gcc.git
6 years ago[OpenACC] C++ reference mapping
Cesar Philippidis [Wed, 12 Sep 2018 15:21:19 +0000 (08:21 -0700)]
[OpenACC] C++ reference mapping

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.

libgomp/
* testsuite/libgomp.oacc-c++/non-scalar-data.C: Remove XFAIL.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
From-SVN: r264244

6 years agoi386.md (sqrt_extend<mode>xf3_i387): Remove.
Uros Bizjak [Wed, 12 Sep 2018 15:04:41 +0000 (17:04 +0200)]
i386.md (sqrt_extend<mode>xf3_i387): Remove.

* config/i386/i386.md (sqrt_extend<mode>xf3_i387): Remove.
(sqrt<mode>2): Extend operand 1 to XFmode and generate sqrtxf3 insn.

From-SVN: r264243

6 years agore PR tree-optimization/87280 (ICE in set_ssa_val_to, at tree-ssa-sccvn.c:3647)
Richard Biener [Wed, 12 Sep 2018 14:03:21 +0000 (14:03 +0000)]
re PR tree-optimization/87280 (ICE in set_ssa_val_to, at tree-ssa-sccvn.c:3647)

2018-09-12  Richard Biener  <rguenther@suse.de>

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.

* gcc.dg/torture/pr87280.c: New testcase.

From-SVN: r264241

6 years agoFix interleaving of Fortran STOP messages
Kwok Cheung Yeung [Wed, 12 Sep 2018 13:53:34 +0000 (13:53 +0000)]
Fix interleaving of Fortran STOP messages

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).

2018-09-12  Kwok Cheung Yeung  <kcy@codesourcery.com>

libgfortran/
* runtime/minimal.c (estr_write): Define in terms of write.

From-SVN: r264239

6 years agoFix Fortran STOP.
Andrew Stubbs [Wed, 12 Sep 2018 13:53:32 +0000 (13:53 +0000)]
Fix Fortran STOP.

The minimal libgfortran setup was created for NVPTX, but will also be used by
AMD GCN.

This patch simply removes an assumption that NVPTX is the only user.
Specifically, NVPTX exit is broken, but AMD GCN exit works just fine.

2018-09-12  Andrew Stubbs  <ams@codesourcery.com>

libgfortran/
* runtime/minimal.c (exit): Only work around nvptx bugs on nvptx.

From-SVN: r264238

6 years agore PR tree-optimization/87266 (ICE: Segmentation fault (in useless_type_conversion_p...
Richard Biener [Wed, 12 Sep 2018 12:14:27 +0000 (12:14 +0000)]
re PR tree-optimization/87266 (ICE: Segmentation fault (in useless_type_conversion_p, tree_nop_conversion_p, or is_gimple_reg_type))

2018-09-12  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87266
* tree-ssa-sccvn.c (do_rpo_vn): Always iterate to not yet
visited blocks.

* gcc.dg/torture/pr87266-1.c: New testcase.
* gcc.dg/torture/pr87266-2.c: Likewise.
* gcc.dg/torture/pr87266-3.c: Likewise.
* gcc.dg/torture/pr87266-4.c: Likewise.

From-SVN: r264237

6 years agoS/390: Use proper rounding mode for DFP to BFD conversions
Andreas Krebbel [Wed, 12 Sep 2018 10:58:42 +0000 (10:58 +0000)]
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.

From-SVN: r264234

6 years agoAdd quotes for -fconstexpr-depth= in an error message.
Martin Liska [Wed, 12 Sep 2018 10:39:45 +0000 (12:39 +0200)]
Add quotes for -fconstexpr-depth= in an error message.

From-SVN: r264233

6 years agore PR tree-optimization/86844 (wrong code caused by store merging pass)
Jakub Jelinek [Wed, 12 Sep 2018 09:25:07 +0000 (11:25 +0200)]
re PR tree-optimization/86844 (wrong code caused by store merging pass)

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.

Co-Authored-By: Andreas Krebbel <krebbel@linux.ibm.com>
From-SVN: r264232

6 years agore PR middle-end/87248 (Bad code for masked operations involving signed ints)
Jakub Jelinek [Wed, 12 Sep 2018 09:18:55 +0000 (11:18 +0200)]
re PR middle-end/87248 (Bad code for masked operations involving signed ints)

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.

* c-c++-common/torture/pr87248.c: New test.

From-SVN: r264230

6 years ago[debug] Add -gdescribe-dies
Tom de Vries [Wed, 12 Sep 2018 07:27:26 +0000 (07:27 +0000)]
[debug] Add -gdescribe-dies

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.

From-SVN: r264229

6 years agotree-vrp.c (vrp_shift_undefined_p): Remove.
Aldy Hernandez [Wed, 12 Sep 2018 07:04:19 +0000 (07:04 +0000)]
tree-vrp.c (vrp_shift_undefined_p): Remove.

* 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.

From-SVN: r264228

6 years ago* gimple-ssa-warn-alloca.c
Aldy Hernandez [Wed, 12 Sep 2018 06:50:34 +0000 (06:50 +0000)]
* gimple-ssa-warn-alloca.c
(alloca_type_and_limit::alloca_type_and_limit): Initialize limit
field for ALLOCA_BOUND_*_LARGE.

From-SVN: r264227

6 years agoDaily bump.
GCC Administrator [Wed, 12 Sep 2018 00:16:47 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264226

6 years ago* gcc.c (load_specs, execute, run_attempt): Use %qs not '%s'.
Nathan Sidwell [Tue, 11 Sep 2018 21:34:14 +0000 (21:34 +0000)]
* gcc.c (load_specs, execute, run_attempt): Use %qs not '%s'.

From-SVN: r264217

6 years agoreg-stack.c (subst_asm_stack_regs): Call replace_reg also for clobbers.
Uros Bizjak [Tue, 11 Sep 2018 19:36:35 +0000 (21:36 +0200)]
reg-stack.c (subst_asm_stack_regs): Call replace_reg also for clobbers.

* reg-stack.c (subst_asm_stack_regs): Call replace_reg also
for clobbers.  Remove obsolete comment.

From-SVN: r264216

6 years agore PR fortran/87172 (Spurious "Derived type 'c_funptr' at (1) has not been declared...
Janus Weil [Tue, 11 Sep 2018 17:44:04 +0000 (19:44 +0200)]
re PR fortran/87172 (Spurious "Derived type 'c_funptr' at (1) has not been declared" error after r263782)

fix PR 87172

2018-09-11  Janus Weil  <janus@gcc.gnu.org>

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.

From-SVN: r264214

6 years agoi386.md (define_attr "type"): Remove mpxmov, mpxmk, mpxchk, mpxld and mpxst types.
Uros Bizjak [Tue, 11 Sep 2018 16:53:15 +0000 (18:53 +0200)]
i386.md (define_attr "type"): Remove mpxmov, mpxmk, mpxchk, mpxld and mpxst types.

* config/i386/i386.md (define_attr "type"): Remove mpxmov, mpxmk,
mpxchk, mpxld and mpxst types.
(define_attr length_immediate): Remove all processing of mpx types.
(define_attr prefix_0f): Ditto.
(define_attr memory): Ditto.

From-SVN: r264212

6 years agoi386.md (fyl2x_extend<mode>xf3_i387): Remove.
Uros Bizjak [Tue, 11 Sep 2018 16:42:35 +0000 (18:42 +0200)]
i386.md (fyl2x_extend<mode>xf3_i387): Remove.

