does not support many of the features
that the target supports test is expecting.

gcc/testsuite/ChangeLog:

* lib/target-supports.exp: Exclude cctools assembler based on
GAS 1.38.

2 years agoconfigure: Adjust several assembler checks to remove an unused parm.
Iain Sandoe [Thu, 26 Aug 2021 10:32:52 +0000 (11:32 +0100)]
configure: Adjust several assembler checks to remove an unused parm.

In r12-3048-ge0b6d0b39c6, the GAS version parameter was removed from
the gcc_GAS_CHECK_FEATURE macro.  It seems that overlapping comit/test
cycles resulted in several AMDGCN and one Darwin commit with the now
extra parameter still present.

This causes wrong configure code to be generated when autoreconf is
used in the gcc directory.

Fixed by removing the extraneous parm from the AMDGCN and Darwin cases.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* configure.ac (darwin2[[0-9]]* | darwin19*): Alter use of
gcc_GAS_CHECK_FEATURE to remove an extraneous parameter.
(amdgcn-* | gcn-*) Likewise.

2 years agocall_summary: add missing template keyword
Anthony Sharp [Fri, 27 Aug 2021 14:02:42 +0000 (10:02 -0400)]
call_summary: add missing template keyword

Without the 'template', this function template compares 'traverse' to 'f',
and then compares the result to 'a'.  Evidently it hasn't been instantiated
yet.

gcc/ChangeLog:

* symbol-summary.h: Added missing template keyword.

2 years agotree-optimization/45178 - DCE of dead control flow in infinite loop
Richard Biener [Fri, 27 Aug 2021 07:47:49 +0000 (09:47 +0200)]
tree-optimization/45178 - DCE of dead control flow in infinite loop

This fixes DCE to be able to elide dead control flow in an
infinite loop without an exit edge.  This special situation is
handled well by the code finding an edge to preserve since there's
no chance it will find the exit edge and make the loop finite.

2021-08-27  Richard Biener  <rguenther@suse.de>

PR tree-optimization/45178
* tree-ssa-dce.c (find_obviously_necessary_stmts): For
infinite loops without exit do not mark control dependent
edges of the latch necessary.

* gcc.dg/tree-ssa/ssa-dce-3.c: Adjust testcase.

2 years agoi386: Fix wrong optimization for consecutive masked scatters [PR 101472]
konglin1 [Mon, 9 Aug 2021 03:37:52 +0000 (11:37 +0800)]
i386: Fix wrong optimization for consecutive masked scatters [PR 101472]

gcc/ChangeLog:

PR target/101472
* config/i386/sse.md: (<avx512>scattersi<mode>): Add mask operand to
UNSPEC_VSIBADDR.
(<avx512>scattersi<mode>): Likewise.
(*avx512f_scattersi<VI48F:mode>): Merge mask operand to set_dest.
(*avx512f_scatterdi<VI48F:mode>): Likewise

gcc/testsuite/ChangeLog:

PR target/101472
* gcc.target/i386/avx512f-pr101472.c: New test.
* gcc.target/i386/avx512vl-pr101472.c: New test.

2 years agors6000: Make some BIFs vectorized on P10
Kewen Lin [Fri, 27 Aug 2021 01:23:58 +0000 (20:23 -0500)]
rs6000: Make some BIFs vectorized on P10

This patch is to add the support to make vectorizer able to
vectorize some built-in function scalar versions on Power10.

gcc/ChangeLog:

* config/rs6000/rs6000.c (rs6000_builtin_md_vectorized_function): Add
support for built-in functions MISC_BUILTIN_DIVWE, MISC_BUILTIN_DIVWEU,
MISC_BUILTIN_DIVDE, MISC_BUILTIN_DIVDEU, P10_BUILTIN_CFUGED,
P10_BUILTIN_CNTLZDM, P10_BUILTIN_CNTTZDM, P10_BUILTIN_PDEPD and
P10_BUILTIN_PEXTD on Power10.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/dive-vectorize-1.c: New test.
* gcc.target/powerpc/dive-vectorize-1.h: New test.
* gcc.target/powerpc/dive-vectorize-2.c: New test.
* gcc.target/powerpc/dive-vectorize-2.h: New test.
* gcc.target/powerpc/dive-vectorize-run-1.c: New test.
* gcc.target/powerpc/dive-vectorize-run-2.c: New test.
* gcc.target/powerpc/p10-bifs-vectorize-1.c: New test.
* gcc.target/powerpc/p10-bifs-vectorize-1.h: New test.
* gcc.target/powerpc/p10-bifs-vectorize-run-1.c: New test.

2 years agors6000: Add missing unsigned info for some P10 bifs
Kewen Lin [Fri, 27 Aug 2021 01:23:17 +0000 (20:23 -0500)]
rs6000: Add missing unsigned info for some P10 bifs

This patch is to make prototypes of some Power10 built-in
functions consistent with what's in the documentation, as
well as the vector version.  Otherwise, useless conversions
can be generated in gimple IR, and the vectorized versions
will have inconsistent types.

gcc/ChangeLog:

* config/rs6000/rs6000-call.c (builtin_function_type): Add unsigned
signedness for some Power10 bifs.

2 years agoaix: packed struct alignment [PR102068]
David Edelsohn [Thu, 26 Aug 2021 21:14:18 +0000 (17:14 -0400)]
aix: packed struct alignment [PR102068]

Further fixes to structure alignment when the structure is packed
and contains double.  This patch checks for packed attribute
at the top level.

gcc/ChangeLog:

PR target/102068
* config/rs6000/rs6000.c (rs6000_adjust_field_align): Use
computed alignment if the entire struct has attribute packed.

2 years agoFold more shuffle builtins to VEC_PERM_EXPR.
liuhongt [Fri, 11 Dec 2020 11:02:43 +0000 (19:02 +0800)]
Fold more shuffle builtins to VEC_PERM_EXPR.

A follow-up to https://gcc.gnu.org/pipermail/gcc-patches/2019-May/521983.html

gcc/
PR target/98167
PR target/43147
* config/i386/i386.c (ix86_gimple_fold_builtin): Fold
IX86_BUILTIN_SHUFPD512, IX86_BUILTIN_SHUFPS512,
IX86_BUILTIN_SHUFPD256, IX86_BUILTIN_SHUFPS,
IX86_BUILTIN_SHUFPS256.
(ix86_masked_all_ones): New function.

gcc/testsuite/
* gcc.target/i386/avx512f-vshufpd-1.c: Adjust testcase.
* gcc.target/i386/avx512f-vshufps-1.c: Adjust testcase.
* gcc.target/i386/pr43147.c: New test.

2 years agoDaily bump.
GCC Administrator [Fri, 27 Aug 2021 00:16:22 +0000 (00:16 +0000)]
Daily bump.

2 years agolibstdc:: Use markdown for Doxygen comments in std::function
Jonathan Wakely [Thu, 26 Aug 2021 15:55:31 +0000 (16:55 +0100)]
libstdc:: Use markdown for Doxygen comments in std::function

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/std_function.h (function): Adjust doxygen
comments.
* include/bits/unique_ptr.h (make_unique_for_overwrite):
Change parameter name to match doxygen comment.

2 years agolibstdc++: Avoid a move in std::function construction (LWG 2447)
Jonathan Wakely [Thu, 26 Aug 2021 13:01:36 +0000 (14:01 +0100)]
libstdc++: Avoid a move in std::function construction (LWG 2447)

This makes the std::function constructor use perfect forwarding, to
avoid an unnecessary move-construction of the target. This means we need
to rewrite the _Function_base::_Base_manager::_M_init_functor function
to use a forwarding reference, and so can reuse it for the clone
operation.

Also simplify the SFINAE constraints on the constructor, by combining
the !is_same_v<remove_cvref_t<F>, function> constraint into the
_Callable trait.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/std_function.h (_function_base::_Base_manager):
Replace _M_init_functor with a function template using a
forwarding reference, and a pair of _M_create function
templates. Reuse _M_create for the clone operation.
(function::_Decay_t): New alias template.
(function::_Callable): Simplify by using _Decay.
(function::function(F)): Change parameter to forwarding
reference, as per LWG 2447. Add noexcept-specifier. Simplify
constraints.
(function::operator=(F&&)): Add noexcept-specifier.
* testsuite/20_util/function/cons/lwg2774.cc: New test.
* testsuite/20_util/function/cons/noexcept.cc: New test.

2 years agolibstdc++: Add static assertions for std::function requirements
Jonathan Wakely [Thu, 26 Aug 2021 12:13:27 +0000 (13:13 +0100)]
libstdc++: Add static assertions for std::function requirements

Add static assertions to std::function, so that more user-friendly
diagnostics are given when trying to store a non-copyable target object.

