PR gcov-profile/93623
* libgcov-interface.c (__gcov_fork): Do not flush
and reset only in child process.
(__gcov_execl): Dump counters only and reset them
only if exec* fails.
(__gcov_execlp): Likewise.
(__gcov_execle): Likewise.
(__gcov_execv): Likewise.
(__gcov_execvp): Likewise.
(__gcov_execve): Likewise.
gcc/ChangeLog:
2020-04-17 Martin Liska <mliska@suse.cz>
PR gcov-profile/94636
* gcov.c (main): Print total lines summary at the end.
(generate_results): Expect file_name always being non-null.
Print newline after intermediate file is printed in order to align with
what we do for normal files.
libstdc++-v3/ChangeLog:
2020-02-04 Martin Liska <mliska@suse.cz>
PR c/92472
* include/parallel/multiway_merge.h:
Use const for _Compare template argument.
This rewrites hybrid SLP detection to be simpler and cope with
group size changes in the SLP graph. In particular detection
works starting from non-SLP stmts following use->def chains
rather than walking the SLP graph and following def->use chains.
It's all temporary of course since non-SLP and thus hybrid SLP
will go away.
2020-05-05 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (struct vdhs_data): New.
(vect_detect_hybrid_slp): New walker.
(vect_detect_hybrid_slp): Rewrite.
We now always vectorize two BBs, adjust the selector to also scan
for integer multiplication vectorization explicitely.
2020-05-05 Richard Biener <rguenther@suse.de>
PR testsuite/92177
* gcc.dg/vect/bb-slp-22.c: Adjust.
This fixes lack of an escape point of externally declared variables.
2020-05-05 Richard Biener <rguenther@suse.de>
PR ipa/94947
* tree-ssa-structalias.c (ipa_pta_execute): Use
varpool_node::externally_visible_p ().
(refered_from_nonlocal_var): Likewise.
* gcc.dg/torture/pr94947-1.c: New testcase.
* gcc.dg/torture/pr94947-2.c: Likewise.
The link phase is always partial (-r) for VxWorks in kernel mode, which
means that it uses incremental LTO linking by default (-flinker-output=rel).
But in this mode the LTO plugin outputs a warning if one of the object files
involved in the link does not contain LTO bytecode, before switching to
nolto-rel mode. We do not do repeated incremental linking for VxWorks so
silence the warning.
lto-plugin/
* lto-plugin.c: Document -linker-output-auto-notlo-rel option.
(linker_output_set): Change type to bool.
(linker_output_known): Likewise.
(linker_output_auto_nolto_rel): New variable.
(all_symbols_read_handler): Take it into account.
<LDPO_REL>: Do not issue the warning if it is set.
(process_option): Process -linker-output-auto-notlo-rel.
(cleanup_handler): Remove unused variable.
(onload) <LDPT_LINKER_OUTPUT>: Adjust to above type change.
gcc/
* gcc.c (LTO_PLUGIN_SPEC): Define if not already.
(LINK_PLUGIN_SPEC): Execute LTO_PLUGIN_SPEC.
* config/vxworks.h (LTO_PLUGIN_SPEC): Define.
The CONSTRUCTOR_NO_CLEARING flag was invented to avoid generating a memset
for CONSTRUCTORS that lack elements, but it turns out that the gimplifier
can generate a memcpy for them instead, which is worse performance-wise,
so this prevents it from doing that for them.
* gimplify.c (gimplify_init_constructor): Do not put the constructor
into static memory if it is not complete.
This fixes the case of not using the multithreaded model when
only conditionally storing to the destination. We cannot elide
the load in this case.
2020-05-05 Richard Biener <rguenther@suse.de>
PR tree-optimization/94949
* tree-ssa-loop-im.c (execute_sm): Check whether we use
the multithreaded model or always compute the stored value
before eliding a load.
* gcc.dg/torture/pr94949.c: New testcase.
