reassoc: Improve maybe_optimize_range_tests [PR96480]

On the following testcase, if the IL before reassoc would be:
...
  <bb 4> [local count: 354334800]:
  if (x_3(D) == 2)
    goto <bb 7>; [34.00%]
  else
    goto <bb 5>; [66.00%]

  <bb 5> [local count: 233860967]:
  if (x_3(D) == 3)
    goto <bb 7>; [34.00%]
  else
    goto <bb 6>; [66.00%]

  <bb 6> [local count: 79512730]:

  <bb 7> [local count: 1073741824]:
  # prephitmp_7 = PHI <1(3), 1(4), 1(5), 1(2), 0(6)>
then we'd optimize it properly, but as bb 5-7 is instead:
  <bb 5> [local count: 233860967]:
  if (x_3(D) == 3)
    goto <bb 6>; [34.00%]
  else
    goto <bb 7>; [66.00%]

  <bb 6> [local count: 79512730]:

  <bb 7> [local count: 1073741824]:
  # prephitmp_7 = PHI <1(3), 1(4), 0(5), 1(2), 1(6)>
(i.e. the true/false edges on the last bb with condition swapped
and ditto for the phi args), we don't recognize it.  If bb 6
is empty, there should be no functional difference between the two IL
representations.

This patch handles those special cases.

2020-08-06  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/96480
* tree-ssa-reassoc.c (suitable_cond_bb): Add TEST_SWAPPED_P argument.
If TEST_BB ends in cond and has one edge to *OTHER_BB and another
through an empty bb to that block too, if PHI args don't match, retry
them through the other path from TEST_BB.
(maybe_optimize_range_tests): Adjust callers.  Handle such LAST_BB
through inversion of the condition.

* gcc.dg/tree-ssa/pr96480.c: New test.

bpf: more flexible support for kernel helpers

This patch changes the existing support for BPF kernel helpers to be
more flexible, in two main ways.

First, there is no longer a hardcoded list of kernel helpers defined
in the compiler internals.  This is replaced by a new target-specific
attribute `kernel_helper' that the user can use to define her own
helpers, annotating function prototypes.

Second, following feedback from the kernel hackers, the pre-defined
helpers in the distributed bpf-helpers.h are no longer available
conditionally depending on the kernel version used in -mkernel.  The
command-line option stays for now, as it may be useful for other
things.

Target tests and documentation updated.

2020-08-06  Jose E. Marchesi  <jose.marchesi@oracle.com>

gcc/
* config/bpf/bpf-helpers.h (KERNEL_HELPER): Define.
(KERNEL_VERSION): Remove.
* config/bpf/bpf-helpers.def: Delete.
* config/bpf/bpf.c (bpf_handle_fndecl_attribute): New function.
(bpf_attribute_table): Define.
(bpf_helper_names): Delete.
(bpf_helper_code): Likewise.
(enum bpf_builtins): Adjust to new helpers mechanism.
(bpf_output_call): Likewise.
(bpf_init_builtins): Likewise.
(bpf_init_builtins): Likewise.
* doc/extend.texi (BPF Function Attributes): New section.
(BPF Kernel Helpers): Delete section.

gcc/testsuite/
* gcc.target/bpf/helper-bind.c: Adjust to new kernel helpers
mechanism.
* gcc.target/bpf/helper-bpf-redirect.c: Likewise.
* gcc.target/bpf/helper-clone-redirect.c: Likewise.
* gcc.target/bpf/helper-csum-diff.c: Likewise.
* gcc.target/bpf/helper-csum-update.c: Likewise.
* gcc.target/bpf/helper-current-task-under-cgroup.c: Likewise.
* gcc.target/bpf/helper-fib-lookup.c: Likewise.
* gcc.target/bpf/helper-get-cgroup-classid.c: Likewise.
* gcc.target/bpf/helper-get-current-cgroup-id.c: Likewise.
* gcc.target/bpf/helper-get-current-comm.c: Likewise.
* gcc.target/bpf/helper-get-current-pid-tgid.c: Likewise.
* gcc.target/bpf/helper-get-current-task.c: Likewise.
* gcc.target/bpf/helper-get-current-uid-gid.c: Likewise.
* gcc.target/bpf/helper-get-hash-recalc.c: Likewise.
* gcc.target/bpf/helper-get-listener-sock.c: Likewise.
* gcc.target/bpf/helper-get-local-storage.c: Likewise.
* gcc.target/bpf/helper-get-numa-node-id.c: Likewise.
* gcc.target/bpf/helper-get-prandom-u32.c: Likewise.
* gcc.target/bpf/helper-get-route-realm.c: Likewise.
* gcc.target/bpf/helper-get-smp-processor-id.c: Likewise.
* gcc.target/bpf/helper-get-socket-cookie.c: Likewise.
* gcc.target/bpf/helper-get-socket-uid.c: Likewise.
* gcc.target/bpf/helper-get-stack.c: Likewise.
* gcc.target/bpf/helper-get-stackid.c: Likewise.
* gcc.target/bpf/helper-getsockopt.c: Likewise.
* gcc.target/bpf/helper-ktime-get-ns.c: Likewise.
* gcc.target/bpf/helper-l3-csum-replace.c: Likewise.
* gcc.target/bpf/helper-l4-csum-replace.c: Likewise.
* gcc.target/bpf/helper-lwt-push-encap.c: Likewise.
* gcc.target/bpf/helper-lwt-seg6-action.c: Likewise.
* gcc.target/bpf/helper-lwt-seg6-adjust-srh.c: Likewise.
* gcc.target/bpf/helper-lwt-seg6-store-bytes.c: Likewise.
* gcc.target/bpf/helper-map-delete-elem.c: Likewise.
* gcc.target/bpf/helper-map-lookup-elem.c: Likewise.
* gcc.target/bpf/helper-map-peek-elem.c: Likewise.
* gcc.target/bpf/helper-map-pop-elem.c: Likewise.
* gcc.target/bpf/helper-map-push-elem.c: Likewise.
* gcc.target/bpf/helper-map-update-elem.c: Likewise.
* gcc.target/bpf/helper-msg-apply-bytes.c: Likewise.
* gcc.target/bpf/helper-msg-cork-bytes.c: Likewise.
* gcc.target/bpf/helper-msg-pop-data.c: Likewise.
* gcc.target/bpf/helper-msg-pull-data.c: Likewise.
* gcc.target/bpf/helper-msg-push-data.c: Likewise.
* gcc.target/bpf/helper-msg-redirect-hash.c: Likewise.
* gcc.target/bpf/helper-msg-redirect-map.c: Likewise.
* gcc.target/bpf/helper-override-return.c: Likewise.
* gcc.target/bpf/helper-perf-event-output.c: Likewise.
* gcc.target/bpf/helper-perf-event-read-value.c: Likewise.
* gcc.target/bpf/helper-perf-event-read.c: Likewise.
* gcc.target/bpf/helper-perf-prog-read-value.c: Likewise.
* gcc.target/bpf/helper-probe-read-str.c: Likewise.
* gcc.target/bpf/helper-probe-read.c: Likewise.
* gcc.target/bpf/helper-probe-write-user.c: Likewise.
* gcc.target/bpf/helper-rc-keydown.c: Likewise.
* gcc.target/bpf/helper-rc-pointer-rel.c: Likewise.
* gcc.target/bpf/helper-rc-repeat.c: Likewise.
* gcc.target/bpf/helper-redirect-map.c: Likewise.
* gcc.target/bpf/helper-set-hash-invalid.c: Likewise.
* gcc.target/bpf/helper-set-hash.c: Likewise.
* gcc.target/bpf/helper-setsockopt.c: Likewise.
* gcc.target/bpf/helper-sk-fullsock.c: Likewise.
* gcc.target/bpf/helper-sk-lookup-tcp.c: Likewise.
* gcc.target/bpf/helper-sk-lookup-upd.c: Likewise.
* gcc.target/bpf/helper-sk-redirect-hash.c: Likewise.
* gcc.target/bpf/helper-sk-redirect-map.c: Likewise.
* gcc.target/bpf/helper-sk-release.c: Likewise.
* gcc.target/bpf/helper-sk-select-reuseport.c: Likewise.
* gcc.target/bpf/helper-sk-storage-delete.c: Likewise.
* gcc.target/bpf/helper-sk-storage-get.c: Likewise.
* gcc.target/bpf/helper-skb-adjust-room.c: Likewise.
* gcc.target/bpf/helper-skb-cgroup-id.c: Likewise.
* gcc.target/bpf/helper-skb-change-head.c: Likewise.
* gcc.target/bpf/helper-skb-change-proto.c: Likewise.
* gcc.target/bpf/helper-skb-change-tail.c: Likewise.
* gcc.target/bpf/helper-skb-change-type.c: Likewise.
* gcc.target/bpf/helper-skb-ecn-set-ce.c: Likewise.
* gcc.target/bpf/helper-skb-get-tunnel-key.c: Likewise.
* gcc.target/bpf/helper-skb-get-tunnel-opt.c: Likewise.
* gcc.target/bpf/helper-skb-get-xfrm-state.c: Likewise.
* gcc.target/bpf/helper-skb-load-bytes-relative.c: Likewise.
* gcc.target/bpf/helper-skb-load-bytes.c: Likewise.
* gcc.target/bpf/helper-skb-pull-data.c: Likewise.
* gcc.target/bpf/helper-skb-set-tunnel-key.c: Likewise.
* gcc.target/bpf/helper-skb-set-tunnel-opt.c: Likewise.
* gcc.target/bpf/helper-skb-store-bytes.c: Likewise.
* gcc.target/bpf/helper-skb-under-cgroup.c: Likewise.
* gcc.target/bpf/helper-skb-vlan-pop.c: Likewise.
* gcc.target/bpf/helper-skb-vlan-push.c: Likewise.
* gcc.target/bpf/helper-skc-lookup-tcp.c: Likewise.
* gcc.target/bpf/helper-sock-hash-update.c: Likewise.
* gcc.target/bpf/helper-sock-map-update.c: Likewise.
* gcc.target/bpf/helper-sock-ops-cb-flags-set.c: Likewise.
* gcc.target/bpf/helper-spin-lock.c: Likewise.
* gcc.target/bpf/helper-spin-unlock.c: Likewise.
* gcc.target/bpf/helper-strtol.c: Likewise.
* gcc.target/bpf/helper-strtoul.c: Likewise.
* gcc.target/bpf/helper-sysctl-get-current-value.c: Likewise.
* gcc.target/bpf/helper-sysctl-get-name.c: Likewise.
* gcc.target/bpf/helper-sysctl-get-new-value.c: Likewise.
* gcc.target/bpf/helper-sysctl-set-new-value.c: Likewise.
* gcc.target/bpf/helper-tail-call.c: Likewise.
* gcc.target/bpf/helper-tcp-check-syncookie.c: Likewise.
* gcc.target/bpf/helper-tcp-sock.c: Likewise.
* gcc.target/bpf/helper-trace-printk.c: Likewise.
* gcc.target/bpf/helper-xdp-adjust-head.c: Likewise.
* gcc.target/bpf/helper-xdp-adjust-meta.c: Likewise.
* gcc.target/bpf/helper-xdp-adjust-tail.c: Likewise.
* gcc.target/bpf/skb-ancestor-cgroup-id.c: Likewise.

tree-optimization/96491 - avoid store commoning across abnormal edges

This avoids store commoning across abnormal edges since that easily
can disrupt abnormal coalescing because it might create overlapping
lifetime of variables.

2020-08-06  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96491
* tree-ssa-sink.c (sink_common_stores_to_bb): Avoid
sinking across abnormal edges.

* gcc.dg/torture/pr96491.c: New testcase.

tree-optimization/96483 - fix ICE in PRE with POLY_INT_CST

This adds a missing case for PRE expression re-materialization.

2020-08-06  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96483
* tree-ssa-pre.c (create_component_ref_by_pieces_1): Handle
POLY_INT_CST.

Remove std::map use from graphite

This replaces the use of std::map with hash_map for mapping
ISL ids to SSA names.