These preconditions were added as "Mandates:" by LWG 2447, but I'm
committing them separately from implementing that, to allow just this
change to be backported more easily.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/std_function.h (function::function(F)): Add
static assertions to check constructibility requirements.

2 years agors6000: inline ldouble __gcc_qsub
David Edelsohn [Wed, 25 Aug 2021 21:41:50 +0000 (17:41 -0400)]
rs6000: inline ldouble __gcc_qsub

While performing some tests of IEEE 128 float for PPC64LE, Michael
Meissner noticed that __gcc_qsub is substantially slower than
__gcc_qadd.  __gcc_qsub calls __gcc_add with the second operand
negated.  Because the functions normally are invoked through
libgcc shared object, the extra PLT overhead has a large impact
on the overall time of the function.  This patch converts
__gcc_qadd to a static inline function invoked by __gcc_qadd
and __gcc_qsub.

libgcc/ChangeLog:

* config/rs6000/ibm-ldouble.c (ldouble_qadd_internal): Rename from
__gcc_qadd.
(__gcc_qadd): Call ldouble_qadd_internal.
(__gcc_qsub): Call ldouble_qadd_internal with second long double
argument negated.

2 years ago[i386] Call force_reg unconditionally.
Uros Bizjak [Thu, 26 Aug 2021 20:08:25 +0000 (22:08 +0200)]
[i386] Call force_reg unconditionally.

There is no point to check RTXes before calling force_reg,
force_reg checks for REG RTX by itself.

2021-08-26  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
* config/i386/i386.md (*btr<mode>_1): Call force_reg unconditionally.
(conditional moves with memory inputs splitters): Ditto.
* config/i386/sse.md (one_cmpl<mode>2): Simplify.

2 years agoFix ipa-modref verification ices
Jan Hubicka [Thu, 26 Aug 2021 19:37:46 +0000 (21:37 +0200)]
Fix ipa-modref verification ices

* ipa-modref-tree.h (modref_access_node::try_merge_with): Restart
search after merging.

2 years agors6000: Add remaining overloads
Bill Schmidt [Thu, 26 Aug 2021 19:34:00 +0000 (14:34 -0500)]
rs6000: Add remaining overloads

2021-08-26  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-overload.def: Add remaining overloads.

2 years agors6000: Add Cell builtins
Bill Schmidt [Mon, 7 Jun 2021 21:09:15 +0000 (16:09 -0500)]
rs6000: Add Cell builtins

2021-06-07  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add cell stanza.

2 years agors6000: Add miscellaneous builtins
Bill Schmidt [Tue, 15 Jun 2021 13:43:11 +0000 (08:43 -0500)]
rs6000: Add miscellaneous builtins

2021-06-15  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add ieee128-hw, dfp,
crypto, and htm stanzas.

2 years agors6000: Add MMA builtins
Bill Schmidt [Thu, 17 Jun 2021 14:05:24 +0000 (09:05 -0500)]
rs6000: Add MMA builtins

2021-06-16  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add mma stanza.

2 years agoRefactor warn_uninit() code.
Martin Sebor [Thu, 26 Aug 2021 19:15:38 +0000 (13:15 -0600)]
Refactor warn_uninit() code.

gcc/ChangeLog:

* tree-ssa-uninit.c (warn_uninit): Refactor and simplify.
(warn_uninit_phi_uses): Remove argument from calls to warn_uninit.
(warn_uninitialized_vars): Same.  Reduce visibility of locals.
(warn_uninitialized_phi): Same.

2 years agoImproved handling of shifts/rotates in bit CCP.
Roger Sayle [Thu, 26 Aug 2021 17:57:00 +0000 (18:57 +0100)]
Improved handling of shifts/rotates in bit CCP.

This patch is the next in the series to improve bit bounds in tree-ssa's
bit CCP pass, this time: bounds for shifts and rotates by unknown amounts.
This allows us to optimize expressions such as ((x&15)<<(y&24))&64.
In this case, the expression (y&24) contains only two unknown bits,
and can therefore have only four possible values: 0, 8, 16 and 24.
From this (x&15)<<(y&24) has the nonzero bits 0x0f0f0f0f, and from
that ((x&15)<<(y&24))&64 must always be zero.

One clever use of computer science in this patch is the use of XOR
to efficiently enumerate bit patterns in Gray code order.  As the
order in which we generate values is not significant, it's faster
and more convenient to enumerate values by flipping one bit at a
time, rather than in numerical order [which would require carry
bits and additional logic].

There's a pre-existing ??? comment in tree-ssa-ccp.c that we should
eventually be able to optimize (x<<(y|8))&255, but this patch takes the
conservatively paranoid approach of only optimizing cases where the
shift/rotate is guaranteed to be less than the target precision, and
therefore avoids changing any cases that potentially might invoke
undefined behavior.  This patch does optimize (x<<((y&31)|8))&255.

2021-08-26  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* tree-ssa-ccp.c (get_individual_bits): Helper function to
extract the individual bits from a widest_int constant (mask).
(gray_code_bit_flips): New read-only table for effiently
enumerating permutations/combinations of bits.
(bit_value_binop) [LROTATE_EXPR, RROTATE_EXPR]: Handle rotates
by unknown counts that are guaranteed less than the target
precision and four or fewer unknown bits by enumeration.
[LSHIFT_EXPR, RSHIFT_EXPR]: Likewise, also handle shifts by
enumeration under the same conditions.  Handle remaining
shifts as a mask based upon the minimum possible shift value.

gcc/testsuite/ChangeLog
* gcc.dg/tree-ssa/ssa-ccp-41.c: New test case.

2 years ago[Committed] Tidy up !POINTER_TYPE_P test in match.pd LSHIFT_EXPR folding
Roger Sayle [Thu, 26 Aug 2021 17:51:46 +0000 (18:51 +0100)]
[Committed] Tidy up !POINTER_TYPE_P test in match.pd LSHIFT_EXPR folding

As suggested by Richard Biener in the comments of PR middle-end/102029,
the new test "INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type) ..." is
redundant, and just "INTEGRAL_TYPE_P (type)" is the preferred form.

2021-08-26  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
* match.pd (shift transformations): Remove a redundant
!POINTER_TYPE_P check.

2 years ago[i386] Set all_regs to true in the call to replace_rtx [PR102057]
Uros Bizjak [Thu, 26 Aug 2021 15:26:06 +0000 (17:26 +0200)]
[i386] Set all_regs to true in the call to replace_rtx [PR102057]

We want to replace all REGs equal to FROM.

2021-08-26  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/102057
* config/i386/i386.md (cmove reg-reg move elimination peephole2s):
Set all_regs to true in the call to replace_rtx.

2 years agoImprove handling of modref params.
Jan Hubicka [Thu, 26 Aug 2021 14:57:46 +0000 (16:57 +0200)]
Improve handling of modref params.

this patch makes insertion to modref access tree smarter when --param
modref-max-bases and moredref-max-refs are hit.  Instead of giving up
we either give up on base alias set (make it equal to ref) or turn the
alias set to 0.  This lets us to track useful info on quite large
functions, such as ggc_free.

gcc/ChangeLog:

* ipa-modref-tree.c (test_insert_search_collapse): Update test.
* ipa-modref-tree.h (modref_base_node::insert): Be smarter when
hiting --param modref-max-refs limit.
(modref_tree:insert_base): Be smarter when hitting
--param modref-max-bases limit. Add new parameter REF.
(modref_tree:insert): Update.
(modref_tree:merge): Update.
* ipa-modref.c (read_modref_records): Update.

2 years agoAdd full stop to params.opt.
Jan Hubicka [Thu, 26 Aug 2021 14:01:53 +0000 (16:01 +0200)]
Add full stop to params.opt.

gcc/ChangeLog:

* params.opt: (modref-max-adjustments): Add full stop.

2 years agoFix off-by-one error in try_merge_with
Jan Hubicka [Thu, 26 Aug 2021 13:33:56 +0000 (15:33 +0200)]
Fix off-by-one error in try_merge_with

gcc/ChangeLog:

* ipa-modref-tree.h (modref_ref_node::verify): New member
functoin.
(modref_ref_node::insert): Use it.
(modref_ref_node::try_mere_with): Fix off by one error.

