My upcoming work replacing the VRP threaders with a fully resolving
backward threader has tripped over various corner cases in the path
sensitive relation oracle. This patch kills second order relations when
we kill a relation.
Tested on x86-64 and ppc64le Linux.
Co-authored-by: Andrew MacLeod <amacleod@redhat.com>
gcc/ChangeLog:
* value-relation.cc (path_oracle::killing_def): Kill second
order relations.
The file fptr.c is specific to 32-bit hppa-linux and should not be
included in LIB2ADD on hppa64-linux.
There is a builtin type mismatch in linux-atomic.c using the type
long long unsigned int for 64-bit atomic operations on hppa64-linux.
2021-10-27 John David Anglin <danglin@gcc.gnu.org>
libgcc/ChangeLog:
* config.host (hppa*64*-*-linux*): Don't add pa/t-linux to
tmake_file.
* config/pa/linux-atomic.c: Define u8, u16 and u64 types.
Use them in FETCH_AND_OP_2, OP_AND_FETCH_2, COMPARE_AND_SWAP_2,
SYNC_LOCK_TEST_AND_SET_2 and SYNC_LOCK_RELEASE_1 macros.
* config/pa/t-linux64 (LIB1ASMSRC): New define.
(LIB1ASMFUNCS): Revise.
(HOST_LIBGCC2_CFLAGS): Add "-DLINUX=1".
This is a follow-up small patch to address Honza's review of my
previous patch to select saner profile count to base heuristics on.
Currently the IPA-CP heuristics switch to PGO-mode only if there are
PGO counters available for any part of the call graph. This change
makes it to switch to the PGO mode only if any of the incoming edges
bringing in the constant in question had any ipa-quality counts on
them. Consequently, if a part of the program is built with
-fprofile-use and another part without, IPA-CP will use
estimated-frequency-based heuristics for the latter.
I still wonder whether this should only happen with
flag_profile_partial_training on. It seems like we're behaving as if
it was always on.
gcc/ChangeLog:
2021-10-18 Martin Jambor <mjambor@suse.cz>
* ipa-cp.c (good_cloning_opportunity_p): Decide whether to use
profile feedback depending on their local availability.
When profile feedback is available, IPA-CP takes the count of the
hottest node and then evaluates all call contexts relative to it.
This means that typically almost no clones for specialized contexts
are ever created because the maximum is some special function, called
from everywhere (that is likely to get inlined anyway) and all the
examined edges look cold compared to it.
This patch changes the selection. It simply sorts counts of all edges
eligible for cloning in a vector and then picks the count in 90th
percentile (the actual number is configurable via a parameter).
I also tried more complex approaches which were summing the counts and
picking the edge which together with all hotter edges accounted for a
given portion of the total sum of all edge counts. But first it was
not apparently clear to me that they make more logical sense that the
simple method and practically I always also had to ignore a few
percent of the hottest edges with really extreme counts (looking at
bash and python). And when I had to do that anyway, it seemed simpler
to just "ignore" more and take the first non-ignored count as the
base.
Nevertheless, if people think some more sophisticated method should be
used anyway, I am willing to be persuaded. But this patch is a clear
improvement over the current situation.
gcc/ChangeLog:
2021-10-26 Martin Jambor <mjambor@suse.cz>
* params.opt (param_ipa_cp_profile_count_base): New parameter.
* doc/invoke.texi (Optimize Options): Add entry for
ipa-cp-profile-count-base.
* ipa-cp.c (max_count): Replace with base_count, replace all
occurrences too, unless otherwise stated.
(ipcp_cloning_candidate_p): identify mostly-directly called
functions based on their counts, not max_count.
(compare_edge_profile_counts): New function.
(ipcp_propagate_stage): Instead of setting max_count, find the
appropriate edge count in a sorted vector of counts of eligible
edges and make it the base_count.
IPA-CP does not do a reasonable job when it is updating profile counts
after it has created clones of recursive functions. This patch
addresses that by:
1. Only updating counts for special-context clones. When a clone is
created for all contexts, the original is going to be dead and the
cgraph machinery has copied counts to the new node which is the right
thing to do. Therefore updating counts has been moved from
create_specialized_node to decide_about_value and
decide_whether_version_node.
2. The current profile updating code artificially increased the assumed
old count when the sum of counts of incoming edges to both the
original and new node were bigger than the count of the original
node. This always happened when self-recursive edge from the clone
was also redirected to the clone because both the original edge and
its clone had original high counts. This clutch was removed and
replaced by the next point.
3. When cloning also redirects a self-recursive clone to the clone
itself, new logic has been added to divide the counts brought by such
recursive edges between the original node and the clone. This is
impossible to do well without special knowledge about the function and
which non-recursive entry calls are responsible for what portion of
recursion depth, so the approach taken is rather crude.