2020-08-06  Richard Biener  <rguenther@suse.de>

* graphite-isl-ast-to-gimple.c (ivs_params): Use hash_map instead
of std::map.
(ivs_params_clear): Adjust.
(gcc_expression_from_isl_ast_expr_id): Likewise.
(graphite_create_new_loop): Likewise.
(add_parameters_to_ivs_params): Likewise.

x86_64: Integer min/max improvements.

This patch tweaks the way that min and max are expanded, so that the
semantics of these operations are visible to the early RTL optimization
passes, until split into explicit comparison and conditional move
instructions. The good news is that i386.md already contains all of
the required logic (many thanks to Richard Biener and Uros Bizjak),
but this is currently only enabled to scalar-to-vector (STV) synthesis
of min/max instructions.  This change enables this functionality for
all TARGET_CMOVE architectures for SImode, HImode and DImode.

2020-08-06  Roger Sayle  <roger@nextmovesoftware.com>
    Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog
* config/i386/i386.md (MAXMIN_IMODE): No longer needed.
(<maxmin><mode>3):  Support SWI248 and general_operand for
second operand, when TARGET_CMOVE.
(<maxmin><mode>3_1 splitter): Optimize comparisons against
0, 1 and -1 to use "test" instead of "cmp".
(*<maxmin>di3_doubleword): Likewise, allow general_operand
and enable on TARGET_CMOVE.
(peephole2): Convert clearing a register after a flag setting
instruction into an xor followed by the original flag setter.

gcc/testsuite/ChangeLog
* gcc.target/i386/minmax-8.c: New test.
* gcc.target/i386/minmax-9.c: New test.
* gcc.target/i386/minmax-10.c: New test.
* gcc.target/i386/minmax-11.c: New test.

ipa-fnsummary: Include <vector> the proper way

This fixes a bootstrap error with clang 10 that would complain
  /usr/include/c++/v1/typeinfo:346:5: error: no member named
  'fancy_abort' in namespace 'std::__1'; did you mean simply
  'fancy_abort'?
It mirrors how this is handled in gcov.c and indirectly includes
<vector> via system.h.

gcc/ChangeLog:

* ipa-fnsummary.c (INCLUDE_VECTOR): Define.
Remove direct inclusion of <vector>.

vect/rs6000: Support vector with length cost modeling

This patch is to add the cost modeling for vector with length,
it mainly follows what we generate for vector with length in
functions vect_set_loop_controls_directly and vect_gen_len
at the worst case.

For Power, the length is expected to be in bits 0-7 (high bits),
we have to model the cost of shifting bits, which is implemented
in adjust_vect_cost_per_loop.

Bootstrapped/regtested on powerpc64le-linux-gnu (P9) with explicit
param vect-partial-vector-usage=1.

gcc/ChangeLog:

* config/rs6000/rs6000.c (rs6000_adjust_vect_cost_per_loop): New
function.
(rs6000_finish_cost): Call rs6000_adjust_vect_cost_per_loop.
* tree-vect-loop.c (vect_estimate_min_profitable_iters): Add cost
modeling for vector with length.
(vect_rgroup_iv_might_wrap_p): New function, factored out from...
* tree-vect-loop-manip.c (vect_set_loop_controls_directly): ...this.
Update function comment.
* tree-vect-stmts.c (vect_gen_len): Update function comment.
* tree-vectorizer.h (vect_rgroup_iv_might_wrap_p): New declare.

vect: Skip epilogue loops for dbgcnt check [PR96451]

As the PR shows, commit r11-2453 exposed one issue that vectorizer
wants to vectorize the epilogue loop and leaves the if-cvt body there,
but later dbgcnt check disables it, the left scalar mask_store
statement causes ICE.

As Richard pointed out in that PR, the dbgcnt is to count original
scalar loops, so this fix is to make it skip the epilogue loops.

gcc/ChangeLog:

* tree-vectorizer.c (try_vectorize_loop_1): Skip the epilogue loops
for dbgcnt check.

Daily bump.

libstdc++: Break long lines to fit in 80 columns

libstdc++-v3/ChangeLog:

* include/std/atomic (atomic<T>::store): Reformat.

libstdc++: Change URL for PSTL again

libstdc++-v3/ChangeLog:

* doc/xml/manual/status_cxx2017.xml: Replace oneAPI DPC++ link
with LLVM repo for PSTL.
* doc/html/manual/status.html: Regenerate.

libstdc++: Replace operator>>(istream&, char*) [LWG 2499]

P0487R1 resolved LWG 2499 for C++20 by removing the operator>> overloads
that have high risk of buffer overflows. They were replaced by
equivalents that only accept a reference to an array, and so can
guarantee not to write past the end of the array.

In order to support both the old and new functionality, this patch
introduces a new overloaded __istream_extract function which takes a
maximum length. The new operator>> overloads use the array size as the
maximum length. The old overloads now use __builtin_object_size to
determine the available buffer size if available (which requires -O2) or
use numeric_limits<streamsize>::max()/sizeof(char_type) otherwise. This
is a change in behaviour, as the old overloads previously always used
numeric_limits<streamsize>::max(), without considering sizeof(char_type)
and without attempting to prevent overflows.

Because they now do little more than call __istream_extract, the old
operator>> overloads are very small inline functions. This means there
is no advantage to explicitly instantiating them in the library (in fact
that would prevent the __builtin_object_size checks from ever working).
As a result, the explicit instantiation declarations can be removed from
the header. The explicit instantiation definitions are still needed, for
backwards compatibility with existing code that expects to link to the
definitions in the library.

While working on this change I noticed that src/c++11/istream-inst.cc
has the following explicit instantiation definition:
  template istream& operator>>(istream&, char*);
This had no effect (and so should not have been present in that file),
because there was an explicit specialization declared in <istream> and
defined in src/++98/istream.cc. However, this change removes the
explicit specialization, and now the explicit instantiation definition
is necessary to ensure the symbol gets defined in the library.

libstdc++-v3/ChangeLog:

* config/abi/pre/gnu.ver (GLIBCXX_3.4.29): Export new symbols.
* include/bits/istream.tcc (__istream_extract): New function
template implementing both of operator>>(istream&, char*) and
operator>>(istream&, char(&)[N]). Add explicit instantiation
declaration for it. Remove explicit instantiation declarations
for old function templates.
* include/std/istream (__istream_extract): Declare.
(operator>>(basic_istream<C,T>&, C*)): Define inline and simply
call __istream_extract.
(operator>>(basic_istream<char,T>&, signed char*)): Likewise.
(operator>>(basic_istream<char,T>&, unsigned char*)): Likewise.
(operator>>(basic_istream<C,T>&, C(7)[N])): Define for LWG 2499.
(operator>>(basic_istream<char,T>&, signed char(&)[N])):
Likewise.
(operator>>(basic_istream<char,T>&, unsigned char(&)[N])):
Likewise.
* include/std/streambuf (basic_streambuf): Declare char overload
of __istream_extract as a friend.
* src/c++11/istream-inst.cc: Add explicit instantiation
definition for wchar_t overload of __istream_extract. Remove
explicit instantiation definitions of old operator>> overloads
for versioned-namespace build.
* src/c++98/istream.cc (operator>>(istream&, char*)): Replace
with __istream_extract(istream&, char*, streamsize).
* testsuite/27_io/basic_istream/extractors_character/char/3.cc:
Do not use variable-length array.
* testsuite/27_io/basic_istream/extractors_character/char/4.cc:
Do not run test for C++20.
* testsuite/27_io/basic_istream/extractors_character/char/9555-ic.cc:
Do not test writing to pointers for C++20.
* testsuite/27_io/basic_istream/extractors_character/char/9826.cc:
Use array instead of pointer.
* testsuite/27_io/basic_istream/extractors_character/wchar_t/3.cc:
Do not use variable-length array.
* testsuite/27_io/basic_istream/extractors_character/wchar_t/4.cc:
Do not run test for C++20.
* testsuite/27_io/basic_istream/extractors_character/wchar_t/9555-ic.cc:
Do not test writing to pointers for C++20.
* testsuite/27_io/basic_istream/extractors_character/char/lwg2499.cc:
New test.
* testsuite/27_io/basic_istream/extractors_character/char/lwg2499_neg.cc:
New test.
* testsuite/27_io/basic_istream/extractors_character/char/overflow.cc:
New test.
* testsuite/27_io/basic_istream/extractors_character/wchar_t/lwg2499.cc:
New test.
* testsuite/27_io/basic_istream/extractors_character/wchar_t/lwg2499_neg.cc:
New test.

c++: cxx_eval_vec_init after zero-initialization [PR96282]

In the first testcase below, expand_aggr_init_1 sets up t's default
constructor such that the ctor first zero-initializes the entire base b,
followed by calling b's default constructor, the latter of which just
default-initializes the array member b::m via a VEC_INIT_EXPR.

So upon constexpr evaluation of this latter VEC_INIT_EXPR, ctx->ctor is
nonempty due to the prior zero-initialization, and we proceed in
cxx_eval_vec_init to append new constructor_elts to the end of ctx->ctor
without first checking if a matching constructor_elt already exists.
This leads to ctx->ctor having two matching constructor_elts for each
index.

This patch fixes this issue by truncating a zero-initialized array
CONSTRUCTOR in cxx_eval_vec_init_1 before we begin appending array
elements to it.  We propagate its zeroed out state during evaluation by
clearing CONSTRUCTOR_NO_CLEARING on each new appended aggregate element.

gcc/cp/ChangeLog:

PR c++/96282
* constexpr.c (cxx_eval_vec_init_1): Truncate ctx->ctor and
then clear CONSTRUCTOR_NO_CLEARING on each appended element
initializer if we're initializing a previously zero-initialized
array object.

gcc/testsuite/ChangeLog:

PR c++/96282
* g++.dg/cpp0x/constexpr-array26.C: New test.
* g++.dg/cpp0x/constexpr-array27.C: New test.
* g++.dg/cpp2a/constexpr-init18.C: New test.

Co-authored-by: Jason Merrill <jason@redhat.com>

Added test case to make sure that legal cases still pass.

gcc/testsuite/ChangeLog:

PR fortran/96469
* gfortran.dg/do_check_14.f90: New test.

Static analysis for definition of DO index variables in contained procedures.

When encountering a procedure call in a DO loop, this patch checks if
the call is to a contained procedure, and if it is, check for
changes in the index variable.

gcc/fortran/ChangeLog:

PR fortran/96469
* frontend-passes.c (doloop_contained_function_call): New
function.
(doloop_contained_procedure_code): New function.
(CHECK_INQ): Macro for inquire checks.
(doloop_code): Invoke doloop_contained_procedure_code and
doloop_contained_function_call if appropriate.
(do_intent): Likewise.

gcc/testsuite/ChangeLog:

PR fortran/96469
* gfortran.dg/do_check_4.f90: Hide change in index variable
from compile-time analysis.
* gfortran.dg/do_check_13.f90: New test.

VEC_COND_EXPR optimizations

When vector comparisons were forced to use vec_cond_expr, we lost a number of optimizations (my fault for not adding enough testcases to
prevent that). This patch tries to unwrap vec_cond_expr a bit so some optimizations can still happen.

I wasn't planning to add all those transformations together, but adding one caused a regression, whose fix introduced a second regression,
etc.

Restricting to constant folding would not be sufficient, we also need at least things like X|0 or X&X. The transformations are quite
conservative with :s and folding only if everything simplifies, we may want to relax this later. And of course we are going to miss things
like a?b:c + a?c:b -> b+c.

In terms of number of operations, some transformations turning 2 VEC_COND_EXPR into VEC_COND_EXPR + BIT_IOR_EXPR + BIT_NOT_EXPR might not look
like a gain... I expect the bit_not disappears in most cases, and VEC_COND_EXPR looks more costly than a simpler BIT_IOR_EXPR.

2020-08-05  Marc Glisse  <marc.glisse@inria.fr>

PR tree-optimization/95906
PR target/70314
* match.pd ((c ? a : b) op d, (c ? a : b) op (c ? d : e),
(v ? w : 0) ? a : b, c1 ? c2 ? a : b : b): New transformations.
(op (c ? a : b)): Update to match the new transformations.