2 years agolibstdc++: Make Networking TS headers more portable [PR100285]
Jonathan Wakely [Thu, 26 Aug 2021 11:06:55 +0000 (12:06 +0100)]
libstdc++: Make Networking TS headers more portable [PR100285]

Add more preprocessor conditions to check for constants being defined
before using them, so that the Networking TS headers can be compiled on
a wider range of platforms.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/100285
* configure.ac: Check for O_NONBLOCK.
* configure: Regenerate.
* include/experimental/internet: Include <ws2tcpip.h> for
Windows.  Use preprocessor conditions around more constants.
* include/experimental/socket: Use preprocessor conditions
around more constants.
* testsuite/experimental/net/internet/resolver/base.cc: Only use
constants when the corresponding C macro is defined.
* testsuite/experimental/net/socket/basic_socket.cc: Likewise.
* testsuite/experimental/net/socket/socket_base.cc: Likewise.
Make preprocessor checks more fine-grained.

2 years agomklog: fix renaming with version 0.7.0.
Martin Liska [Thu, 26 Aug 2021 08:20:20 +0000 (10:20 +0200)]
mklog: fix renaming with version 0.7.0.

contrib/ChangeLog:

* mklog.py: Use file.{source,target}_file for proper rename
handling.

2 years agoUse non-numbered clones for target_clones.
Martin Liska [Fri, 20 Aug 2021 14:35:18 +0000 (16:35 +0200)]
Use non-numbered clones for target_clones.

gcc/ChangeLog:

* cgraph.h (create_version_clone_with_body): Add new parameter.
* cgraphclones.c: Likewise.
* multiple_target.c (create_dispatcher_calls): Do not use
numbered suffixes.
(create_target_clone): Likewise here.

gcc/testsuite/ChangeLog:

* gcc.target/i386/mvc5.c: Scan assembly names.
* gcc.target/i386/mvc7.c: Likewise.
* gcc.target/i386/pr95778-1.c: Update scanned patterns.
* gcc.target/i386/pr95778-2.c: Likewise.

Co-Authored-By: Stefan Kneifel <stefan.kneifel@bluewin.ch>
2 years agoextend.texi: add note about reserved ctor/dtor priorities
Jonathan Yong [Wed, 25 Aug 2021 16:36:14 +0000 (16:36 +0000)]
extend.texi: add note about reserved ctor/dtor priorities

gcc/Changelog:

* doc/extend.texi: Add note about reserved priorities
to the constructor attribute.

Signed-off-by: Jonathan Yong <10walls@gmail.com>
2 years agocygming-crtend.c: fix build warnings
Jonathan Yong [Wed, 25 Aug 2021 16:33:36 +0000 (16:33 +0000)]
cygming-crtend.c: fix build warnings

libgcc/Changelog:

* config/i386/cygming-crtend.c: Fix register_frame_ctor
and register_frame_dtor warnings.

Signed-off-by: Jonathan Yong <10walls@gmail.com>
2 years agoDaily bump.
GCC Administrator [Thu, 26 Aug 2021 00:17:03 +0000 (00:17 +0000)]
Daily bump.

2 years agoAdd -details to dump option needed after r12-3144.
Martin Sebor [Wed, 25 Aug 2021 23:25:08 +0000 (17:25 -0600)]
Add -details to dump option needed after r12-3144.

gcc/testsuite:
* gcc.dg/tree-ssa/evrp1.c: Add -details to dump option.
* gcc.dg/tree-ssa/evrp2.c: Same.
* gcc.dg/tree-ssa/evrp3.c: Same.
* gcc.dg/tree-ssa/evrp4.c: Same.
* gcc.dg/tree-ssa/evrp6.c: Same.
* gcc.dg/tree-ssa/pr64130.c: Same.

2 years agoFix tests that require IBM 128-bit long double
Michael Meissner [Wed, 25 Aug 2021 23:25:12 +0000 (19:25 -0400)]
Fix tests that require IBM 128-bit long double

This patch adds 3 more selections to target-supports.exp to see if we can
specify to use a particular long double format (IEEE 128-bit, IBM extended
double, 64-bit), and the library support will track the changes for the long
double.  This is needed because two of the tests in the test suite use long
double, and they are actually testing IBM extended double.

This patch also forces the two tests that explicitly require long double
to use the IBM double-double encoding to explicitly run the test.  This
requires GLIBC 2.32 or greater in order to do the switch.

I have run tests on a little endian power9 system with 3 compilers.  There were
no regressions with these patches, and the two tests in the following patches
now work if the default long double is not IBM 128-bit:

 * One compiler used the default IBM 128-bit format;
 * One compiler used the IEEE 128-bit format; (and)
 * One compiler used 64-bit long doubles.

I have also tested compilers on a big endian power8 system with a compiler
defaulting to power8 code generation and another with the default cpu
set.  There were no regressions.

2021-08-25  Michael Meissner  <meissner@linux.ibm.com>

gcc/testsuite/
PR target/94630
* gcc.target/powerpc/pr70117.c: Specify that we need the long double
type to be IBM 128-bit.  Remove the code to use __ibm128.
* c-c++-common/dfp/convert-bfp-11.c: Specify that we need the long
double type to be IBM 128-bit.  Run the test at -O2 optimization.
* lib/target-supports.exp (add_options_for_long_double_ibm128): New
function.
(check_effective_target_long_double_ibm128): New function.
(add_options_for_long_double_ieee128): New function.
(check_effective_target_long_double_ieee128): New function.
(add_options_for_long_double_64bit): New function.
(check_effective_target_long_double_64bit): New function.

2 years agolibstdc++: Add another non-reserved name to tests
Jonathan Wakely [Wed, 25 Aug 2021 22:19:25 +0000 (23:19 +0100)]
libstdc++: Add another non-reserved name to tests

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/17_intro/names.cc: Check 'sz'.

2 years agolibstdc++: Fix names.cc test failures on Windows
Jonathan Wakely [Wed, 25 Aug 2021 21:27:22 +0000 (22:27 +0100)]
libstdc++: Fix names.cc test failures on Windows

The Windows CRT headers define structs with members called f, x, y etc
so don't check those. There are also lots of unnecessary function
parameters in mingw headers using non-reserved names, e.g.

<time.h> uses p and z as parameters of mingw_gettimeofday
<inttypes.h> uses j as a parameter of imaxabs
<pthread.h> uses l, o and func as parameter names

Those should be fixed in the headers instead.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/17_intro/names.cc: Adjust for Windows.

2 years agolibstdc++: Fix non-reserved names in <valarray>
Jonathan Wakely [Wed, 25 Aug 2021 21:24:54 +0000 (22:24 +0100)]
libstdc++: Fix non-reserved names in <valarray>

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/valarray: Uglify 'func' parameters.
* testsuite/17_intro/names.cc: Add 'func' to checks.

2 years agolibstdc++: Fix conditions for optimizing uninitialized algos [PR102064]
Jonathan Wakely [Wed, 25 Aug 2021 20:10:48 +0000 (21:10 +0100)]
libstdc++: Fix conditions for optimizing uninitialized algos [PR102064]

While laying some groundwork for constexpr std::vector, I noticed some
bugs in the std::uninitialized_xxx algorithms. The conditions being
checked for optimizing trivial cases were not quite right, as shown in
the examples in the PR.

This consolidates the checks into a single macro. The macro has
appropriate definitions for C++98 or for later standards, to avoid a #if
everywhere the checks are used. For C++11 and later the check makes a
call to a new function doing a static_assert to ensure we don't use
assignment in cases where construction would have been invalid.
Extracting that check to a separate function will be useful for
constexpr std::vector, as that can't use std::uninitialized_copy
directly because it isn't constexpr).

The consolidated checks mean that some slight variations in static
assert message are gone, as there is only one place that does the assert
now. That required adjusting some tests. As part of that the redundant
89164_c++17.cc test was merged into 89164.cc which is compiled as C++17
by default now, but can also use other -std options if the
C++17-specific error is made conditional with a target selector.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/102064
* include/bits/stl_uninitialized.h (_GLIBCXX_USE_ASSIGN_FOR_INIT):
Define macro to check conditions for optimizing trivial cases.
(__check_constructible): New function to do static assert.
(uninitialized_copy, uninitialized_fill, uninitialized_fill_n):
Use new macro.
* testsuite/20_util/specialized_algorithms/uninitialized_copy/1.cc:
Adjust dg-error pattern.
* testsuite/23_containers/vector/cons/89164.cc: Likewise. Add
C++17-specific checks from 89164_c++17.cc.
* testsuite/23_containers/vector/cons/89164_c++17.cc: Removed.
* testsuite/20_util/specialized_algorithms/uninitialized_copy/102064.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_copy_n/102064.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_fill/102064.cc:
New test.
* testsuite/20_util/specialized_algorithms/uninitialized_fill_n/102064.cc:
New test.

2 years agolibstdc++: Remove __gnu_cxx::rope::erase(size_type) [PR102048]
Jonathan Wakely [Wed, 25 Aug 2021 15:42:49 +0000 (16:42 +0100)]
libstdc++: Remove __gnu_cxx::rope::erase(size_type) [PR102048]

This function claims to remove a single character at index p, but it
actually removes p+1 characters beginning at p. So r.erase(0) removes
the first character, but r.erase(1) removes the second and third, and
r.erase(2) removes the second, third and fourth. This is not a useful
API.

