Record when a local pointer variable is set to a value such that
indirecting through the pointer does not require a write barrier. Use
that to eliminate write barriers when indirecting through that local
pointer variable. Only keep this information per-block, so it's not
all that applicable.
This reduces the number of write barriers generated when compiling the
runtime package from 553 to 524.
The point of this is to eliminate a bad write barrier in the bytes
function in runtime/print.go. Mark that function nowritebarrier so
that the problem does not recur.
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/191581
From-SVN: r274890
We used to implement darn with unspecs, not unspec_volatiles, which
means two darn instructions could be CSEd together.
This testcase tests it by adding together four random numbers. If all
is well that means we get four darn instructions, because such a small
loop is unrolled fine at -O2 already. If things go bad, combine will
combine it all to one darn and a shift left by two.
gcc/testsuite/
PR target/91481
* gcc.target/powerpc/darn-3.c: New testcase.
From-SVN: r274889
Building for i686-mingw32 target (with some local changes) produced an
error
error: ISO C++ forbids declaration of 'lstat' with no type [-fpermissive]
in libstdc++-v3/src/filesystem/ops-common.h. This patch adds the
missing return type. OK to commit (trunk and GCC 9 branch)?
Note 1: I haven't run the testsuite with this patch, but it fixes the
build failure I see (in sources with other local changes).
Note 2: I don't know why this hasn't produced build failures for other
people, but the missing return type looks wrong in any case even if
other local changes are for some reason needed for it to produce an
error.
* src/filesystem/ops-common.h [_GLIBCXX_FILESYSTEM_IS_WINDOWS]
(std::filesystem::__gnu_posix::lstat): Add return type.
From-SVN: r274885
* config/mips/mips.md (mips_get_fcsr, *mips_get_fcsr): Use SI
machine mode for unspec_volatile operand.
* gcc.target/mips/get-fcsr-3.c: New test.
From-SVN: r274863
Prevent lines overlapping the right hand margin, this occurred for
references in the "See also:" section, it is prevented by placing
references on their own lines. Two example programs overlapped the
right hand margin: fixed by adjusting the position and of the comments.
Some tables had overlapping colums: the column widths were adjusted
accordingly.
From-SVN: r274857
When we are parsing a source file, the very first token might
be a PRAGMA_GCC_PCH_PREPROCESS. This indicates that we are going
read in a PCH file (named as the value of the pragma). If we don't
see this pragma, then we know that it's OK to release any resources
that the host might have set aside for the PCH file.
This fixes a thinko in the current implementation, in that the decision
to release resources was happening unconditionally right after the first
token is extracted but before it's been checked or acted upon.
This leads to the pch bug (seen on Darwin), because we actually do release
resources - which are subsequently (reasonably) assumed to be available
when reading a PCH file. We then get random crashes or hangs depending
on the interaction between unmmap and malloc.
The bug is present everywhere but doesn't show on (say) Linux, since
the release of PCH resources is a NOP there.
This effects all the c-family front ends, because they all use c_lex_with_flags ()
to implement this.
The solution is to check for the PRAGMA_GCC_PCH_PREPROCESS and only call
c_common_no_more_pch () when that is not the first token.
A secondary effect of the collection is that the name of the PCH file
can be collected during the ggc_pch_read() reset of state. Therefore
we should issue any diagnostic that might name the file before the
collections are triggered.
gcc/c-family/
2019-08-23 Iain Sandoe <iain@sandoe.co.uk>
PR pch/61250
* c-lex.c (c_lex_with_flags): Don't call
c_common_no_more_pch () from here.
gcc/c/
2019-08-23 Iain Sandoe <iain@sandoe.co.uk>
PR pch/61250
* c-parser.c (c_parse_file): Call c_common_no_more_pch ()
after determining that the first token is not
PRAGMA_GCC_PCH_PREPROCESS.
gcc/cp/
2019-08-23 Iain Sandoe <iain@sandoe.co.uk>
PR pch/61250
* parser.c (cp_parser_initial_pragma): Call c_common_no_more_pch ()
after determining that the first token is not
PRAGMA_GCC_PCH_PREPROCESS.
gcc/
2019-08-23 Iain Sandoe <iain@sandoe.co.uk>
PR pch/61250
* ggc-page.c (ggc_pch_read): Read the ggc_pch_ondisk structure
and issue any diagnostics needed before collecting the pre-PCH
state.