For local nodes, we detect the case when the original node is never
called (in the training run at least) with another value and if so,
steal all its counts like if it was dead. If that is not the case, we
try to divide the count brought by recursive edges (or rather not
brought by direct edges) proportionally to the counts brought by
non-recursive edges - but with artificial limits in place so that we
do not take too many or too few, because that was happening with
detrimental effect in mcf_r.
4. When cloning creates extra clones for values brought by a formerly
self-recursive edge with an arithmetic pass-through jump function on
it, such as it does in exchange2_r, all such clones are processed at
once rather than one after another. The counts of all such nodes are
distributed evenly (modulo even-formerly-non-recursive-edges) and the
whole situation is then fixed up so that the edge counts fit. This is
what new function update_counts_for_self_gen_clones does.
5. When values brought by a formerly self-recursive edge with an
arithmetic pass-through jump function on it are evaluated by
heuristics which assumes vast majority of node counts are result of
recursive calls and so we simply divide those with the number of
clones there would be if we created another one.
6. The mechanisms in init_caller_stats and gather_caller_stats and
get_info_about_necessary_edges was enhanced to gather data required
for the above and a missing check not to count dead incoming edges was
also added.
gcc/ChangeLog:
2021-10-15 Martin Jambor <mjambor@suse.cz>
* ipa-cp.c (struct caller_statistics): New fields rec_count_sum,
n_nonrec_calls and itself, document all fields.
(init_caller_stats): Initialize the above new fields.
(gather_caller_stats): Gather self-recursive counts and calls number.
(get_info_about_necessary_edges): Gather counts of self-recursive and
other edges bringing in the requested value separately.
(dump_profile_updates): Rework to dump info about a single node only.
(lenient_count_portion_handling): New function.
(struct gather_other_count_struct): New type.
(gather_count_of_non_rec_edges): New function.
(struct desc_incoming_count_struct): New type.
(analyze_clone_icoming_counts): New function.
(adjust_clone_incoming_counts): Likewise.
(update_counts_for_self_gen_clones): Likewise.
(update_profiling_info): Rewritten.
(update_specialized_profile): Adjust call to dump_profile_updates.
(create_specialized_node): Do not update profiling info.
(decide_about_value): New parameter self_gen_clones, either push new
clones into it or updat their profile counts. For self-recursively
generated values, use a portion of the node count instead of count
from self-recursive edges to estimate goodness.
(decide_whether_version_node): Gather clones for self-generated values
in a new vector, update their profiles at once at the end.
This refactors epilogue loop handling in try_vectorize_loop_1 to not
suggest we're analyzing those there by splitting out the transform
phase which then can handle the epilogues.
2021-10-27 Richard Biener <rguenther@suse.de>
* tree-vectorizer.c (vect_transform_loops): New function,
split out from ...
(try_vectorize_loop_1): ... here. Simplify as epilogues
are now fully handled in the split part.
The testcase FAILs on i686-linux due to:
FAIL: gcc.dg/pr102897.c (test for excess errors)
Excess errors:
.../gcc/gcc/testsuite/gcc.dg/pr102897.c:11:1: warning: MMX vector return without MMX enabled changes the ABI [-Wpsabi]
.../gcc/gcc/testsuite/gcc.dg/pr102897.c:10:10: warning: MMX vector argument without MMX enabled changes the ABI [-Wpsabi]
Fixed by adding -Wno-psabi.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/102897
* gcc.dg/pr102897.c: Add -Wno-psabi to dg-options.
I've found we claim to support non-rectangular loops, but don't actually
support those in Fortran, as can be seen on:
integer i, j
!$omp parallel do collapse(2)
do i = 0, 10
do j = 0, i
end do
end do
end
To support this, the Fortran FE needs to allow the valid forms of
non-rectangular loops and disallow others, so mainly it needs its
updated version of c-omp.c c_omp_check_loop_iv etc., plus for non-rectangular
lb or ub expressions emit a TREE_VEC instead of normal expression as the C/C++ FE
do, plus testsuite coverage.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (OpenMP 5.0): Mention that Non-rectangular loop nests
aren't implemented for Fortran yet.
This patch handles pointer iterators for non-rectangular loops. They are
more limited than integral iterators of non-rectangular loops, in particular
only var-outer, var-outer + a2, a2 + var-outer or var-outer - a2 can appear
in lb or ub where a2 is some integral loop invariant expression, so no e.g.
multiplication etc.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-expand.c (expand_omp_for_init_counts): Handle non-rectangular
iterators with pointer types.
(expand_omp_for_init_vars, extract_omp_for_update_vars): Likewise.
gcc/c-family/
* c-omp.c (c_omp_check_loop_iv_r): Don't clear 3rd bit for
POINTER_PLUS_EXPR.
(c_omp_check_nonrect_loop_iv): Handle POINTER_PLUS_EXPR.
(c_omp_check_loop_iv): Set kind even if the iterator is non-integral.
gcc/testsuite/
* c-c++-common/gomp/loop-8.c: New test.