The attached patch eliminates a redundant zero extend from the AArch64 backend. Given the following C code:
unsigned long long foo(unsigned a)
{
return ~a;
}
prior to this patch, AArch64 GCC at -O2 generates:
foo:
mvn w0, w0
uxtw x0, w0
ret
but the uxtw is redundant, since the mvn clears the upper half of the x0 register. After applying this patch, GCC at -O2 gives:
foo:
mvn w0, w0
ret
Testing:
Added regression test which passes after applying the change to aarch64.md.
Full bootstrap and regression on aarch64-linux with no additional failures.
* config/aarch64/aarch64.md (*one_cmpl_zero_extend): New.
* gcc.target/aarch64/mvn_zero_ext.c: New test.
The popcount* testcases show yet another creative way to write popcount,
but rather than adjusting the popcount matcher to deal with it, I think
we just should canonicalize those (X + (X << C) to X * (1 + (1 << C))
and (X << C1) + (X << C2) to X * ((1 << C1) + (1 << C2)), because for
multiplication we already have simplification rules that can handle nested
multiplication (X * CST1 * CST2), while the the shifts and adds we have
nothing like that. And user could have written the multiplication anyway,
so if we don't emit the fastest or smallest code for the multiplication by
constant, we should improve that. At least on the testcases seems the
emitted code is reasonable according to cost, except that perhaps we could
in some cases try to improve expansion of vector multiplication by
uniform constant.
2020-05-05 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/94800
* match.pd (X + (X << C) to X * (1 + (1 << C)),
(X << C1) + (X << C2) to X * ((1 << C1) + (1 << C2))): New
canonicalizations.
* gcc.dg/tree-ssa/pr94800.c: New test.
* gcc.dg/tree-ssa/popcount5.c: New test.
* gcc.dg/tree-ssa/popcount5l.c: New test.
* gcc.dg/tree-ssa/popcount5ll.c: New test.
This insn and split splits into HI->V?HImode broadcast for avx2 and later,
but either the operands need to be %xmm0-%xmm15 (i.e. VEX encoded insn), or
the insn needs both AVX512BW and AVX512VL.
Now, Yv constraint is v for AVX512VL and x otherwise, so for -mavx512vl -mno-avx512bw
we ICE if we end up with a %xmm16+ register from RA.
Yw constraint is v for AVX512VL and AVX512BW and nothing otherwise, so
in this pattern we actually need xYw.
2020-05-05 Jakub Jelinek <jakub@redhat.com>
PR target/94942
* config/i386/mmx.md (*vec_dupv4hi): Use xYw constraints instead of Yv.
* gcc.target/i386/pr94942.c: New test.
On x86 (the only target with umulv4_optab) one can use mull; seto to check
for overflow instead of performing wider multiplication and performing
comparison on the high bits.
2020-05-05 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/94914
* match.pd ((((type)A * B) >> prec) != 0 to .MUL_OVERFLOW(A, B) != 0):
New simplification.
* gcc.target/i386/pr94914.c: New test.
Pattern explosing and manual mode checks can be avoided by using
int_nonimmediate_operand special predicate.
While there, rewrite *x86_mov<SWI48:mode>cc_0_m1_neg_leu<SWI:mode>
to a combine pass splitter.
* config/i386/i386.md (*testqi_ext_3): Use
int_nonimmediate_operand instead of manual mode checks.
(*x86_mov<SWI48:mode>cc_0_m1_neg_leu<SWI:mode>):
Use int_nonimmediate_operand predicate. Rewrite
define_insn_and_split pattern to a combine pass splitter.
C++ makes mismatched prototype and implementation OK.
2020-05-05 Richard Biener <rguenther@suse.de>
* targhooks.h (default_add_stmt_cost): Add vec_info * parameter.