* gcc.dg/tree-ssa/andnot-2.c: New file.
* gcc.dg/tree-ssa/pr95906.c: Likewise.
* gcc.target/i386/pr70314.c: Likewise.

aarch64: Clear canary value after stack_protect_test [PR96191]

The stack_protect_test patterns were leaving the canary value in the
temporary register, meaning that it was often still in registers on
return from the function.  An attacker might therefore have been
able to use it to defeat stack-smash protection for a later function.

gcc/
PR target/96191
* config/aarch64/aarch64.md (stack_protect_test_<mode>): Set the
CC register directly, instead of a GPR.  Replace the original GPR
destination with an extra scratch register.  Zero out operand 3
after use.
(stack_protect_test): Update accordingly.

gcc/testsuite/
PR target/96191
* gcc.target/aarch64/stack-protector-1.c: New test.
* gcc.target/aarch64/stack-protector-2.c: Likewise.

aarch64: Add missing %z prefixes to LDP/STP patterns

For LDP/STP Q, the memory operand might not be valid for "m",
so we need to use %z<N> instead of %<N> in the asm template.
This patch does that for all Ump LDP/STP patterns, regardless
of whether it's strictly needed.

This is needed to unbreak bootstrap.

2020-08-05  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.md (load_pair_sw_<SX:mode><SX2:mode>)
(load_pair_dw_<DX:mode><DX2:mode>, load_pair_dw_tftf)
(store_pair_sw_<SX:mode><SX2:mode>)
(store_pair_dw_<DX:mode><DX2:mode>, store_pair_dw_tftf)
(*load_pair_extendsidi2_aarch64)
(*load_pair_zero_extendsidi2_aarch64): Use %z for the memory operand.
* config/aarch64/aarch64-simd.md (load_pair<DREG:mode><DREG2:mode>)
(vec_store_pair<DREG:mode><DREG2:mode>, load_pair<VQ:mode><VQ2:mode>)
(vec_store_pair<VQ:mode><VQ2:mode>): Likewise.

refactor LIM a bit

This refactors LIM to eschew alloc_aux_for_edges and re-uses the RPO
order of the move_computations walk for invariantness computation as well.
It also removes one unnecessary sorting (but retaining it as checking
code because we bsearch the vector) and moves edge insert commit code
to the place where it doesn't have to scan all the functions edges.

This was all done when investigating whether LIM can be refactored
to work on a specific loop for on-demand processing (but we're not
there yet).

2020-08-05  Richard Biener  <rguenther@suse.de>

* tree-ssa-loop-im.c (invariantness_dom_walker): Remove.
(invariantness_dom_walker::before_dom_children): Move to ...
(compute_invariantness): ... this function.
(move_computations): Inline ...
(tree_ssa_lim): ... here, share RPO order and avoid some
cfun references.
(analyze_memory_references): Remove sorting of location
lists, instead assert they are sorted already when checking.
(prev_flag_edges): Remove.
(execute_sm_if_changed): Pass down and adjust prev edge state.
(execute_sm_exit): Likewise.
(hoist_memory_references): Likewise.  Commit edge insertions
of each processed exit.
(store_motion_loop): Do not commit edge insertions on all
edges in the function.
(tree_ssa_lim_initialize): Do not call alloc_aux_for_edges.
(tree_ssa_lim_finalize): Do not call free_aux_for_edges.

Make genmatch transform failure handling more consistent

Currently whether a fail during the transform stage is fatal or
whether following patterns are still considers is a bit random
depending on whether the pattern is wrapped in a for for example.
The follwing makes it consistent by replacing early returns with
gotos to the end of the pattern processing.

2020-08-05  Richard Biener  <rguenther@suse.de>

* genmatch.c (fail_label): New global.
(expr::gen_transform): Branch to fail_label instead of
returning.  Fix indent of call argument checking.
(dt_simplify::gen_1): Compute and emit fail_label, branch
to it instead of returning early.

openmp: Handle even some combined non-rectangular loops

The number of loops computation and logical iteration -> actual iterator values
computations can now be done separately even on composite constructs (though
for triangular loops it would still be more efficient to propagate a few values
through, will handle that incrementally).
simd and taskloop are still unhandled.

2020-08-05  Jakub Jelinek  <jakub@redhat.com>

* omp-expand.c (expand_omp_for): Don't disallow combined non-rectangular
loops.

* testsuite/libgomp.c/loop-22.c: New test.
* testsuite/libgomp.c/loop-23.c: New test.

openmp: Handle reduction clauses on host teams construct [PR96459]

As the new testcase shows, we weren't actually performing reductions on
host teams construct.  And fixing that revealed a flaw in the for-14.c testcase.
The problem is that the tests perform also initialization and checking around the
calls to the functions with the OpenMP constructs.  In that testcase, all the
tests have been spawned from a teams construct but only the tested loops were
distribute, which means the initialization and checking has been performed
redundantly and racily in each team.  Fixed by performing the initialization
and checking outside of host teams and only do the calls to functions with
the tested constructs inside of host teams.

2020-08-05  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/96459
* omp-low.c (lower_omp_taskreg): Call lower_reduction_clauses even in
for host teams.

* testsuite/libgomp.c/teams-3.c: New test.
* testsuite/libgomp.c-c++-common/for-2.h (OMPTEAMS): Define to nothing
if not defined yet.
(N(test)): Use it before all N(f*) calls.
* testsuite/libgomp.c-c++-common/for-14.c (DO_PRAGMA, OMPTEAMS): Define.
(main): Don't call all test_* functions from within
#pragma omp teams reduction(|:err), call them directly.

openmp: Use more efficient logical -> actual computation even if # iterations is computed at runtime

For triangular loops use more efficient logical iteration number
to actual iterator values computation even for non-rectangular loops
where number of loop iterations could not be computed at compile time.

2020-08-05  Jakub Jelinek  <jakub@redhat.com>

* omp-expand.c (expand_omp_for_init_counts): Remember
first_inner_iterations, factor and n1o from the number of iterations
computation in *fd.
(expand_omp_for_init_vars): Use more efficient logical iteration number
to actual iterator values computation even for non-rectangular loops
where number of loop iterations could not be computed at compile time.

rs6000 Add vector blend, permute builtin support

GCC maintainers:

The following patch adds support for the vec_blendv and vec_permx
builtins.

The patch has been compiled and tested on

  powerpc64le-unknown-linux-gnu (Power 8 LE)
  powerpc64le-unknown-linux-gnu (Power 9 LE)

with no regression errors.

The test cases were compiled on a Power 9 system and then tested on
Mambo.

                         Carl Love

rs6000 RFC2609 vector blend, permute instructions

gcc/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h (vec_blendv, vec_permx): Add define.
* config/rs6000/altivec.md (UNSPEC_XXBLEND, UNSPEC_XXPERMX.): New
unspecs.
(VM3): New define_mode.
(VM3_char): New define_attr.
(xxblend_<mode> mode VM3): New define_insn.
(xxpermx): New define_expand.
(xxpermx_inst): New define_insn.
* config/rs6000/rs6000-builtin.def (VXXBLEND_V16QI, VXXBLEND_V8HI,
VXXBLEND_V4SI, VXXBLEND_V2DI, VXXBLEND_V4SF, VXXBLEND_V2DF): New
BU_P10V_3 definitions.
(XXBLEND): New BU_P10_OVERLOAD_3 definition.
(XXPERMX): New BU_P10_OVERLOAD_4 definition.
* config/rs6000/rs6000-c.c (altivec_resolve_overloaded_builtin):
(P10_BUILTIN_VXXPERMX): Add if statement.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VXXBLEND_V16QI,
P10_BUILTIN_VXXBLEND_V8HI, P10_BUILTIN_VXXBLEND_V4SI,
P10_BUILTIN_VXXBLEND_V2DI, P10_BUILTIN_VXXBLEND_V4SF,
P10_BUILTIN_VXXBLEND_V2DF, P10_BUILTIN_VXXPERMX): Define
overloaded arguments.
(rs6000_expand_quaternop_builtin): Add if case for CODE_FOR_xxpermx.
(builtin_quaternary_function_type): Add v16uqi_type and xxpermx_type
variables, add case statement for P10_BUILTIN_VXXPERMX.
(builtin_function_type): Add case statements for
P10_BUILTIN_VXXBLEND_V16QI, P10_BUILTIN_VXXBLEND_V8HI,
P10_BUILTIN_VXXBLEND_V4SI, P10_BUILTIN_VXXBLEND_V2DI.
* doc/extend.texi: Add documentation for vec_blendv and vec_permx.

gcc/testsuite/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>
* gcc.target/powerpc/vec-blend-runnable.c: New test.
* gcc.target/powerpc/vec-permute-ext-runnable.c: New test.

rs6000, Add vector splat builtin support

GCC maintainers:

The following patch adds support for the vec_splati, vec_splatid and
vec_splati_ins builtins.

This patch adds support for instructions that take a 32-bit immediate
value that represents a floating point value.  This support adds new
predicates and a support function to properly handle the immediate value.

The patch has been compiled and tested on

  powerpc64le-unknown-linux-gnu (Power 8 LE)
  powerpc64le-unknown-linux-gnu (Power 9 LE)

with no regression errors.

The test case was compiled on a Power 9 system and then tested on
Mambo.

Please let me know if this patch is acceptable for the mainline
branch.  Thanks.

                         Carl Love
--------------------------------------------------------
gcc/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h (vec_splati, vec_splatid, vec_splati_ins):
Add defines.
* config/rs6000/altivec.md (UNSPEC_XXSPLTIW, UNSPEC_XXSPLTID,
UNSPEC_XXSPLTI32DX): New.
(vxxspltiw_v4si, vxxspltiw_v4sf_inst, vxxspltidp_v2df_inst,
vxxsplti32dx_v4si_inst, vxxsplti32dx_v4sf_inst): New define_insn.
(vxxspltiw_v4sf, vxxspltidp_v2df, vxxsplti32dx_v4si,
vxxsplti32dx_v4sf.): New define_expands.
* config/rs6000/predicates.md (u1bit_cint_operand,
s32bit_cint_operand, c32bit_cint_operand): New predicates.
* config/rs6000/rs6000-builtin.def (VXXSPLTIW_V4SI, VXXSPLTIW_V4SF,
VXXSPLTID): New definitions.
(VXXSPLTI32DX_V4SI, VXXSPLTI32DX_V4SF): New BU_P10V_3
definitions.
(XXSPLTIW, XXSPLTID): New definitions.
(XXSPLTI32DX): Add definitions.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VEC_XXSPLTIW,
P10_BUILTIN_VEC_XXSPLTID, P10_BUILTIN_VEC_XXSPLTI32DX):
New definitions.
* config/rs6000/rs6000-protos.h (rs6000_constF32toI32): New extern
declaration.
* config/rs6000/rs6000.c (rs6000_constF32toI32): New function.
* doc/extend.texi: Add documentation for vec_splati,
vec_splatid, and vec_splati_ins.

gcc/testsuite/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* gcc.target/powerpc/vec-splati-runnable.c: New test.

rs6000, Add vector shift double builtin support

GCC maintainers:

The following patch adds support for the vector shift double builtins.

The patch has been compiled and tested on

  powerpc64le-unknown-linux-gnu (Power 8 LE)
  powerpc64le-unknown-linux-gnu (Power 9 LE)

and Mambo with no regression errors.

Please let me know if this patch is acceptable for the mainline branch.

Thanks.