The overload is present in the SGI STL <stl_rope.h> header that we
imported, but it isn't documented in the API reference. The erase
overloads that are documented are:

erase(const iterator& p)
erase(const iterator& f, const iterator& l)
erase(size_type i, size_type n);

Having an erase(size_type p) overload that erases a single character (as
the comment says it does) might be useful, but would be inconsistent
with std::basic_string::erase(size_type p = 0, size_type n = npos),
which erases from p to the end of the string when called with a single
argument.

Since the function isn't part of the documented API, doesn't do what it
claims to do (or anything useful) and "fixing" it would leave it
inconsistent with basic_string, I'm just removing that overload.

libstdc++-v3/ChangeLog:

PR libstdc++/102048
* include/ext/rope (rope::erase(size_type)): Remove broken
function.

2 years agoFix PR c++/66590: incorrect warning "reaches end of non-void function" for switch
Andrew Pinski [Tue, 10 Aug 2021 01:33:17 +0000 (18:33 -0700)]
Fix PR c++/66590: incorrect warning "reaches end of non-void function" for switch

So the problem here is there is code in the C++ front-end not to add a
break statement (to the IR) if the previous block does not fall through.
The problem is the code which does the check to see if the block
may fallthrough does not check a CLEANUP_STMT; it assumes it is always
fall through.  Anyways this adds the code for the case of a CLEANUP_STMT
that is only for !CLEANUP_EH_ONLY (the try/finally case).

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

gcc/cp/ChangeLog:

PR c++/66590
* cp-objcp-common.c (cxx_block_may_fallthru): Handle
CLEANUP_STMT for the case which will be try/finally.

gcc/testsuite/ChangeLog:

PR c++/66590
* g++.dg/warn/Wreturn-5.C: New test.

2 years agoAvoid printing range table header alone.
Martin Sebor [Wed, 25 Aug 2021 20:36:13 +0000 (14:36 -0600)]
Avoid printing range table header alone.

gcc/ChangeLog:
* gimple-range-cache.cc (ssa_global_cache::dump): Avoid printing
range table header alone.
* gimple-range.cc (gimple_ranger::export_global_ranges): Same.

2 years agoc++: Fix up value initialization of structs with zero width bitfields [PR102019]
Jakub Jelinek [Wed, 25 Aug 2021 20:35:21 +0000 (22:35 +0200)]
c++: Fix up value initialization of structs with zero width bitfields [PR102019]

The removal of remove_zero_width_bit_fields, in addition to triggering
some ABI issues that need solving anyway (ABI incompatibility between
C and C++) also resulted in UB inside of gcc, we now call build_zero_init
which calls build_int_cst on an integral type with TYPE_PRECISION of 0.

Fixed by ignoring the zero width bitfields.  I understand
build_value_init_noctor wants to initialize to 0 even unnamed bitfields
(of non-zero width), at least until we have some CONSTRUCTOR flag that says
that even all the padding bits should be cleared.

2021-08-25  Jakub Jelinek  <jakub@redhat.com>

PR c++/102019
* init.c (build_value_init_noctor): Ignore unnamed zero-width
bitfields.

2 years agoMerge load/stores in ipa-modref summaries
Jan Hubicka [Wed, 25 Aug 2021 19:43:07 +0000 (21:43 +0200)]
Merge load/stores in ipa-modref summaries

this patch adds logic needed to merge neighbouring accesses in ipa-modref
summaries.  This helps analyzing array initializers and similar code.  It is
bit of work, since it breaks the fact that modref tree makes a good lattice for
dataflow: the access ranges can be extended indefinitely.  For this reason I
added counter tracking number of adjustments and a cap to limit them during the
dataflow.

gcc/ChangeLog:

* doc/invoke.texi: Document --param modref-max-adjustments.
* ipa-modref-tree.c (test_insert_search_collapse): Update.
(test_merge): Update.
* ipa-modref-tree.h (struct modref_access_node): Add adjustments;
(modref_access_node::operator==): Fix handling of access ranges.
(modref_access_node::contains): Constify parameter; handle also
mismatched parm offsets.
(modref_access_node::update): New function.
(modref_access_node::merge): New function.
(unspecified_modref_access_node): Update constructor.
(modref_ref_node::insert_access): Add record_adjustments parameter;
handle merging.
(modref_ref_node::try_merge_with): New private function.
(modref_tree::insert): New record_adjustments parameter.
(modref_tree::merge): New record_adjustments parameter.
(modref_tree::copy_from): Update.
* ipa-modref.c (dump_access): Dump adjustments field.
(get_access): Update constructor.
(record_access): Update call of insert.
(record_access_lto): Update call of insert.
(merge_call_side_effects): Add record_adjustments parameter.
(get_access_for_fnspec): Update.
(process_fnspec): Update.
(analyze_call): Update.
(analyze_function): Update.
(read_modref_records): Update.
(ipa_merge_modref_summary_after_inlining): Update.
(propagate_unknown_call): Update.
(modref_propagate_in_scc): Update.
* params.opt (param-max-modref-adjustments=): New.

gcc/testsuite/ChangeLog:

* gcc.dg/ipa/modref-1.c: Update testcase.
* gcc.dg/tree-ssa/modref-4.c: Update testcase.
* gcc.dg/tree-ssa/modref-8.c: New test.

2 years agoMake xxsplti*, xpermx, xxeval be vecperm type.
Michael Meissner [Wed, 25 Aug 2021 19:11:47 +0000 (15:11 -0400)]
Make xxsplti*, xpermx, xxeval be vecperm type.

I noticed that the built-functions for xxspltiw, xxspltidp, xxsplti32dx,
xxpermx, and xxeval all used the 'vecsimple' type.  These instructions are
permute instructions (3 cycle latency) and should use 'vecperm' instead.

While I was at it, I changed the UNSPEC name for xxspltidp to be
UNSPEC_XXSPLTIDP instead of UNSPEC_XXSPLTID.

2021-08-25  Michael Meissner  <meissner@linux.ibm.com>

gcc/
* config/rs6000/vsx.md (UNSPEC_XXSPLTIDP): Rename from
UNSPEC_XXSPLTID.
(xxspltiw_v4si): Use vecperm type attribute.
(xxspltiw_v4si_inst): Use vecperm type attribute.
(xxspltiw_v4sf_inst): Likewise.
(xxspltidp_v2df): Use vecperm type attribute.  Use
UNSPEC_XXSPLTIDP instead of UNSPEC_XXSPLTID.
(xxspltidp_v2df_inst): Likewise.
(xxsplti32dx_v4si): Use vecperm type attribute.
(xxsplti32dx_v4si_inst): Likewise.
(xxsplti32dx_v4sf_inst): Likewise.
(xxblend_<mode>): Likewise.
(xxpermx): Likewise.
(xxpermx_inst): Likewise.
(xxeval): Likewise.

2 years agodiagnostics: Support for -finput-charset [PR93067]
Lewis Hyatt [Tue, 24 Aug 2021 23:30:44 +0000 (19:30 -0400)]
diagnostics: Support for -finput-charset [PR93067]

Adds the logic to handle -finput-charset in layout_get_source_line(), so that
source lines are converted from their input encodings prior to being output by
diagnostics machinery. Also adds the ability to strip a UTF-8 BOM similarly.

gcc/c-family/ChangeLog:

PR other/93067
* c-opts.c (c_common_input_charset_cb): New function.
(c_common_post_options): Call new function
diagnostic_initialize_input_context().

gcc/d/ChangeLog:

PR other/93067
* d-lang.cc (d_input_charset_callback): New function.
(d_init): Call new function
diagnostic_initialize_input_context().

gcc/fortran/ChangeLog:

PR other/93067
* cpp.c (gfc_cpp_post_options): Call new function
diagnostic_initialize_input_context().

gcc/ChangeLog:

PR other/93067
* coretypes.h (typedef diagnostic_input_charset_callback): Declare.
* diagnostic.c (diagnostic_initialize_input_context): New function.
* diagnostic.h (diagnostic_initialize_input_context): Declare.
* input.c (default_charset_callback): New function.
(file_cache::initialize_input_context): New function.
(file_cache_slot::create): Added ability to convert the input
according to the input context.
(file_cache::file_cache): Initialize the new input context.
(class file_cache_slot): Added new m_alloc_offset member.
(file_cache_slot::file_cache_slot): Initialize the new member.
(file_cache_slot::~file_cache_slot): Handle potentially offset buffer.
(file_cache_slot::maybe_grow): Likewise.
(file_cache_slot::needs_read_p): Handle NULL fp, which is now possible.
(file_cache_slot::get_next_line): Likewise.
* input.h (class file_cache): Added input context member.

libcpp/ChangeLog:

PR other/93067
* charset.c (init_iconv_desc): Adapt to permit PFILE argument to
be NULL.
(_cpp_convert_input): Likewise. Also move UTF-8 BOM logic to...
(cpp_check_utf8_bom): ...here.  New function.
(cpp_input_conversion_is_trivial): New function.
* files.c (read_file_guts): Allow PFILE argument to be NULL.  Add
INPUT_CHARSET argument as an alternate source of this information.
(read_file): Pass the new argument to read_file_guts.
(cpp_get_converted_source): New function.
* include/cpplib.h (struct cpp_converted_source): Declare.
(cpp_get_converted_source): Declare.
(cpp_input_conversion_is_trivial): Declare.
(cpp_check_utf8_bom): Declare.

gcc/testsuite/ChangeLog:

PR other/93067
* gcc.dg/diagnostic-input-charset-1.c: New test.
* gcc.dg/diagnostic-input-utf8-bom.c: New test.