* config/i386/i386.md (fyl2x_extend<mode>xf3_i387): Remove.
(log<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate logxf3 insn.
(log10<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate log10xf3 insn.
(log2<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate log2xf3 insn.
(fyl2xp1_extend<mode>xf3_i387): Remove.
(log1p<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate log1pxf3 insn.
(fxtract_extend<mode>xf3_i387): Remove.
(logb<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate logbxf3 insn.
(ilogb<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate fxtractxf3_i387 insn.
(significand<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate significandxf3 insn.

From-SVN: r264211

6 years agore PR fortran/87277 (Segfault on using array component of class scalar pointer as...
Paul Thomas [Tue, 11 Sep 2018 15:59:07 +0000 (15:59 +0000)]
re PR fortran/87277 (Segfault on using array component of class scalar pointer as an actual argument)

2018-09-11  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87277
* expr.c (is_subref_array): Add the check of dimensionality for
class, dummy, pointer arrays.

2018-09-11  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/87277
* gfortran.dg/select_type_43.f90: New test.

From-SVN: r264210

6 years ago[driver] Use regular error routines
Nathan Sidwell [Tue, 11 Sep 2018 15:03:05 +0000 (15:03 +0000)]
[driver] Use regular error routines

https://gcc.gnu.org/ml/gcc-patches/2018-09/msg00545.html
* gcc.c (perror_with_name, pfatal_with_name): Delete.
(load_specs): Use fatal_error.
(DELETE_IF_ORDINARY, process_command): Use error.
(execute, run_attempt): Use fatal_error.

* gcc.dg/driver-specs.c: New.

From-SVN: r264209

6 years agoPR libstdc++/87278 restore support for std::make_shared<volatile T>()
Jonathan Wakely [Tue, 11 Sep 2018 11:50:39 +0000 (12:50 +0100)]
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.

From-SVN: r264207

6 years agoImplement LWG 2905 changes to constrain unique_ptr constructors
Jonathan Wakely [Tue, 11 Sep 2018 10:55:49 +0000 (11:55 +0100)]
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.

From-SVN: r264206

6 years agoFix typo in my last ChangeLog entry.
Aldy Hernandez [Tue, 11 Sep 2018 10:36:20 +0000 (10:36 +0000)]
Fix typo in my last ChangeLog entry.

From-SVN: r264205

6 years agoAdd sorry_at diagnostic function.
Andrew Stubbs [Tue, 11 Sep 2018 10:26:57 +0000 (10:26 +0000)]
Add sorry_at diagnostic function.

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.

From-SVN: r264204

6 years agotree-vrp (extract_range_from_binary_expr_1): Treat all divisions by zero as VR_UNDEFINED.
Aldy Hernandez [Tue, 11 Sep 2018 10:19:45 +0000 (10:19 +0000)]
tree-vrp (extract_range_from_binary_expr_1): Treat all divisions by zero as VR_UNDEFINED.

* tree-vrp (extract_range_from_binary_expr_1): Treat all divisions
by zero as VR_UNDEFINED.

From-SVN: r264203

6 years agogcov.exp (verify-intermediate): Add missing close.
Joey Ye [Tue, 11 Sep 2018 08:55:25 +0000 (08:55 +0000)]
gcov.exp (verify-intermediate): Add missing close.

2018-09-11  Joey Ye  <joey.ye@arm.com>

    * lib/gcov.exp (verify-intermediate): Add missing close.

From-SVN: r264202

6 years agore PR fortran/86830 (Contiguous array pointer function result not recognized as conti...
Janus Weil [Tue, 11 Sep 2018 06:33:39 +0000 (08:33 +0200)]
re PR fortran/86830 (Contiguous array pointer function result not recognized as contiguous)

fix PR 86830

2018-09-11  Janus Weil  <janus@gcc.gnu.org>

PR fortran/86830
* expr.c (gfc_is_simply_contiguous): Handle type-bound procedure calls
with non-polymorphic objects.

2018-09-11  Janus Weil  <janus@gcc.gnu.org>

PR fortran/86830
* gfortran.dg/typebound_call_30.f90: New test case.

From-SVN: r264201

6 years agoDaily bump.
GCC Administrator [Tue, 11 Sep 2018 00:16:33 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264200

6 years agore PR fortran/85395 ([F03] private clause contained in derived type acquires spurious...
Janus Weil [Mon, 10 Sep 2018 21:25:33 +0000 (23:25 +0200)]
re PR fortran/85395 ([F03] private clause contained in derived type acquires spurious scope)

fix PR 85395

2018-09-10  Janus Weil  <janus@gcc.gnu.org>

PR fortran/85395
* decl.c (match_binding_attributes): Use correct default accessibility
for procedure pointer components.

2018-09-10  Janus Weil  <janus@gcc.gnu.org>

PR fortran/85395
* gfortran.dg/proc_ptr_comp_52.f90: New test case.

From-SVN: r264196

6 years agoi386.md (<sincos>xf2): Rename from *<sincos>xf2_i387.
Uros Bizjak [Mon, 10 Sep 2018 19:13:49 +0000 (21:13 +0200)]
i386.md (<sincos>xf2): Rename from *<sincos>xf2_i387.

* config/i386/i386.md (<sincos>xf2): Rename from *<sincos>xf2_i387.
(*<sincos>_extend<mode>xf2_i387): Remove insn pattern.
(<sincos>mode2): New expander.
(sincos_extend<mode>xf3_i387): Remove insn pattern.
(sincos -> sin, cos splitters): Remove splitter patterns.
(sincos<mode>3): Change operand 2 predicate to general_operand.
Extend operand 2 to XFmode and generate sincosxf3 insn.
(fptanxf4_i387): Change mode of operands 0 and 3 to SFmode.
Change operand 3 predicate to const1_operand.
(fptan_extend<mode>xf4_i387): Remove insn pattern.
(tanxf2): Update operands in the call to fptanxf4_i387.
(tan<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate tanxf3 insn.
(atan2xf3): Rename from *fpatanxf3_i387.
(fpatan_extend<mode>xf3_i387): Remove insn pattern.
(atan2xf3): Remove expander.
(atan2<mode<3):  Change operand 1 and 2 predicates to general_operand.
Extend operands 1 and 2 to XFmode and generate atan2xf3 insn.
(atan<mode>2): Change operand 1 predicate to general_operand.
Extend operand 1 to XFmode and generate atanxf3 insn.

From-SVN: r264195

6 years agoi386.md (x87/SSE constant load splitter): Use memory_operand instead of nonimmediate_...
Uros Bizjak [Mon, 10 Sep 2018 18:46:11 +0000 (20:46 +0200)]
i386.md (x87/SSE constant load splitter): Use memory_operand instead of nonimmediate_operand for input operand predicate.

* config/i386/i386.md (x87/SSE constant load splitter): Use
memory_operand instead of nonimmediate_operand for input operand
predicate.

From-SVN: r264194

6 years agoclass.c (build_vtbl_ref): Remove.
Marek Polacek [Mon, 10 Sep 2018 14:49:18 +0000 (14:49 +0000)]
class.c (build_vtbl_ref): Remove.

* class.c (build_vtbl_ref): Remove.
(build_vtbl_ref_1): Rename to build_vtbl_ref.
(build_vfn_ref): Call build_vtbl_ref instead of build_vtbl_ref_1.

