gcc/analyzer/ChangeLog:
* state-purge.cc (state_purge_annotator::add_node_annotations):
Avoid duplicate before-supernode annotations when returning from
an interprocedural call. Show after-supernode annotations.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
Implementations of the x87 floating point instruction set have always
had some pretty strange characteristics. For example on the original
Intel Pentium the FLDPI instruction (to load 3.14159... into a register)
took 5 cycles, and the FLDZ instruction (to load 0.0) took 2 cycles,
when a regular FLD (load from memory) took just 1 cycle!? Given that
back then memory latencies were much lower (relatively) than they are
today, these instructions were all but useless except when optimizing
for size (impressively FLDZ/FLDPI require only two bytes).
Such was the world back in 2006 when Uros Bizjak first added support for
fldz https://gcc.gnu.org/pipermail/gcc-patches/2006-November/202589.html
and then shortly after sensibly disabled them for !optimize_size with
https://gcc.gnu.org/pipermail/gcc-patches/2006-November/204405.html
Alas this vestigial logic still persists in the compiler today,
so for example on x86_64 for the following function:
double foo(double x) { return x + 0.0; }
generates with -O2
foo: addsd .LC0(%rip), %xmm0
ret
.LC0: .long 0
.long 0
preferring to read the constant 0.0 from memory [the constant pool],
except when optimizing for size. With -Os we get:
foo: xorps %xmm1, %xmm1
addsd %xmm1, %xmm0
ret
Which is not only smaller (the two instructions require seven bytes vs.
eight for the original addsd from mem, even without considering the
constant pool) but is also faster on modern hardware. The latter code
sequence is generated by both clang and msvc with -O2. Indeed Agner
Fogg documents the set of floating point/SSE constants that it's
cheaper to materialize than to load from memory.
This patch shuffles the conditions on the i386 backend's *movtf_internal,
*movdf_internal and *movsf_internal define_insns to untangle the newer
TARGET_SSE_MATH clauses from the historical standard_80387_constant_p
conditions. Amongst the benefits of this are that it improves the code
generated for PR tree-optimization/90356 and resolves PR target/86722.
2022-03-17 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/86722
PR tree-optimization/90356
* config/i386/i386.md (*movtf_internal): Don't guard
standard_sse_constant_p clause by optimize_function_for_size_p.
(*movdf_internal): Likewise.
(*movsf_internal): Likewise.
gcc/testsuite/ChangeLog
PR target/86722
PR tree-optimization/90356
* gcc.target/i386/pr86722.c: New test case.
* gcc.target/i386/pr90356.c: New test case.
This patch adjusts range_from_dom to follow the dominator tree through the
cache until value is found, then apply any outgoing ranges encountered
along the way. This reduces the amount of cache storage required.
PR tree-optimization/102943
* gimple-range-cache.cc (ranger_cache::range_from_dom): Find range via
dominators and apply intermediary outgoing edge ranges.
This only affects Windows, but reduces the preprocessed size of
<filesystem> significantly.
libstdc++-v3/ChangeLog:
PR libstdc++/92546
* include/bits/fs_path.h (path::make_preferred): Use
handwritten loop instead of std::replace.
GCC thinks the following can lead to a buffer overflow when __ns.size()
equals zero:
const basic_string<_CharT>& __ns = __mp.negative_sign();
_M_negative_sign_size = __ns.size();
__negative_sign = new _CharT[_M_negative_sign_size];
__ns.copy(__negative_sign, _M_negative_sign_size);
This happens because operator new might be replaced with something that
writes to this->_M_negative_sign_size and so the basic_string::copy call
could use a non-zero size to write to a zero-length buffer.
The solution suggested by Richi is to cache the size in a local variable
so that the compiler knows it won't be changed between the allocation
and the copy.
This commit goes further and rewrites the whole function to use RAII and
delay all modifications of *this until after all allocations have
succeeded. The RAII helper type caches the size and copies the string
and owns the memory until told to release it.
libstdc++-v3/ChangeLog:
PR middle-end/104966
* include/bits/locale_facets_nonio.tcc
(__moneypunct_cache::_M_cache): Replace try-catch with RAII and
make all string copies before any stores to *this.
As mentioned in the PR, the latest Intel SDM has added:
"Processors that enumerate support for Intel® AVX (by setting the feature flag CPUID.01H:ECX.AVX[bit 28])
guarantee that the 16-byte memory operations performed by the following instructions will always be
carried out atomically:
• MOVAPD, MOVAPS, and MOVDQA.
• VMOVAPD, VMOVAPS, and VMOVDQA when encoded with VEX.128.
• VMOVAPD, VMOVAPS, VMOVDQA32, and VMOVDQA64 when encoded with EVEX.128 and k0 (masking disabled).
(Note that these instructions require the linear addresses of their memory operands to be 16-byte
aligned.)"