2 years agoanalyzer: Impose recursion limit on indirect calls.
Ankur Saini [Wed, 25 Aug 2021 07:03:06 +0000 (12:33 +0530)]
analyzer: Impose recursion limit on indirect calls.

2021-08-25  Ankur Saini  <arsenic@sourceware.org>

gcc/analyzer/ChangeLog:
PR analyzer/101980
* engine.cc (exploded_graph::maybe_create_dynamic_call): Don't create
calls if max recursion limit is reached.

2 years agotree-optimization/102046 - fix SLP build from scalars with patterns
Richard Biener [Wed, 25 Aug 2021 08:06:01 +0000 (10:06 +0200)]
tree-optimization/102046 - fix SLP build from scalars with patterns

When we swap operands for SLP builds we lose track where exactly
pattern defs are - but we fail to update the any_pattern member
of the operands info.  Do so conservatively.

2021-08-25  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102046
* tree-vect-slp.c (vect_build_slp_tree_2): Conservatively
update ->any_pattern when swapping operands.

* gcc.dg/vect/pr102046.c: New testcase.

2 years agoi386: Optimize lea with zero-extend. [PR 101716]
Hongyu Wang [Tue, 17 Aug 2021 08:53:46 +0000 (16:53 +0800)]
i386: Optimize lea with zero-extend. [PR 101716]

For ASHIFT + ZERO_EXTEND pattern, combine pass failed to
match it to lea since it will generate non-canonical
zero-extend. Adjust predicate and cost_model to allow combine
for lea.

gcc/ChangeLog:

PR target/101716
* config/i386/i386.c (ix86_live_on_entry): Adjust comment.
(ix86_decompose_address): Remove retval check for ASHIFT,
allow non-canonical zero extend if AND mask covers ASHIFT
count.
(ix86_legitimate_address_p): Adjust condition for decompose.
(ix86_rtx_costs): Adjust cost for lea with non-canonical
zero-extend.

Co-Authored by: Uros Bizjak <ubizjak@gmail.com>

gcc/testsuite/ChangeLog:

PR target/101716
* gcc.target/i386/pr101716.c: New test.

2 years agoAnalyze niter for until-wrap condition [PR101145]
Jiufu Guo [Wed, 7 Jul 2021 05:41:01 +0000 (13:41 +0800)]
Analyze niter for until-wrap condition [PR101145]

For code like:
unsigned foo(unsigned val, unsigned start)
{
  unsigned cnt = 0;
  for (unsigned i = start; i > val; ++i)
    cnt++;
  return cnt;
}

The number of iterations should be about UINT_MAX - start.

There is function adjust_cond_for_loop_until_wrap which
handles similar work for const bases.
Like adjust_cond_for_loop_until_wrap, this patch enhance
function number_of_iterations_cond/number_of_iterations_lt
to analyze number of iterations for this kind of loop.

gcc/ChangeLog:

2021-08-25  Jiufu Guo  <guojiufu@linux.ibm.com>

PR tree-optimization/101145
* tree-ssa-loop-niter.c (number_of_iterations_until_wrap):
New function.
(number_of_iterations_lt): Invoke above function.
(adjust_cond_for_loop_until_wrap):
Merge to number_of_iterations_until_wrap.
(number_of_iterations_cond): Update invokes for
adjust_cond_for_loop_until_wrap and number_of_iterations_lt.

gcc/testsuite/ChangeLog:

2021-08-25  Jiufu Guo  <guojiufu@linux.ibm.com>

PR tree-optimization/101145
* gcc.dg/vect/pr101145.c: New test.
* gcc.dg/vect/pr101145.inc: New test.
* gcc.dg/vect/pr101145_1.c: New test.
* gcc.dg/vect/pr101145_2.c: New test.
* gcc.dg/vect/pr101145_3.c: New test.
* gcc.dg/vect/pr101145inf.c: New test.
* gcc.dg/vect/pr101145inf.inc: New test.
* gcc.dg/vect/pr101145inf_1.c: New test.

2 years agoi386: Fix _mm512_fpclass_ps_mask in O0 [PR 101471]
konglin1 [Mon, 9 Aug 2021 02:58:24 +0000 (10:58 +0800)]
i386: Fix _mm512_fpclass_ps_mask in O0 [PR 101471]

gcc/ChangeLog:

PR target/101471
* config/i386/avx512dqintrin.h (_mm512_fpclass_ps_mask): Fix
macro define in O0.
(_mm512_mask_fpclass_ps_mask): Ditto.

gcc/testsuite/ChangeLog:

PR target/101471
* gcc.target/i386/avx512f-pr101471.c: New test.

2 years agors6000: Add vec_unpacku_{hi,lo}_v4si
Kewen Lin [Wed, 25 Aug 2021 02:58:14 +0000 (21:58 -0500)]
rs6000: Add vec_unpacku_{hi,lo}_v4si

The existing vec_unpacku_{hi,lo} supports emulated unsigned
unpacking for short and char but misses the support for int.
This patch adds the support of vec_unpacku_{hi,lo}_v4si.

Meanwhile, the current implementation uses vector permutation
way, which requires one extra customized constant vector as
the permutation control vector.  It's better to use vector
merge high/low with zero constant vector, to save the space
in constant area as well as the cost to initialize pcv in
prologue.  This patch updates it with vector merging and
simplify it with iterators.

gcc/ChangeLog:

* config/rs6000/altivec.md (vec_unpacku_hi_v16qi): Remove.
(vec_unpacku_hi_v8hi): Likewise.
(vec_unpacku_lo_v16qi): Likewise.
(vec_unpacku_lo_v8hi): Likewise.
(vec_unpacku_hi_<VP_small_lc>): New define_expand.
(vec_unpacku_lo_<VP_small_lc>): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/unpack-vectorize-1.c: New test.
* gcc.target/powerpc/unpack-vectorize-1.h: New test.
* gcc.target/powerpc/unpack-vectorize-2.c: New test.
* gcc.target/powerpc/unpack-vectorize-2.h: New test.
* gcc.target/powerpc/unpack-vectorize-3.c: New test.
* gcc.target/powerpc/unpack-vectorize-3.h: New test.
* gcc.target/powerpc/unpack-vectorize-run-1.c: New test.
* gcc.target/powerpc/unpack-vectorize-run-2.c: New test.
* gcc.target/powerpc/unpack-vectorize-run-3.c: New test.
* gcc.target/powerpc/unpack-vectorize.h: New test.

2 years agoAdjust testcases to avoid new failures brought by r12-3108 when compiled w -march...
liuhongt [Wed, 25 Aug 2021 01:45:25 +0000 (09:45 +0800)]
Adjust testcases to avoid new failures brought by r12-3108 when compiled w -march=cascadelake.

gcc/testsuite/ChangeLog:

PR target/101989
* gcc.target/i386/avx2-shiftqihi-constant-1.c: Add -mno-avx512f.
* gcc.target/i386/sse2-shiftqihi-constant-1.c: Add -mno-avx

2 years agoDaily bump.
GCC Administrator [Wed, 25 Aug 2021 00:16:57 +0000 (00:16 +0000)]
Daily bump.

2 years agoaix: SYSTEM_IMPLICIT_EXTERN_C
David Edelsohn [Tue, 24 Aug 2021 20:48:57 +0000 (16:48 -0400)]
aix: SYSTEM_IMPLICIT_EXTERN_C

AIX 7.3 system headers are C++ safe and GCC no longer needs to define
SYSTEM_IMPLICIT_EXTERN_C for AIX 7.3.  This patch moves the definition
from aix.h to the individual OS-level configuration files and does not
define the macro for AIX 7.3.