From-SVN: r274856
2019-08-23 Martin Liska <mliska@suse.cz>
* lto-wrapper.c (run_gcc): When setting jobserver
set also parallel to 1. This was done so before r273908.
From-SVN: r274849
Documentation of all intrinsics now have their sections in a consistent
order. Empty sections have been removed. Stray words have been removed.
Text in a wrong section has been moved. A missing standard was inserted.
From-SVN: r274848
2019-08-23 Martin Liska <mliska@suse.cz>
* profile.c (instrument_values): Do not set
0 as last argument.
* tree-profile.c (gimple_gen_interval_profiler): Remove
last argument.
(gimple_gen_pow2_profiler): Likewise.
(gimple_gen_topn_values_profiler): Likewise.
(gimple_gen_ic_profiler): Likewise.
(gimple_gen_time_profiler): Likewise.
(gimple_gen_average_profiler): Likewise.
(gimple_gen_ior_profiler): Likewise.
* value-prof.c (dump_histogram_value): Use default
in switch statement instead of HIST_TYPE_MAX.
(stream_in_histogram_value): Likewise.
(gimple_duplicate_stmt_histograms): Do not
use NULL for implicitly set arguments.
(gimple_divmod_values_to_profile): Do not use
reserve+quick_push.
(gimple_indirect_call_to_profile): Likewise.
(gimple_find_values_to_profile): Use implicit
function call arguments.
* value-prof.h (gimple_alloc_histogram_value):
Set default values.
(gimple_gen_interval_profiler): Remove last argument.
(gimple_gen_pow2_profiler): Likewise.
(gimple_gen_topn_values_profiler): Likewise.
(gimple_gen_ic_profiler): Likewise.
(gimple_gen_time_profiler): Likewise.
(gimple_gen_average_profiler): Likewise.
(gimple_gen_ior_profiler): Likewise.
From-SVN: r274844
2019-08-22 Steven G. Kargl <kargl@gcc.gnu.org>
* intrinsic.c (add_subroutines): ERRMSG is INTENT(INOUT) in
co_broadcast, co_max, co_min, co_reduce, and co_sum.
From-SVN: r274838
gcc/c-family/ChangeLog:
PR middle-end/91490
* c-common.c (braced_list_to_string): Add argument and overload.
Handle flexible length arrays and unions.
gcc/testsuite/ChangeLog:
PR middle-end/91490
* c-c++-common/Warray-bounds-7.c: New test.
* gcc.dg/Warray-bounds-39.c: Expect either -Warray-bounds or
-Wstringop-overflow.
* gcc.dg/strlenopt-78.c: New test.
gcc/ChangeLog:
PR middle-end/91490
* builtins.c (c_strlen): Rename argument and introduce new local.
Set no-warning bit on original argument.
* expr.c (string_constant): Pass argument type to fold_ctor_reference.
Fold empty and zero constructors into empty strings.
* gimple-fold.c (fold_nonarray_ctor_reference): Return a STRING_CST
for missing initializers.
* tree.c (build_string_literal): Handle optional argument.
* tree.h (build_string_literal): Add defaulted argument.
* gimple-ssa-warn-restrict.c (maybe_diag_access_bounds): Check
no-warning bit on original expression.
From-SVN: r274837
Every call to darn should deliver a *new* random number; such calls
should not be CSEd together. So they should be unspec_volatile, not
plain unspec.
PR target/91481
* config/rs6000/rs6000.md (unspec): Delete UNSPEC_DARN, UNSPEC_DARN_32,
and UNSPEC_DARN_RAW.