* c-c++-common/gomp/loop-9.c: New test.
libgomp/
* testsuite/libgomp.c/loop-26.c: New test.
* testsuite/libgomp.c/loop-27.c: New test.
In C++, if an iterator has or might have (e.g. dependent type) class type we
remember the original init expressions and check those separately for presence
of iterators, because for class iterators we turn those into expressions that
always do contain reference to the current iterator. But this resulted in
rejecting valid non-rectangular loop where the dependent type is later instantiated
to an integral type.
Non-rectangular loops with class random access iterators remain broken, that is something
to be fixed incrementally.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
PR c++/102854
gcc/c-family/
* c-common.h (c_omp_check_loop_iv_exprs): Add enum tree_code argument.
* c-omp.c (c_omp_check_loop_iv_r): For trees other than decls,
TREE_VEC, PLUS_EXPR, MINUS_EXPR, MULT_EXPR, POINTER_PLUS_EXPR or
conversions temporarily clear the 3rd bit from d->kind while walking
subtrees.
(c_omp_check_loop_iv_exprs): Add CODE argument. Or in 4 into data.kind
if possibly non-rectangular.
gcc/cp/
* semantics.c (handle_omp_for_class_iterator,
finish_omp_for): Adjust c_omp_check_loop_iv_exprs caller.
gcc/testsuite/
* g++.dg/gomp/loop-3.C: Don't expect some errors.
* g++.dg/gomp/loop-7.C: New test.
Another thing that wasn't in the previous patch, but I'm wondering whether we don't
handle it incorrectly. constexpr.c has:
/* Check that immediate invocation does not return an expression referencing
any immediate function decls. They need to be allowed while parsing
immediate functions, but can't leak outside of them. */
if (is_consteval
&& t != r
&& (current_function_decl == NULL_TREE
|| !DECL_IMMEDIATE_FUNCTION_P (current_function_decl)))
as condition for the discovery of embedded immediate FUNCTION_DECLs
(or now PTRMEM_CSTs). If I remove the && (current... ..._decl))
then g++.dg/cpp2a/consteval7.C's
struct S { int b; int (*c) (); };
consteval S baz () { return { 5, foo }; }
consteval int qux () { S s = baz (); return s.b + s.c (); }
consteval int quux () { constexpr S s = baz (); return s.b + s.c (); }
quux line fails, but based on
http://eel.is/c++draft/expr.const#11
I wonder if it shouldn't fail (clang++ -std=c++20 rejects it),
and be only accepted without the constexpr keyword before S s.
Also wonder about e.g.
consteval int foo () { return 42; }
consteval int
bar ()
{
auto fn1 = foo; // This must be ok
constexpr auto fn2 = foo; // Isn't this an error?
return fn1 () + fn2 ();
}
constexpr int
baz ()
{
if consteval {
auto fn1 = foo; // This must be ok
constexpr auto fn2 = foo; // Isn't this an error?
return fn1 () + fn2 ();
}
return 0;
}
auto a = bar ();
static_assert (bar () == 84);
static_assert (baz () == 84);
(again, clang++ -std=c++20 rejects the fn2 = foo; case,
but doesn't implement consteval if, so can't test the other one).
For taking address of an immediate function or method if it is taken
outside of immediate function context we already have diagnostics
about it, but shouldn't the immediate FUNCTION_DECL discovery in
cxx_eval_outermost_constant_expression be instead guarded with something
like
if (is_consteval || in_immediate_context ())
and be done regardless of whether t != r?
2021-10-27 Jakub Jelinek <jakub@redhat.com>
PR c++/102753
* constexpr.c (cxx_eval_outermost_constant_expr): Perform
find_immediate_fndecl discovery if is_consteval or
in_immediate_context () rather than if is_consteval, t != r
and not in immediate function's body.
* g++.dg/cpp2a/consteval7.C: Expect diagnostics on quux.
* g++.dg/cpp2a/consteval24.C: New test.
* g++.dg/cpp23/consteval-if12.C: New test.
The consteval20.C testcase ICEs, because while we have in cp_build_addr_expr_1
diagnostics for taking address of an immediate function (and as an exception
deal with build_address from immediate invocation), I forgot to diagnose
taking address of a member function which is done in a different place.
I hope (s.*&S::foo) () is not an immediate invocation like
(*&foo) () is not, so this patch just diagnoses taking address of a member
function when not in immediate context.
On Mon, Oct 18, 2021 at 12:42:00PM -0400, Jason Merrill wrote:
> > --- gcc/cp/typeck.c.jj 2021-10-05 09:53:55.382734051 +0200
> > +++ gcc/cp/typeck.c 2021-10-15 19:28:38.034213437 +0200
> > @@ -6773,9 +6773,21 @@ cp_build_addr_expr_1 (tree arg, bool str
> > return error_mark_node;
> > }
> > + if (TREE_CODE (t) == FUNCTION_DECL
> > + && DECL_IMMEDIATE_FUNCTION_P (t)
> > + && cp_unevaluated_operand == 0
> > + && (current_function_decl == NULL_TREE
> > + || !DECL_IMMEDIATE_FUNCTION_P (current_function_decl)))
>
> This doesn't cover some of the other cases of immediate context; we should
> probably factor most of immediate_invocation_p out into a function called
> something like in_immediate_context and use it here, and in several other
> places as well.