From-SVN: r264193

6 years agoDaily bump.
GCC Administrator [Mon, 10 Sep 2018 00:16:42 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264189

6 years agoi386.md (float partial SSE register stall splitter): Move splitter near its instructi...
Uros Bizjak [Sun, 9 Sep 2018 21:32:08 +0000 (23:32 +0200)]
i386.md (float partial SSE register stall splitter): Move splitter near its instruction pattern.

* config/i386/i386.md (float partial SSE register stall splitter): Move
splitter near its instruction pattern.
(float_extend partial SSE register stall splitter): Ditto.
(float_truncate partial SSE register stall splitter): Ditto.

From-SVN: r264185

6 years agore PR target/86794 (mmix port needs updating for CVE-2017-5753)
Hans-Peter Nilsson [Sun, 9 Sep 2018 18:13:18 +0000 (18:13 +0000)]
re PR target/86794 (mmix port needs updating for CVE-2017-5753)

PR target/86794
* config/mmix/mmix.c (TARGET_HAVE_SPECULATION_SAFE_VALUE): Redefine
to speculation_safe_value_not_needed.

From-SVN: r264184

6 years agore PR target/85666 (gcc-8.0.1 fails to build mmix target: gcc/libgcc/libgcc2.h:203...
Hans-Peter Nilsson [Sun, 9 Sep 2018 18:12:14 +0000 (18:12 +0000)]
re PR target/85666 (gcc-8.0.1 fails to build mmix target: gcc/libgcc/libgcc2.h:203:20: internal compiler error: in leaf_function_p, at final.c:4488)

PR target/85666
* config/mmix/mmix.c (mmix_assemble_integer): Handle byte-size
non-CONST_INT rtx:es using assemble_integer_with_op ".byte".

From-SVN: r264183

6 years agore PR target/85666 (gcc-8.0.1 fails to build mmix target: gcc/libgcc/libgcc2.h:203...
Hans-Peter Nilsson [Sun, 9 Sep 2018 18:05:48 +0000 (18:05 +0000)]
re PR target/85666 (gcc-8.0.1 fails to build mmix target: gcc/libgcc/libgcc2.h:203:20: internal compiler error: in leaf_function_p, at final.c:4488)

PR target/85666
* config/mmix/mmix.c (MMIX_CFUN_NEEDS_SAVED_EH_RETURN_ADDRESS): Don't
call leaf_function_p, instead use has_hard_reg_initial_val.

From-SVN: r264182

6 years agocompiler: make sure type descriptor initializers go in .rodata
Ian Lance Taylor [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.

    Reviewed-on: https://go-review.googlesource.com/132596

From-SVN: r264181

6 years ago[PATCH] internalize a driver fn
Nathan Sidwell [Sun, 9 Sep 2018 07:42:17 +0000 (07:42 +0000)]
[PATCH] internalize a driver fn

https://gcc.gnu.org/ml/gcc-patches/2018-09/msg00462.html
* gcc.h (pfatal_with_name): Don't declare here.
* gcc.c (pfatal_with_name): Make static.

From-SVN: r264178

6 years agocsky.md (*cskyv2_adddi3, [...]): Add earlyclobber.
Xianmiao Qu [Sun, 9 Sep 2018 04:06:16 +0000 (04:06 +0000)]
csky.md (*cskyv2_adddi3, [...]): Add earlyclobber.

2018-09-09  Xianmiao Qu  <xianmiao_qu@c-sky.com>

* config/csky/csky.md (*cskyv2_adddi3, *cskyv2_subdi3): Add
earlyclobber.

From-SVN: r264177

6 years agoDaily bump.
GCC Administrator [Sun, 9 Sep 2018 00:16:52 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264176

6 years agoPR c++/87150 - wrong ctor with maybe-rvalue semantics.
Marek Polacek [Sat, 8 Sep 2018 17:36:08 +0000 (17:36 +0000)]
PR c++/87150 - wrong ctor with maybe-rvalue semantics.

* call.c (struct conversion): Update commentary.
(standard_conversion): Set rvaluedness_matches_p if LOOKUP_PREFER_RVALUE
for ck_base.

* g++.dg/cpp0x/move-return2.C: New test.

From-SVN: r264172

6 years agoPR c++/86678 - constexpr function with non-constant after return.
Jason Merrill [Sat, 8 Sep 2018 16:00:02 +0000 (12:00 -0400)]
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.

From-SVN: r264171

6 years agocxx-pretty-print.c (cxx_pretty_printer::statement): Handle RANGE_FOR_INIT_STMT.
Marek Polacek [Sat, 8 Sep 2018 14:31:24 +0000 (14:31 +0000)]
cxx-pretty-print.c (cxx_pretty_printer::statement): Handle RANGE_FOR_INIT_STMT.

* cxx-pretty-print.c (cxx_pretty_printer::statement) <case
RANGE_FOR_SMT>: Handle RANGE_FOR_INIT_STMT.

From-SVN: r264170

6 years agoarray-init.c: Add dg-prune-output.
Marek Polacek [Sat, 8 Sep 2018 14:27:25 +0000 (14:27 +0000)]
array-init.c: Add dg-prune-output.

* c-c++-common/array-init.c: Add dg-prune-output.
* g++.dg/cpp0x/lambda/lambda-const-this.C: Add dg-warning.
* g++.dg/cpp0x/lambda/lambda-in-class-neg.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-in-class.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-nested.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-nsdmi1.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-nsdmi4.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-this.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-this17.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-this18.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-this2.C: Likewise.
* g++.dg/cpp0x/lambda/lambda-this8.C: Likewise.
* g++.dg/cpp1y/pr64382.C: Likewise.
* g++.dg/cpp1y/pr77739.C: Likewise.
* g++.dg/cpp1z/lambda-this1.C: Likewise.
* g++.dg/cpp1z/lambda-this2.C: Likewise.
* g++.dg/template/crash84.C: Adjust dg-error.

From-SVN: r264169

6 years agore PR rtl-optimization/85458 (FAIL: gcc.dg/pr83480.c (internal compiler error))
John David Anglin [Sat, 8 Sep 2018 13:20:23 +0000 (13:20 +0000)]
re PR rtl-optimization/85458 (FAIL: gcc.dg/pr83480.c (internal compiler error))

PR rtl-optimization/85458
* sel-sched.c (sel_target_adjust_priority): Allow backend adjust
priority hook to reduce the priority of EXPR.

From-SVN: r264168

6 years agoDaily bump.
GCC Administrator [Sat, 8 Sep 2018 00:16:42 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264167

6 years agoRemove unused init_unsigned_integer function.
Janne Blomqvist [Fri, 7 Sep 2018 18:59:50 +0000 (21:59 +0300)]
Remove unused init_unsigned_integer function.

As pointed out by Bernhard Reutner-Fischer, this function is unused
since the fix for PR 53796 in November 2017.

2018-09-07  Janne Blomqvist  <jb@gcc.gnu.org>

        * runtime/environ.c (init_unsigned_integer): Remove.