The following patch deals with it just on the libatomic library side so far,
currently (since ~ 2017) we emit all the __atomic_* 16-byte builtins as
library calls since and this is something that we can hopefully backport.
The patch simply introduces yet another ifunc variant that takes priority
over the pure CMPXCHG16B one, one that checks AVX and CMPXCHG16B bits and
on non-Intel clears the AVX bit during detection for now (if AMD comes
with the same guarantee, we could revert the config/x86/init.c hunk),
which implements 16-byte atomic load as vmovdqa and 16-byte atomic store
as vmovdqa followed by mfence.
2022-03-17 Jakub Jelinek <jakub@redhat.com>
PR target/104688
* Makefile.am (IFUNC_OPTIONS): Change on x86_64 to -mcx16 -mcx16.
(libatomic_la_LIBADD): Add $(addsuffix _16_2_.lo,$(SIZEOBJS)) for
x86_64.
* Makefile.in: Regenerated.
* config/x86/host-config.h (IFUNC_COND_1): For x86_64 define to
both AVX and CMPXCHG16B bits.
(IFUNC_COND_2): Define.
(IFUNC_NCOND): For x86_64 define to 2 * (N == 16).
(MAYBE_HAVE_ATOMIC_CAS_16, MAYBE_HAVE_ATOMIC_EXCHANGE_16,
MAYBE_HAVE_ATOMIC_LDST_16): Define to IFUNC_COND_2 rather than
IFUNC_COND_1.
(HAVE_ATOMIC_CAS_16): Redefine to 1 whenever IFUNC_ALT != 0.
(HAVE_ATOMIC_LDST_16): Redefine to 1 whenever IFUNC_ALT == 1.
(atomic_compare_exchange_n): Define whenever IFUNC_ALT != 0
on x86_64 for N == 16.
(__atomic_load_n, __atomic_store_n): Redefine whenever IFUNC_ALT == 1
on x86_64 for N == 16.
(atomic_load_n, atomic_store_n): New functions.
* config/x86/init.c (__libat_feat1_init): On x86_64 clear bit_AVX
if CPU vendor is not Intel.
Something went wrong when testing the earlier patch to move the
late sinking to before the late phiopt for PR102008. The following
makes sure to unsplit edges after the late sinking since the split
edges confuse the following phiopt leading to missed optimizations.
I've went for a new pass parameter for this to avoid changing the
CFG after the early sinking pass at this point.
2022-03-17 Richard Biener <rguenther@suse.de>
PR tree-optimization/104960
* passes.def: Add pass parameter to pass_sink_code, mark
last one to unsplit edges.
* tree-ssa-sink.cc (pass_sink_code::set_pass_param): New.
(pass_sink_code::execute): Always execute TODO_cleanup_cfg
when we need to unsplit edges.
* gcc.dg/gimplefe-37.c: Adjust to allow either the true
or false edge to have a forwarder.
As mentioned in the PR, we emit a bogus uninitialized warning but
easily could emit wrong-code for it or similar testcases too.
The bug is that we emit clobber for a TARGET_EXPR_SLOT too early:
D.2499.e = B::qux (&h); [return slot optimization]
D.2516 = 1;
try
{
B::B (&D.2498, &h);
try
{
_2 = baz (&D.2498);
D.2499.f = _2;
D.2516 = 0;
try
{
try
{
bar (&D.2499);
}
finally
{
C::~C (&D.2499);
}
}
finally
{
D.2499 = {CLOBBER(eol)};
}
}
finally
{
D.2498 = {CLOBBER(eol)};
}
}
catch
{
if (D.2516 != 0) goto <D.2517>; else goto <D.2518>;
<D.2517>:
A::~A (&D.2499.e);
goto <D.2519>;
<D.2518>:
<D.2519>:
}
The CLOBBER for D.2499 is essentially only emitted on the non-exceptional
path, if B::B or baz throws, then there is no CLOBBER for it but there
is a conditional destructor A::~A (&D.2499.e). Now, ehcleanup1
sink_clobbers optimization assumes that clobbers in the EH cases are
emitted after last use and so sinks the D.2499 = {CLOBBER(eol)}; later,
so we then have
# _3 = PHI <1(3), 0(9)>
<L2>:
D.2499 ={v} {CLOBBER(eol)};
D.2498 ={v} {CLOBBER(eol)};
if (_3 != 0)
goto <bb 11>; [INV]
else
goto <bb 15>; [INV]
<bb 11> :
_35 = D.2499.a;
if (&D.2499.b != _35)
where that _35 = D.2499.a comes from inline expansion of the A::~A dtor,
and that is a load from a clobbered memory.
Now, what the gimplifier sees in this case is a CLEANUP_POINT_EXPR with
somewhere inside of it a TARGET_EXPR for D.2499 (with the C::~C (&D.2499)
cleanup) which in its TARGET_EXPR_INITIAL has another TARGET_EXPR for
D.2516 bool flag which has CLEANUP_EH_ONLY which performs that conditional
A::~A (&D.2499.e) call.
