ian [Fri, 14 Sep 2018 16:55:41 +0000 (16:55 +0000)]
compiler: don't use address of temporary for deferred delete
This CL corrects the handling of a deferred delete in a loop, to not
use a temporary whose value will, at deferred execution time, wind up
being the last value in the loop.
The test for this is TestDeferDeleteSlow in the 1.11 runtime package.
This test started failing because some of the functions in the combine dump that it scans uses a different pattern to match the same instruction: insv_regsi rather than aarch64_bfi.
The code generation is still the same.
The patch changes the scan to look for the actual instruction we want in the assembly.
This fixes the test.
Committing to trunk as obvious.
* gcc.target/aarch64/combine_bfi_1.c: Scan for bfi instruction
rather than pattern name in combine dump.
* reg-stack.c: Include regs.h.
(replace_reg): Assert that mode is MODE_FLOAT or MODE_COMPLEX_FLOAT.
(emit_pop_insn): Default pop insn mode to the reg_raw_mode of
FIRST_STACK_REG, not DFmode.
(emit_swap_insn): Default swap insn mode to the reg_raw_mode of
FIRST_STACK_REG, not XFmode. Explicitly construct swap RTX.
(change stack): Default register mode to the reg_raw_mode of
FIRST_STACK_REG, not DFmode.
* config/i386/i386.md (*swap<mode>): Remove insn pattern.
(*swapxf): Rename from swapxf.
[tree-ssa-mathopts] PR tree-optimization/87259: Call execute_cse_reciprocals_1 before trying optimize_recip_sqrt
PR tree-optimization/87259
PR lto/87283
(pass_cse_reciprocals::execute): Run optimize_recip_sqrt after
execute_cse_reciprocals_1 has tried transforming.
PR tree-optimization/87259
* gcc.dg/pr87259.c: New test.
[libgfortran] Fix uninitialized variable use in fallback_access
I've been tracking down a bug in a Fortran program on a newlib target and it boils down to fallback_access doing something bad.
The unconditional calls to close cause havoc when open doesn't get called due to the short-circuiting in the if-statement above
because the fd is uninitialised. In my environment GCC ends up calling close on file descriptor 0, thus trying to close stdin.
This patch tightens up the calling so that close is called only when the corresponding open call succeeded.
With this my runtime failure disappears.
Bootstrapped and tested on aarch64-none-linux-gnu.
Though that doesn't exercise this call I hope it's an obviously correct change.
* io/unix.c (fallback_access): Avoid calling close on
uninitialized file descriptor.
law [Fri, 14 Sep 2018 03:34:19 +0000 (03:34 +0000)]
* builtins.h (c_srlen): Add argument.
* builtins.c (warn_string_no_nul): New function.
(c_strlen): Add argument and use it. Update recursive calls.
Pass DECL argument to string_constant to get info on non
terminated strings. Update *NONSTR as needed.
(fold_builtin_strlen): Add argument to calls to c_strlen.
Warn for unterminated arrays.
(warn_string_no_null): Add prototype.
* expr.c (string_constant): Update arguments. Update recursive
calls appropriately. Detect missing NUL terminator and outermost
declaration its missing in.
Improve checks for arrays with nonzero lower bound or elements
that are not a single byte. Simplify offset computation.
Simplify checks for non-NUL terminated strings.
* gimple-fold.c (get_range_strlen): Add argument to c_strlen call.
* gimple-ssa-sprintf.c (get_string_length): Remove unnecessary code.
law [Fri, 14 Sep 2018 01:49:38 +0000 (01:49 +0000)]
* builtins.c (c_strlen): Handle not zero terminated STRING_CSTs
correctly.
* fold-const.c (c_getstr): Fix function comment. Remove unused third
argument. Fix range checks.
* fold-const.h (c_getstr): Adjust protoype.
* gimple-fold.c (gimple_fold_builtin_memory_op): Avoid folding when
string is constant but contains no NUL byte.
ian [Thu, 13 Sep 2018 22:25:58 +0000 (22:25 +0000)]
compiler, runtime: call gcWriteBarrier instead of writebarrierptr
In 1.11 writebarrierptr is going away, so change the compiler to call
gcWriteBarrier instead. We weren't using gcWriteBarrier before;
adjust the implementation to use the putFast method.