From-SVN: r264163

6 years ago* config/i386/i386.md (float<SWI48x:mode><MODEF:mode>2) Enable
Uros Bizjak [Fri, 7 Sep 2018 16:01:04 +0000 (18:01 +0200)]
* config/i386/i386.md (float<SWI48x:mode><MODEF:mode>2) Enable
DImode for x87 on 32bit targets.  Conditionally disable x87 modes
with X87_ENABLE_FLOAT.  Remove preparation code.
(*float<SWI48:mode><MODEF:mode>2): Rename from
*float<SWI48:mode><MODEF:mode>2_mixed.  Handle x87, SSE and mixed
math using "enabled" attribute.
(*floatdi<MODEF:mode>2_i387): Rename from
*float<SWI48x:mode><MODEF:mode>2_i387.  Handle only DImode and
enable for 32bit targets only.
(floatdi<X87MODEF:mode>2_i387_with_xmm pre-reload splitter): New
splitter.
(floatdi<X87MODEF:mode>2_i387_with_xmm): Use register_operand
as operand 1 predicate.  Rewrite as define_insn_and_split.
(floatdi<X87MODEF:mode>2_i387_with_xmm memory input splitter): Remove.

From-SVN: r264160

6 years agodirect-enum-init1.C: Remove "inside" from diagnostic messages.
Marek Polacek [Fri, 7 Sep 2018 15:52:54 +0000 (15:52 +0000)]
direct-enum-init1.C: Remove "inside" from diagnostic messages.

* g++.dg/cpp1z/direct-enum-init1.C: Remove "inside" from diagnostic
messages.

From-SVN: r264159

6 years agoPR c++/87152 - range-based for loops with initializer broken in templates.
Marek Polacek [Fri, 7 Sep 2018 14:12:48 +0000 (14:12 +0000)]
PR c++/87152 - range-based for loops with initializer broken in templates.

* constexpr.c (potential_constant_expression_1) <case RANGE_FOR_STMT>:
Recur into RANGE_FOR_INIT_STMT.
* cp-tree.def: Add RANGE_FOR_INIT_STMT to RANGE_FOR_STMT.
* cp-tree.h (RANGE_FOR_INIT_STMT): Define.
* dump.c (cp_dump_tree) <case RANGE_FOR_STMT>: Also dump
RANGE_FOR_INIT_STMT.
* pt.c (tsubst_expr) <case RANGE_FOR_STMT>: Recur into
RANGE_FOR_INIT_STMT.
* semantics.c (begin_range_for_stmt): Adjust call to build_stmt.
Do put the init statement in RANGE_FOR_INIT_STMT.
(finish_range_for_decl): Pop it for templates.

* g++.dg/cpp2a/range-for11.C: New test.
* g++.dg/cpp2a/range-for12.C: New test.
* g++.dg/cpp2a/range-for13.C: New test.
* g++.dg/cpp2a/range-for14.C: New test.
* g++.dg/cpp2a/range-for15.C: New test.
* g++.dg/cpp2a/range-for16.C: New test.
* g++.dg/cpp2a/range-for17.C: New test.
* g++.dg/cpp2a/range-for18.C: New test.
* g++.dg/parse/error61.C (foo): Adjust dg-error.

From-SVN: r264158

6 years agoFix C++17 expected errors for pointer to member function template args.
Jason Merrill [Fri, 7 Sep 2018 13:12:24 +0000 (09:12 -0400)]
Fix C++17 expected errors for pointer to member function template args.

* g++.dg/template/ptrmem20.C: Adjust expected diagnostic.
* g++.dg/template/ptrmem8.C: Likewise.

From-SVN: r264157

6 years agoDaily bump.
GCC Administrator [Fri, 7 Sep 2018 00:16:44 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264155

6 years agoreg-stack.c (get_true_reg): Reorder to fallthru to FLOAT case.
Uros Bizjak [Thu, 6 Sep 2018 21:39:47 +0000 (23:39 +0200)]
reg-stack.c (get_true_reg): Reorder to fallthru to FLOAT case.

* reg-stack.c (get_true_reg) <case FLOAT_TRUNCATE>: Reorder
to fallthru to FLOAT case.

From-SVN: r264151

6 years agore PR target/86731 (Miscompiles vec_sl at -O3 with -fwrapv on ppc64el)
Will Schmidt [Thu, 6 Sep 2018 19:47:06 +0000 (19:47 +0000)]
re PR target/86731 (Miscompiles vec_sl at -O3 with -fwrapv on ppc64el)

[gcc]

2018-09-06  Will Schmidt  <will_schmidt@vnet.ibm.com>

PR target/86731
* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Update logic
around folding of vec_sl to handle out of range shift values.

[testsuite]

2018-08-15  Will Schmidt  <will_schmidt@vnet.ibm.com>

PR target/86731
* gcc.target/powerpc/pr86731.c: New test.
* gcc.target/powerpc/pr86731-longlong.c: New test.
* gcc.target/powerpc/pr86731-fwrapv.c: New test.
* gcc.target/powerpc/pr86731-fwrapv-longlong.c: New test.
* gcc.target/powerpc/pr86731-nogimplefold.c: New test.
* gcc.target/powerpc/pr86731-nogimplefold-longlong.c: New test.

From-SVN: r264150

6 years agoi386.md (fix_trunc<mode>_fisttp_i387_1): Remove.
Uros Bizjak [Thu, 6 Sep 2018 18:53:01 +0000 (20:53 +0200)]
i386.md (fix_trunc<mode>_fisttp_i387_1): Remove.

* config/i386/i386.md (fix_trunc<mode>_fisttp_i387_1): Remove.
Update callers to gen_fix_trunc<mode>_i387_fisttp
(fix_trunc<mode>_i387_fisttp): Change operand 0 predicate to
nonimmediate_operand.
(fix_trunc<mode>_i387_fisttp_with_temp): Remove insn pattern
and corresponding splitters.
(*fix_trunc<mode>_i387_1): Always expand with fix_trunc<mode>_i387.
(fix_truncdi_i387): Change operand 0 predicate to nonimmediate_operand.
(fix_truncdi_i387_with_temp): Remove insn pattern
and corresponding splitters.
(fix_trunc<mode>_i387): Change operand 0 predicate to
nonimmediate_operand.
(fix_trunc<mode>_i387_with_temp): Remove insn pattern
and corresponding splitters.
(*fistdi2_1): Remove.
(fistdi2): Ditto.
(fistdi2_with_temp): Remove insn pattern and corresponding splitters.
(lrintxfdi2): Remove expander.  Reimplement as define_insn.
(*fist<mode>2_1): Remove.
(fist<mode>2): Ditto.
(fist<mode>2_with_temp): Remove insn pattern and corresponding
splitters.
(lrintxf<mode>2): Remove expander.  Reimplement as define_insn.
(*fist<mode>2_<rounding>_1): Always expand with fist<mode>2_<rounding>.
(fistdi2_<rounding>): Change operand 0 predicate to
nonimmediate_operand.
(fistdi2_<rounding>_with_temp): Remove insn pattern
and corresponding splitters.
(fist<mode>2_<rounding>): Change operand 0 predicate to
nonimmediate_operand.
(fist<mode>2_<rounding>_with_temp): Remove insn pattern
and corresponding splitters.

(*fixuns_trunc<mode>si2_avx512f_zext): Depend on TARGET_SSE_MATH.

From-SVN: r264148

6 years agovarasm.c (output_constructor_regular_field): Check TYPE_SIZE_UNIT of the init value.
Bernd Edlinger [Thu, 6 Sep 2018 15:39:48 +0000 (15:39 +0000)]
varasm.c (output_constructor_regular_field): Check TYPE_SIZE_UNIT of the init value.

* varasm.c (output_constructor_regular_field): Check TYPE_SIZE_UNIT of
the init value.

* c-common.c (complete_flexible_array_elts): New helper function.
* c-common.h (complete_flexible_array_elts): Declare.