The following patch ensures that CLOBBERs (and asan poisoning) are emitted
after even those gimple_push_cleanup pushed cleanups from within the
TARGET_EXPR_INITIAL gimplification (i.e. the last point where the slot could
be in theory used). In my first version of the patch I've done it by just
moving the
/* Add a clobber for the temporary going out of scope, like
gimplify_bind_expr. */
if (gimplify_ctxp->in_cleanup_point_expr
&& needs_to_live_in_memory (temp))
{
...
}
block earlier in gimplify_target_expr, but that regressed a couple of tests
where temp is marked TREE_ADDRESSABLE only during (well, very early during
that) the gimplification of TARGET_EXPR_INITIAL, so we didn't emit e.g. on
pr80032.C or stack2.C tests any clobbers for the slots and thus stack slot
reuse wasn't performed.
So that we don't regress those tests, this patch gimplifies
TARGET_EXPR_INITIAL as before, but doesn't emit it directly into pre_p,
emits it into a temporary sequence. Then emits the CLOBBER cleanup
into pre_p, then asan poisoning if needed, then appends the
TARGET_EXPR_INITIAL temporary sequence and finally adds TARGET_EXPR_CLEANUP
gimple_push_cleanup. The earlier a GIMPLE_WCE appears in the sequence, the
outer try/finally or try/catch it is.
So, with this patch the part of the testcase in gimple dump cited above
looks instead like:
try
{
D.2499.e = B::qux (&h); [return slot optimization]
D.2516 = 1;
try
{
try
{
B::B (&D.2498, &h);
_2 = baz (&D.2498);
D.2499.f = _2;
D.2516 = 0;
try
{
bar (&D.2499);
}
finally
{
C::~C (&D.2499);
}
}
finally
{
D.2498 = {CLOBBER(eol)};
}
}
catch
{
if (D.2516 != 0) goto <D.2517>; else goto <D.2518>;
<D.2517>:
A::~A (&D.2499.e);
goto <D.2519>;
<D.2518>:
<D.2519>:
}
}
finally
{
D.2499 = {CLOBBER(eol)};
}
2022-03-17 Jakub Jelinek <jakub@redhat.com>
PR middle-end/103984
* gimplify.cc (gimplify_target_expr): Gimplify type sizes and
TARGET_EXPR_INITIAL into a temporary sequence, then push clobbers
and asan unpoisioning, then append the temporary sequence and
finally the TARGET_EXPR_CLEANUP clobbers.
* g++.dg/opt/pr103984.C: New test.
This is Christophe Lyon's fix to PR c/98198, an ICE-on-invalid-code
regression affecting mainline, and a suitable testcase.
Tested on x86_64-pc-linux-gnu with make bootstrap and make -k check
with no new failures. Ok for mainline?
2022-03-16 Christophe Lyon <christophe.lyon@arm.com>
Roger Sayle <roger@nextmovesoftware.com>
gcc/c-family/ChangeLog
PR c/98198
* c-attribs.cc (decl_or_type_attrs): Add error_mark_node check.
gcc/testsuite/ChangeLog
PR c/98198
* gcc.dg/pr98198.c: New test case.
This simple i386 patch unblocks a more significant change. The testcase
gcc.target/i386/sse2-pr94680.c isn't quite testing what's intended, and
alas the fix for PR target/94680 doesn't (yet) handle V2DF mode.
For the first test from sse2-pr94680.c, below
v2df foo_v2df (v2df x) {
return __builtin_shuffle (x, (v2df) { 0, 0 }, (v2di) { 0, 2 });
}
GCC on x86_64-pc-linux-gnu with -O2 currently generates:
movhpd .LC0(%rip), %xmm0
ret
.LC0:
.long 0
.long 0
which passes the test as it contains a mov insn and no xor.
Alas reading a zero from the constant pool isn't quite the
desired implementation. With this patch we now generate:
movq %xmm0, %xmm0
ret
The same code as we generate for V2DI, and add a stricter
test case. This implementation generalizes the sse2_movq128
to V2DI and V2DF modes using a VI8F_128 mode iterator and
renames it *sse2_movq128_<mode>. A new define_expand is
introduced for sse2_movq128 so that the exisiting builtin
interface (CODE_FOR_sse2_movq128) remains the same.
2022-03-16 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/94680
* config/i386/sse.md (sse2_movq128): New define_expand to
preserve previous named instruction.
(*sse2_movq128_<mode>): Renamed from sse2_movq128, and
generalized to VI8F_128 (both V2DI and V2DF).
gcc/testsuite/ChangeLog
PR target/94680
* gcc.target/i386/sse2-pr94680-2.c: New stricter V2DF test case.