This revealed a problem in the kickoff function. When using cgo,
kickoff can be called on the g0 of an m allocated by newExtraM. In
that case the m will generally have a p, but systemstack may be called
by wbBufFlush as part of flushing the write barrier buffer. At that
point the buffer is full, so we can not do a write barrier. So adjust
the existing code in kickoff so that in the case where we are g0,
don't do any write barrier at all.
ian [Thu, 13 Sep 2018 22:06:16 +0000 (22:06 +0000)]
runtime: correct counters in sweep
In the sweep code we can sometimes see incorrect counts when
conservative stack scanning causes us to grey an object that we
earlier decided could be freed. We already ignored this check, but
adjust this case to maintain correct span counts when it happens.
This gives us slightly more correct numbers in MemStats, and helps
avoid a rare failure in TestReadMemStats.
Also fix the free index, and cope with finding a full span when
allocating a new one.
law [Thu, 13 Sep 2018 21:43:16 +0000 (21:43 +0000)]
* varasm.c (compare_constant): Compare type size of STRING_CSTs.
(get_constant_size): Don't make STRING_CSTs larger than they are.
(check_string_literal): New check function for STRING_CSTs.
(output_constant): Use it.
ian [Thu, 13 Sep 2018 21:32:24 +0000 (21:32 +0000)]
compiler, runtime: open code select
This is the gofrontend version of https://golang.org/cl/37933,
https://golang.org/cl/37934, and https://golang.org/cl/37935.
Open code the initialization of select cases.
This is a step toward updating libgo to the 1.11 release.
ian [Thu, 13 Sep 2018 17:30:00 +0000 (17:30 +0000)]
runtime: avoid write barriers with traceback info
Unlike the gc runtime, libgo stores traceback information in location
structs, which contain strings. Therefore, copying location structs
around appears to require write barriers, although in fact write
barriers are never important because the strings are never allocated
in Go memory. They come from libbacktrace.
Some of the generated write barriers come at times when write barriers
are not permitted. For example, exitsyscall, marked
nowritebarrierrec, calls exitsyscallfast which calls traceGoSysExit
which calls traceEvent which calls traceStackID which calls
trace.stackTab.put which copies location values into memory allocated
by tab.newStack. This write barrier can be invoked when there is no
p, causing a crash.
This change fixes the problem by ensuring that location values are
copied around in the tracing code with no write barriers.
This was found by fixing the compiler to fully implement
//go:nowritebarrierrec; CL to follow.
ian [Thu, 13 Sep 2018 16:44:43 +0000 (16:44 +0000)]
libgo: build roots index to speed up bulkBarrierPreWrite
To reduce the amount of time spent in write barrier processing
(specifically runtime.bulkBarrierPreWrite), add support for building a
'GC roots index', basically a sorted list of all roots, so as to
allow more efficient lookups of gcdata structures for globals. The
previous implementation worked on the raw (unsorted) roots list
itself, which did not scale well.
PR tree-optimization/87263
* tree-ssa-sccvn.c (visit_phi): Revert some earlier changes.
(struct unwind_state): Add max_rpo field.
(do_rpo_vn): Allow up-to-date loop state to be used when not iterating.
Compute max_rpo, the max RPO number a block can be backwards reached
from. Re-write non-iterating mode to a RPO ordered worklist approach,
separating it from the iterating mode.
PR bootstrap/87134
* tree-ssa-sccvn.c (vn_nary_op_insert_into): Fix assert.
(vn_nary_op_insert_pieces_predicated): Do not write useless
valid_dominated_by_p entry outside of the allocated storage.
jakub [Thu, 13 Sep 2018 07:43:29 +0000 (07:43 +0000)]
* configure.ac: Only append
" : (reconfigured) $TOPLEVEL_CONFIGURE_ARGUMENTS" to
gcc_config_arguments if it was never reconfigured or last reconfigure
was with different arguments.
* configure: Regenerated.
jakub [Thu, 13 Sep 2018 07:36:50 +0000 (07:36 +0000)]
PR tree-optimization/87287
* fold-const.c (fold_binary_loc) <case EQ_EXPR>: Move signed modulo
X % C == 0 to X % (unsigned) C == 0 optimization to ...