                         Carl Love

-------------------------------------------------------

gcc/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h (vec_sldb, vec_srdb): New defines.
* config/rs6000/altivec.md (UNSPEC_SLDB, UNSPEC_SRDB): New.
(SLDB_lr): New attribute.
(VSHIFT_DBL_LR): New iterator.
(vs<SLDB_lr>db_<mode>): New define_insn.
* config/rs6000/rs6000-builtin.def (VSLDB_V16QI, VSLDB_V8HI,
VSLDB_V4SI, VSLDB_V2DI, VSRDB_V16QI, VSRDB_V8HI, VSRDB_V4SI,
VSRDB_V2DI): New BU_P10V_3 definitions.
(SLDB, SRDB): New BU_P10_OVERLOAD_3 definitions.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VEC_SLDB,
P10_BUILTIN_VEC_SRDB): New definitions.
(rs6000_expand_ternop_builtin) [CODE_FOR_vsldb_v16qi,
CODE_FOR_vsldb_v8hi, CODE_FOR_vsldb_v4si, CODE_FOR_vsldb_v2di,
CODE_FOR_vsrdb_v16qi, CODE_FOR_vsrdb_v8hi, CODE_FOR_vsrdb_v4si,
CODE_FOR_vsrdb_v2di]: Add clauses.
* doc/extend.texi: Add description for vec_sldb and vec_srdb.

gcc/testsuite/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* gcc.target/powerpc/vec-shift-double-runnable.c:  New test file.

rs6000, Add vector replace builtin support GCC maintainers:

The following patch adds support for builtins vec_replace_elt and
vec_replace_unaligned.

The patch has been compiled and tested on

  powerpc64le-unknown-linux-gnu (Power 8 LE)
  powerpc64le-unknown-linux-gnu (Power 9 LE)

and mambo with no regression errors.

Please let me know if this patch is acceptable for the mainline
branch.  Thanks.

                         Carl Love

-------------------------------------------------------

gcc/ChangeLog

    2020-08-04 Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h: Add define for vec_replace_elt and
vec_replace_unaligned.
* config/rs6000/vsx.md (UNSPEC_REPLACE_ELT, UNSPEC_REPLACE_UN): New
unspecs.
(REPLACE_ELT): New mode iterator.
(REPLACE_ELT_char, REPLACE_ELT_sh, REPLACE_ELT_max): New mode attributes.
(vreplace_un_<mode>, vreplace_elt_<mode>_inst): New.
* config/rs6000/rs6000-builtin.def (VREPLACE_ELT_V4SI,
VREPLACE_ELT_UV4SI, VREPLACE_ELT_V4SF, VREPLACE_ELT_UV2DI,
VREPLACE_ELT_V2DF, VREPLACE_UN_V4SI, VREPLACE_UN_UV4SI,
VREPLACE_UN_V4SF, VREPLACE_UN_V2DI, VREPLACE_UN_UV2DI,
VREPLACE_UN_V2DF, (REPLACE_ELT, REPLACE_UN, VREPLACE_ELT_V2DI): New builtin
entries.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VEC_REPLACE_ELT,
P10_BUILTIN_VEC_REPLACE_UN): New builtin argument definitions.
(rs6000_expand_quaternop_builtin): Add 3rd argument checks for
CODE_FOR_vreplace_elt_v4si, CODE_FOR_vreplace_elt_v4sf,
CODE_FOR_vreplace_un_v4si, CODE_FOR_vreplace_un_v4sf.
(builtin_function_type) [P10_BUILTIN_VREPLACE_ELT_UV4SI,
P10_BUILTIN_VREPLACE_ELT_UV2DI, P10_BUILTIN_VREPLACE_UN_UV4SI,
P10_BUILTIN_VREPLACE_UN_UV2DI]: New cases.
* doc/extend.texi: Add description for vec_replace_elt and
vec_replace_unaligned builtins.

gcc/testsuite/ChangeLog

    2020-08-04 Carl Love  <cel@us.ibm.com>

* gcc.target/powerpc/vec-replace-word-runnable.c: New test.

rs6000 Add vector insert builtin support

GCC maintainers:

This patch adds support for vec_insertl and vec_inserth builtins.

The patch has been compiled and tested on

  powerpc64le-unknown-linux-gnu (Power 8 LE)
  powerpc64le-unknown-linux-gnu (Power 9 LE)

and mambo with no regression errors.

Please let me know if this patch is acceptable for the mainline branch.

Thanks.

                         Carl Love

--------------------------------------------------------------
gcc/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h (vec_insertl, vec_inserth): New defines.
* config/rs6000/rs6000-builtin.def (VINSERTGPRBL, VINSERTGPRHL,
VINSERTGPRWL, VINSERTGPRDL, VINSERTVPRBL, VINSERTVPRHL, VINSERTVPRWL,
VINSERTGPRBR, VINSERTGPRHR, VINSERTGPRWR, VINSERTGPRDR, VINSERTVPRBR,
VINSERTVPRHR, VINSERTVPRWR): New builtins.
(INSERTL, INSERTH): New builtins.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VEC_INSERTL,
P10_BUILTIN_VEC_INSERTH): New overloaded definitions.
(P10_BUILTIN_VINSERTGPRBL, P10_BUILTIN_VINSERTGPRHL,
P10_BUILTIN_VINSERTGPRWL, P10_BUILTIN_VINSERTGPRDL,
P10_BUILTIN_VINSERTVPRBL, P10_BUILTIN_VINSERTVPRHL,
P10_BUILTIN_VINSERTVPRWL): Add case entries.
* config/rs6000/vsx.md (define_c_enum): Add UNSPEC_INSERTL,
UNSPEC_INSERTR.
(define_expand): Add vinsertvl_<mode>, vinsertvr_<mode>,
vinsertgl_<mode>, vinsertgr_<mode>, mode is VI2.
(define_ins): vinsertvl_internal_<mode>, vinsertvr_internal_<mode>,
vinsertgl_internal_<mode>, vinsertgr_internal_<mode>, mode VEC_I.
* doc/extend.texi: Add documentation for vec_insertl, vec_inserth.

gcc/testsuite/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* gcc.target/powerpc/vec-insert-word-runnable.c: New test case.

rs6000, Update support for vec_extract

GCC maintainers:

Move the existing vector extract support in altivec.md to vsx.md
so all of the vector insert and extract support is in the same file.

The patch also updates the name of the builtins and descriptions for the
builtins in the documentation file so they match the approved builtin
names and descriptions.

The patch does not make any functional changes.

Please let me know if the changes are acceptable for mainline.  Thanks.

                  Carl Love

------------------------------------------------------

gcc/ChangeLog

    2020-08-04  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.md: (UNSPEC_EXTRACTL, UNSPEC_EXTRACTR)
(vextractl<mode>, vextractr<mode>)
(vextractl<mode>_internal, vextractr<mode>_internal for mode VI2)
(VI2): Move to ...
* config/rs6000/vsx.md: (UNSPEC_EXTRACTL, UNSPEC_EXTRACTR)
(vextractl<mode>, vextractr<mode>)
(vextractl<mode>_internal, vextractr<mode>_internal for mode VI2)
(VI2):  ..here.
* doc/extend.texi: Update documentation for vec_extractl.
Replace builtin name vec_extractr with vec_extracth.  Update
description of vec_extracth.

Daily bump.

aarch64: Delete duplicated option docs.

Noticed while reviewing the RISC-V -mstack-protector-guard docs.  The
AArch64 section has two identical copies of the docs for this option.

gcc/
* doc/invoke.texi (AArch64 Options): Delete duplicate
-mstack-protector-guard docs.

[PATCH] nvptx: Add support for PTX highpart multiplications (HI/SI)

This patch adds support for signed and unsigned, HImode and SImode highpart
multiplications to the nvptx backend.

This patch has been tested on nvptx-none hosted on x86_64-pc-linux-gnu
with a "make" and "make -k check" with no new failures with the
above patch.

2020-08-04  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog:

* config/nvptx/nvptx.md (smulhi3_highpart, smulsi3_highpart)
(umulhi3_highpart, umulsi3_highpart): New instructions.

gcc/testsuite/ChangeLog:

* gcc.target/nvptx/mul-hi.c: New test.
* gcc.target/nvptx/umul-hi.c: New test.

c++: Template keyword following :: [PR96082]

In r9-4235 I tried to make sure that the template keyword follows
a nested-name-specifier.  :: is a valid nested-name-specifier, so
I also have to check 'globalscope' before giving the error.

gcc/cp/ChangeLog:

PR c++/96082
* parser.c (cp_parser_elaborated_type_specifier): Allow
'template' following ::.

gcc/testsuite/ChangeLog:

PR c++/96082
* g++.dg/template/template-keyword3.C: New test.

compiler: delete lowered constant strings

If we lower a constant string operation in a Binary_expression,
delete the strings.  This is safe because constant strings are always
newly allocated.

This is a hack to use much less memory when compiling the new
time/tzdata package, which has a file that contains the sum of over
13,000 constant strings.  We don't do this for numeric expressions
because that could cause us to delete an Iota_expression.

We should have a cleaner approach to memory usage some day.

Fixes PR go/96450

amdgcn: Remove dead defines from gcn-run

Nothing uses these since the switch to HSACOv3.

gcc/ChangeLog:

* config/gcn/gcn-run.c (R_AMDGPU_NONE): Delete.
(R_AMDGPU_ABS32_LO): Delete.
(R_AMDGPU_ABS32_HI): Delete.
(R_AMDGPU_ABS64): Delete.
(R_AMDGPU_REL32): Delete.
(R_AMDGPU_REL64): Delete.
(R_AMDGPU_ABS32): Delete.
(R_AMDGPU_GOTPCREL): Delete.
(R_AMDGPU_GOTPCREL32_LO): Delete.
(R_AMDGPU_GOTPCREL32_HI): Delete.
(R_AMDGPU_REL32_LO): Delete.
(R_AMDGPU_REL32_HI): Delete.
(reserved): Delete.
(R_AMDGPU_RELATIVE64): Delete.

[Arm] Modify default tuning of armv8.1-m.main to use Cortex-M55

Previously, compiling with -march=armv8.1-m.main would tune for
Cortex-M7.
However, the Cortex-M7 only supports up to Armv7e-M. The Cortex-M55 is
the earliest CPU that supports Armv8.1-M Mainline so is more appropriate.
This also has the effect of changing the branch cost function used, which
will be necessary to correctly prioritise conditional instructions over branches
in the rest of this patch series.

Regression tested on arm-none-eabi.

gcc/ChangeLog

2020-08-04  Omar Tahir  <omar.tahir@arm.com>

* config/arm/arm-cpus.in (armv8.1-m.main): Tune for Cortex-M55.

aarch64: Delete unnecessary code

gcc/
* config/aarch64/aarch64.c (aarch64_if_then_else_costs): Delete
redundant extra_cost variable.

c++: fix template parm count leak

I noticed that we could leak parser->num_template_parameter_lists with
erroneous specializations.  We'd increment, notice a problem and then
bail out.  This refactors cp_parser_explicit_specialization to avoid
that code path.  A couple of tests get different diagnostics because
of the fix.  pr39425 then goes to unbounded template instantiation
and exceeds the implementation limit.

gcc/cp/
* parser.c (cp_parser_explicit_specialization): Refactor
to avoid leak of num_template_parameter_lists value.
gcc/testsuite/
* g++.dg/template/pr39425.C: Adjust errors, (unbounded
template recursion).
* g++.old-deja/g++.pt/spec20.C: Remove fallout diagnostics.

AArch64: Use FLOAT_MODE_P macro and add FLAG_AUTO_FP [PR94442]

Since all FP intrinsics are set by FLAG_FP by default, but not all FP intrinsics
raise FP exceptions or read FPCR register.  So we add a global flag FLAG_AUTO_FP
to suppress the flag FLAG_FP.

2020-08-04  Zhiheng Xie  <xiezhiheng@huawei.com>

gcc/ChangeLog:

* config/aarch64/aarch64-builtins.c (aarch64_call_properties):
Use FLOAT_MODE_P macro instead of enumerating all floating-point
modes and add global flag FLAG_AUTO_FP.

Fortran/OpenMP: Fix detecting not perfectly nested loops

gcc/fortran/ChangeLog:

* openmp.c (resolve_omp_do): Detect not perfectly
nested loop with innermost collapse.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/collapse1.f90: Add dg-error.
* gfortran.dg/gomp/collapse2.f90: New test.

doc: Add @cindex to symver attribute

When looking at the symver attr documentation in html, I found there is no
name to refer to for it.

2020-08-04  Jakub Jelinek  <jakub@redhat.com>

* doc/extend.texi (symver): Add @cindex for symver function attribute.