* c-decl.c (finish_decl): Call complete_flexible_array_elts.

* decl.c (check_initializer): Call cp_complete_array_type.

From-SVN: r264147

6 years agors6000.c (rs6000_gimple_fold_builtin): Add support for early gimple folding of vec_sp...
Will Schmidt [Thu, 6 Sep 2018 15:34:58 +0000 (15:34 +0000)]
rs6000.c (rs6000_gimple_fold_builtin): Add support for early gimple folding of vec_splat().

[gcc]

2018-09-06  Will Schmidt  <will_schmidt@vnet.ibm.com>

* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add support for
early gimple folding of vec_splat().
* tree-vect-generic.c: Remove static from tree_vec_extract() definition.
* gimple-fold.h: Add an extern define for tree_vec_extract().

From-SVN: r264146

6 years agors6000.c (fold_mergehl_helper): Add types_compatible_p wrappers around TREE_TYPE...
Will Schmidt [Thu, 6 Sep 2018 13:52:14 +0000 (13:52 +0000)]
rs6000.c (fold_mergehl_helper): Add types_compatible_p wrappers around TREE_TYPE comparisons.

[gcc]

2018-08-20  Will Schmidt  <will_schmidt@vnet.ibm.com>

* config/rs6000/rs6000.c (fold_mergehl_helper): Add types_compatible_p
wrappers around TREE_TYPE comparisons.

From-SVN: r264145

6 years agoS/390: Prohibit SYMBOL_REF in UNSPECV_CAS
Ilya Leoshkevich [Thu, 6 Sep 2018 07:38:42 +0000 (07:38 +0000)]
S/390: Prohibit SYMBOL_REF in UNSPECV_CAS

Inhibit constant propagation inlining SYMBOL_REF loads into
UNSPECV_CAS.  Even though reload can later undo it, the resulting
code will be less efficient.

gcc/ChangeLog:

2018-09-06  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/80080
* config/s390/predicates.md: Add nonsym_memory_operand.
* config/s390/s390.c (s390_legitimize_cs_operand): If operand
contains a SYMBOL_REF, load it into an intermediate pseudo.
(s390_emit_compare_and_swap): Legitimize operand.
* config/s390/s390.md: Use the new nonsym_memory_operand
with UNSPECV_CAS patterns.

gcc/testsuite/ChangeLog:

2018-09-06  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/80080
* gcc.target/s390/pr80080-3.c: New test.
* gcc.target/s390/s390.exp: Make sure the new test passes
on all optimization levels.

From-SVN: r264143

6 years agoS/390: Register pass_s390_early_mach statically
Ilya Leoshkevich [Thu, 6 Sep 2018 07:35:35 +0000 (07:35 +0000)]
S/390: Register pass_s390_early_mach statically

The dump file used to come at the end of the sorted dump file list,
because the pass was registered dynamically. This did not reflect the
order in which passes are executed. Static registration fixes this:

* foo4.c.277r.split2
* foo4.c.281r.early_mach
* foo4.c.282r.pro_and_epilogue

gcc/ChangeLog:

2018-09-06  Ilya Leoshkevich  <iii@linux.ibm.com>

PR target/80080
* config/s390/s390-passes.def: New file.
* config/s390/s390-protos.h (class rtl_opt_pass): Add forward
declaration.
(make_pass_s390_early_mach): Add declaration.
* config/s390/s390.c (make_pass_s390_early_mach):
(s390_option_override): Remove dynamic registration.
* config/s390/t-s390: Add s390-passes.def.

From-SVN: r264142

6 years agoS/390: Remove UNSPEC_LTREL_BASE
Ilya Leoshkevich [Thu, 6 Sep 2018 07:31:23 +0000 (07:31 +0000)]
S/390: Remove UNSPEC_LTREL_BASE

It was needed only for g5/g6 machines, which are now gone.

gcc/ChangeLog:

2018-09-06  Ilya Leoshkevich  <iii@linux.ibm.com>

* config/s390/s390.c (s390_decompose_constant_pool_ref):
Remove UNSPEC_LTREL_BASE check.
(annotate_constant_pool_refs): Likewise.
(find_constant_pool_ref): Likewise.
(find_ltrel_base): Removed.
(replace_ltrel_base): Removed.
(s390_mainpool_finish): Remove replace_ltrel_base call.
(s390_chunkify_start): Remove pending LTREL_BASE logic.
(s390_chunkify_finish): Remove replace_ltrel_base call.
* config/s390/s390.md: Remove UNSPEC_LTREL_BASE.

From-SVN: r264141

6 years agoPR c++/86982, -Wreturn-local-addr and std::move and std::forward.
Marek Polacek [Thu, 6 Sep 2018 00:38:13 +0000 (00:38 +0000)]
PR c++/86982, -Wreturn-local-addr and std::move and std::forward.

* typeck.c (maybe_warn_about_returning_address_of_local): Handle calls
to std::move or std::forward.
(is_std_forward_p): New function.

* g++.dg/warn/Wreturn-local-addr-5.C: New test.

From-SVN: r264139

6 years agoDaily bump.
GCC Administrator [Thu, 6 Sep 2018 00:16:49 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264138

6 years agore PR target/86779 (Cris port needs updating for CVE-2017-5753)
Hans-Peter Nilsson [Wed, 5 Sep 2018 23:14:42 +0000 (23:14 +0000)]
re PR target/86779 (Cris port needs updating for CVE-2017-5753)

PR target/86779
* config/cris/cris.c (TARGET_HAVE_SPECULATION_SAFE_VALUE): Redefine
to speculation_safe_value_not_needed.

From-SVN: r264134

6 years ago[nvptx] Basic -misa support for nvptx
Cesar Philippidis [Wed, 5 Sep 2018 22:27:31 +0000 (15:27 -0700)]
[nvptx] Basic -misa support for nvptx

gcc/
* config/nvptx/nvptx-opts.h: New file.
* config/nvptx/nvptx.c (nvptx_file_start): Print the correct .target.
* config/nvptx/nvptx.h: Include "nvptx-opts.h".
(ASM_SPEC): Define.
(TARGET_SM35): New macro.
* config/nvptx/nvptx.md (atomic_fetch_<logic><mode>): Enable with the
correct predicate.
* config/nvptx/nvptx.opt (ptx_isa, sm_30, sm_35): New enum and its
values.
(misa=): New option.
* doc/invoke.texi (Nvidia PTX Options): Document -misa.

gcc/testsuite/
* gcc.target/nvptx/atomic_fetch-1.c: New test.
* gcc.target/nvptx/atomic_fetch-1.c: New test.

Co-Authored-By: Bernd Schmidt <bernds_cb1@t-online.de>
From-SVN: r264133

6 years agoPR c++/87109, wrong overload with ref-qualifiers.
Marek Polacek [Wed, 5 Sep 2018 21:17:47 +0000 (21:17 +0000)]
PR c++/87109, wrong overload with ref-qualifiers.

* call.c (build_user_type_conversion_1): Use NULL instead of 0.  Bail
out if performing the maybe-rvalue overload resolution and a conversion
function is getting called.

* g++.dg/cpp0x/ref-qual19.C: New test.

From-SVN: r264132

6 years agoi386.md (truncdfsf2): Remove expander.
Uros Bizjak [Wed, 5 Sep 2018 18:01:42 +0000 (20:01 +0200)]
i386.md (truncdfsf2): Remove expander.