The patch also corrects the definition of TARGET_AIX_VERSION to 73.

gcc/ChangeLog:
* config/rs6000/aix.h (SYSTEM_IMPLICIT_EXTERN_C): Delete.
* config/rs6000/aix71.h (SYSTEM_IMPLICIT_EXTERN_C): Define.
* config/rs6000/aix72.h (SYSTEM_IMPLICIT_EXTERN_C): Define.
* config/rs6000/aix73.h (TARGET_AIX_VERSION): Increase to 73.

2 years ago[Committed] PR middle-end/102031: Fix typo/mistake in simplify_truncation patch.
Roger Sayle [Tue, 24 Aug 2021 21:07:41 +0000 (22:07 +0100)]
[Committed] PR middle-end/102031: Fix typo/mistake in simplify_truncation patch.

My apologies again.  My patch to simplify truncations of SUBREGs in
simplify-rtx.c contained an error where I'd accidentally compared
against a mode instead of the precision of that mode.  Grr!  It even
survived regression testing on two platforms.  Fixed below, and
committed as obvious, after a full "make bootstrap" and "make -k check"
on x86_64-pc-linux-gnu with no new regressions.

2021-08-24  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
PR middle-end/102031
* simplify-rtx.c (simplify_truncation): When comparing precisions
use "subreg_prec" variable, not "subreg_mode".

2 years agoFortran: fix pointless warning for static variables
Harald Anlauf [Tue, 24 Aug 2021 19:07:50 +0000 (21:07 +0200)]
Fortran: fix pointless warning for static variables

gcc/fortran/ChangeLog:

PR fortran/98411
* trans-decl.c (gfc_finish_var_decl): Adjust check to handle
implicit SAVE as well as variables in the main program.  Improve
warning message text.

gcc/testsuite/ChangeLog:

PR fortran/98411
* gfortran.dg/pr98411.f90: Adjust testcase options to restrict to
F2008, and verify case of implicit SAVE.

2 years agors6000: Add Power10 builtins
Bill Schmidt [Wed, 28 Jul 2021 17:22:57 +0000 (13:22 -0400)]
rs6000: Add Power10 builtins

2021-07-28  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add power10 and power10-64
stanzas.

2 years agors6000: Add more type nodes to support builtin processing
Bill Schmidt [Tue, 24 Aug 2021 16:50:09 +0000 (11:50 -0500)]
rs6000: Add more type nodes to support builtin processing

2021-08-24  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-call.c (rs6000_init_builtins): Initialize
various pointer type nodes.
* config/rs6000/rs6000.h (rs6000_builtin_type_index): Add enum
values for various pointer types.
(ptr_V16QI_type_node): New macro.
(ptr_V1TI_type_node): New macro.
(ptr_V2DI_type_node): New macro.
(ptr_V2DF_type_node): New macro.
(ptr_V4SI_type_node): New macro.
(ptr_V4SF_type_node): New macro.
(ptr_V8HI_type_node): New macro.
(ptr_unsigned_V16QI_type_node): New macro.
(ptr_unsigned_V1TI_type_node): New macro.
(ptr_unsigned_V8HI_type_node): New macro.
(ptr_unsigned_V4SI_type_node): New macro.
(ptr_unsigned_V2DI_type_node): New macro.
(ptr_bool_V16QI_type_node): New macro.
(ptr_bool_V8HI_type_node): New macro.
(ptr_bool_V4SI_type_node): New macro.
(ptr_bool_V2DI_type_node): New macro.
(ptr_bool_V1TI_type_node): New macro.
(ptr_pixel_type_node): New macro.
(ptr_intQI_type_node): New macro.
(ptr_uintQI_type_node): New macro.
(ptr_intHI_type_node): New macro.
(ptr_uintHI_type_node): New macro.
(ptr_intSI_type_node): New macro.
(ptr_uintSI_type_node): New macro.
(ptr_intDI_type_node): New macro.
(ptr_uintDI_type_node): New macro.
(ptr_intTI_type_node): New macro.
(ptr_uintTI_type_node): New macro.
(ptr_long_integer_type_node): New macro.
(ptr_long_unsigned_type_node): New macro.
(ptr_float_type_node): New macro.
(ptr_double_type_node): New macro.
(ptr_long_double_type_node): New macro.
(ptr_dfloat64_type_node): New macro.
(ptr_dfloat128_type_node): New macro.
(ptr_ieee128_type_node): New macro.
(ptr_ibm128_type_node): New macro.
(ptr_vector_pair_type_node): New macro.
(ptr_vector_quad_type_node): New macro.
(ptr_long_long_integer_type_node): New macro.
(ptr_long_long_unsigned_type_node): New macro.

2 years agors6000: Add Power9 builtins
Bill Schmidt [Tue, 24 Aug 2021 15:01:47 +0000 (10:01 -0500)]
rs6000: Add Power9 builtins

2021-08-24  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add power9-vector, power9,
and power9-64 stanzas.

2 years agonvptx: Add a __PTX_SM__ predefined macro based on target ISA.
Roger Sayle [Tue, 24 Aug 2021 17:02:18 +0000 (18:02 +0100)]
nvptx: Add a __PTX_SM__ predefined macro based on target ISA.

This patch adds a __PTX_SM__ predefined macro to the nvptx backend that
allows code to check the compute model being targeted by the compiler.
This is equivalent to the __CUDA_ARCH__ macro defined by CUDA's nvcc
compiler, but to avoid causing problems for source code that checks
for that compiler, this macro uses GCC's nomenclature; it's easy
enough for users to "#define __CUDA_ARCH__ __PTX_SM__".

What might have been a four line patch is actually a little more
complicated, as this patch takes the opportunity to upgrade the
nvptx backend to use the now preferred nvptx-c.c idiom.

2021-08-24  Roger Sayle  <roger@nextmovesoftware.com>
    Tom de Vries  <tdevries@suse.de>

gcc/ChangeLog
* config.gcc (nvptx-*-*): Define {c,c++}_target_objs.
* config/nvptx/nvptx-protos.h (nvptx_cpu_cpp_builtins): Prototype.
* config/nvptx/nvptx.h (TARGET_CPU_CPP_BUILTINS): Implement with
a call to the new nvptx_cpu_cpp_builtins function in nvptx-c.c.
* config/nvptx/t-nvptx (nvptx-c.o): New rule.
* config/nvptx/nvptx-c.c: New source file.
(nvptx_cpu_cpp_builtins): Move implementation here.

2 years agoReset PHI base0 flag if it's clear in any argument [PR101977, ...]
Martin Sebor [Tue, 24 Aug 2021 16:49:11 +0000 (10:49 -0600)]
Reset PHI base0 flag if it's clear in any argument [PR101977, ...]

Resolves:
PR middle-end/101600 - Spurious -Warray-bounds downcasting a polymorphic pointer
PR middle-end/101977 - bogus -Warray-bounds on a negative index into a parameter in conditional with null

gcc/ChangeLog:

PR middle-end/101600
PR middle-end/101977
* gimple-ssa-warn-access.cc (maybe_warn_for_bound): Tighten up
the phrasing of a warning.
(check_access): Use the remaining size after subtracting any offset
rather than the whole object size.
* pointer-query.cc (access_ref::get_ref): Clear BASE0 flag if it's
clear for any nonnull PHI argument.
(compute_objsize): Clear argument.

gcc/testsuite/ChangeLog:

PR middle-end/101600
PR middle-end/101977
* g++.dg/pr100574.C: Prune out valid warning.
* gcc.dg/pr20126.c: Same.
* gcc.dg/Wstringop-overread.c: Adjust text of expected warnings.
Add new instances.
* gcc.dg/warn-strnlen-no-nul.c: Same.
* g++.dg/warn/Warray-bounds-26.C: New test.
* gcc.dg/Warray-bounds-88.c: New test.

2 years agolibstdc++: Update C++20 status table for layout-compatibility traits
Jonathan Wakely [Tue, 24 Aug 2021 15:15:48 +0000 (16:15 +0100)]
libstdc++: Update C++20 status table for layout-compatibility traits

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* doc/xml/manual/status_cxx2020.xml: Update table.
* doc/html/manual/status.html: Regenerate.

2 years agolibstdc++: Add std::is_layout_compatible trait for C++20
Jonathan Wakely [Tue, 24 Aug 2021 13:42:37 +0000 (14:42 +0100)]
libstdc++: Add std::is_layout_compatible trait for C++20

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/type_traits (is_layout_compatible): Define.
(is_corresponding_member): Define.
* include/std/version (__cpp_lib_is_layout_compatible): Define.
* testsuite/20_util/is_layout_compatible/is_corresponding_member.cc:
New test.
* testsuite/20_util/is_layout_compatible/value.cc: New test.
* testsuite/20_util/is_layout_compatible/version.cc: New test.
* testsuite/20_util/is_pointer_interconvertible/with_class.cc:
New test.
* testsuite/23_containers/span/layout_compat.cc: Do not use real
std::is_layout_compatible trait if available.