The new std::from_chars implementation means that those symbols are now
defined on Solaris 11.3, which lacks uselocale. They were not present in
gcc-11, but the linker script gives them the GLIBCXX_3.4.29 symbol
version because that is the version where they appeared for systems with
uselocale.
This makes the version for those symbols depend on whether uselocale is
available or not, so that they get version GLIBCXX_3.4.30 on targets
where they weren't defined in gcc-11.
In order to avoid needing separate ABI baseline files for Solaris 11.3
and 11.4, the ABI checker program now treats the floating-point
std::from_chars overloads as undesignated if they are not found in the
baseline symbols file. This means they can be left out of the SOlaris
baseline without causing the check-abi target to fail.
libstdc++-v3/ChangeLog:
PR libstdc++/103407
* config/abi/pre/gnu.ver: Make version for std::from_chars
depend on HAVE_USELOCALE macro.
* testsuite/util/testsuite_abi.cc (compare_symbols): Treat
std::from_chars for floating-point types as undesignated if
not found in the baseline symbols file.
Avoid generating execution paths for warnings that are ultimately
rejected due to -Wno-analyzer-* flags.
This improves the test case from taking at least several minutes
(before I killed it) to taking under a second.
This doesn't fix the slowdown seen in PR analyzer/104955 with large
numbers of warnings when the warnings are still enabled.
gcc/analyzer/ChangeLog:
PR analyzer/104955
* diagnostic-manager.cc (get_emission_location): New.
(diagnostic_manager::diagnostic_manager): Initialize
m_num_disabled_diagnostics.
(diagnostic_manager::add_diagnostic): Reject diagnostics that
will eventually be rejected due to being disabled.
(diagnostic_manager::emit_saved_diagnostics): Log the number
of disabled diagnostics.
(diagnostic_manager::emit_saved_diagnostic): Split out logic for
determining emission location to get_emission_location.
* diagnostic-manager.h
(diagnostic_manager::m_num_disabled_diagnostics): New field.
* engine.cc (stale_jmp_buf::get_controlling_option): New.
(stale_jmp_buf::emit): Use it.
* pending-diagnostic.h
(pending_diagnostic::get_controlling_option): New vfunc.
* region-model.cc
(poisoned_value_diagnostic::get_controlling_option): New.
(poisoned_value_diagnostic::emit): Use it.
(shift_count_negative_diagnostic::get_controlling_option): New.
(shift_count_negative_diagnostic::emit): Use it.
(shift_count_overflow_diagnostic::get_controlling_option): New.
(shift_count_overflow_diagnostic::emit): Use it.
(dump_path_diagnostic::get_controlling_option): New.
(dump_path_diagnostic::emit): Use it.
(write_to_const_diagnostic::get_controlling_option): New.
(write_to_const_diagnostic::emit): Use it.
(write_to_string_literal_diagnostic::get_controlling_option): New.
(write_to_string_literal_diagnostic::emit): Use it.
* sm-file.cc (double_fclose::get_controlling_option): New.
(double_fclose::emit): Use it.
(file_leak::get_controlling_option): New.
(file_leak::emit): Use it.
* sm-malloc.cc (mismatching_deallocation::get_controlling_option):
New.
(mismatching_deallocation::emit): Use it.
(double_free::get_controlling_option): New.
(double_free::emit): Use it.
(possible_null_deref::get_controlling_option): New.
(possible_null_deref::emit): Use it.
(possible_null_arg::get_controlling_option): New.
(possible_null_arg::emit): Use it.
(null_deref::get_controlling_option): New.
(null_deref::emit): Use it.
(null_arg::get_controlling_option): New.
(null_arg::emit): Use it.
(use_after_free::get_controlling_option): New.
(use_after_free::emit): Use it.
(malloc_leak::get_controlling_option): New.
(malloc_leak::emit): Use it.
(free_of_non_heap::get_controlling_option): New.
(free_of_non_heap::emit): Use it.
* sm-pattern-test.cc (pattern_match::get_controlling_option): New.
(pattern_match::emit): Use it.
* sm-sensitive.cc
(exposure_through_output_file::get_controlling_option): New.
(exposure_through_output_file::emit): Use it.
* sm-signal.cc (signal_unsafe_call::get_controlling_option): New.
(signal_unsafe_call::emit): Use it.
* sm-taint.cc (tainted_array_index::get_controlling_option): New.
(tainted_array_index::emit): Use it.
(tainted_offset::get_controlling_option): New.
(tainted_offset::emit): Use it.
(tainted_size::get_controlling_option): New.
(tainted_size::emit): Use it.
(tainted_divisor::get_controlling_option): New.
(tainted_divisor::emit): Use it.
(tainted_allocation_size::get_controlling_option): New.
(tainted_allocation_size::emit): Use it.
gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/many-disabled-diagnostics.c: New test.
* gcc.dg/plugin/analyzer_gil_plugin.c
(gil_diagnostic::get_controlling_option): New.