You're right, I've done that for the two spots in cp_build_addr_expr_1
and added testsuite coverage for where it changed behavior.
While doing that I've discovered further issues.
One is that we weren't diagnosing PMFs referring to immediate methods
returned from immediate functions (either directly or embedded in
aggregates). I'm not sure if it can only appear as PTRMEM_CST which
I've handled (cp_walk_subtree only walks the type and not the
PTRMEM_CST_MEMBER) or something else.
Another issue is that while default arg in immediate function
containing &immediate_fn works properly, if it is immediate_fn
instead, we were incorrectly rejecting it.
I've handled this in build_over_call, though with this usage
in_consteval_if_p is slightly misnamed, it stands for in consteval
if or some other reason why we are currently in immediate function context.
Though, that flag alone can't be all the reasons for being in immediate
function contexts, as I've tried the other reasons can't be handled in such
a bool and need to be tested too.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
PR c++/102753
* cp-tree.h (saved_scope): Document that consteval_if_p member
is also set while processing immediate invocation.
(in_immediate_context): Declare.
* call.c (in_immediate_context): New function.
(immediate_invocation_p): Use it.
(struct in_consteval_if_p_temp_override): New class.
(build_over_call): Temporarily set in_consteval_if_p for processing
immediate invocation arguments.
* typeck.c (cp_build_addr_expr_1): Diagnose taking address of
an immediate method. Use t instead of TREE_OPERAND (arg, 1).
Use in_immediate_context function.
* constexpr.c (find_immediate_fndecl): Handle PTRMEM_CST
which refers to immediate function decl.
* g++.dg/cpp2a/consteval13.C: Don't expect errors.
* g++.dg/cpp2a/consteval20.C: New test.
* g++.dg/cpp2a/consteval21.C: New test.
* g++.dg/cpp2a/consteval22.C: New test.
* g++.dg/cpp2a/consteval23.C: New test.
* g++.dg/cpp23/consteval-if11.C: New test.
The original compiler directive support was only for //line at the
start of a line and for //go: comments before function declarations.
When support was added for //go:notinheap for types and //go:embed for
variables the code did not adapt to permit spaces before the comment
or to permit the comments in var() or type() groups. This change
corrects those omissions.
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/358699
Resolves:
PR tree-optimization/102238 - alias_offset in gimple-ssa-sprintf.c is broken
PR tree-optimization/102919 - spurious -Wrestrict warning for sprintf into the same member array as argument plus offset
gcc/ChangeLog:
PR tree-optimization/102238
PR tree-optimization/102919
* gimple-ssa-sprintf.c (get_string_length): Add an argument.
(array_elt_at_offset): Move to pointer-query.
(set_aggregate_size_and_offset): New function.
(field_at_offset): Move to pointer-query.
(get_origin_and_offset): Rename...
(get_origin_and_offset_r): this. Add an argument. Make aggregate
handling more robust.
(get_origin_and_offset): New.
(alias_offset): Add an argument.
(format_string): Use subobject size determined by get_origin_and_offset.
* pointer-query.cc (field_at_offset): Move from gimple-ssa-sprintf.c.
Improve/correct handling of aggregates.
(array_elt_at_offset): Same.
* pointer-query.h (field_at_offset): Declare.
(array_elt_at_offset): Declare.
gcc/testsuite/ChangeLog:
PR tree-optimization/102238
PR tree-optimization/102919
* gcc.dg/tree-ssa/builtin-sprintf-warn-23.c: Remove warnings.
* gcc.dg/Wrestrict-23.c: New test.
gcc/fortran/ChangeLog:
PR fortran/102956
* symbol.c (gfc_check_conflict): Add conflict check for PDT KIND
and LEN type parameters.
gcc/testsuite/ChangeLog:
PR fortran/102956
* gfortran.dg/pdt_32.f03: New test.
Without considering all output insn operands (not only processed
before), in rare cases LRA can use the same hard register for
different outputs of the insn on different assignment subpasses. The
patch fixes the problem.
gcc/ChangeLog:
PR rtl-optimization/102842
* lra-constraints.c (match_reload): Ignore out in checking values
of outs.
(curr_insn_transform): Collect outputs before doing reloads of operands.
gcc/testsuite/ChangeLog:
PR rtl-optimization/102842
* g++.target/arm/pr102842.C: New test.
The allocated memory is not freed when returning early due to an error.
c++tools/ChangeLog:
* resolver.cc (module_resolver::read_tuple_file): Use unique_ptr
to ensure memory is freed before returning.
Tests which only include <x86intrin.h> expect many other include files
to be brought in, but not enough are.