Test case for PR rtl-optimization/60473

PR rtl-optimization/60473 is code quality regression that has
been cured by improvements to register allocation.  For the function
in the test case, GCC 4.4, 4.5 and 4.6 generated very poor code
requiring two mov instructions, and GCC 4.7 and 4.8 (when the PR was
filed) produced better but still poor code with one mov instruction.
Since GCC 4.9 (including current mainline), it generates optimal
code with no mov instructions, matching what used to be generated
in GCC 4.1.

2020-08-04  Roger Sayle  <roger@nextmovesoftware.com>

gcc/testsuite/ChangeLog
PR rtl-optimization/60473
* gcc.target/i386/pr60473.c: New test.

Simplify X * C1 == C2 with undefined overflow

this transformation is quite straightforward, without overflow, 3*X==15 is
the same as X==5 and 3*X==5 cannot happen. Adding a single_use restriction
for the first case didn't seem necessary, although of course it can
slightly increase register pressure in some cases.

2020-08-04  Marc Glisse  <marc.glisse@inria.fr>

PR tree-optimization/95433
* match.pd (X * C1 == C2): New transformation.

* gcc.c-torture/execute/pr23135.c: Add -fwrapv to avoid
undefined behavior.
* gcc.dg/tree-ssa/pr95433.c: New file.

Adjust gimple-ssa-sprintf.c for irange API.

gcc/ChangeLog:

* gimple-ssa-sprintf.c (get_int_range): Adjust for irange API.
(format_integer): Same.
(handle_printf_call): Same.

d: Fix struct literals that have non-deterministic hash values (PR96153)

Adds code generation for generating a temporary for, and pre-filling
struct and array literals with zeroes before assigning, so that
alignment holes don't cause objects to produce a non-deterministic hash
value.  A new field has been added to the expression visitor to track
whether the result is being generated for another literal, so that
memset() is only called once on the top-level literal expression, and
not for nesting struct or arrays.

gcc/d/ChangeLog:

PR d/96153
* d-tree.h (build_expr): Add literalp argument.
* expr.cc (ExprVisitor): Add literalp_ field.
(ExprVisitor::ExprVisitor): Initialize literalp_.
(ExprVisitor::visit (AssignExp *)): Call memset() on blits where RHS
is a struct literal.  Elide assignment if initializer is all zeroes.
(ExprVisitor::visit (CastExp *)): Forward literalp_ to generation of
subexpression.
(ExprVisitor::visit (AddrExp *)): Likewise.
(ExprVisitor::visit (ArrayLiteralExp *)): Use memset() to pre-fill
object with zeroes.  Set literalp in subexpressions.
(ExprVisitor::visit (StructLiteralExp *)): Likewise.
(ExprVisitor::visit (TupleExp *)): Set literalp in subexpressions.
(ExprVisitor::visit (VectorExp *)): Likewise.
(ExprVisitor::visit (VectorArrayExp *)): Likewise.
(build_expr): Forward literal_p to ExprVisitor.

gcc/testsuite/ChangeLog:

PR d/96153
* gdc.dg/pr96153.d: New test.

amdgcn: TImode shifts

Implement TImode shifts in the backend.

The middle-end support that does it for other architectures doesn't work for
GCN because BITS_PER_WORD==32, meaning that TImode is quad-word, not
double-word.

gcc/ChangeLog:

* config/gcn/gcn.md ("<expander>ti3"): New.

c++: Member initializer list diagnostic locations [PR94024]

This patch preserves the source locations of each node in a member
initializer list so that during processing of the list we can set
input_location appropriately for generally more accurate diagnostic
locations.  Since TREE_LIST nodes are tcc_exceptional, they can't have
source locations, so we instead store the location in a dummy
tcc_expression node within the TREE_TYPE of the list node.

gcc/cp/ChangeLog:

PR c++/94024
* init.c (sort_mem_initializers): Preserve TREE_TYPE of the
member initializer list node.
(emit_mem_initializers): Set input_location when performing each
member initialization.
* parser.c (cp_parser_mem_initializer): Attach the source
location of this initializer to a dummy EMPTY_CLASS_EXPR
within the TREE_TYPE of the list node.
* pt.c (tsubst_initializer_list): Preserve TREE_TYPE of the
member initializer list node.

gcc/testsuite/ChangeLog:

PR c++/94024
* g++.dg/diagnostic/mem-init1.C: New test.

tree-optimization/88240 - stopgap for floating point code-hoisting issues

This adds a stopgap measure to avoid performing code-hoisting
on mixed type loads when the load we'd insert in the hoisting
position would be a floating point one.  This is because certain
targets (hello x87) cannot perform floating point loads without
possibly altering the bit representation and thus cannot be used
in place of integral loads.

2020-08-04  Richard Biener  <rguenther@suse.de>

PR tree-optimization/88240
* tree-ssa-sccvn.h (vn_reference_s::punned): New flag.
* tree-ssa-sccvn.c (vn_reference_insert): Initialize punned.
(vn_reference_insert_pieces): Likewise.
(visit_reference_op_call): Likewise.
(visit_reference_op_load): Track whether a ref was punned.
* tree-ssa-pre.c (do_hoist_insertion): Refuse to perform hoist
insertion on punned floating point loads.

* gcc.target/i386/pr88240.c: New testcase.

Fortran: Fix for OpenMP's 'lastprivate(conditional:'

gcc/fortran/ChangeLog:

* trans-openmp.c (gfc_trans_omp_do): Fix 'lastprivate(conditional:'.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/lastprivate-conditional-3.f90: Enable some
previously disabled 'lastprivate(conditional:' dg-warnings.

aarch64: Use Q-reg loads/stores in movmem expansion

This is my attempt at reviving the old patch
https://gcc.gnu.org/pipermail/gcc-patches/2019-January/514632.html

I have followed on Kyrill's comment upstream on the link above and I
am using the recommended option iii that he mentioned.
"1) Adjust the copy_limit to 256 bits after checking
    AARCH64_EXTRA_TUNE_NO_LDP_STP_QREGS in the tuning.
 2) Adjust aarch64_copy_one_block_and_progress_pointers to handle
    256-bit moves. by iii:
   iii) Emit explicit V4SI (or any other 128-bit vector mode) pairs
ldp/stps. This wouldn't need any adjustments to MD patterns,
but would make aarch64_copy_one_block_and_progress_pointers
more complex as it would now have two paths, where one
handles two adjacent memory addresses in one calls."

gcc/ChangeLog:

* config/aarch64/aarch64.c (aarch64_gen_store_pair): Add case
for E_V4SImode.
(aarch64_gen_load_pair): Likewise.
(aarch64_copy_one_block_and_progress_pointers): Handle 256 bit copy.
(aarch64_expand_cpymem): Expand copy_limit to 256bits where
appropriate.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/cpymem-q-reg_1.c: New test.
* gcc.target/aarch64/large_struct_copy_2.c: Update for ldp q regs.

aarch64: Add missing clobber for fjcvtzs

gcc/ChangeLog

2020-07-30  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/aarch64.md (aarch64_fjcvtzs): Add missing
clobber.
* doc/sourcebuild.texi (aarch64_fjcvtzs_hw) Document new
target supports option.

gcc/testsuite/ChangeLog

2020-07-30  Andrea Corallo  <andrea.corallo@arm.com>

* gcc.target/aarch64/acle/jcvt_2.c: New testcase.
* lib/target-supports.exp
(check_effective_target_aarch64_fjcvtzs_hw): Add new check for
FJCVTZS hw.

[nvptx] Handle V2DI/V2SI mode in nvptx_gen_shuffle

With the pr96628-part1.f90 source and -ftree-slp-vectorize, we run into an
ICE due to the fact that V2DI mode is not handled in nvptx_gen_shuffle.

Fix this by adding handling of V2DI as well as V2SI mode in
nvptx_gen_shuffle.

Build and reg-tested on x86_64 with nvptx accelerator.

gcc/ChangeLog:

PR target/96428
* config/nvptx/nvptx.c (nvptx_gen_shuffle): Handle V2SI/V2DI.

libgomp/ChangeLog:

PR target/96428
* testsuite/libgomp.oacc-fortran/pr96628-part1.f90: New test.
* testsuite/libgomp.oacc-fortran/pr96628-part2.f90: New test.

veclower: Don't ICE on .VEC_CONVERT calls with no lhs [PR96426]

.VEC_CONVERT is a const internal call, so normally if the lhs is not used,
we'd DCE it far before getting to veclower, but with -O0 (or perhaps
-fno-tree-dce and some other -fno-* options) it can happen.
But as the internal fn needs the lhs to know the type to which the
conversion is done (and I think that is a reasonable representation, having
some magic another argument and having to create constants with that type
looks overkill to me), we just should DCE those calls ourselves.
During veclower, we can't really remove insns, as the callers would be
upset, so this just replaces it with a GIMPLE_NOP.

2020-08-04  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/96426
* tree-vect-generic.c (expand_vector_conversion): Replace .VEC_CONVERT
call with GIMPLE_NOP if there is no lhs.

* gcc.c-torture/compile/pr96426.c: New test.

gimple-fold: Fix ICE in maybe_canonicalize_mem_ref_addr on debug stmt [PR96354]

In debug stmts, we are less strict about what is and what is not accepted
there, so this patch just punts on optimization of a debug stmt rather than
ICEing.

2020-08-04  Jakub Jelinek  <jakub@redhat.com>

PR debug/96354
* gimple-fold.c (maybe_canonicalize_mem_ref_addr): Add IS_DEBUG
argument.  Return false instead of gcc_unreachable if it is true and
get_addr_base_and_unit_offset returns NULL.
(fold_stmt_1) <case GIMPLE_DEBUG>: Adjust caller.

* g++.dg/opt/pr96354.C: New test.

Add is_gimple_min_invariant dropped from previous patch.

gcc/ChangeLog:

* vr-values.c (simplify_using_ranges::vrp_evaluate_conditional):
Call is_gimple_min_invariant dropped from previous patch.

openmp: Compute number of collapsed loop iterations more efficiently for some non-rectangular loops

2020-08-04  Jakub Jelinek  <jakub@redhat.com>

* omp-expand.c (expand_omp_for_init_counts): For triangular loops
compute number of iterations at runtime more efficiently.
(expand_omp_for_init_vars): Adjust immediate dominators.
(extract_omp_for_update_vars): Likewise.

d: Fix PR96429: Pointer subtraction uses TRUNC_DIV_EXPR

gcc/d/ChangeLog:

PR d/96429
* expr.cc (ExprVisitor::visit (BinExp*)): Use EXACT_DIV_EXPR for
pointer diff expressions.

gcc/testsuite/ChangeLog:

PR d/96429
* gdc.dg/pr96429.d: New test.

Change testcase for pr96325 from run to compile.

2020-08-04  Paul Thomas  <pault@gcc.gnu.org>

gcc/testsuite/
PR fortran/96325
* gfortran.dg/pr96325.f90: Change from run to compile.

Adjust two_valued_val_range_p for irange API.

gcc/ChangeLog:

* vr-values.c (simplify_using_ranges::two_valued_val_range_p):
Use irange API.

Adjust simplify_conversion_using_ranges for irange API.

gcc/ChangeLog:

* vr-values.c (simplify_conversion_using_ranges): Convert to irange API.

Use irange API in test_for_singularity.

gcc/ChangeLog:

* vr-values.c (test_for_singularity): Use irange API.
(simplify_using_ranges::simplify_cond_using_ranges_1): Do not
special case VR_RANGE.

Adjust vrp_evaluate_conditional for irange API.

gcc/ChangeLog:

* vr-values.c (simplify_using_ranges::vrp_evaluate_conditional): Adjust
for irange API.

Adjust op_with_boolean_value_range_p for irange API.

gcc/ChangeLog:

* vr-values.c (simplify_using_ranges::op_with_boolean_value_range_p): Adjust
for irange API.

Adjust get_range_info to use the base irange class.

gcc/ChangeLog:

* tree-ssanames.c (get_range_info): Use irange instead of value_range.
* tree-ssanames.h (get_range_info): Same.

Adjust expr_not_equal_to to use irange API.

gcc/ChangeLog:

* fold-const.c (expr_not_equal_to): Adjust for irange API.