I've recently tested i386-pc-solaris2.11 bootstrap on Solaris 11/x86
with only the bundled tools (using /usr/gnu/bin/as from binutils 2.30 in
this case). It failed compiling libgo/runtime/proc.c, creating invalid
assembly:
proc.s: Assembler messages:
proc.s:2092: Error: junk at end of line, first unrecognized character is `*'
.globl __emutls_v.*runtime.g
and several more errors. This is completely unexpected since Solaris
does support TLS. It turned out that 32-bit TLS detection in
gcc/configure had failed:
configure:25145: checking assembler for thread-local storage support
configure:25158: /usr/gnu/bin/as --fatal-warnings -o conftest.o conftest.s >&5
conftest.s: Assembler messages:
conftest.s:6: Error: relocated field and relocation type differ in signedness
conftest.s:7: Error: @TLSLDM reloc is not supported with 64-bit output format
conftest.s:7: Error: junk `@tlsldm' after expression
which isn't unexpected given that the bundled gas has been configured
for x86_64-pc-solaris2.11, i.e. 64-bit-default.
This is easily fixed by explicitly passing --32 for the 32-bit case,
matching what is done for the 64-bit test.
Tested on i386-pc-solaris2.11 with 32-bit-default and 64-bit-default gas
as well as with /usr/bin/as, always correctly detecting TLS support.
* configure.ac <i[34567]86-*-*>: Add --32 to tls_as_opt on Solaris.
* configure: Regenerate.
Soonish we'll get SLP nodes which have no corresponding scalar
stmt and thus not stmt_vec_info and thus no way to get back to
the associated vec_info. This patch makes the vec_info available
as part of the APIs instead of putting in that back-pointer into
the leaf data structures.
2020-05-05 Richard Biener <rguenther@suse.de>
* tree-vectorizer.h (_stmt_vec_info::vinfo): Remove.
(STMT_VINFO_LOOP_VINFO): Likewise.
(STMT_VINFO_BB_VINFO): Likewise.
* tree-vect-data-refs.c: Adjust for the above, adding vec_info *
parameters and adjusting calls.
* tree-vect-loop-manip.c: Likewise.
* tree-vect-loop.c: Likewise.
* tree-vect-patterns.c: Likewise.
* tree-vect-slp.c: Likewise.
* tree-vect-stmts.c: Likewise.
* tree-vectorizer.c: Likewise.
* target.def (add_stmt_cost): Add vec_info * parameter.
* target.h (stmt_in_inner_loop_p): Likewise.
* targhooks.c (default_add_stmt_cost): Adjust.
* doc/tm.texi: Re-generate.
* config/aarch64/aarch64.c (aarch64_extending_load_p): Add
vec_info * parameter and adjust.
(aarch64_sve_adjust_stmt_cost): Likewise.
(aarch64_add_stmt_cost): Likewise.
* config/arm/arm.c (arm_add_stmt_cost): Likewise.
* config/i386/i386.c (ix86_add_stmt_cost): Likewise.
* config/rs6000/rs6000.c (rs6000_add_stmt_cost): Likewise.
As mentioned in the previous PR94460 patch, the RTL patterns look too
large/complicated, we can simplify them by just performing two 2 arg
permutations to move the arguments into the right spots and then just
doing the plus/minus (or signed saturation version thereof).
2020-05-05 Jakub Jelinek <jakub@redhat.com>
PR target/94460
* config/i386/sse.md (avx2_ph<plusminus_mnemonic>wv16hi3,
ssse3_ph<plusminus_mnemonic>wv8hi3, ssse3_ph<plusminus_mnemonic>wv4hi3,
avx2_ph<plusminus_mnemonic>dv8si3, ssse3_ph<plusminus_mnemonic>dv4si3,
ssse3_ph<plusminus_mnemonic>dv2si3): Simplify RTL patterns.
When folding a CALL_EXPR, we can avoid copying it until folding changes
one of its arguments. And when folding a TREE_VEC, we can avoid using
an intermediate releasing_vec by copying the TREE_VEC as soon as folding
changes one of its arguments, like we do in the CALL_EXPR case.
Incidentally, the CALL_EXPR change also fixes the testcase in PR94038.