* match.pd (X % C == 0): ... here. New optimization.
ian [Thu, 13 Sep 2018 00:45:55 +0000 (00:45 +0000)]
compiler: omit a couple of write barriers
Omit a write barrier for
s = s[0:]
for a slice s. In this case the pointer is not changing and no write
barrier is required.
Omit a write barrier for
s = append(s, v)
in the case where len(s) < cap(s) (and similarly when appending more
values). When the slice has enough capacity the pointer is not
changing and no write barrier is required.
These changes are required to avoid write barriers in the method
randomOrder.reset in the runtime package. That method is called from
procresize, at a point where we do not want to allocate memory.
Otherwise that method can use a write barrier, allocate memory, and
break TestReadMemStats.
C++: special-case single non-viable candidate (more PR c++/85110)
I broke out the "no viable candidates" case in build_new_method_call_1
into a subroutine, and added special-case handling for when there's
a single non-viable candidate where there's an argument conversion
error. I turned the error-handling from convert_for_assignment into
a subroutine, calling it from this new special-case.
This converts:
demo.cc: In function 'int test_4(int, const char*, float)':
demo.cc:5:44: error: no matching function for call to 's4::member_1(int&, const char*&, float&)'
5 | return s4::member_1 (first, second, third);
| ^
demo.cc:1:24: note: candidate: 'static int s4::member_1(int, const char**, float)'
1 | struct s4 { static int member_1 (int one, const char **two, float three); };
| ^~~~~~~~
demo.cc:1:56: note: no known conversion for argument 2 from 'const char*' to 'const char**'
1 | struct s4 { static int member_1 (int one, const char **two, float three); };
| ~~~~~~~~~~~~~^~~
thus highlighting the problematic argument at the callsite (and its type).
gcc/cp/ChangeLog:
PR c++/85110
* call.c (struct conversion_info): Add "loc" field.
(arg_conversion_rejection): Add "loc" param, using it to
initialize the new field.
(bad_arg_conversion_rejection): Likewise.
(explicit_conversion_rejection): Initialize the new field to
UNKNOWN_LOCATION.
(template_conversion_rejection): Likewise.
(add_function_candidate): Pass on the argument location to the new
param of arg_conversion_rejection.
(add_conv_candidate): Likewise.
(build_builtin_candidate): Likewise.
(build_user_type_conversion_1): Likewise.
(single_z_candidate): New function.
(maybe_get_bad_conversion_for_unmatched_call): New function.
(complain_about_bad_argument): New function, based on part of
convert_for_assignment.
(build_new_method_call_1): Split out handling of the "no viable
candidates" case into...
(complain_about_no_candidates_for_method_call): ...this new
function, and use the new functions above to special-case the
handling of a single non-viable candidate due to a bad argument.
* cp-tree.h (complain_about_bad_argument): New decl.
* typeck.c (convert_for_assignment): Split out one error-handling
case into complain_about_bad_argument.
gcc/testsuite/ChangeLog:
PR c++/85110
* g++.dg/cpp0x/explicit4.C: Update expected output to reflect
special-casing of diagnostic for a single non-viable candidate due
to a bad argument.
* g++.dg/diagnostic/param-type-mismatch-2.C: Likewise.
Add test coverage for an unmatched overloaded operator.
* g++.dg/expr/pmf-1.C: Likewise.
* g++.old-deja/g++.bugs/900330_02.C: Likewise.
* g++.old-deja/g++.jason/conversion11.C: Likewise.
* g++.old-deja/g++.law/arg11.C: Likewise.
* g++.old-deja/g++.law/arm9.C: Likewise.
* g++.old-deja/g++.robertl/eb131.C: Likewise.
PR fortran/87284
* trans-expr.c (gfc_trans_class_init_assign): Access to
to array elements of the dynamic type requires that the array
reference be added to the class expression and not the _data
component, unlike scalar expressions.
2018-09-12 Paul Thomas <pault@gcc.gnu.org>
PR fortran/87284
* gfortran.dg/allocate_with_mold_2.f90: New test.