Try to increase compatibility with x86 headers by:
- Create new immintrin.h, including the analogous subset of intrinsics
headers available for powerpc.
- Create new x86gprintrin.h, serving exclusively as the umbrella for
bmiintrin.h and bmi2intrin.h.
- Modify x86intrin.h:
- Include new immintrin.h.
- Remove mmintrin.h, xmmintrin.h, emmintrin.h, now included indirectly
from immintrin.h.
- Remove bmiintrin.h, bmi2intrin.h, now included indirectly from
x86gprintrin.h (which is now included from immintrin.h).
Add the new files to gcc/config.gcc.
Also, fix up the testcase that provoked PR102719, which requires
Power8 vector support.
Fixes commit 29fb1e831bf1c25e4574bf2f98a9f534e5c67665.
2021-10-25 Paul A. Clarke <pc@us.ibm.com>
gcc
PR target/102719
* config/rs6000/x86intrin.h: Move some included headers to new
headers. Include new immintrin.h instead of those headers.
* config/rs6000/immintrin.h: New.
* config/rs6000/x86gprintrin.h: New.
* config.gcc (powerpc*-*-*): Add new headers to extra_headers.
gcc/testsuite
* gcc.target/powerpc/pr78102.c: Fix dg directives to require Power8
vector support. Also, add -DNO_WARN_X86_INTRINSICS.
The following patch implements C++23 P2360R0. This proposal merely
extends init-statement to contain alias-declaration. init-statement
is used in if/for/switch. It also removes the unsightly duplication
of code by calling cp_parser_init_statement twice.
PR c++/102617
gcc/cp/ChangeLog:
* parser.c (cp_parser_for): Maybe call cp_parser_init_statement
twice. Warn about range-based for loops with initializer here.
(cp_parser_init_statement): Don't duplicate code. Allow
alias-declaration in init-statement.
gcc/testsuite/ChangeLog:
* g++.dg/cpp23/init-stmt1.C: New test.
* g++.dg/cpp23/init-stmt2.C: New test.
gcc.target/i386/avx512f-pr96891-3.c currently FAILs on 32-bit Solaris/x86:
FAIL: gcc.target/i386/avx512f-pr96891-3.c scan-assembler-times
(?n)vpcmp[bwdq][ \\\\t]*\\\\\$7 4
There are only 3 instances of the expected pattern because Solaris/x86
defaults to -mno-stv. Fixed by compiling with -mstv and
-mno-stackrealign. Tested on i386-pc-solaris2.11 and
x86_64-pc-linux-gnu.
2021-10-20 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
gcc/testsuite:
PR testsuite/102834
* gcc.target/i386/avx512f-pr96891-3.c: Add -mstv -mno-stackrealign
to dg-options.
The gcc.target/i386/avx512fp16-trunchf.c test FAILs on 32-bit Solaris/x86:
FAIL: gcc.target/i386/avx512fp16-trunchf.c scan-assembler-times vcvttsh2si[
\\\\t]+[^{\\n]*(?:%xmm[0-9]|\\\\(%esp\\\\))+, %eax(?:\\n|[ \\\\t]+#) 3
FAIL: gcc.target/i386/avx512fp16-trunchf.c scan-assembler-times
vcvttsh2usi[ \\\\t]+[^{\\n]*(?:%xmm[0-9]|\\\\(%esp\\\\))+, %eax(?:\\n|[
\\\\t]+#) 2
This happens because Solaris defaults to -fno-omit-frame-pointer, so it
uses %ebp instead of the expected %esp. As Hongyu Wang suggested in the
PR, this can be fixed by accepting both forms, which this patch does.
Tested on i386-pc-solaris2.11 and x86_64-pc-linux-gnu.
2021-10-20 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
gcc/testsuite:
PR testsuite/102835
* gcc.target/i386/avx512fp16-trunchf.c: Allow for %esp instead of
%ebp.
28_regex/basic_regex/84110.cc currently FAILs on Solaris:
FAIL: 28_regex/basic_regex/84110.cc (test for excess errors)
UNRESOLVED: 28_regex/basic_regex/84110.cc compilation failed to produce executable
Excess errors:
/vol/gcc/src/hg/master/local/libstdc++-v3/testsuite/28_regex/basic_regex/84110.cc:14: error: reference to 'extended' is ambiguous
The issue is seen in the full output:
/vol/gcc/src/hg/master/local/libstdc++-v3/testsuite/28_regex/basic_regex/84110.cc: In function ‘void test01()’:
/vol/gcc/src/hg/master/local/libstdc++-v3/testsuite/28_regex/basic_regex/84110.cc:14: error: reference to ‘extended’ is ambiguous
In file included from /var/gcc/regression/master/11.4-gcc-gas/build/gcc/include-fixed/math.h:391,
from /var/gcc/regression/master/11.4-gcc-gas/build/i386-pc-solaris2.11/libstdc++-v3/include/cmath:45,
from /vol/gcc/src/hg/master/local/libstdc++-v3/include/precompiled/stdc++.h:41:
/usr/include/floatingpoint.h:73: note: candidates are: ‘typedef unsigned int extended [3]’
Fixed by disambiguating extended. Tested on i386-pc-solaris2.11,
sparc-sun-solaris2.11, and x86_64-pc-linux-gnu.