(double_save_thread::emit): Use it.
(fncall_without_gil::emit): Likewise.
(pyobject_usage_without_gil::emit): Likewise.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
This adjusts the declarations in <charconv> to match when the definition
is present. This solves the issue that std::from_chars is present on
Solaris 11.3 (using fast_float) but was not declared in the header
(because the declarations were guarded by _GLIBCXX_HAVE_USELOCALE).
Additionally, do not define __cpp_lib_to_chars unless both from_chars
and to_chars are supported (which is only true for IEEE float and
double). We might still provide from_chars (via strtold) but if to_chars
isn't provided, we shouldn't define the feature test macro.
Finally, this simplifies some of the preprocessor checks in the bodies
of std::from_chars in src/c++17/floating_from_chars.cc and hoists the
repeated code for the strtod version into a new function template.
N.B. the long double overload of std::from_chars will always be defined
if the float and double overloads are defined. We can always use one of
strtold or fast_float's binary64 routines (although the latter might
produce errors for some long double values if they are not representable
as binary64).
libstdc++-v3/ChangeLog:
* include/std/charconv (__cpp_lib_to_chars): Only define when
both from_chars and to_chars are supported for floating-point
types.
(from_chars, to_chars): Adjust preprocessor conditions guarding
declarations.
* include/std/version (__cpp_lib_to_chars): Adjust condition to
match <charconv> definition.
* src/c++17/floating_from_chars.cc (from_chars_strtod): New
function template.
(from_chars): Simplify preprocessor checks and use
from_chars_strtod when appropriate.
Assign the result of fold_convert to offset. Also make the useless
conversion check lighter since the two way check is not needed here.
gcc/ChangeLog:
PR tree-optimization/104941
* tree-object-size.cc (size_for_offset): Make useless conversion
check lighter and assign result of fold_convert to OFFSET.
gcc/testsuite/ChangeLog:
PR tree-optimization/104941
* gcc.dg/builtin-dynamic-object-size-0.c (S1, S2): New structs.
(test_alloc_nested_structs, g): New functions.
(main): Call test_alloc_nested_structs.
Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>
This patch fixes a small bug in the omp_set_num_teams implementation.
libgomp/ChangeLog:
* fortran.c (omp_set_num_teams_8_): Call omp_set_num_teams instead of
omp_set_max_active_levels.
* testsuite/libgomp.fortran/icv-8.f90: New test.
Push target("general-regs-only") in <x86gprintrin.h> if x87 is enabled.
gcc/
PR target/104890
* config/i386/x86gprintrin.h: Also check _SOFT_FLOAT before
pushing target("general-regs-only").
gcc/testsuite/
PR target/104890
* gcc.target/i386/pr104890.c: New test.
We just expand `zk`, `zkn` and `zks` before, but need version for
combine them back.
gcc/ChangeLog:
* common/config/riscv/riscv-common.cc (riscv_ext_version_table):
Add version info for zk, zks and zkn.
The crypto extension have several shorthand extensions that don't consist of any extra instructions.
Take zk for example, while the extension would imply zkn, zkr, zkt.
The 3 extensions should also combine back into zk to maintain the canonical order in isa strings.
This patch addresses the above.
And if the other extension has the same situation, you can add them in riscv_combine_info[]
gcc/ChangeLog:
* common/config/riscv/riscv-common.cc
(riscv_combine_info): New.
(riscv_subset_list::handle_combine_ext): Combine back into zk to
maintain the canonical order in isa strings.
(riscv_subset_list::parse): Ditto.
* config/riscv/riscv-subset.h (handle_combine_ext): New.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/predef-17.c: New test.
The following re-orders the newly added code sinking pass before
the last phiopt pass which performs hoisting of adjacent loads
with the intent to enable if-conversion on those.
I've added the aarch64 specific testcase from the PR.
2022-03-16 Richard Biener <rguenther@suse.de>
PR tree-optimization/102008
* passes.def: Move the added code sinking pass before the
preceeding phiopt pass.
* gcc.target/aarch64/pr102008.c: New testcase.
As a minor followup to r12-7656-gffe9c0a0d3564a, this condenses the
handling of ambiguity and access w.r.t. the value of 'protect' so that
the logic is more clear.
gcc/cp/ChangeLog:
* search.cc (lookup_member): Simplify by handling all values
of protect together in the ambiguous case. Don't modify protect.
A well-formed call to std::move/forward is equivalent to a cast, but the
former being a function call means the compiler generates debug info,
which persists even after the call gets inlined, for an operation that's
never interesting to debug.
This patch addresses this problem by folding calls to std::move/forward
and other cast-like functions into simple casts as part of the frontend's
general expression folding routine. This behavior is controlled by a
new flag -ffold-simple-inlines, and otherwise by -fno-inline, so that
users can enable this folding with -O0 (which implies -fno-inline).