* config/i386/i386.md (truncdfsf2): Remove expander.
(truncdfsf2_with_temp): Ditto.
(truncxf<mode>2): Ditto.
(*truncdfsf_fast_mixed): Remove insn pattern.
(*truncdfsf_fast_i387): Ditto.
(*truncdfsf_mixed): Ditto.
(*truncdfsf_i387): Ditto.
(*truncdfsf2_i387_1): Ditto.
(*truncxfsf2_mixed): Ditto.
(*truncxfdf2_mixed): Ditto.
(*truncxf<mode>2_i387_noop): Ditto. Update callers
to call gen_truncxf<mode>2 instead.
(*truncxf<mode>2_i387): Remove.
(reg->reg splitters): Remove splitter pattern.
(reg->mem splitters): Ditto.

(truncdfsf2): New insn pattern.
(truncxf<mode>2): Ditto.

From-SVN: r264130

6 years agoPR52665 do not let .ident confuse assembler scan tests
Bernhard Reutner-Fischer [Wed, 5 Sep 2018 15:31:13 +0000 (17:31 +0200)]
PR52665 do not let .ident confuse assembler scan tests

A branch with a name matching scan-assembler pattern triggers
inappropriate FAIL.

E.g. branch fixups-testsuite and
- gcc.target/i386/pr65871-?.c (scan-assembler-not "test")
- gcc.target/i386/pr41442.c (scan-assembler-times "test|cmp" 2)
etc.

This is a recurring problem as can be seen by some -fno-ident additions
by commits from e.g. Michael Meissner over the years: builtins-58.c,
powerpc/pr46728-?.c

The patch below adds -fno-ident if a testcase contains one of
scan-assembler, scan-assembler-not or scan-assembler-times.

Regression tested on x86_64-unknown-linux on a fixups-testsuite branch
where it fixes several false FAILs without regressions.

gcc/testsuite/ChangeLog

2016-06-18  Bernhard Reutner-Fischer  <aldot@gcc.gnu.org>

        PR testsuite/52665
        * lib/gcc-dg.exp (gcc-dg-test-1): Iterate over _required_options.
        * lib/target-supports.exp (scan-assembler_required_options,
        scan-assembler-not_required_options,
        scan-assembler-times_required_options): Add -fno-ident.
        * lib/scanasm.exp (scan-assembler-times): Fix error message.
        * c-c++-common/ident-0a.c: New test.
        * c-c++-common/ident-0b.c: New test.
        * c-c++-common/ident-1a.c: New test.
        * c-c++-common/ident-1b.c: New test.
        * c-c++-common/ident-2a.c: New test.
        * c-c++-common/ident-2b.c: New test.

From-SVN: r264128

6 years ago* libsupc++/cxxabi.h (__cxa_demangle): Clarify doxygen comment.
Jonathan Wakely [Wed, 5 Sep 2018 15:18:16 +0000 (16:18 +0100)]
* libsupc++/cxxabi.h (__cxa_demangle): Clarify doxygen comment.

From-SVN: r264127

6 years agoOptimise sqrt reciprocal multiplications
Kyrylo Tkachov [Wed, 5 Sep 2018 13:39:38 +0000 (13:39 +0000)]
Optimise sqrt reciprocal multiplications

This patch aims to optimise sequences involving uses of 1.0 / sqrt (a) under -freciprocal-math and -funsafe-math-optimizations.
In particular consider:

x = 1.0 / sqrt (a);
r1 = x * x;  // same as 1.0 / a
r2 = a * x; // same as sqrt (a)

If x, r1 and r2 are all used further on in the code, this can be transformed into:
tmp1 = 1.0 / a
tmp2 = sqrt (a)
tmp3 = tmp1 * tmp2
x = tmp3
r1 = tmp1
r2 = tmp2

A bit convoluted, but this saves us one multiplication and, more importantly, the sqrt and division are now independent.
This also allows optimisation of a subset of these expressions.
For example:
x = 1.0 / sqrt (a)
r1 = x * x

can be transformed to r1 = 1.0 / a, eliminating the sqrt if x is not used anywhere else.
And similarly:
x = 1.0 / sqrt (a)
r1 = a * x

can be transformed to sqrt (a) eliminating the division.

For the testcase:
double res, res2, tmp;
void
foo (double a, double b)
{
  tmp = 1.0 / __builtin_sqrt (a);
  res = tmp * tmp;
  res2 = a * tmp;
}

We now generate for aarch64 with -Ofast:
foo:
        fmov    d2, 1.0e+0
        adrp    x2, res2
        fsqrt   d1, d0
        adrp    x1, res
        fdiv    d0, d2, d0
        adrp    x0, tmp
        str     d1, [x2, #:lo12:res2]
        fmul    d1, d1, d0
        str     d0, [x1, #:lo12:res]
        str     d1, [x0, #:lo12:tmp]
        ret

where before it generated:
foo:
        fsqrt   d2, d0
        fmov    d1, 1.0e+0
        adrp    x1, res2
        adrp    x2, tmp
        adrp    x0, res
        fdiv    d1, d1, d2
        fmul    d0, d1, d0
        fmul    d2, d1, d1
        str     d1, [x2, #:lo12:tmp]
        str     d0, [x1, #:lo12:res2]
        str     d2, [x0, #:lo12:res]
        ret

As you can see, the new sequence has one fewer multiply and the fsqrt and fdiv are independent.

* tree-ssa-math-opts.c (is_mult_by): New function.
(is_square_of): Use the above.
(optimize_recip_sqrt): New function.
(pass_cse_reciprocals::execute): Use the above.

* gcc.dg/recip_sqrt_mult_1.c: New test.
* gcc.dg/recip_sqrt_mult_2.c: Likewise.
* gcc.dg/recip_sqrt_mult_3.c: Likewise.
* gcc.dg/recip_sqrt_mult_4.c: Likewise.
* gcc.dg/recip_sqrt_mult_5.c: Likewise.
* g++.dg/recip_sqrt_mult_1.C: Likewise.
* g++.dg/recip_sqrt_mult_2.C: Likewise.

From-SVN: r264126

6 years agore PR bootstrap/87134 (SEGV in cc1 caused by r263875)
Richard Biener [Wed, 5 Sep 2018 11:44:13 +0000 (11:44 +0000)]
re PR bootstrap/87134 (SEGV in cc1 caused by r263875)

2018-09-05  Richard Biener  <rguenther@suse.de>

PR bootstrap/87134
* tree-ssa-sccvn.c (rpo_elim::eliminate_push_avail): Make sure
to zero-init the emplaced vec.

From-SVN: r264125

6 years agoGroup switch cases in switch lowering (PR tree-optimization/87205).
Martin Liska [Wed, 5 Sep 2018 11:28:49 +0000 (13:28 +0200)]
Group switch cases in switch lowering (PR tree-optimization/87205).

2018-09-05  Martin Liska  <mliska@suse.cz>

PR tree-optimization/87205
* tree-switch-conversion.c (pass_lower_switch::execute):
Group cases for switch statements.
2018-09-05  Martin Liska  <mliska@suse.cz>

PR tree-optimization/87205
* gcc.dg/tree-ssa/pr87205-2.c: New test.
* gcc.dg/tree-ssa/pr87205.c: New test.