2018-09-09 Cesar Philippidis <cesar@codesourcery.com>
Julian Brown <julian@codesourcery.com>
PR middle-end/86336
gcc/cp/
* semantics.c (finish_omp_clauses): Treat C++ references the same in
OpenACC as OpenMP.
gcc/
* gimplify.c (gimplify_scan_omp_clauses): Set
target_firstprivatize_array_bases in OpenACC parallel and kernels
region contexts. Remove GOMP_MAP_FIRSTPRIVATE_REFERENCE clauses from
OpenACC data regions.
PR tree-optimization/87280
* tree-ssa-sccvn.c (process_bb): Handle the case of executable
edge but unreachable target.
(do_rpo_vn): For conservatively handling a PHI only mark
the backedge executable but not the block reachable.
Fortran STOP and ERROR STOP use a different function to print the "STOP" string
and the message string. On GCN this results in out-of-order output, such as
"<msg>ERROR STOP ".
This patch fixes the problem by making estr_write use the proper Fortran write,
not C printf, so both parts are now output the same way. This also ensures
that both parts are output to STDERR (not that that means anything on GCN).
S/390: Use proper rounding mode for DFP to BFD conversions
According to IEEE 754 2008 4.3 'Rounding-direction attributes' the
rounding mode of the target format needs to be used. By not setting
the value so far we have always used the DFP rounding mode.
gcc/ChangeLog:
2018-09-12 Andreas Krebbel <krebbel@linux.ibm.com>
* config/s390/s390.md (PFPO_RND_MODE_DFP, PFPO_RND_MODE_BFP): New
constants.
("trunc<BFP:mode><DFP_ALL:mode>2")
("trunc<DFP_ALL:mode><BFP:mode>2")
("extend<BFP:mode><DFP_ALL:mode>2")
("extend<DFP_ALL:mode><BFP:mode>2"): Set proper rounding mode
according to the target operand type.
gcc/testsuite/ChangeLog:
2018-09-12 Andreas Krebbel <krebbel@linux.ibm.com>
* gcc.target/s390/dfp_to_bfp_rounding.c: New test.
jakub [Wed, 12 Sep 2018 09:25:07 +0000 (09:25 +0000)]
PR tree-optimization/86844
* gimple-ssa-store-merging.c
(imm_store_chain_info::coalesce_immediate): For overlapping stores, if
there are any overlapping stores in between them, make sure they are
also coalesced or we give up completely.
* gcc.c-torture/execute/pr86844.c: New test.
* gcc.dg/store_merging_22.c: New test.
* gcc.dg/store_merging_23.c: New test.
jakub [Wed, 12 Sep 2018 09:18:55 +0000 (09:18 +0000)]
PR middle-end/87248
* fold-const.c (fold_ternary_loc) <case COND_EXPR>: Verify also that
BIT_AND_EXPR's second operand is a power of two. Formatting fix.
This patch adds option -gdescribe-dies. It sets the DW_AT_description
attribute of dies that do not get a DW_AT_name attribute, to make it easier
to figure out what the die is describing.
The option exports the names of artificial variables:
...
DIE 0: DW_TAG_variable (0x7fa934dd54b0)
+ DW_AT_description: "D.1922"
DW_AT_type: die -> 0 (0x7fa934dd0d70)
DW_AT_artificial: 1
...
which can be traced back to gimple dumps:
...
char a[0:D.1922] [value-expr: *a.0];
...
Furthermore, it adds names to external references:
...
DIE 0: DW_TAG_subprogram (0x7fa88b9650f0)
+DW_AT_description: "main"
DW_AT_abstract_origin: die -> label: vla_1.c.6719312a + 29 (0x7fa88b965140)
...
and likewise to DW_TAG_call_site_parameter DIEs.
Bootstrapped and reg-tested on x86_64.
2018-09-12 Tom de Vries <tdevries@suse.de>
* common.opt (gdescribe-dies): Add option.
* dwarf2out.c (add_name_and_src_coords_attributes): Add description
attribute for artifical and nameless decls.
(dwarf2out_register_external_die): Add description attribute to
external reference die.
(add_desc_attribute): New functions.
(gen_subprogram_die): Add description attribute to
DW_TAG_call_site_parameter.
* tree-pretty-print.c (print_generic_expr_to_str): New function.