The reason is that the call to maybe_constant_value from cp_fold on
the call 'bar<int>()' now reuses the result of the earlier call to
maybe_constant_value from fold_for_warn, via the cv_cache. This earlier
call passes uid_sensitive=true, whereas the call from cp_fold passes
uid_sensitive=false, and so by reusing the cached result of the earlier
call we now avoid instantiating bar<int> at all.
gcc/cp/ChangeLog:
PR c++/94038
* cp-gimplify.c (cp_fold) <case CALL_EXPR>: Move some variable
declarations closer to their uses. Copy the CALL_EXPR only
when one of its arguments has changed.
<case TREE_VEC>: Instead of first collecting the folded
arguments into a releasing_vec, just make a copy of the TREE_VEC
as soon as folding changes one of its arguments.
gcc/testsuite/ChangeLog:
PR c++/94038
* g++.dg/warn/pr94038.C: New test.
This should return void according to the Itanium C++ ABI.
2020-05-04 Fangrui Song <maskray@google.com>
* libsupc++/cxxabi.h (__cxa_finalize): Fix return type.
The previous URL to an entry in the wayback machine now redirects to a
page saying "SGI.com Tech Archive Resources now retired" so use an older
entry from the archive.
* doc/xml/faq.xml: Use working link for SGI STL FAQ.
* doc/html/*: Regenerate.
Calculating the size of a chunk being returned to the upstream allocator
was done with a 32-bit type, so it wrapped if the chunk was 4GB or
larger.
I don't know how to test this without allocating 4GB, so there's no test
in the testsuite. It has been tested manually with allocations sizes and
alignments exceeding 4GB.
PR libstdc++/94906
* src/c++17/memory_resource.cc
(monotonic_buffer_resource::_Chunk::release): Use size_t for shift
operands.
This fixes two compilation errors preventing bootstrap with Ada
on x86_64-pc-cygwin.
2020-05-04 Mikael Pettersson <mikpelinux@gmail.com>
PR bootstrap/94918
* mingw32.h: Prevent windows.h from including emmintrin.h on Cygw64.
* s-oscons-tmplt.c (Serial_Port_Descriptor): Use System.Win32.HANDLE
also on Cygwin.
When long doubles are 64 bit, the AIX C library overrides the definitions
but GCC builtins point to 128 bit names. This patch overrides the
builtins for fmodl, frexpl, ldexpl and modfl to refer to the 64 bit symbols.
2020-05-04 Clement Chigot <clement.chigot@atos.net>
David Edelsohn <dje.gcc@gmail.com>
* config/rs6000/rs6000-call.c (rs6000_init_builtins): Override explicit
for fmodl, frexpl, ldexpl and modfl builtins.
create_output_operand coerces an output operand to the insn's
predicates, using a suggested rtx location if convenient.
But if that rtx location is actually required rather than
optional, the builder of the insn has to emit a move afterwards.
(We could instead add a new interface that does this automatically,
but that's future work.)
This PR shows that we were failing to emit the move for some of the
vector load internal functions. I think there are other routines in
internal-fn.c that potentially have the same problem, but this patch is
supposed to be a conservative subset suitable for backporting to GCC 10.
2020-05-04 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR middle-end/94941
* internal-fn.c (expand_load_lanes_optab_fn): Emit a move if the
chosen lhs is different from the gcall lhs.
(expand_mask_load_optab_fn): Likewise.
(expand_gather_load_optab_fn): Likewise.
gcc/testsuite/
PR middle-end/94941
* gcc.target/aarch64/sve/acle/general/unoptimized_1.c: New test.
This corrects an oversight, the coro.gro object is a
a compiler-generated entity and should be marked as
artificial and ignored.
gcc/cp/ChangeLog:
2020-05-04 Iain Sandoe <iain@sandoe.co.uk>
* coroutines.cc (morph_fn_to_coro): Mark the coro.gro variable
as artificial and ignored.
Although the code here is well formed, it doesn't show intent well.
The reason checkers trigger on this is that it is a cause of real
bugs. So, negate a ptrdiff_t instead.