From-SVN: r264124

6 years agore PR tree-optimization/87217 (ICE in in check_loop_closed_ssa_def, at tree-ssa-loop...
Richard Biener [Wed, 5 Sep 2018 10:33:41 +0000 (10:33 +0000)]
re PR tree-optimization/87217 (ICE in in check_loop_closed_ssa_def, at tree-ssa-loop-manip.c:709 when compiling SPEC2000 starting with r264069)

2018-09-05  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87217
* tree-ssa-sccvn.c (vuse_valueize): New.
(vn_reference_lookup_pieces): Use it.
(vn_reference_lookup): Likewise.

* gfortran.dg/pr87217.f: New testcase.

From-SVN: r264121

6 years ago* io/async.h: Use __gthread_mutex_t, not pthread_mutex_t.
Hans-Peter Nilsson [Wed, 5 Sep 2018 10:28:45 +0000 (10:28 +0000)]
* io/async.h: Use __gthread_mutex_t, not pthread_mutex_t.

From-SVN: r264120

6 years agoPR c++/87137] GCC-8 Fix
Nathan Sidwell [Wed, 5 Sep 2018 10:04:58 +0000 (10:04 +0000)]
PR c++/87137] GCC-8 Fix

https://gcc.gnu.org/ml/gcc-patches/2018-08/msg01966.html
PR c++/87137
* stor-layout.c (place_field): Scan forwards to check last
bitfield when ms_bitfield_placement is in effect.
gcc/testsuite/
* g++.dg/abi/pr87137.C: New.

From-SVN: r264119

6 years ago[PR c++/87185] ICE in prune-lambdas
Pádraig Brady [Wed, 5 Sep 2018 09:59:08 +0000 (09:59 +0000)]
[PR c++/87185] ICE in prune-lambdas

https://gcc.gnu.org/ml/gcc-patches/2018-09/msg00249.html
cp/
PR c++/87185
* lambda.c (prune_lambda_captures): Protect against const_vars.get
returning NULL.
testsuite/
PR c++/87185
* g++.dg/pr87185.C: New.

From-SVN: r264118

6 years agore PR bootstrap/87225 (tree-vect-stmts.c:3748 error: converting to 'bool' from 'std...
Richard Biener [Wed, 5 Sep 2018 09:39:34 +0000 (09:39 +0000)]
re PR bootstrap/87225 (tree-vect-stmts.c:3748 error: converting to 'bool' from 'std::nullptr_t' requires direct-initialization [-fpermissive])

2018-09-05  Richard Biener  <rguenther@suse.de>

PR bootstrap/87225
* tree-vect-stmts.c (vectorizable_simd_clone_call): Fix bogus
return.

From-SVN: r264116

6 years ago[aarch64] Avoid tag collisions for loads falkor
Siddhesh Poyarekar [Wed, 5 Sep 2018 09:37:27 +0000 (09:37 +0000)]
[aarch64] Avoid tag collisions for loads falkor

This is a rewrite of the tag collision avoidance patch that Kugan had
written as a machine reorg pass back in February.

The falkor hardware prefetching system uses a combination of the
source, destination and offset to decide which prefetcher unit to
train with the load.  This is great when loads in a loop are
sequential but sub-optimal if there are unrelated loads in a loop that
tag to the same prefetcher unit.

This pass attempts to rename the desination register of such colliding
loads using routines available in regrename.c so that their tags do
not collide.  This shows some performance gains with mcf and xalancbmk
(~5% each) and will be tweaked further.  The pass is placed near the
fag end of the pass list so that subsequent passes don't inadvertantly
end up undoing the renames.

2018-07-02  Siddhesh Poyarekar  <siddhesh@sourceware.org>
            Kugan Vivekanandarajah  <kugan.vivekanandarajah@linaro.org>

* config/aarch64/falkor-tag-collision-avoidance.c: New file.
* config.gcc (extra_objs): Build it.
* config/aarch64/t-aarch64 (falkor-tag-collision-avoidance.o):
Likewise.
* config/aarch64/aarch64-passes.def
(pass_tag_collision_avoidance): New pass.
* config/aarch64/aarch64.c (qdf24xx_tunings): Add
AARCH64_EXTRA_TUNE_RENAME_LOAD_REGS to tuning_flags.
(aarch64_classify_address): Remove static qualifier.
(aarch64_address_info, aarch64_address_type): Move to...
* config/aarch64/aarch64-protos.h: ... here.
(make_pass_tag_collision_avoidance): New function.
* config/aarch64/aarch64-tuning-flags.def (rename_load_regs):
New tuning flag.

Co-Authored-By: Kugan Vivekanandarajah <kuganv@linaro.org>
From-SVN: r264115

6 years agoFix typo in ChangeLog.
Martin Liska [Wed, 5 Sep 2018 09:09:46 +0000 (09:09 +0000)]
Fix typo in ChangeLog.

From-SVN: r264114

6 years agoFix a scan in test for Darwin target (PR testsuite/87216).
Martin Liska [Wed, 5 Sep 2018 09:08:07 +0000 (11:08 +0200)]
Fix a scan in test for Darwin target (PR testsuite/87216).

2018-09-05  Martin Liska  <mliska@suse.cz>

PR testsuite/87216
* gcc.dg/tree-prof/pr59521-3.c: Update scanned pattern
to support Dawring names.

From-SVN: r264113

6 years agoGCOV: Print one decimal place in human readable mode.
Martin Liska [Wed, 5 Sep 2018 08:08:45 +0000 (10:08 +0200)]
GCOV: Print one decimal place in human readable mode.

2018-09-05  Martin Liska  <mliska@suse.cz>

* doc/gcov.texi: Update documentation of humar
readable mode.
* gcov.c (format_count): Print one decimal place, it provides
more fine number of situations like '1G' vs. '1.4G'.
2018-09-05  Martin Liska  <mliska@suse.cz>

* g++.dg/gcov/loop.C: Update test to support new format.

From-SVN: r264112

6 years agoIgnore properly -mdirect-move (PR target/87164).
Martin Liska [Wed, 5 Sep 2018 08:07:48 +0000 (10:07 +0200)]
Ignore properly -mdirect-move (PR target/87164).

2018-09-05  Martin Liska  <mliska@suse.cz>

PR target/87164
* config/rs6000/rs6000.opt: Mark the option as Deprecated.
* optc-gen.awk: Allow 'Var' for Deprecated options in order
to generate a MASK value.

From-SVN: r264111

6 years agocp-tree.h (treat_lvalue_as_rvalue_p): Declare.
Marek Polacek [Wed, 5 Sep 2018 02:08:19 +0000 (02:08 +0000)]
cp-tree.h (treat_lvalue_as_rvalue_p): Declare.

* cp-tree.h (treat_lvalue_as_rvalue_p): Declare.
* except.c (build_throw): Use it.  Use CP_TYPE_VOLATILE_P.
* typeck.c (treat_lvalue_as_rvalue_p): No longer static.  Add PARM_OK
parameter.
(maybe_warn_pessimizing_move): Adjust treat_lvalue_as_rvalue_p call.
(check_return_expr): Likewise.