(unspecv): New enumerator values UNSPECV_DARN, UNSPECV_DARN_32, and
UNSPECV_DARN_RAW.
(darn_32): Use an unspec_volatile, and UNSPECV_DARN_32.
(darn_raw): Use an unspec_volatile, and UNSPECV_DARN_RAW.
(darn): Use an unspec_volatile, and UNSPECV_DARN.
From-SVN: r274835
* c-parser.c (c_parser_declaration_or_fndef): Set DECL_ARGUMENTS of a
FUNCTION_DECL to the right value in the presence of nested declarators.
From-SVN: r274828
We currently have a nasty error when trying to use the __crc* intrinsics
with an -mfloat-abi=hard.
That is because the target pragma guarding them uses armv8-a+crc that
does not include fp by default.
So we get errors like:
error: '-mfloat-abi=hard': selected processor lacks an FPU
This patch fixes that by using an FP-enabled arch target pragma to guard
these intrinsics when floating-point is available.
That way both the softfloat and hardfloat variants work.
* config/arm/arm_acle.h: Use arch=armv8-a+crc+simd pragma for CRC32
intrinsics if __ARM_FP.
Use __ARM_FEATURE_CRC32 ifdef guard.
* gcc.target/arm/acle/crc_hf_1.c: New test.
From-SVN: r274827
Remove the remaining Neon adddi3, subdi3 and negdi2 patterns. As a result
adddi3, subdi3 and negdi2 can now always be expanded early irrespectively of
whether Neon is available. Also expand the extenddi patterns at the same
time. Several Neon arch attributes are no longer used and removed.
Code generation is improved in all cases, saving another 400-500 instructions
from the PR77308 testcase (total improvement is over 1700 instructions with
-mcpu=cortex-a57 -O2).
Bootstrap & regress OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57
gcc/
* config/arm/arm.md (neon_for_64bits): Remove.
(avoid_neon_for_64bits): Remove.
(arm_adddi3): Always split early.
(arm_subdi3): Always split early.
(negdi2): Remove Neon expansion.
(split zero_extend): Split before reload.
(split sign_extend): Split before reload.
From-SVN: r274825
Like the logical operations, expand all shifts early rather than only
sometimes. The Neon shift expansions are never emitted (not even with
-fneon-for-64bits), so they are not useful. So all the late expansions
and Neon shift patterns can be removed, and shifts are more optimized
as a result. Since some extend patterns use Neon DImode shifts, remove
the Neon extend variants and related splits.
A simple example now generates the same efficient code after this
patch with -mfpu=neon and -mfpu=vfp (previously just the fact of
having Neon enabled resulted inefficient code for no reason).
unsigned long long f(unsigned long long x, unsigned long long y)
{ return x & (y >> 33); }
Before:
strd r4, r5, [sp, #-8]!
lsr r4, r3, #1
mov r5, #0
and r1, r1, r5
and r0, r0, r4
ldrd r4, r5, [sp]
add sp, sp, #8
bx lr
After:
and r0, r0, r3, lsr #1
mov r1, #0
bx lr
Bootstrap and regress OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57
gcc/
* config/arm/iterators.md (qhs_extenddi_cstr): Update.
(qhs_extenddi_cstr): Likewise.
* config/arm/arm.md (ashldi3): Always expand early.
(ashlsi3): Likewise.
(ashrsi3): Likewise.
(zero_extend<mode>di2): Remove Neon variants.
(extend<mode>di2): Likewise.
* config/arm/neon.md (ashldi3_neon_noclobber): Remove.
(signed_shift_di3_neon): Likewise.
(unsigned_shift_di3_neon): Likewise.
(ashrdi3_neon_imm_noclobber): Likewise.
(lshrdi3_neon_imm_noclobber): Likewise.
(<shift>di3_neon): Likewise.
(split extend): Remove DI extend split patterns.
gcc/testsuite/
* gcc.target/arm/neon-extend-1.c: Remove test.
* gcc.target/arm/neon-extend-2.c: Remove test.