* libsupc++/dyncast.cc (__dynamic_cast): Cast offsetof to
ptrdiff_t before negation, to show intent more clearly.
When returning 0 or -1, "SBB reg,reg" instruction that borrows carry
flag can be used. Carry flag can be generated by converting compare
with zero to a LTU compare with one, so e.g.
return -(x == 0)
generates:
cmpq $1, %rdi
sbbq %rax, %rax
instead of:
xorl %eax, %eax
testq %rdi, %rdi
sete %al
negq %rax
A similar conversion can be used for
return -(x != 0)
where NEG insn can be used instead of compare. According to x86 ISA,
NEG insn sets carry flag when the source operand is != 0, resulting in:
negq %rdi
sbbq %rax, %rax
The conversion avoids partial register stall with SETcc instructions.
PR target/94795
* config/i386/i386.md (*neg<mode>_ccc): New insn pattern.
(EQ compare->LTU compare splitter): New splitter.
(NE compare->NEG splitter): Ditto.
testsuite/ChangeLog:
PR target/94795
* gcc.target/i386/pr94795-1.c: New test.
* gcc.target/i386/pr94795-2.c: New test.
Process_template_parm ends up walking the parameter list twice.
There's not need to do this. Just rember the final node and modify
its CHAIN directly. Also comment on why end_template_parm_list does a
pop and a push, rather than modifying the header in place.
pt.c (process_template_parm): Don't walk the template list twice,
remember the final node instead.
(end_template_parm_list): Refactor. Comment on why we do a pop
and a push.
This followup patch just removes some stashing that we never made use of.
* constraint.cc (tsubst_nested_requirement): TYPE directly holds
notmalized requirement.
(finish_nested_requirement): Don't stash current tpl parms into
the requirement.
(diagnose_nested_requirement): TYPE directly holds notmalized
requirement.
This reverts commit 6318fe7739.
Revert:
2020-04-30 Marek Polacek <polacek@redhat.com>
PR c++/94775
* tree.c (check_base_type): Return true only if TYPE_USER_ALIGN match.
(check_aligned_type): Check if TYPE_USER_ALIGN match.
We need a reference to assess alignment, fall back to the original
reference tree if available.
2020-05-04 Richard Biener <rguenther@suse.de>
PR tree-optimization/93891
* tree-ssa-sccvn.c (vn_reference_lookup_3): Fall back to
the original reference tree for assessing access alignment.
I implicitly assumed that programs using pmr::synchronized_pool_resource
would also be using multiple threads, and so the weak symbols in
gthr-posix.h would be resolved by linking to libpthread. If that isn't
true then it crashes when trying to use pthread_key_create.
This commit makes the pool resource check __gthread_active_p() before
using thread-specific data, and just use a single set of memory pools
when there's only a single thread.
PR libstdc++/94936
* src/c++17/memory_resource.cc (synchronized_pool_resource::_TPools):
Add comment about single-threaded behaviour.
(synchronized_pool_resource::_TPools::move_nonempty_chunks()): Hoist
class member access out of loop.
(synchronized_pool_resource::synchronized_pool_resource())
(synchronized_pool_resource::~synchronized_pool_resource())
(synchronized_pool_resource::release()): Check __gthread_active_p
before creating and/or deleting the thread-specific data key.
(synchronized_pool_resource::_M_thread_specific_pools()): Adjust
assertions.
(synchronized_pool_resource::do_allocate(size_t, size_t)): Add fast
path for single-threaded case.
(synchronized_pool_resource::do_deallocate(void*, size_t, size_t)):
Likewise. Return if unable to find a pool that owns the allocation.
* testsuite/20_util/synchronized_pool_resource/allocate_single.cc:
New test.
* testsuite/20_util/synchronized_pool_resource/cons_single.cc: New
test.
* testsuite/20_util/synchronized_pool_resource/release_single.cc: New
test.
Currently store-motion emits a load of the value in the loop
preheader even when the original loop does not contain any read
of the reference. This avoids doing this. In the conditional
store-motion case we need to mark the sunk stores with no-warning
since the control dependence is too tricky to figure out for
the uninit warning.