From-SVN: r264101

6 years agoDaily bump.
GCC Administrator [Wed, 5 Sep 2018 00:16:36 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264100

6 years agoDWARF: Allow hard frame pointer even if frame pointer isn't used
H.J. Lu [Tue, 4 Sep 2018 20:42:06 +0000 (20:42 +0000)]
DWARF: Allow hard frame pointer even if frame pointer isn't used

r251028

commit cd557ff63f388ad27c376d0a225e74d3594a6f9d
Author: hjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Thu Aug 10 15:29:05 2017 +0000

    i386: Don't use frame pointer without stack access

    When there is no stack access, there is no need to use frame pointer
    even if -fno-omit-frame-pointer is used and caller's frame pointer is
    unchanged.

frame pointer may not be available even if -fno-omit-frame-pointer is
used.  When this happened, arg pointer may be eliminated by hard frame
pointer.  Since hard frame pointer is encoded with DW_OP_fbreg which
uses the DW_AT_frame_base attribute, not hard frame pointer directly,
we should allow hard frame pointer when generating DWARF info even if
frame pointer isn't used.

gcc/

PR debug/86593
* dwarf2out.c (based_loc_descr): Allow hard frame pointer even
if frame pointer isn't used.
(compute_frame_pointer_to_fb_displacement): Likewise.

gcc/testsuite/

PR debug/86593
* g++.dg/pr86593.C: New test.

From-SVN: r264096

6 years agore PR target/87198 (ICE in extract_insn, at recog.c:2304)
Jakub Jelinek [Tue, 4 Sep 2018 17:49:57 +0000 (19:49 +0200)]
re PR target/87198 (ICE in extract_insn, at recog.c:2304)

PR target/87198
* common/config/i386/i386-common.c (OPTION_MASK_ISA_XSAVEOPT_SET,
OPTION_MASK_ISA_XSAVES_SET, OPTION_MASK_ISA_XSAVEC_SET): Use
OPTION_MASK_ISA_XSAVE_SET instead of OPTION_MASK_ISA_XSAVE.
(OPTION_MASK_ISA_XSAVE_UNSET): Add OPTION_MASK_ISA_XSAVES_UNSET
and OPTION_MASK_ISA_XSAVEC_UNSET.

* gcc.target/i386/pr87198.c: New test.

From-SVN: r264088

6 years agogcc: xtensa: fix NAND code in xtensa_expand_atomic
Max Filippov [Tue, 4 Sep 2018 17:43:20 +0000 (17:43 +0000)]
gcc: xtensa: fix NAND code in xtensa_expand_atomic

NAND is ~(a1 & a2), but xtensa_expand_atomic does ~a1 & a2.
That fixes libatomic tests atomic-op-{1,2}.

gcc/
2018-09-04  Max Filippov  <jcmvbkbc@gmail.com>

* config/xtensa/xtensa.c (xtensa_expand_atomic): Reorder AND and
XOR operations in NAND case.

From-SVN: r264087

6 years agoDon't xfail gcc.target/i386/addr-sel-1.c (PR target/86744)
Rainer Orth [Tue, 4 Sep 2018 15:00:41 +0000 (15:00 +0000)]
Don't xfail gcc.target/i386/addr-sel-1.c (PR target/86744)

PR target/86744
* gcc.target/i386/addr-sel-1.c: Don't xfail "b\\+1" scan.

From-SVN: r264086

6 years agowide-int-range.cc (wide_int_range_convert): New.
Aldy Hernandez [Tue, 4 Sep 2018 14:52:41 +0000 (14:52 +0000)]
wide-int-range.cc (wide_int_range_convert): New.

* wide-int-range.cc (wide_int_range_convert): New.
* wide-int-range.h (wide_int_range_convert): New.
* tree-vrp.c (extract_range_from_unary_expr): Abstract wide int
code into wide_int_range_convert.
(extract_range_into_wide_ints): Do not munge anti range constants
into the entire domain.  Just return the range back.

From-SVN: r264085

6 years agogenmatch: put reporting on a cold path
Martin Liska [Tue, 4 Sep 2018 13:59:38 +0000 (15:59 +0200)]
genmatch: put reporting on a cold path

2018-09-04  Martin Liska  <mliska@suse.cz>

* genmatch.c (output_line_directive): Add new argument
fnargs.
(dt_simplify::gen_1): Encapsulate dump within __builtin_expect.

From-SVN: r264084

6 years agoAdd whitespace before warning option added in previous commmit
Jonathan Wakely [Tue, 4 Sep 2018 13:34:47 +0000 (14:34 +0100)]
Add whitespace before warning option added in previous commmit

* doc/invoke.texi (Option Summary): Add whitespace.

From-SVN: r264083

6 years agoAdd -Waligned-new to Option Summary
Jonathan Wakely [Tue, 4 Sep 2018 13:25:42 +0000 (14:25 +0100)]
Add -Waligned-new to Option Summary

* doc/invoke.texi (Option Summary): Add -Waligned-new.

From-SVN: r264080

6 years agore PR tree-optimization/87211 (gcc ICE at O2: in set_ssa_val_to, at tree-ssa-sccvn...
Richard Biener [Tue, 4 Sep 2018 12:04:42 +0000 (12:04 +0000)]
re PR tree-optimization/87211 (gcc ICE at O2: in set_ssa_val_to, at tree-ssa-sccvn.c:3628)

2018-09-04  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87211
* tree-ssa-sccvn.c (visit_phi): When value-numbering to a
backedge value we're supposed to treat as VARYING also number
the PHI to VARYING in case it got a different value-number already.

* gcc.dg/torture/pr87211.c: New testcase.

From-SVN: r264079

6 years agotree-vrp.c (vrp_can_optimize_bit_op): Remove.
Aldy Hernandez [Tue, 4 Sep 2018 11:58:14 +0000 (11:58 +0000)]
tree-vrp.c (vrp_can_optimize_bit_op): Remove.

* tree-vrp.c (vrp_can_optimize_bit_op): Remove.
(extract_range_from_binary_expr_1): Do not call
vrp_can_optimize_bit_op.
* wide-int-range.cc (wide_int_range_can_optimize_bit_op): Make
static.
(wide_int_range_get_mask_and_bounds): New.
(wide_int_range_optimize_bit_op): New.
(wide_int_range_bit_ior): Call wide_int_range_optimize_bit_op.
(wide_int_range_bit_and): Same.
* wide-int-range.h (wide_int_range_can_optimize_bit_op): Remove.
(wide_int_range_optimize_bit_op): New.
(wide_int_range_get_mask_and_bounds): New.

From-SVN: r264078

6 years agore PR tree-optimization/87176 (wrong code at -Os and above on x86-64-linux-gnu)
Richard Biener [Tue, 4 Sep 2018 10:55:46 +0000 (10:55 +0000)]
re PR tree-optimization/87176 (wrong code at -Os and above on x86-64-linux-gnu)

2018-09-04  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87176
* tree-ssa-sccvn.c (visit_phi): Remove redundant allsame
variable.  When value-numbering a virtual PHI node make sure
to not value-number to the backedge value.

* gcc.dg/torture/pr87176.c: New testcase.
* gcc.dg/torture/ssa-fre-1.c: Likewise.

From-SVN: r264077

6 years agoUpdate C Extensions docs for support in latest C++
Jonathan Wakely [Tue, 4 Sep 2018 10:27:14 +0000 (11:27 +0100)]
Update C Extensions docs for support in latest C++

* doc/extend.texi (Long Long, Hex Floats): Document support for
long long and hex floats in more recent versions of ISO C++.

From-SVN: r264076

6 years ago* MAINTAINERS: Update my email address.
Xuepeng Guo [Tue, 4 Sep 2018 03:57:03 +0000 (03:57 +0000)]
* MAINTAINERS: Update my email address.

From-SVN: r264075

6 years agoDaily bump.
GCC Administrator [Tue, 4 Sep 2018 00:16:57 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r264074