After this patch with -O2 and a non-checking compiler, debug info size
for some testcases from range-v3 and cmcstl2 decreases by as much as ~10%
and overall compile time and memory usage decreases by ~2%.
PR c++/96780
gcc/ChangeLog:
* doc/invoke.texi (C++ Dialect Options): Document
-ffold-simple-inlines.
gcc/c-family/ChangeLog:
* c.opt: Add -ffold-simple-inlines.
gcc/cp/ChangeLog:
* cp-gimplify.cc (cp_fold) <case CALL_EXPR>: Fold calls to
std::move/forward and other cast-like functions into simple
casts.
gcc/testsuite/ChangeLog:
* g++.dg/opt/pr96780.C: New test.
Retain the sizetype alloc_object_size to guarantee the assertion in
size_for_offset and to avoid adding a conversion there. nop conversions
are eliminated at the end anyway in dynamic object size computation.
gcc/ChangeLog:
PR tree-optimization/104942
* tree-object-size.cc (alloc_object_size): Remove STRIP_NOPS.
gcc/testsuite/ChangeLog:
PR tree-optimization/104942
* gcc.dg/builtin-dynamic-object-size-0.c (alloc_func_long,
test_builtin_malloc_long): New functions.
(main): Use it.
Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>
We unshare all RTL created during expansion, but when
aarch64_load_symref_appropriately is called after expansion like in the
following testcases, we use imm in both HIGH and LO_SUM operands.
If imm is some RTL that shouldn't be shared like a non-sharable CONST,
we get at least with --enable-checking=rtl a checking ICE, otherwise might
just get silently wrong code.
The following patch fixes that by copying it if it can't be shared.
2022-03-16 Jakub Jelinek <jakub@redhat.com>
PR target/104910
* config/aarch64/aarch64.cc (aarch64_load_symref_appropriately): Copy
imm rtx.
* gcc.dg/pr104910.c: New test.
This patch improves the implementation of single_use as used in code
generated from match.pd for patterns using :s. The current implementation
contains the logic "has_zero_uses (t) || has_single_use (t)" which
performs a loop over the uses to first check if there are zero non-debug
uses [which is rare], then another loop over these uses to check if there
is exactly one non-debug use. This can be better implemented using a
single loop.
This function is currently inlined over 800 times in gimple-match.cc,
whose .o on x86_64-pc-linux-gnu is now up to 30 Mbytes, so speeding up
and shrinking this function should help offset the growth in match.pd
for GCC 12.
I've also done an analysis of the stage3 sizes of gimple-match.o on
x86_64-pc-linux-gnu, which I believe is dominated by debug information,
the .o file is 30MB in stage3, but only 4.8M in stage2. Before my
proposed patch gimple-match.o is 31385160 bytes. The patch as proposed
yesterday (using a single loop in single_use) reduces that to 31105040
bytes, saving 280120 bytes. The suggestion to remove the "inline"
keyword saves only 56 more bytes, but annotating ATTRIBUTE_PURE on a
function prototype was curiously effective, saving 1888 bytes.
before: 31385160
after: 31105040 saved 280120
-inline: 31104984 saved 56
+pure: 31103096 saved 1888
2022-03-16 Roger Sayle <roger@nextmovesoftware.com>
Richard Biener <rguenther@suse.de>
gcc/ChangeLog
* gimple-match-head.cc (single_use): Implement inline using a
single loop.
Tweak the constant folding of X CMP X in when X can't be a NaN.
2022-03-16 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* match.pd (X CMP X -> true): Test tree_expr_maybe_nan_p
instead of HONOR_NANS.
(X LTGT X -> false): Enable if X is not tree_expr_maybe_nan_p, as
this can't trap/signal.
It's an orthogonal concern why these diagnostics do appear at all for
non-offloaded OpenACC constructs (where they're not relevant at all); PR90115.
Depending on how 'assert' is implemented, it may cause temporaries to be
created, and/or may lower into 'COND_EXPR's, and
'gcc/gimplify.cc:gimplify_cond_expr' uses 'create_tmp_var (type, "iftmp")'.
Fix-up for commit 11b8286a83
"[OpenACC privatization] Largely extend diagnostics and
corresponding testsuite coverage [PR90115]".
PR testsuite/102841
libgomp/
* testsuite/libgomp.oacc-c-c++-common/host_data-7.c: Adjust.
__builtin_ia32_blendvpd is defined under sse4.1 and gimple folded
to ((v2di) c) < 0 ? b : a where vec_cmpv2di is under sse4.2 w/o which
it's veclowered to scalar operations and not combined back in rtl.
gcc/ChangeLog:
PR target/104946
* config/i386/i386-builtin.def (BDESC): Add
CODE_FOR_sse4_1_blendvpd for IX86_BUILTIN_BLENDVPD.