2020-05-04 Richard Biener <rguenther@suse.de>
PR tree-optimization/39612
* tree-ssa-loop-im.c (im_mem_ref::loaded): New member.
(set_ref_loaded_in_loop): New.
(mark_ref_loaded): Likewise.
(gather_mem_refs_stmt): Call mark_ref_loaded for loads.
(execute_sm): Avoid issueing a load when it was not there.
(execute_sm_if_changed): Avoid issueing warnings for the
conditional store.
* gcc.dg/tree-ssa/pr39612.c: New testcase.
when callers and callees do not quite agree on the type of a
parameter, usually with ill-defined K&R or with smilarly wrong LTO
input, IPA-CP can attempt to try and substitute a wrong type for a
parameter (see e.g. gcc.dg/torture/pr48063.c). Function
tree_function_versioning attempts to handle this in order not to
create invalid gimple but it then creates the mapping using
setup_one_parameter which also handles the same situation to avoid
similar problems when inlining and in more defined way.
So this patch simply removes the conversion attempts in
tree_function_versioning itself. It is helpful for my upcoming fix of
PR 93385 because then I do not need to teach
ipa_param_body_adjustments to distinguish between successful and
unsuccessful mappings - setup_one_parameter uses force_value_to_type
for conversions which simly maps the worst cases to zero.
2020-05-04 Martin Jambor <mjambor@suse.cz>
PR ipa/93385
* tree-inline.c (tree_function_versioning): Leave any type conversion
of replacements to setup_one_parameter and its friend
force_value_to_type.
Convert unsigned compares where
m >= LARGE_POWER_OF_TWO
and LARGE_POWER_OF_TWO represent an immediate where bit 33+ is set to use
a SHR instruction and compare the result to 0. This avoids loading a
large immediate with MOVABS insn.
movabsq $1099511627775, %rax
cmpq %rax, %rdi
ja .L5
gets converted to:
shrq $40, %rdi
jne .L5
PR target/94650
* config/i386/predicates.md (shr_comparison_operator): New predicate.
* config/i386/i386.md (compare->shr splitter): New splitters.
testsuite/ChangeLog:
PR target/94650
* gcc.targeti/i386/pr94650.c: New test.
The following patch (on top of the two other PR94718 patches) performs the
actual optimization requested in the PR.
2020-05-04 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/94718
* match.pd ((X < 0) != (Y < 0) into (X ^ Y) < 0): New simplification.
* gcc.dg/tree-ssa/pr94718-4.c: New test.
* gcc.dg/tree-ssa/pr94718-5.c: New test.
On the following testcase, there are in *.optimized dump 14 nop conversions
(from signed to unsigned and back), while this patch decreases that number
to just 4; for bitwise ops it really doesn't matter if they are performed in
signed or unsigned, so the patch (in GIMPLE only, there are some comments
about it being undesirable during GENERIC earlier), if it sees both
bitop operands nop converted from the same types performs the bitop in their
non-converted type and converts the result (i.e. 2 conversions into 1),
similarly, if a bitop has one operand nop converted from something, the
other not and the result is converted back to the type of the nop converted
operand before conversion, it is possible to replace those 2 conversions
with just a single conversion of the other operand.
2020-05-04 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/94718
* match.pd (bitop (convert @0) (convert? @1)): For GIMPLE, if we can,
replace two nop conversions on bit_{and,ior,xor} argument
and result with just one conversion on the result or another argument.
* gcc.dg/tree-ssa/pr94718-3.c: New test.
Martin Liska
Richard Biener
Eric Botcazou
Alex Coplan
Jakub Jelinek
Uros Bizjak
Rainer Orth
GCC Administrator
Patrick Palka
Fangrui Song
Jonathan Wakely
David Edelsohn
Richard Sandiford
Iain Sandoe
Joseph Myers
Nathan Sidwell
Marek Polacek
Martin Jambor