Remove ad-hoc range canonicalization from determine_block_size.

Anti ranges of ~[MIN,X] are automatically canonicalized to [X+1,MAX],
at creation time.  There is no need to handle them specially.

Tested by adding a gcc_unreachable and bootstrapping/testing.

gcc/ChangeLog:

* builtins.c (determine_block_size): Remove ad-hoc range canonicalization.

dse: Remove partial load after full store for high part access[PR71309]

v5 update as comments:
1. Move const_rhs out of loop;
2. Iterate from int size for read_mode.

This patch could optimize(works for char/short/int/void*):

6: r119:TI=[r118:DI+0x10]
7: [r118:DI]=r119:TI
8: r121:DI=[r118:DI+0x8]

=>

6: r119:TI=[r118:DI+0x10]
16: r122:DI=r119:TI#8

Final ASM will be as below without partial load after full store(stxv+ld):
  ld 10,16(3)
  mr 9,3
  ld 3,24(3)
  std 10,0(9)
  std 3,8(9)
  blr

It could achieve ~25% performance improvement for typical cases on
Power9.  Bootstrap and regression tested on Power9-LE.

For AArch64, one ldr is replaced by mov with this patch:

ldp     x2, x3, [x0, 16]
stp     x2, x3, [x0]
ldr     x0, [x0, 8]

=>

mov     x1, x0
ldp     x2, x0, [x0, 16]
stp     x2, x0, [x1]

gcc/ChangeLog:

2020-08-04  Xionghu Luo  <luoxhu@linux.ibm.com>

PR rtl-optimization/71309
* dse.c (find_shift_sequence): Use subreg of shifted from high part
register to avoid loading from address.

gcc/testsuite/ChangeLog:

2020-08-04  Xionghu Luo  <luoxhu@linux.ibm.com>

PR rtl-optimization/71309
* gcc.target/powerpc/pr71309.c: New test.

Daily bump.

c++: Remove unused declaration.

gcc/cp/ChangeLog:

* cp-tree.h (after_nsdmi_defaulted_late_checks): Remove.

libgcc: increase required stack space for x86_64 -fsplit-stack

This accomodates increased space required by use of the xsavec
instruction in the dynamic linker trampoline.

libgcc/ChangeLog:

* config/i386/morestack.S (BACKOFF) [x86_64]: Add 2048 bytes.

rs6000: Fix vector_float.c testcase for -m32

It should be skipped then.

2020-08-03  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/testsuite/
* gcc.target/powerpc/vector_float.c: Skip if not lp64.

c++: Variable template and template parameter pack [PR96218]

This is DR 2032 which says that the restrictions regarding template
parameter packs and default arguments apply to variable templates as
well, but we weren't detecting that.

gcc/cp/ChangeLog:

DR 2032
PR c++/96218
* pt.c (check_default_tmpl_args): Also consider variable
templates.

gcc/testsuite/ChangeLog:

DR 2032
PR c++/96218
* g++.dg/cpp1y/var-templ67.C: New test.

aarch64: Fix up __aarch64_cas16_acq_rel fallback

As mentioned in the PR, the fallback path when LSE is unavailable writes
incorrect registers to the memory if the previous content compares equal
to x0, x1 - it writes copy of x0, x1 from the start of function, but it
should write x2, x3.

2020-08-03  Jakub Jelinek  <jakub@redhat.com>

PR target/96402
* config/aarch64/lse.S (__aarch64_cas16_acq_rel): Use x2, x3 instead
of x(tmp0), x(tmp1) in STXP arguments.

* gcc.target/aarch64/pr96402.c: New test.

cpp: Do not use @dots for ... tokens in code examples

This prevents a ... token in code examples from being turned into a
single HORIZONTAL ELLIPSIS glyph (e.g. via the HTML &hellip; entity).

gcc/ChangeLog:

* doc/cpp.texi (Variadic Macros): Use the exact ... token in
code examples.

Refer to C++20

I noticed a bunch of references to c++2a.

gcc/
* doc/invoke.texi: Refer to c++20

openacc: No attach/detach present/release mappings for array descriptors

Standalone attach and detach clauses should not create present/release
mappings for Fortran array descriptors (e.g. used when we have a pointer
to an array), both because it is unnecessary and because those mappings
will be incorrectly subject to reference counting. Simply omitting the
mappings means we just use GOMP_MAP_TO_PSET and GOMP_MAP_{ATTACH,DETACH}
mappings for array descriptors.

That requires a tweak in gimplify.c, since we may now see GOMP_MAP_TO_PSET
without a preceding data-movement mapping.

2020-08-03  Julian Brown  <julian@codesourcery.com>
    Thomas Schwinge  <thomas@codesourcery.com>

gcc/fortran/
* trans-openmp.c (gfc_trans_omp_clauses): Don't create present/release
mappings for array descriptors.

gcc/
* gimplify.c (gimplify_omp_target_update): Allow GOMP_MAP_TO_PSET
without a preceding data-movement mapping.

gcc/testsuite/
* gfortran.dg/goacc/attach-descriptor.f90: Update pattern output. Add
scanning of gimplify dump.

libgomp/
* testsuite/libgomp.oacc-fortran/attach-descriptor-1.f90: Don't run for
shared-memory devices.  Extend with further checking.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>

Darwin: Guard two macros in darwin.h.

Work on the Arm64 port shows that these two macros can be declared
ahead of the version in darwin.h which needs to override (for X86
and PPC this wasn't needed).

gcc/ChangeLog:

* config/darwin.h (ASM_DECLARE_FUNCTION_NAME): UNDEF before
use.
(DEF_MIN_OSX_VERSION): Only define if there's no existing
def.

Darwin: Handle poly_int machine_modes.

The common code that selects suitable sections for literals needs
to inspect the machine_mode.  For some sub-targets that might be
represented as a poly-int.

There was a workaround in place that allowed for cases where the poly
int had only one component.  This removes the workaround and handles
the cases where we care about the machine_mode size.

gcc/ChangeLog:

* config/darwin.c (IN_TARGET_CODE): Remove.
(darwin_mergeable_constant_section): Handle poly-int machine modes.
(machopic_select_rtx_section): Likewise.

internal/syscall/unix: correct ia64 syscall numbers

Per Andreas Schwab.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/246264

internal/syscall/unix: define copyFileRangeTrap for all architectures

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/246363

Avoid shifting by amounts larger than target int in irange self-tests.

gcc/ChangeLog:

PR tree-optimization/96430
* range-op.cc (operator_tests): Do not shift by 31 on targets with
integer's smaller than 32 bits.

Removal of HSA offloading from gcc and libgomp

This patch removes the generation of HSAIL from the compiler, the HSA
offloading plugin from libgomp and the associated testsuite tests and
infrastructure bits from the respective testsuites.

Apart from removal of the obvious files, I removed bits that I found
by searching for HSA related terms and by re-tracing my steps and
looking at the patches that introduced HSA in the first place.  I did
not remove everything these patches brought in, for example:

  - the mechanism to pass offload-target specific info from the application to
    the offloading plugin - but the same mechanism is also used to
    communicate number of teams and the thread limit to all offload targets.

  - run_func hook in gomp_device_descr stays too, although now it is
    not used.  If some future offload target would like the ability to
    refuse to offload some functions, it can use it.  It is easy to
    remove as a follow-up if it is considered clutter, though.

  - configure options --with-hsa-runtime=PATH, -with-hsa-runtime-include=PATH
    and --with-hsa-runtime-lib=PATH rmeain because GCN uses them too.

  - Surprisingly, GOMP_TARGET_ARG_HSA_KERNEL_ATTRIBUTES (a constant
    from gomp-constants.h) appears in the source of the amdgcn libgomp
    plugin, although I tend to think that code path is not ever used
    and this patch certainly removes it from the compiler.
    Nevertheless, it seems it has potential value beyond HSAIL and so
    I've kept it, it can of course always be easily removed in the
    future of GCN folk abandon it too.

  - I assume constants OFFLOAD_TARGET_TYPE_HSA and GOMP_DEVICE_HSA
    need to stay indefinitely too just so that no future offload
    target picks that number.

  - I have kept dg-require-effective-target
    offload_device_nonshared_as requirement of thests which have it.

It is quite probable I missed some small HSA artifacts but those
should be easy to remove later as we find them.

include/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* gomp-constants.h (GOMP_VERSION_HSA): Remove.

gcc/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* hsa-brig-format.h: Moved to brig/brigfrontend.
* hsa-brig.c: Removed.
* hsa-builtins.def: Likewise.
* hsa-common.c: Likewise.
* hsa-common.h: Likewise.
* hsa-dump.c: Likewise.
* hsa-gen.c: Likewise.
* hsa-regalloc.c: Likewise.
* ipa-hsa.c: Likewise.
* omp-grid.c: Likewise.
* omp-grid.h: Likewise.
* Makefile.in (BUILTINS_DEF): Remove hsa-builtins.def.
(OBJS): Remove hsa-common.o, hsa-gen.o, hsa-regalloc.o, hsa-brig.o,
hsa-dump.o, ipa-hsa.c and omp-grid.o.
(GTFILES): Removed hsa-common.c and omp-expand.c.
* builtins.def: Remove processing of hsa-builtins.def.
(DEF_HSA_BUILTIN): Remove.
* common.opt (flag_disable_hsa): Remove.
(-Whsa): Ignore.
* config.in (ENABLE_HSA): Removed.
* configure.ac: Removed handling configuration for hsa offloading.
(ENABLE_HSA): Removed.
* configure: Regenerated.
* doc/install.texi (--enable-offload-targets): Remove hsa from the
example.
(--with-hsa-runtime): Reword to reference any HSA run-time, not
specifically HSA offloading.
* doc/invoke.texi (Option Summary): Remove -Whsa.
(Warning Options): Likewise.
(Optimize Options): Remove hsa-gen-debug-stores.
* doc/passes.texi (Regular IPA passes): Remove section on IPA HSA
pass.
* gimple-low.c (lower_stmt): Remove GIMPLE_OMP_GRID_BODY case.
* gimple-pretty-print.c (dump_gimple_omp_for): Likewise.
(dump_gimple_omp_block): Likewise.
(pp_gimple_stmt_1): Likewise.
* gimple-walk.c (walk_gimple_stmt): Likewise.
* gimple.c (gimple_build_omp_grid_body): Removed function.
(gimple_copy): Remove GIMPLE_OMP_GRID_BODY case.
* gimple.def (GIMPLE_OMP_GRID_BODY): Removed.
* gimple.h (gf_mask): Removed GF_OMP_PARALLEL_GRID_PHONY,
OMP_FOR_KIND_GRID_LOOP, GF_OMP_FOR_GRID_PHONY,
GF_OMP_FOR_GRID_INTRA_GROUP, GF_OMP_FOR_GRID_GROUP_ITER and
GF_OMP_TEAMS_GRID_PHONY.  Renumbered GF_OMP_FOR_KIND_SIMD and
GF_OMP_TEAMS_HOST.
(gimple_build_omp_grid_body): Removed declaration.
(gimple_has_substatements): Remove GIMPLE_OMP_GRID_BODY case.
(gimple_omp_for_grid_phony): Removed.
(gimple_omp_for_set_grid_phony): Likewise.
(gimple_omp_for_grid_intra_group): Likewise.
(gimple_omp_for_grid_intra_group): Likewise.
(gimple_omp_for_grid_group_iter): Likewise.
(gimple_omp_for_set_grid_group_iter): Likewise.
(gimple_omp_parallel_grid_phony): Likewise.
(gimple_omp_parallel_set_grid_phony): Likewise.
(gimple_omp_teams_grid_phony): Likewise.
(gimple_omp_teams_set_grid_phony): Likewise.
(CASE_GIMPLE_OMP): Remove GIMPLE_OMP_GRID_BODY case.
* lto-section-in.c (lto_section_name): Removed hsa.
* lto-streamer.h (lto_section_type): Removed LTO_section_ipa_hsa.
* lto-wrapper.c (compile_images_for_offload_targets): Remove special
handling of hsa.
* omp-expand.c: Do not include hsa-common.h and gt-omp-expand.h.
(parallel_needs_hsa_kernel_p): Removed.
(grid_launch_attributes_trees): Likewise.
(grid_launch_attributes_trees): Likewise.
(grid_create_kernel_launch_attr_types): Likewise.
(grid_insert_store_range_dim): Likewise.
(grid_get_kernel_launch_attributes): Likewise.
(get_target_arguments): Remove code passing HSA grid sizes.
(grid_expand_omp_for_loop): Remove.
(grid_arg_decl_map): Likewise.
(grid_remap_kernel_arg_accesses): Likewise.
(grid_expand_target_grid_body): Likewise.
(expand_omp): Remove call to grid_expand_target_grid_body.
(omp_make_gimple_edges): Remove GIMPLE_OMP_GRID_BODY case.
* omp-general.c: Do not include hsa-common.h.
(omp_maybe_offloaded): Do not check for HSA offloading.
(omp_context_selector_matches): Likewise.
* omp-low.c: Do not include hsa-common.h and omp-grid.h.
(build_outer_var_ref): Remove handling of GIMPLE_OMP_GRID_BODY.
(scan_sharing_clauses): Remove handling of OMP_CLAUSE__GRIDDIM_.
(scan_omp_parallel): Remove handling of the phoney variant.
(check_omp_nesting_restrictions): Remove handling of
GIMPLE_OMP_GRID_BODY and GF_OMP_FOR_KIND_GRID_LOOP.
(scan_omp_1_stmt): Remove handling of GIMPLE_OMP_GRID_BODY.
(lower_omp_for_lastprivate): Remove handling of gridified loops.
(lower_omp_for): Remove phony loop handling.
(lower_omp_taskreg): Remove phony construct handling.
(lower_omp_teams): Likewise.
(lower_omp_grid_body): Removed.
(lower_omp_1): Remove GIMPLE_OMP_GRID_BODY case.
(execute_lower_omp): Do not call omp_grid_gridify_all_targets.
* opts.c (common_handle_option): Do not handle hsa when processing
OPT_foffload_.
* params.opt (hsa-gen-debug-stores): Remove.
* passes.def: Remove pass_ipa_hsa and pass_gen_hsail.
* timevar.def: Remove TV_IPA_HSA.
* toplev.c: Do not include hsa-common.h.
(compile_file): Do not call hsa_output_brig.
* tree-core.h (enum omp_clause_code): Remove OMP_CLAUSE__GRIDDIM_.
(tree_omp_clause): Remove union field dimension.
* tree-nested.c (convert_nonlocal_omp_clauses): Remove the
OMP_CLAUSE__GRIDDIM_ case.
(convert_local_omp_clauses): Likewise.
* tree-pass.h (make_pass_gen_hsail): Remove declaration.
(make_pass_ipa_hsa): Likewise.
* tree-pretty-print.c (dump_omp_clause): Remove GIMPLE_OMP_GRID_BODY
case.
* tree.c (omp_clause_num_ops): Remove the element corresponding to
OMP_CLAUSE__GRIDDIM_.
(omp_clause_code_name): Likewise.
(walk_tree_1): Remove GIMPLE_OMP_GRID_BODY case.
* tree.h (OMP_CLAUSE__GRIDDIM__DIMENSION): Remove.
(OMP_CLAUSE__GRIDDIM__SIZE): Likewise.
(OMP_CLAUSE__GRIDDIM__GROUP): Likewise.