* config/i386/i386.cc (ix86_gimple_fold_builtin): Don't fold
__builtin_ia32_blendvpd w/o sse4.2
gcc/testsuite/ChangeLog:
* gcc.target/i386/sse4_1-blendvpd-1.c: New test.
Testing cc1 on pr93032-mztools-unsigned-char.c
Benchmark #1: (without patch)
Time (mean ± σ): 338.8 ms ± 13.6 ms [User: 323.2 ms, System: 14.2 ms]
Range (min … max): 326.7 ms … 363.1 ms 10 runs
Benchmark #2: (with patch)
Time (mean ± σ): 332.3 ms ± 12.8 ms [User: 316.6 ms, System: 14.3 ms]
Range (min … max): 322.5 ms … 357.4 ms 10 runs
Summary
./cc1.new ran 1.02 ± 0.06 times faster than ./cc1.old
gcc/analyzer/ChangeLog:
* store.cc (store::store): Presize m_cluster_map.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
The mma_disassemble_output_operand predicate is too lenient on the types
of addresses it will accept, leading to combine creating invalid address
that eventually lead to ICEs in LRA. The solution is to restrict the
addresses to indirect, indexed or those valid for quad memory accesses.
2022-03-15 Peter Bergner <bergner@linux.ibm.com>
gcc/
PR target/104923
* config/rs6000/predicates.md (mma_disassemble_output_operand): Restrict
acceptable MEM addresses.
gcc/testsuite/
PR target/104923
* gcc.target/powerpc/pr104923.c: New test.
When a lookup is ambiguous, lookup_member still attempts to check
access of the first member found before diagnosing the ambiguity and
propagating the error, and this may cause us to issue an extraneous
access error as in the testcase below (for B1::foo).
This patch fixes this by swapping the order of the ambiguity and access
checks within lookup_member. In passing, since the only thing that could
go wrong during lookup_field_r is ambiguity, we might as well hardcode
that in lookup_member and get rid of lookup_field_info::errstr.
PR c++/103177
gcc/cp/ChangeLog:
* search.cc (lookup_field_info::errstr): Remove this data
member.
(lookup_field_r): Don't set errstr.
(lookup_member): Check ambiguity before checking access.
Simplify accordingly after errstr removal. Exit early upon
error or empty result.
gcc/testsuite/ChangeLog:
* g++.dg/lookup/ambig6.C: New test.
While checking if all targets honor -Wno-psabi for ABI related warnings
or messages, I found that almost all do, except for riscv.
In the testsuite when we want to ignore ABI related messages we
typically use -Wno-psabi -w, but it would be nice to get rid of those
-w uses eventually.
The following allows silencing those warnings with -Wno-psabi rather than
just -w even on riscv.
2022-03-15 Jakub Jelinek <jakub@redhat.com>
PR target/91229
* config/riscv/riscv.cc (riscv_pass_aggregate_in_fpr_pair_p,
riscv_pass_aggregate_in_fpr_and_gpr_p): Pass OPT_Wpsabi instead of 0
to warning calls.
Regardless of the outcome of the general-regs-only stuff in x86gprintrin.h,
apparently general-regs-only is much bigger hammer than no-sse, and e.g.
using 387 instructions in the unwinder isn't a big deal, it never needs
to realign the stack because of it.
So, the following patch uses no-sse (and adds no-mmx to it, even when not
strictly needed).
2022-03-15 Jakub Jelinek <jakub@redhat.com>
PR target/104890
* config/i386/i386.h (LIBGCC2_UNWIND_ATTRIBUTE): Use no-mmx,no-sse
instead of general-regs-only.
This patch resolves PR tree-optimization/101895 a missed optimization
regression, by adding a costant folding simplification to match.pd to
simplify the transform "mult; vec_perm; plus" into "vec_perm; mult; plus"
with the aim that keeping the multiplication and addition next to each
other allows them to be recognized as fused-multiply-add on suitable
targets. This transformation requires a tweak to match.pd's
vec_same_elem_p predicate to handle CONSTRUCTOR_EXPRs using the same
SSA_NAME_DEF_STMT idiom used for constructors elsewhere in match.pd.
The net effect is that the following code example:
void foo(float * __restrict__ a, float b, float *c) {
a[0] = c[0]*b + a[0];
a[1] = c[2]*b + a[1];
a[2] = c[1]*b + a[2];
a[3] = c[3]*b + a[3];
}
when compiled on x86_64-pc-linux-gnu with -O2 -march=cascadelake
currently generates:
vbroadcastss %xmm0, %xmm0
vmulps (%rsi), %xmm0, %xmm0
vpermilps $216, %xmm0, %xmm0
vaddps (%rdi), %xmm0, %xmm0
vmovups %xmm0, (%rdi)
ret
but with this patch now generates the improved:
vpermilps $216, (%rsi), %xmm1
vbroadcastss %xmm0, %xmm0
vfmadd213ps (%rdi), %xmm0, %xmm1
vmovups %xmm1, (%rdi)
ret
2022-03-15 Roger Sayle <roger@nextmovesoftware.com>
Marc Glisse <marc.glisse@inria.fr>
Richard Biener <rguenther@suse.de>
gcc/ChangeLog
PR tree-optimization/101895
* match.pd (vec_same_elem_p): Handle CONSTRUCTOR_EXPR def.