* tree-pretty-print.h (print_generic_expr_to_str): Declare.
* doc/invoke.texi (@item Debugging Options): Add -gdescribe-dies and
-gno-describe-dies.
(@item -gdescribe-dies): Add.
* tree-vrp.c (vrp_shift_undefined_p): Remove.
(extract_range_from_binary_expr_1: Call
wide_int_range_shift_undefined_p instead of vrp_shift_undefined_p.
* wide-int-range.h (wide_int_range_shift_undefined_p): Do not
depend on sign.
PR fortran/87172
* resolve.c (resolve_fl_derived): If a type has the 'use_assoc'
attribute, then it was declared in another module, so there should be
no error that it has not been declared.
2018-09-11 Janus Weil <janus@gcc.gnu.org>
PR fortran/87172
* gfortran.dg/iso_c_binding_only_2.f90: New test case.
PR libstdc++/87278 restore support for std::make_shared<volatile T>()
PR libstdc++/87278
* include/bits/shared_ptr.h (make_shared): Use remove_cv instead of
remove_const.
* testsuite/20_util/shared_ptr/creation/87278.cc: New test.
Implement LWG 2905 changes to constrain unique_ptr constructors
LWG DR 2905 says that is_constructible_v<unique_ptr<P, D>, P, D const &>
should be false when D is not copy constructible. This commit implements
the changes from the DR and simplifies the signatures as per
https://github.com/cplusplus/draft/issues/1530
* include/bits/unique_ptr.h (__uniq_ptr_impl): Add assertions to
check deleter type.
(unique_ptr::unique_ptr(pointer, const deleter_type&)): Add copy
constructible constraint.
(unique_ptr::unique_ptr(pointer, deleter_type&&)): Disable for
deleters of reference type and add move constructible constraint.
(unique_ptr::unique_ptr(pointer, remove_reference_t<deleter_type>&&)):
Disable for deleters of non-reference type. Define as deleted.
(unique_ptr<T[], D>): Likewise.
* testsuite/20_util/unique_ptr/assign/48635_neg.cc: Replace dg-error
directives with unstable line numbers with dg-prune-output.
* testsuite/20_util/unique_ptr/cons/cv_qual_neg.cc: Likewise.
* testsuite/20_util/unique_ptr/cons/lwg2905.cc: New test.
* testsuite/20_util/unique_ptr/specialized_algorithms/swap_cxx17.cc:
Make deleter types invocable.
The plain "sorry" diagnostic only gives the "current" location, which is
typically the last line of the function or translation unit by time we get to
the back end.
GCN uses "sorry" to report unsupported language features, such as static
constructors, so it's useful to have a "sorry_at" variant.
This patch implements "sorry_at" according to the pattern of the other "at"
variants.
2018-09-11 Andrew Stubbs <ams@codesourcery.com>
gcc/
* diagnostic-core.h (sorry_at): New prototype.
* diagnostic.c (sorry_at): New function.
ian [Sun, 9 Sep 2018 13:26:17 +0000 (13:26 +0000)]
compiler: make sure type descriptor initializers go in .rodata
When creating type descriptor variables, the compiler (specifically
Type::make_type_descriptor_var) invokes the immutable_struct() and
immutable_struct_set_init() back end methods, so as to insure that
these items go into the ".rodata" section of the generate object file.
The expression initializers for these variables can contain nested
subexpressions, however, and these subexpressions were not always
being placed into .rodata. This patch changes the backend-gen code for
slice initializers to emit implicit variables into .rodata if those
initializers are part of a type descriptor init.
jason [Sat, 8 Sep 2018 16:00:02 +0000 (16:00 +0000)]
PR c++/86678 - constexpr function with non-constant after return.
In this testcase, the call to f() can never be a constant
expression, but that's not a problem because it isn't always
reached by calls to g. We were wrongly rejecting this because
potential_constant_expression_1 lacked the jump tracking that
cxx_eval_constant_expression has. So this patch adds a simpler
version of that tracking.
* constexpr.c (potential_constant_expression_1): Add jump_target.
(breaks): Check for BREAK_STMT.
(continues): Check for CONTINUE_STMT.
projects / thirdparty / gcc.git / log