2021-10-20 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
libstdc++-v3:
* testsuite/28_regex/basic_regex/84110.cc (test01)
[__cpp_exceptions]: Disambiguate extended.
17_intro/names.cc and experimental/names.cc currently FAIL on Solaris
FAIL: 17_intro/names.cc (test for excess errors)
FAIL: experimental/names.cc (test for excess errors)
Excess errors:
/usr/include/sys/timespec_util.h:22: error: expected ')' before ';' token
/usr/include/stdlib.h:157: error: expected unqualified-id before '[' token
/usr/include/stdlib.h:157: error: expected ')' before '[' token
<sys/timespec_util.h> has
extern int timespeccompare(const struct timespec *l, const struct timespec *r);
while <stdlib.h> has
typedef struct drand48_data {
unsigned int _initialised;
unsigned short int x[3];
unsigned short int a[3];
unsigned int c;
unsigned short lastx[3];
} drand48_data;
both of which are broken by defining r resp. x to ( in the testcase.
Fixed by undoing the defines. Tested on i386-pc-solaris2.11,
sparc-sun-solaris2.11, and x86_64-pc-linux-gnu.
2021-10-20 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
libstdc++-v3:
* testsuite/17_intro/names.cc [__sun__] (r, x): Undef.
Now that both are measured in bytes we can unify the two parameters.
2021-10-26 Richard Biener <rguenther@suse.de>
* tree-vectorizer.h (vect_create_addr_base_for_vector_ref):
Remove byte_offset parameter.
(vect_create_data_ref_ptr): Likewise.
* tree-vect-data-refs.c (vect_create_addr_base_for_vector_ref):
Likewise.
(vect_create_data_ref_ptr): Likewise.
* tree-vect-stmts.c (vectorizable_store): Adjust.
(vectorizable_load): Likewise.
This moves applying of a bias for negative stride accesses out of
dr_misalignment in favor of a more general optional offset argument.
The negative bias is now computed by get_load_store_type and applied
accordingly to determine the alignment support scheme. Likewise
the peeling/versioning code is adjusted albeit that still assumes
we'll end up with VMAT_CONTIGUOUS_DOWN or VMAT_CONTIGUOUS_REVERSE
but at least when not so (VMAT_STRIDED_SLP is one possibility) then
get_load_store_type will _not_ falsely report an aligned access but
instead an access with known misalignment.
This fixes PR96109.
2021-10-25 Richard Biener <rguenther@suse.de>
PR tree-optimization/96109
* tree-vectorizer.h (dr_misalignment): Add optional offset
parameter.
* tree-vect-data-refs.c (dr_misalignment): Likewise. Remove
offset applied for negative stride accesses.
(vect_enhance_data_refs_alignment): Compute negative stride
access offset and pass it to dr_misalignment.
* tree-vect-stmts.c (get_negative_load_store_type): Pass
negative offset to dr_misalignment.
(get_group_load_store_type): Likewise.
(get_load_store_type): Likewise.
(vectorizable_store): Remove asserts about alignment.
(vectorizable_load): Likewise.
As PR102897 shows, there is one incorrect assertion in function
simplify_permutation, which is based on the wrong assumption that
all cases with op2_type == tgt_type are handled previously, the
proposed fix is to remove the assertion.
gcc/ChangeLog:
PR tree-optimization/102897
* tree-ssa-forwprop.c (simplify_permutation): Remove a wrong assertion.
gcc/testsuite/ChangeLog:
* gcc.dg/pr102897.c: New test.
This changes the offset in elements for vect_create_addr_base_for_vector_ref
and vect_create_data_ref_ptr to an offset in bytes, easing a following
refactoring.
2021-10-26 Richard Biener <rguenther@suse.de>
* tree-vect-data-refs.c (vect_create_addr_base_for_vector_ref):
Take offset in bytes.
(vect_create_data_ref_ptr): Likewise.
* tree-vect-loop-manip.c (get_misalign_in_elems): Multiply
offset by element size.
(vect_create_cond_for_align_checks): Likewise.
* tree-vect-stmts.c (get_negative_load_store_type): Likewise.
(vectorizable_load): Remove duplicate leftover from merge
conflict.
This patch provides RTL expanders to implement logical shifts and
rotates of 128-bit values (stored in vector integer registers) by
constant bit counts. Previously, GCC would transfer these values
to a pair of integer registers (TImode) via memory to perform the
operation, then transfer the result back via memory. Instead these
operations are now expanded using (between 1 and 5) SSE2 vector
instructions.