gcc/fortran/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* f95-lang.c (gfc_init_builtin_functions): Remove processing of
hsa-builtins.def.

gcc/brig/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* brigfrontend/brig-util.h (hsa_type_packed_p): Declared.
* brigfrontend/brig-util.cc (hsa_type_packed_p): Moved here from
removed gcc/hsa-common.c.

libgomp/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* plugin/Makefrag.am: Remove configuration of HSA plugin.
* aclocal.m4: Regenerated.
* Makefile.in: Regenerated.
* config.h.in: Regenerated.
* configure: Regenerated.
* plugin/configfrag.ac: Likewise.
* plugin/hsa_ext_finalize.h: Removed.
* plugin/plugin-hsa.c: Likewise.
* testsuite/Makefile.in: Regenerated.
* testsuite/lib/libgomp.exp
(offload_target_to_openacc_device_type): Remove hsa case.
(check_effective_target_hsa_offloading_selected_nocache): Removed
(check_effective_target_hsa_offloading_selected): Likewise.
(libgomp_init): Do not add -Wno-hsa to additional_flags.
* testsuite/libgomp.hsa.c/alloca-1.c: Removed test.
* testsuite/libgomp.hsa.c/bitfield-1.c: Likewise.
* testsuite/libgomp.hsa.c/bits-insns.c: Likewise.
* testsuite/libgomp.hsa.c/builtins-1.c: Likewise.
* testsuite/libgomp.hsa.c/c.exp: Likewise.
* testsuite/libgomp.hsa.c/complex-1.c: Likewise.
* testsuite/libgomp.hsa.c/complex-align-2.c: Likewise.
* testsuite/libgomp.hsa.c/formal-actual-args-1.c: Likewise.
* testsuite/libgomp.hsa.c/function-call-1.c: Likewise.
* testsuite/libgomp.hsa.c/get-level-1.c: Likewise.
* testsuite/libgomp.hsa.c/gridify-1.c: Likewise.
* testsuite/libgomp.hsa.c/gridify-2.c: Likewise.
* testsuite/libgomp.hsa.c/gridify-3.c: Likewise.
* testsuite/libgomp.hsa.c/gridify-4.c: Likewise.
* testsuite/libgomp.hsa.c/memory-operations-1.c: Likewise.
* testsuite/libgomp.hsa.c/pr69568.c: Likewise.
* testsuite/libgomp.hsa.c/pr82416.c: Likewise.
* testsuite/libgomp.hsa.c/rotate-1.c: Likewise.
* testsuite/libgomp.hsa.c/staticvar.c: Likewise.
* testsuite/libgomp.hsa.c/switch-1.c: Likewise.
* testsuite/libgomp.hsa.c/switch-branch-1.c: Likewise.
* testsuite/libgomp.hsa.c/switch-sbr-2.c: Likewise.
* testsuite/libgomp.hsa.c/tiling-1.c: Likewise.
* testsuite/libgomp.hsa.c/tiling-2.c: Likewise.

gcc/testsuite/ChangeLog:

2020-07-24  Martin Jambor  <mjambor@suse.cz>

* lib/target-supports.exp (check_effective_target_offload_hsa):
Removed.
* c-c++-common/gomp/gridify-1.c: Removed test.
* c-c++-common/gomp/gridify-2.c: Likewise.
* c-c++-common/gomp/gridify-3.c: Likewise.
* c-c++-common/gomp/hsa-indirect-call-1.c: Likewise.
* gfortran.dg/gomp/gridify-1.f90: Likewise.
* gcc.dg/gomp/gomp.exp: Do not pass -Wno-hsa to tests.
* g++.dg/gomp/gomp.exp: Likewise.
* gfortran.dg/gomp/gomp.exp: Likewise.

aarch64: Add support for unpacked sub [PR96366]

The test case bb-slp-20.c in the gcc testsuit will cause an
ICE in the expand pass due to the lack of a pattern for
subtraction of the VNx2SI mode. This patch solve this problem
by adding support for unpacked sub.

gcc/ChangeLog:

* config/aarch64/aarch64-sve.md (sub<mode>3): Add support for
unpacked vectors.

testsuite: Add -Wno-psabi to gcc.dg/pr96377-[12].c

2020-08-03  Richard Sandiford  <richard.sandiford@arm.com>

gcc/testsuite/
* gcc.dg/pr96377-1.c: Add -Wno-psabi.
* gcc.dg/pr96377-2.c: Likewise.

MSP430: Don't pass redundant -md option to the assembler

The MSP430 GAS option "-md" is supposed to indicate that the CRT startup
code should copy data from ROM to RAM at startup. However, this option
has no effect; GAS handles the related behaviour automatically.

gcc/ChangeLog:

* config/msp430/msp430.h (ASM_SPEC): Don't pass on "-md" option.

regrename: Avoid disrupting SMS schedule [PR95696]

SMS is performed before reload, and each insn in SMS schedule uses
pseudo-register.  After reload, regrename pass try to adjust the hard
registers with def/use chain created by build_def_use.  For now, regrename
pass isn't aware of VLIW bundles created by SMS, it may updated a register
which may not be really unused, which will causes invalid VLIW bundles.
Before the final schedule, we recheck the validation of VLIW bundles and
reschedule the conflicted insns to avoid the above issue.  Rescheduling
the conflicted insns will destroy SMS schedule of the kernel loop, which
would be harmful to performance.

2020-08-03  Yunde Zhong  <zhongyunde@huawei.com>

gcc/
PR rtl-optimization/95696
* regrename.c (regrename_analyze): New param include_all_block_p
with default value TRUE.  If set to false, avoid disrupting SMS
schedule.
* regrename.h (regrename_analyze): Adjust prototype.

lto/96385 - avoid unused global UNDEFs in debug objects

Unused global UNDEFs can have side-effects in some circumstances so
the following patch avoids them by treating them the same as other
to be discarded DEFs - make them local.

2020-08-03  Richard Biener  <rguenther@suse.de>

PR lto/96385
libiberty/
* simple-object-elf.c
(simple_object_elf_copy_lto_debug_sections): Localize global
UNDEFs and reuse the prevailing name.

doc: fix a typo in tm.texi

gcc/
* doc/tm.texi.in (VECTOR_STORE_FLAG_VALUE): Fix a typo.
* doc/tm.texi: Regenerate.

doc: Add missing comma after octeontx2f95mm

gcc/
* doc/invoke.texi: Add missing comma after octeontx2f95mm entry.

aarch64: Add A64FX machine model

This patch add support for Fujitsu A64FX, as the first step of adding
A64FX machine model.

A64FX is used in FUJITSU Supercomputer PRIMEHPC FX1000,
PRIMEHPC FX700, and supercomputer Fugaku.
The official microarchitecture information of A64FX can be read at
https://github.com/fujitsu/A64FX.

2020-08-03  Qian jianhua  <qianjh@cn.fujitsu.com>

gcc/
* config/aarch64/aarch64-cores.def (a64fx): New core.
* config/aarch64/aarch64-tune.md: Regenerated.
* config/aarch64/aarch64.c (a64fx_prefetch_tune, a64fx_tunings): New.
* doc/invoke.texi: Add a64fx to the list.

PR rtl-optimization 61494: Preserve x-0.0 with HONOR_SNANS.

The following patch avoids simplifying x-0.0 to x when -fsignaling-nans
is specified, which resolves PR rtl-optimization 61494.  Indeed, running
the test program attached to that PR now reports no failures.

2020-08-02  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
PR rtl-optimization/61494
* simplify-rtx.c (simplify_binary_operation_1) [MINUS]: Don't
simplify x - 0.0 with -fsignaling-nans.

genmatch: Avoid unused parameter warnings in generated code.

This patch silences a number of unused parameter warnings whilst
compiling both generic-match.c and gimple-match.c.  The problem is
that multiple (polymorphic) functions are generated for generic_simplify
and gimple_simplify, each handling tree codes with a specific number
of children.  Currently, there are no simplifications for tree codes
with four or five children, leading to functions with "empty" bodies
and unused function arguments.  This patch detects those cases, and
generates stub functions (with anonymous arguments) to silence these
warnings.

2020-08-03  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* genmatch.c (decision_tree::gen): Emit stub functions for
tree code operand counts that have no simplifications.
(main): Correct comment typo.

tree-optimization: Fix typos in comments

The only two changes which aren't obvious are s/dirified/specified/ and
s/edirially/especially/ which appear to be caused by a s/spec/dir/ edit
that went too far.

gcc/ChangeLog:

* gimple-ssa-sprintf.c: Fix typos in comments.

AArch64: Fix hwasan failure in readline.

My previous fix added an unchecked call to fgets in the new function readline.
fgets can fail when there's an error reading the file in which case it returns
NULL.  It also returns NULL when the next character is EOF.