2 years agors6000: Add power8-vector builtins
Bill Schmidt [Thu, 1 Apr 2021 18:57:44 +0000 (13:57 -0500)]
rs6000: Add power8-vector builtins

2021-04-01  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add power8-vector stanza.

2 years agors6000: Add power7 and power7-64 builtins
Bill Schmidt [Fri, 2 Apr 2021 21:48:36 +0000 (16:48 -0500)]
rs6000: Add power7 and power7-64 builtins

2021-04-02  Bill Schmidt  <wschmidt@linux.ibm.com>

gcc/
* config/rs6000/rs6000-builtin-new.def: Add power7 and power7-64
stanzas.

2 years agoAdd transitive operations to the relation oracle.
Andrew MacLeod [Wed, 18 Aug 2021 20:36:19 +0000 (16:36 -0400)]
Add transitive operations to the relation oracle.

When registering relations in the oracle, search for other relations which
imply new transitive relations.

gcc/
* value-relation.cc (rr_transitive_table): New.
(relation_transitive): New.
(value_relation::swap): Remove.
(value_relation::apply_transitive): New.
(relation_oracle::relation_oracle): Allocate a new tmp bitmap.
(relation_oracle::register_relation): Call register_transitives.
(relation_oracle::register_transitives): New.
* value-relation.h (relation_oracle): Add new temporary bitmap and
methods.

gcc/testsuite/
* gcc.dg/predict-1.c: Disable evrp.
* gcc.dg/tree-ssa/evrp-trans.c: New.

2 years agolibstdc++: Fix mismatched class-key tags
Jonathan Wakely [Tue, 24 Aug 2021 11:31:06 +0000 (12:31 +0100)]
libstdc++: Fix mismatched class-key tags

Clang warns about this, but GCC doesn't (see PR c++/102036).

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* src/c++11/cxx11-shim_facets.cc: Fix mismatched class-key in
explicit instantiation definitions.

2 years agox86: Broadcast from integer to a pseudo vector register
H.J. Lu [Mon, 23 Aug 2021 21:47:03 +0000 (14:47 -0700)]
x86: Broadcast from integer to a pseudo vector register

Broadcast from integer to a pseudo vector register instead of a hard
vector register to allow LRA to remove redundant move instruction after
broadcast.

gcc/

PR target/102021
* config/i386/i386-expand.c (ix86_expand_vector_move): Broadcast
from integer to a pseudo vector register.

gcc/testsuite/

PR target/102021
* gcc.target/i386/pr100865-10b.c: Expect vzeroupper.
* gcc.target/i386/pr100865-4b.c: Likewise.
* gcc.target/i386/pr100865-6b.c: Expect vmovdqu and vzeroupper.
* gcc.target/i386/pr100865-7b.c: Likewise.
* gcc.target/i386/pr102021.c: New test.

2 years agotree-optimization/100089 - avoid leaving scalar if-converted code around
Richard Biener [Tue, 24 Aug 2021 10:25:25 +0000 (12:25 +0200)]
tree-optimization/100089 - avoid leaving scalar if-converted code around

This avoids leaving scalar if-converted code around for the case
of BB vectorizing an if-converted loop body when using the very-cheap
cost model.  In this case we scan not vectorized scalar stmts in
the basic-block vectorized for COND_EXPRs and force the vectorization
to be marked as not profitable.

The patch also makes sure to always consider all BB vectorization
subgraphs together for costing purposes when vectorizing an
if-converted loop body.

2021-08-24  Richard Biener  <rguenther@suse.de>

PR tree-optimization/100089
* tree-vectorizer.h (vect_slp_bb): Rename to ...
(vect_slp_if_converted_bb): ... this and get the original
loop as new argument.
* tree-vectorizer.c (try_vectorize_loop_1): Revert previous fix,
pass original loop to vect_slp_if_converted_bb.
* tree-vect-slp.c (vect_bb_vectorization_profitable_p):
If orig_loop was passed scan the not vectorized stmts
for COND_EXPRs and force not profitable if found.
(vect_slp_region): Pass down all SLP instances to costing
if orig_loop was specified.
(vect_slp_bbs): Pass through orig_loop.
(vect_slp_bb): Rename to ...
(vect_slp_if_converted_bb): ... this and get the original
loop as new argument.
(vect_slp_function): Adjust.

2 years agoarm: Add tests for VLLDM mitigation [PR102035]
Richard Earnshaw [Tue, 6 Jul 2021 14:10:18 +0000 (15:10 +0100)]
arm: Add tests for VLLDM mitigation [PR102035]

New tests for the erratum mitigation.

gcc/testsuite:
PR target/102035
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-13a.c: New test.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-7a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/soft/cmse-8a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-7a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp-sp/cmse-8a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-13a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-7a.c: Likewise.
* gcc.target/arm/cmse/mainline/8_1m/softfp/cmse-8a.c: Likewise.

2 years agoarm: fix vlldm erratum for Armv8.1-m [PR102035]
Richard Earnshaw [Fri, 18 Jun 2021 16:18:37 +0000 (17:18 +0100)]
arm: fix vlldm erratum for Armv8.1-m [PR102035]

For Armv8.1-m we generate code that emits VLLDM directly and do not
rely on support code in the library, so emit the mitigation directly
as well, when required.  In this case, we can use the compiler options
to determine when to apply the fix and when it is safe to omit it.

gcc:
PR target/102035
* config/arm/arm.md (attribute arch): Add fix_vlldm.
(arch_enabled): Use it.
* config/arm/vfp.md (lazy_store_multiple_insn): Add alternative to
use when erratum mitigation is needed.

2 years agoarm: add erratum mitigation to __gnu_cmse_nonsecure_call [PR102035]
Richard Earnshaw [Fri, 18 Jun 2021 16:16:25 +0000 (17:16 +0100)]
arm: add erratum mitigation to __gnu_cmse_nonsecure_call [PR102035]

Add the recommended erratum mitigation sequence to
__gnu_cmse_nonsecure_call for use on Armv8-m.main devices. Since this
is in the library code we cannot know in advance whether the core we
are running on will be affected by this, so always enable it.

libgcc:
PR target/102035
* config/arm/cmse_nonsecure_call.S (__gnu_cmse_nonsecure_call):
Add vlldm erratum work-around.

2 years agoarm: Add command-line option for enabling CVE-2021-35465 mitigation [PR102035]
Richard Earnshaw [Fri, 11 Jun 2021 15:02:05 +0000 (16:02 +0100)]
arm: Add command-line option for enabling CVE-2021-35465 mitigation [PR102035]

Add a new option, -mfix-cmse-cve-2021-35465 and document it.  Enable it
automatically for cortex-m33, cortex-m35p and cortex-m55.

gcc:
PR target/102035
* config/arm/arm.opt (mfix-cmse-cve-2021-35465): New option.
* doc/invoke.texi (Arm Options): Document it.
* config/arm/arm-cpus.in (quirk_vlldm): New feature bit.
(ALL_QUIRKS): Add quirk_vlldm.
(cortex-m33): Add quirk_vlldm.
(cortex-m35p, cortex-m55): Likewise.
* config/arm/arm.c (arm_option_override): Enable fix_vlldm if
targetting an affected CPU and not explicitly controlled on
the command line.

2 years agoarm: testsuite: improve detection of CMSE hardware.
Richard Earnshaw [Fri, 18 Jun 2021 16:13:04 +0000 (17:13 +0100)]
arm: testsuite: improve detection of CMSE hardware.

The test for CMSE support being available in hardware currently
relies on the compiler not optimizing away a secure gateway operation.
But even that is suspect, because the SG instruction is just a NOP
on armv8-m implementations that do not support the security extension.

Replace the existing test with a new one that reads and checks
the appropriate hardware feature register (memory mapped).  This has
to be run from secure mode, but that shouldn't matter, because if we
can't do that we can't really test the CMSE extensions anyway.  We
retain the SG instruction to ensure the test can't pass accidentally
if run on pre-armv8-m devices.

gcc/testsuite:
* lib/target-supports.exp (check_effective_target_arm_cmse_hw):
Check the CMSE feature register, rather than relying on the
SG operation causing an execution fault.