Logical shifts by multiples of 8 can be implemented using x86_64's
pslldq/psrldq instruction:
ashl_8: pslldq $1, %xmm0
ret
lshr_32:
psrldq $4, %xmm0
ret
Logical shifts by greater than 64 can use pslldq/psrldq $8, followed
by a psllq/psrlq for the remaining bits:
ashl_111:
pslldq $8, %xmm0
psllq $47, %xmm0
ret
lshr_127:
psrldq $8, %xmm0
psrlq $63, %xmm0
ret
The remaining logical shifts make use of the following idiom:
ashl_1:
movdqa %xmm0, %xmm1
psllq $1, %xmm0
pslldq $8, %xmm1
psrlq $63, %xmm1
por %xmm1, %xmm0
ret
lshr_15:
movdqa %xmm0, %xmm1
psrlq $15, %xmm0
psrldq $8, %xmm1
psllq $49, %xmm1
por %xmm1, %xmm0
ret
Rotates by multiples of 32 can use x86_64's pshufd:
rotr_32:
pshufd $57, %xmm0, %xmm0
ret
rotr_64:
pshufd $78, %xmm0, %xmm0
ret
rotr_96:
pshufd $147, %xmm0, %xmm0
ret
Rotates by multiples of 8 (other than multiples of 32) can make
use of both pslldq and psrldq, followed by por:
rotr_8:
movdqa %xmm0, %xmm1
psrldq $1, %xmm0
pslldq $15, %xmm1
por %xmm1, %xmm0
ret
rotr_112:
movdqa %xmm0, %xmm1
psrldq $14, %xmm0
pslldq $2, %xmm1
por %xmm1, %xmm0
ret
And the remaining rotates use one or two pshufd, followed by a
psrld/pslld/por sequence:
rotr_1:
movdqa %xmm0, %xmm1
pshufd $57, %xmm0, %xmm0
psrld $1, %xmm1
pslld $31, %xmm0
por %xmm1, %xmm0
ret
rotr_63:
pshufd $78, %xmm0, %xmm1
pshufd $57, %xmm0, %xmm0
pslld $1, %xmm1
psrld $31, %xmm0
por %xmm1, %xmm0
ret
rotr_111:
pshufd $147, %xmm0, %xmm1
pslld $17, %xmm0
psrld $15, %xmm1
por %xmm1, %xmm0
ret
The new test case, sse2-v1ti-shift.c, is a run-time check to confirm that
the results of V1TImode shifts/rotates by constants, exactly match the
expected results of TImode operations, for various input test vectors.
2021-10-26 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-expand.c (ix86_expand_v1ti_shift): New helper
function to expand V1TI mode logical shifts by integer constants.
(ix86_expand_v1ti_rotate): New helper function to expand V1TI
mode rotations by integer constants.
* config/i386/i386-protos.h (ix86_expand_v1ti_shift,
ix86_expand_v1ti_rotate): Prototype new functions here.
* config/i386/sse.md (ashlv1ti3, lshrv1ti3, rotlv1ti3, rotrv1ti3):
New TARGET_SSE2 expanders to implement V1TI shifts and rotations.
gcc/testsuite/ChangeLog
* gcc.target/i386/sse2-v1ti-shift.c: New test case.
First, ssa-dom-thread-7 was looking at a dump file that was not
being generated. This probably happened in the detangling of the VRP
threader from VRP, and I didn't notice because the test came back as
with UNRESOLVED instead of FAIL.
Second, aarch64 gets far more threads than other architectures (20
versus 12). The difference is sufficiently different to make the
regex awkward.
We already have special casing for aarch64 in other parts of this
test, so perhaps it's simplest to have an arch specific test
for the thread3 count.
I don't know perhaps there's a better way. I wake up with chills in
the middle of the night thinking about this test ;-).
Tested on x86-64 Linux and aarch64 Linux.
gcc/testsuite/ChangeLog:
PR testsuite/102857
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Add -fdump-tree-vrp2-stats.
Tweak for aarch64.
The backward threader keeps a hash of visited blocks to avoid crossing
the same block twice. Interestingly, we haven't been checking it for
the final block out of the path. This may be inherited from the old
code, as it was simple enough that it didn't matter. With the
upcoming changes enabling the fully resolving threader, it gets
tripped often enough to cause wrong code to be generated.
Tested on x86-64 Linux.
gcc/ChangeLog:
* tree-ssa-threadbackward.c (back_threader::maybe_register_path):
Avoid threading circular paths.
The code that threads incoming paths to a PHI is duplicating what we
do generically in find_paths_to_names. This shortcoming is actually
one of the reasons we aren't threading all possible paths into a PHI.
For example, we give up after finding one threadable path, but some
PHIs have multiple threadable paths:
// x_5 = PHI <10(4), 20(5), ...>
// if (x_5 > 5)
Addressing this not only fixes the oversight, but simplifies the
PHI handling code, since we can consider the PHI fully resolved upon
return.