The EOF case is already covered by the existing code but the error case isn't.
This fixes it by returning the empty string on error.

Also I now use strnlen instead of strlen to make sure we never read outside the
buffer.

This was flagged by Matthew Malcomson during his hwasan work.

gcc/ChangeLog:

* config/aarch64/driver-aarch64.c (readline): Check return value fgets.

mark match.pd ! not implemented on GENERIC

This makes us error when the ! operator modifier is encountered
when not targeting GIMPLE.

2020-08-03  Richard Biener  <rguenther@suse.de>

* genmatch.c (parser::gimple): New.
(parser::parser): Initialize gimple flag member.
(parser::parse_expr): Error on ! operator modifier when
not targeting GIMPLE.
(main): Pass down gimple flag to parser ctor.

* doc/match-and-simplify.texi: Amend accordingly.

d: Fix ICE using non-local variable: internal compiler error: Segmentation fault

Moves no frame access error to own function, adding use of it for both
when get_framedecl() cannot find a path to the outer function frame, and
guarding get_decl_tree() from recursively calling itself.

gcc/d/ChangeLog:

PR d/96254
* d-codegen.cc (error_no_frame_access): New.
(get_frame_for_symbol): Use fdparent name in error message.
(get_framedecl): Replace call to assert with error.
* d-tree.h (error_no_frame_access): Declare.
* decl.cc (get_decl_tree): Detect recursion and error.

gcc/testsuite/ChangeLog:

PR d/96254
* gdc.dg/pr96254a.d: New test.
* gdc.dg/pr96254b.d: New test.

libgfortran/caf/single.c: Fix typo.

libgfortran/ChangeLog

* caf/single.c (_gfortran_caf_lock): Fix typo.

Fortran texi: Fix typos

gcc/fortran/ChangeLog

* gfc-internals.texi: Fix typos.
* gfortran.texi: Likewise.
* intrinsic.texi: Likewise.
* invoke.texi: Likewise.

Multi-range implementation for value_range (irange).

Implement class irange, a generic multi-range implementation for
value ranges.  This class is API compatible with value_range, and is meant
to seamlessly coexist with it.

gcc/ChangeLog:

* Makefile.in (GTFILES): Move value-range.h up.
* gengtype-lex.l: Set yylval to handle GTY markers on templates.
* ipa-cp.c (initialize_node_lattices): Call value_range
constructor.
(ipcp_propagate_stage): Use in-place new so value_range construct
is called.
* ipa-fnsummary.c (evaluate_conditions_for_known_args): Use std
vec instead of GCC's vec<>.
(evaluate_properties_for_edge): Adjust for std vec.
(ipa_fn_summary_t::duplicate): Same.
(estimate_ipcp_clone_size_and_time): Same.
* ipa-prop.c (ipa_get_value_range): Use in-place new for
value_range.
* ipa-prop.h (struct GTY): Remove class keyword for m_vr.
* range-op.cc (empty_range_check): Rename to...
(empty_range_varying): ...this and adjust for varying.
(undefined_shift_range_check): Adjust for irange.
(range_operator::wi_fold): Same.
(range_operator::fold_range): Adjust for irange.  Special case
single pairs for performance.
(range_operator::op1_range): Adjust for irange.
(range_operator::op2_range): Same.
(value_range_from_overflowed_bounds): Same.
(value_range_with_overflow): Same.
(create_possibly_reversed_range): Same.
(range_true): Same.
(range_false): Same.
(range_true_and_false): Same.
(get_bool_state):  Adjust for irange and tweak for performance.
(operator_equal::fold_range): Adjust for irange.
(operator_equal::op1_range): Same.
(operator_equal::op2_range): Same.
(operator_not_equal::fold_range): Same.
(operator_not_equal::op1_range): Same.
(operator_not_equal::op2_range): Same.
(build_lt): Same.
(build_le): Same.
(build_gt): Same.
(build_ge): Same.
(operator_lt::fold_range): Same.
(operator_lt::op1_range): Same.
(operator_lt::op2_range): Same.
(operator_le::fold_range): Same.
(operator_le::op1_range): Same.
(operator_le::op2_range): Same.
(operator_gt::fold_range): Same.
(operator_gt::op1_range): Same.
(operator_gt::op2_range): Same.
(operator_ge::fold_range): Same.
(operator_ge::op1_range): Same.
(operator_ge::op2_range): Same.
(operator_plus::wi_fold): Same.
(operator_plus::op1_range): Same.
(operator_plus::op2_range): Same.
(operator_minus::wi_fold): Same.
(operator_minus::op1_range): Same.
(operator_minus::op2_range): Same.
(operator_min::wi_fold): Same.
(operator_max::wi_fold): Same.
(cross_product_operator::wi_cross_product): Same.
(operator_mult::op1_range): New.
(operator_mult::op2_range): New.
(operator_mult::wi_fold): Adjust for irange.
(operator_div::wi_fold): Same.
(operator_exact_divide::op1_range): Same.
(operator_lshift::fold_range): Same.
(operator_lshift::wi_fold): Same.
(operator_lshift::op1_range): New.
(operator_rshift::op1_range): New.
(operator_rshift::fold_range): Adjust for irange.
(operator_rshift::wi_fold): Same.
(operator_cast::truncating_cast_p): Abstract out from
operator_cast::fold_range.
(operator_cast::fold_range): Adjust for irange and tweak for
performance.
(operator_cast::inside_domain_p): Abstract out from fold_range.
(operator_cast::fold_pair): Same.
(operator_cast::op1_range): Use abstracted methods above.  Adjust
for irange and tweak for performance.
(operator_logical_and::fold_range): Adjust for irange.
(operator_logical_and::op1_range): Same.
(operator_logical_and::op2_range): Same.
(unsigned_singleton_p): New.
(operator_bitwise_and::remove_impossible_ranges): New.
(operator_bitwise_and::fold_range): New.
(wi_optimize_and_or):  Adjust for irange.
(operator_bitwise_and::wi_fold): Same.
(set_nonzero_range_from_mask): New.
(operator_bitwise_and::simple_op1_range_solver): New.
(operator_bitwise_and::op1_range): Adjust for irange.
(operator_bitwise_and::op2_range): Same.
(operator_logical_or::fold_range): Same.
(operator_logical_or::op1_range): Same.
(operator_logical_or::op2_range): Same.
(operator_bitwise_or::wi_fold): Same.
(operator_bitwise_or::op1_range): Same.
(operator_bitwise_or::op2_range): Same.
(operator_bitwise_xor::wi_fold): Same.
(operator_bitwise_xor::op1_range): New.
(operator_bitwise_xor::op2_range): New.
(operator_trunc_mod::wi_fold):  Adjust for irange.
(operator_logical_not::fold_range): Same.
(operator_logical_not::op1_range): Same.
(operator_bitwise_not::fold_range): Same.
(operator_bitwise_not::op1_range): Same.
(operator_cst::fold_range): Same.
(operator_identity::fold_range): Same.
(operator_identity::op1_range): Same.
(class operator_unknown): New.
(operator_unknown::fold_range): New.
(class operator_abs): Adjust for irange.
(operator_abs::wi_fold): Same.
(operator_abs::op1_range): Same.
(operator_absu::wi_fold): Same.
(class operator_negate): Same.
(operator_negate::fold_range): Same.
(operator_negate::op1_range): Same.
(operator_addr_expr::fold_range): Same.
(operator_addr_expr::op1_range): Same.
(pointer_plus_operator::wi_fold): Same.
(pointer_min_max_operator::wi_fold): Same.
(pointer_and_operator::wi_fold): Same.
(pointer_or_operator::op1_range): New.
(pointer_or_operator::op2_range): New.
(pointer_or_operator::wi_fold):  Adjust for irange.
(integral_table::integral_table): Add entries for IMAGPART_EXPR
and POINTER_DIFF_EXPR.
(range_cast):  Adjust for irange.
(build_range3): New.
(range3_tests): New.
(widest_irange_tests): New.
(multi_precision_range_tests): New.
(operator_tests): New.
(range_tests): New.
* range-op.h (class range_operator): Adjust for irange.
(range_cast): Same.
* tree-vrp.c (range_fold_binary_symbolics_p): Adjust for irange and
tweak for performance.
(range_fold_binary_expr): Same.
(masked_increment): Change to extern.
* tree-vrp.h (masked_increment): New.
* tree.c (cache_wide_int_in_type_cache): New function abstracted
out from wide_int_to_tree_1.
(wide_int_to_tree_1): Cache 0, 1, and MAX for pointers.
* value-range-equiv.cc (value_range_equiv::deep_copy): Use kind
method.
(value_range_equiv::move): Same.
(value_range_equiv::check): Adjust for irange.
(value_range_equiv::intersect): Same.
(value_range_equiv::union_): Same.
(value_range_equiv::dump): Same.
* value-range.cc (irange::operator=): Same.
(irange::maybe_anti_range): New.
(irange::copy_legacy_range): New.
(irange::set_undefined): Adjust for irange.
(irange::swap_out_of_order_endpoints): Abstract out from set().
(irange::set_varying): Adjust for irange.
(irange::irange_set): New.
(irange::irange_set_anti_range): New.
(irange::set): Adjust for irange.
(value_range::set_nonzero): Move to header file.
(value_range::set_zero): Move to header file.
(value_range::check): Rename to...
(irange::verify_range): ...this.
(value_range::num_pairs): Rename to...
(irange::legacy_num_pairs): ...this, and adjust for irange.
(value_range::lower_bound): Rename to...
(irange::legacy_lower_bound): ...this, and adjust for irange.
(value_range::upper_bound): Rename to...
(irange::legacy_upper_bound): ...this, and adjust for irange.
(value_range::equal_p): Rename to...
(irange::legacy_equal_p): ...this.
(value_range::operator==): Move to header file.
(irange::equal_p): New.
(irange::symbolic_p): Adjust for irange.
(irange::constant_p): Same.
(irange::singleton_p): Same.
(irange::value_inside_range): Same.
(irange::may_contain_p): Same.
(irange::contains_p): Same.
(irange::normalize_addresses): Same.
(irange::normalize_symbolics): Same.
(irange::legacy_intersect): Same.
(irange::legacy_union): Same.
(irange::union_): Same.
(irange::intersect): Same.
(irange::irange_union): New.
(irange::irange_intersect): New.
(subtract_one): New.
(irange::invert): Adjust for irange.
(dump_bound_with_infinite_markers): New.
(irange::dump): Adjust for irange.
(debug): Add irange versions.
(range_has_numeric_bounds_p): Adjust for irange.
(vrp_val_max): Move to header file.
(vrp_val_min): Move to header file.
(DEFINE_INT_RANGE_GC_STUBS): New.
(DEFINE_INT_RANGE_INSTANCE): New.
* value-range.h (class irange): New.
(class int_range): New.
(class value_range): Rename to a instantiation of int_range.
(irange::legacy_mode_p): New.
(value_range::value_range): Remove.
(irange::kind): New.
(irange::num_pairs): Adjust for irange.
(irange::type): Adjust for irange.
(irange::tree_lower_bound): New.
(irange::tree_upper_bound): New.
(irange::type): Adjust for irange.
(irange::min): Same.
(irange::max): Same.
(irange::varying_p): Same.
(irange::undefined_p): Same.
(irange::zero_p): Same.
(irange::nonzero_p): Same.
(irange::supports_type_p): Same.
(range_includes_zero_p): Same.
(gt_ggc_mx): New.
(gt_pch_nx): New.
(irange::irange): New.
(int_range::int_range): New.
(int_range::operator=): New.
(irange::set): Moved from value-range.cc and adjusted for irange.
(irange::set_undefined): Same.
(irange::set_varying): Same.
(irange::operator==): Same.
(irange::lower_bound): Same.
(irange::upper_bound): Same.
(irange::union_): Same.
(irange::intersect): Same.
(irange::set_nonzero): Same.
(irange::set_zero): Same.
(irange::normalize_min_max): New.
(vrp_val_max): Move from value-range.cc.
(vrp_val_min): Same.
* vr-values.c (vr_values::get_lattice_entry): Call value_range
constructor.