From-SVN: r274824
Cleanup the logical DImode operations since the current implementation is way
too complicated. Thumb-1, Thumb-2, VFP/Neon and iwMMXt all work differently,
resulting in a bewildering number of expansions, patterns and splits across
several md files. All this complexity is counterproductive and results in
inefficient code.
A much simpler approach is to split these operations early in the expander
so that optimizations and register allocation are applied on the 32-bit halves.
Codegeneration is unchanged on Thumb-1 and Arm/Thumb-2 without Neon or iwMMXt
(which already expand these instructions early). With Neon these changes save
~1000 instructions from the PR77308 testcase, mostly by significantly reducing
register pressure and spilling.
Bootstrap OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57
gcc/
* config/arm/arm.md (split and/eor/ior): Remove Neon check.
(split not): Add DImode not splitter.
(anddi3): Remove pattern.
(anddi3_insn): Likewise.
(anddi_zesidi_di): Likewise.
(anddi_sesdi_di): Likewise.
(anddi_notdi_di): Likewise.
(anddi_notzesidi_di): Likewise.
(anddi_notsesidi_di): Likewise.
(iordi3): Likewise.
(iordi3_insn): Likewise.
(iordi_zesidi_di): Likewise.
(iordi_sesidi_di): Likewise.
(xordi3): Likewise.
(xordi3_insn): Likewise.
(xordi_sesidi_di): Likewise.
(xordi_zesidi_di): Likewise.
(one_cmpldi2): Likewise.
(one_cmpldi2_insn): Likewise.
* config/arm/constraints.md: Remove De, Df, Dg constraints.
* config/arm/iwmmxt.md (iwmmxt_iordi3): Remove general register
alternative.
(iwmmxt_xordi3): Likewise.
(iwmmxt_anddi3): Likewise.
* config/arm/neon.md (orndi3_neon): Remove pattern.
(anddi_notdi_di): Likewise.
* config/arm/predicates.md (arm_anddi_operand_neon): Remove.
(arm_iordi_operand_neon): Likewise.
(arm_xordi_operand_neon): Likewise.
* config/arm/thumb2.md(iordi_notdi_di): Remove pattern.
(iordi_notzesidi_di): Likewise.
(iordi_notdi_zesidi): Likewise.
(iordi_notsesidi_di): Likewise.
From-SVN: r274823
The iorsi3_compare0 and iorsi3_compare0_scratch patterns can make use
of the 16-bit thumb orrs instruction if suitable registers are
allocated. This patch adds the alternative to allow this to happen.
* config/arm/arm.md (iorsi3_compare0): Add alternative for 16-bit thumb
insn.
(iorsi3_compare0_scratch): Likewise.
From-SVN: r274822
2019-08-21 Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>
PR target/90724
* config/aarch64/aarch64.c (aarch64_gen_compare_reg_maybe_ze): Force y
in reg if it fails aarch64_plus_operand predicate.
From-SVN: r274805
copy_attributes_to_builtin only handles BUILT_IN_NORMAL, but C++ was
calling it immediately after the:
if (DECL_BUILT_IN_CLASS (newdecl) == BUILT_IN_NORMAL)
block. The corresponding C code calls it inside the block instead.
2019-08-21 Richard Sandiford <richard.sandiford@arm.com>
gcc/cp/
PR c++/91505
* decl.c (duplicate_decls): Call copy_attributes_to_builtin inside
the BUILT_IN_NORMAL block rather than afterward.
gcc/testsuite/
PR c++/91505
* g++.target/i386/crc32-4.C: New test.
From-SVN: r274799
* c-ada-spec.c (dump_ada_function_declaration): Be prepared for broken
function declarations where arguments are missing. Rename variables.
From-SVN: r274794
'read' system call may be interrupted by signal with 'errno' is set to
EINTER. In this case, re-try a few times.
2019-08-21 Vadim Godunko <godunko@adacore.com>
gcc/ada/
* libgnat/g-expect.adb (Expect_Internal): Attempt to read
several times when 'read' returns non-positive.
From-SVN: r274791