2 years agoarm: Fix general issues with patterns for VLLDM and VLSTM
Richard Earnshaw [Fri, 11 Jun 2021 16:18:12 +0000 (17:18 +0100)]
arm: Fix general issues with patterns for VLLDM and VLSTM

Both lazy_store_multiple_insn and lazy_load_multiple_insn contain
invalid RTL (eg they contain a post_inc statement outside of a mem).
What's more, the instructions concerned do not modify their input
address register.  We probably got away with this because they are
generated so late in the compilation that no subsequent pass needed to
understand them.  Nevertheless, this could cause problems someday, so
fixed to use a simple legal unspec.

gcc:
* config/arm/vfp.md (lazy_store_multiple_insn): Rewrite as valid RTL.
(lazy_load_multiple_insn): Likewise.

2 years agoEnable avx512 embedde broadcast for vpternlog.
liuhongt [Tue, 24 Aug 2021 10:09:33 +0000 (18:09 +0800)]
Enable avx512 embedde broadcast for vpternlog.

gcc/ChangeLog:

PR target/101989
* config/i386/sse.md (<avx512>_vternlog<mode><sd_maskz_name>):
Enable avx512 embedded broadcast.
(*<avx512>_vternlog<mode>_all): Ditto.
(<avx512>_vternlog<mode>_mask): Ditto.

gcc/testsuite/ChangeLog:

PR target/101989
* gcc.target/i386/pr101989-broadcast-1.c: New test.

2 years agoOptimize (a & b) | (c & ~b) to vpternlog instruction.
liuhongt [Fri, 20 Aug 2021 07:30:40 +0000 (15:30 +0800)]
Optimize (a & b) | (c & ~b) to vpternlog instruction.

Also optimize below 3 forms to vpternlog, op1, op2, op3 are
register_operand or unary_p as (not reg)

A: (any_logic (any_logic op1 op2) op3)
B: (any_logic (any_logic op1 op2) (any_logic op3 op4)) op3/op4 should
be equal to op1/op2
C: (any_logic (any_logic (any_logic:op1 op2) op3) op4) op3/op4 should
be equal to op1/op2

gcc/ChangeLog:

PR target/101989
* config/i386/i386.c (ix86_rtx_costs): Define cost for
UNSPEC_VTERNLOG.
* config/i386/i386.h (STRIP_UNARY): New macro.
* config/i386/predicates.md (reg_or_notreg_operand): New
predicate.
* config/i386/sse.md (*<avx512>_vternlog<mode>_all): New define_insn.
(*<avx512>_vternlog<mode>_1): New pre_reload
define_insn_and_split.
(*<avx512>_vternlog<mode>_2): Ditto.
(*<avx512>_vternlog<mode>_3): Ditto.
(any_logic1,any_logic2): New code iterator.
(logic_op): New code attribute.
(ternlogsuffix): Extend to VNxDF and VNxSF.

gcc/testsuite/ChangeLog:

PR target/101989
* gcc.target/i386/pr101989-1.c: New test.
* gcc.target/i386/pr101989-2.c: New test.
* gcc.target/i386/avx512bw-shiftqihi-constant-1.c: Adjust testcase.

2 years agoAdjust inner loop cost scaling
Richard Biener [Mon, 23 Aug 2021 12:15:14 +0000 (14:15 +0200)]
Adjust inner loop cost scaling

This makes use of the estimated number of iterations of the inner loop
to limit --param vect-inner-loop-cost-factor scaling.  It also reduces
the maximum value of vect-inner-loop-cost-factor to 10000 making it
less likely to cause overflow of costs.

2021-08-23  Richard Biener  <rguenther@suse.de>

* doc/invoke.texi (vect-inner-loop-cost-factor): Adjust.
* params.opt (--param vect-inner-loop-cost-factor): Adjust
maximum value.
* tree-vect-loop.c (vect_analyze_loop_form): Initialize
inner_loop_cost_factor to the minimum of the estimated number
of iterations of the inner loop and vect-inner-loop-cost-factor.

2 years agoFix a few problems with download_prerequisites.
Andrew Pinski [Tue, 24 Aug 2021 05:56:59 +0000 (05:56 +0000)]
Fix a few problems with download_prerequisites.

There are a few problems with download_prerequisites are
described in PR 82704.  The first is on busy-box version of
shasum and md5sum the extended option --check don't exist
so just use -c.  The second issue is the code for which
shasum program to use is included twice and is different.
So move which program to use for the checksum after argument
parsing.  The last issue is --md5 option has been broken for
sometime now as the program is named md5sum and not just md5.
Nobody updated switch table to be correct.

contrib/ChangeLog:

PR other/82704
* download_prerequisites: Fix issues with --md5 and
--sha512 options.

2 years agoTweak -Os costs for scalar-to-vector pass.
Roger Sayle [Tue, 24 Aug 2021 02:04:48 +0000 (03:04 +0100)]
Tweak -Os costs for scalar-to-vector pass.

Back in June I briefly mentioned in one of my gcc-patches posts that
a change that should have always reduced code size, would mysteriously
occasionally result in slightly larger code (according to CSiBE):
https://gcc.gnu.org/pipermail/gcc-patches/2021-June/573233.html

Investigating further, the cause turns out to be that x86_64's
scalar-to-vector (stv) pass is relying on poor estimates of the size
costs/benefits.  This patch tweaks the backend's compute_convert_gain
method to provide slightly more accurate values when compiling with
-Os. Compilation without -Os is (should be) unaffected.  And for
completeness, I'll mention that the stv pass is a net win for code
size so it's much better to improve its heuristics than simply gate
the pass on !optimize_for_size.

The net effect of this change is to save 1399 bytes on the CSiBE
code size benchmark when compiling with -Os.

2021-08-24  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
* config/i386/i386-features.c (compute_convert_gain): Provide
more accurate values for CONST_INT, when optimizing for size.
* config/i386/i386.c (COSTS_N_BYTES): Move definition from here...
* config/i386/i386.h (COSTS_N_BYTES): to here.

2 years ago[Committed] PR middle-end/102029: Stricter typing in LSHIFT_EXPR sign folding.
Roger Sayle [Tue, 24 Aug 2021 01:59:02 +0000 (02:59 +0100)]
[Committed] PR middle-end/102029: Stricter typing in LSHIFT_EXPR sign folding.

My sincere apologies to everyone (again).  As diagnosed by
Jakub Jelinek, my recent patch to fold the signedness of LSHIFT_EXPR
needs to be careful not to attempt transforming a left shift in an
integer type into an invalid left shift of a pointer type.

2021-08-24  Roger Sayle  <roger@nextmovesoftware.com>
    Jakub Jelinek  <jakub@redhat.com>

gcc/ChangeLog
PR middle-end/102029
* match.pd (shift transformations): Add an additional check for
!POINTER_TYPE_P in the recently added left shift transformation.

gcc/testsuite/ChangeLog
PR middle-end/102029
* gcc.dg/fold-convlshift-3.c: New test case.

2 years agoDisable slp in loop vectorizer when cost model is very-cheap.
liuhongt [Wed, 4 Aug 2021 08:39:31 +0000 (16:39 +0800)]
Disable slp in loop vectorizer when cost model is very-cheap.

Performance impact for the commit with option:
-march=x86-64 -O2 -ftree-vectorize -fvect-cost-model=very-cheap

SPEC2017 fprate
503.bwaves_r        BuildSame
507.cactuBSSN_r         -0.04
508.namd_r               0.14
510.parest_r            -0.54
511.povray_r             0.10
519.lbm_r           BuildSame
521.wrf_r                0.64
526.blender_r           -0.32
527.cam4_r               0.17
538.imagick_r            0.09
544.nab_r           BuildSame
549.fotonik3d_r     BuildSame
554.roms_r          BuildSame
997.specrand_fr         -0.09
Geometric mean:  0.02

SPEC2017 intrate
500.perlbench_r          0.26
502.gcc_r                0.21
505.mcf_r               -0.09
520.omnetpp_r       BuildSame
523.xalancbmk_r     BuildSame
525.x264_r              -0.41
531.deepsjeng_r     BuildSame
541.leela_r              0.13
548.exchange2_r     BuildSame
557.xz_r            BuildSame
999.specrand_ir     BuildSame
Geometric mean:  0.02

EEMBC: no regression, only improvement or build the same, the below is
improved benchmarks.

mp2decoddata1       7.59
mp2decoddata2       31.80
mp2decoddata3       12.15
mp2decoddata4       11.16
mp2decoddata5       11.19
mp2decoddata1       7.06
mp2decoddata2       24.12
mp2decoddata3       10.83
mp2decoddata4       10.04
mp2decoddata5       10.07

gcc/ChangeLog:

PR tree-optimization/100089
* tree-vectorizer.c (try_vectorize_loop_1): Disable slp in
loop vectorizer when cost model is very-cheap.