(plus (vec_perm (mult ...) ...) ...): New reordering simplification.
gcc/testsuite/ChangeLog
PR tree-optimization/101895
* gcc.target/i386/pr101895.c: New test case.
We ICE on the following testcase, because we tentatively parse it multiple
times and the erroneous attribute syntax results in
cp_parser_skip_to_end_of_statement, which when seeing CPP_PRAGMA (can be
any deferred one, OpenMP/OpenACC/ivdep etc.) it calls
cp_parser_skip_to_pragma_eol, which calls cp_lexer_purge_tokens_after.
That call purges all the tokens from CPP_PRAGMA until CPP_PRAGMA_EOL,
excluding the initial CPP_PRAGMA though (but including the final
CPP_PRAGMA_EOL). This means the second time we parse this, we see
CPP_PRAGMA with no tokens after it from the pragma, most importantly
not the CPP_PRAGMA_EOL, so either if it is the last pragma in the TU,
we ICE, or if there are other pragmas we treat everything in between
as a pragma.
I've tried various things, including making the CPP_PRAGMA token
itself also purged, or changing the cp_parser_skip_to_end_of_statement
(and cp_parser_skip_to_end_of_block_or_statement) to call it with
NULL instead of token, so that this purging isn't done there,
but each patch resulted in lots of regressions.
But removing the purging altogether surprisingly doesn't regress anything,
and I think it is the right thing, if we e.g. parse tentatively, why can't
we parse the pragma multiple times or at least skip over it?
2022-03-15 Jakub Jelinek <jakub@redhat.com>
PR c++/104623
* parser.cc (cp_parser_skip_to_pragma_eol): Don't purge any tokens.
* g++.dg/gomp/pr104623.C: New test.
find_if_case_{1,2} implicitly assumes conditional jumps and rewrites them,
so if they have extra side-effects or are say asm goto, things don't work
well, either the side-effects are lost or we could ICE.
In particular, the testcase below on s390x has there a doloop instruction
that decrements a register in addition to testing it for non-zero and
conditionally jumping based on that.
The following patch fixes that by punting for !onlyjump_p case, i.e.
if there are side-effects in the jump instruction or it isn't a plain PC
setter.
Also, it assumes BB_END (test_bb) will be always non-NULL, because basic
blocks with 2 non-abnormal successor edges should always have some instruction
at the end that determines which edge to take.
2022-03-15 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/104814
* ifcvt.cc (find_if_case_1, find_if_case_2): Punt if test_bb doesn't
end with onlyjump_p. Assume BB_END (test_bb) is always non-NULL.
* gcc.c-torture/execute/pr104814.c: New test.
This change avoids -Wdangling-pointer for by-value arguments transformed
into by-transparent-reference.
Resolves:
PR middle-end/104436 - spurious -Wdangling-pointer assigning local address to a class passed by value
gcc/ChangeLog:
PR middle-end/104436
* gimple-ssa-warn-access.cc (pass_waccess::check_dangling_stores):
Check for warning suppression. Avoid by-value arguments transformed
into by-transparent-reference.
gcc/testsuite/ChangeLog:
PR middle-end/104436
* c-c++-common/Wdangling-pointer-8.c: New test.
* g++.dg/warn/Wdangling-pointer-5.C: New test.
My sincere apologies for the breakage, but alas handling SImode in the
recently added "xorl;movb -> movzbl" peephole2 turns out to be slightly
more complicated that just using SWI48 as a mode iterator. I'd failed
to check the machine description carefully, but the *zero_extend<mode>si2
define_insn is conditionally defined, based on x86 target tuning using
TARGET_ZERO_EXTEND_WITH_AND, and therefore unavailable on 486 and pentium
unless optimizing the code for size. It turns out that the libitm testsuite
specifies -m486 with make check RUNTESTFLAGS="--target_board='unix{-m32}'"
and therefore encounters/catches oversight.
Fixed by adding the appropriate conditions to the new peephole2 patterns.
2022-03-14 Roger Sayle <roger@nextmovesoftware.com>
Uroš Bizjak <ubizjak@gmail.com>
gcc/ChangeLog
* config/i386/i386.md (peephole2 xorl;movb -> movzbl): Disable
transformation when *zero_extend<mode>si2 is not available.
gcc/testsuite/ChangeLog
* gcc.target/i386/pr98335.c: Skip this test if tuning for i486
or pentium, and not optimizing for size.