Interestingly, for ssa-thread-12.c the main thread everything was
hinging on was unreachable. With this patch, we call
maybe_register_path() earlier. In doing so, the solver realizes
that any path starting with 4->8 is unreachable and can be avoided.
This caused the cascade of threadable paths that depended on this
to no longer happen. Since threadable paths in thread[34] was the only
thing this test was testing, there's no longer anything to test. Neat!
Tested on x86-64 Linux.
gcc/ChangeLog:
* tree-ssa-threadbackward.c (back_threader::resolve_phi):
Attempt to resolve all incoming paths to a PHI.
(back_threader::resolve_def): Always return true for PHIs.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/pr21090.c: Adjust for threading.
* gcc.dg/tree-ssa/ssa-thread-12.c: Removed.
Sometimes we can solve a candidate path without having to recurse
further back. This can mostly happen in fully resolving mode, because
we can ask the ranger what the range on entry to the path is, but
there's no reason this can't always apply. This one-liner removes
the fully-resolving restriction.
I'm tickled pink to see how many things we now get quite early
in the compilation. I actually had to disable jump threading entirely
for a few tests because the early threader was catching things
disturbingly early. Also, as Richi predicted, I saw a lot of pre-VRP
cleanups happening.
I was going to commit this as obvious, but I think the test changes
merit discussion.
We've been playing games with gcc.dg/tree-ssa/ssa-thread-11.c for quite
some time. Every time a threading pass gets smarter, we push the
check further down the pipeline. We've officially run out of dumb
threading passes to disable ;-). In the last year we've gone up from a
handful of threads, to 34 threads with the current combination of
options. I doubt this is testing anything useful anymore, so I've
removed it.
Similarly for gcc.dg/tree-ssa/ssa-dom-thread-4.c. We used to thread 3
jump threads, but they were disallowed because of loop rotation. Then
we started catching more jump threads in VRP2 threading so we tested
there. With this patch though, we triple the number of threads found
from 11 to 31. I believe this test has outlived its usefulness, and
I've removed it. Note that even though we have these outrageous
possibilities for this test, the block copier ultimately chops them
down (23 survive though).
Tested on x86-64 Linux.
gcc/ChangeLog:
* tree-ssa-threadbackward.c (back_threader::find_paths_to_names):
Always try to resolve path without looking back.
* tree-ssa-threadupdate.c (dump_jump_thread): Indidicate whether
edge is a back edge.
gcc/testsuite/ChangeLog:
* gcc.dg/graphite/scop-dsyr2k-2.c: Adjust for jump threading changes.
* gcc.dg/graphite/scop-dsyr2k.c: Same.
* gcc.dg/graphite/scop-dsyrk-2.c: Same.
* gcc.dg/graphite/scop-dsyrk.c: Same.
* gcc.dg/tree-ssa/pr20701.c: Same.
* gcc.dg/tree-ssa/pr20702.c: Same.
* gcc.dg/tree-ssa/pr21086.c: Same.
* gcc.dg/tree-ssa/pr25382.c: Same.
* gcc.dg/tree-ssa/pr58480.c: Same.
* gcc.dg/tree-ssa/ssa-vrp-thread-1.c: Same.
* gcc.dg/tree-ssa/vrp08.c: Same.
* gcc.dg/tree-ssa/vrp55.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-4.c: Removed.
* gcc.dg/tree-ssa/ssa-thread-11.c: Removed.
* gcc.dg/uninit-pr89230-1.c: xfail.
As PR102789 shows, when vectorizer does some peelings for alignment
in prologues, function vect_update_inits_of_drs would update the
inits of some drs. But as the failed case, we shouldn't update the
dr for simd_lane_access, it has the fixed-length storage mainly for
the main loop, the update can make the access out of bound and access
the unexpected element.
gcc/ChangeLog:
PR tree-optimization/102789
* tree-vect-loop-manip.c (vect_update_inits_of_drs): Do not
update inits of simd_lane_access.
This static VRP routine does a simplification with casted conditions. Add it
to the general simplfier, and continue to invoke if from the VRP folder.
* tree-vrp.c (vrp_simplify_cond_using_ranges): Add return type and
move to vr-values.c.
(simplify_casted_conds): Move to vrp_folder class.
(execute_vrp): Call via vrp_folder now.
* vr-values.c (simplify_cond_using_ranges_1): Call simplify_casted_cond.
(simplify_using_ranges::simplify_casted_cond): Relocate from tree-vrp.c.
* vr-values.h (simplify_casted_cond): Add prototype.
Until now, ranger VRP has only simplified statements with ranges. This patch
enables us to fold all statements.
gcc/
* tree-vrp.c (rvrp_folder::fold_stmt): If simplification fails, try
to fold anyway.
gcc/testsuite/
* gcc.dg/tree-ssa/vrp98.c: Disable evrp for vrp1 test.
* gcc.dg/tree-ssa/vrp98-1.c: New. Test for folding in evrp.
