This patch started as work to resole Richard's comment on quadratic lookups
in resolve_speculation. While doing it I however noticed multiple problems
in the new speuclative call code which made the patch quite big. In
particular:
1) Before applying speculation we consider only targets with at lest
probability 1/2.
If profile is sane at most two targets can have probability greater or
equal to 1/2. So the new multi-target speculation code got enabled only
in very special scenario when there ae precisely two target with precise
probability 1/2 (which is tested by the single testcase).
As a conseuqence the multiple target logic got minimal test coverage and
this made us to miss several ICEs.
2) Profile updating in profile merging, tree-inline and indirect call
expansion was wrong which led to inconsistent profiles (as already seen
on the testcase).
3) Code responsible to turn speculative call to direct call was broken for
anything with more than one target.
4) There were multiple cases where call_site_hash went out of sync which
eventually leads to an ICE..
5) Some code expects that all speculative call targets forms a sequence in
the callee linked list but there is no code to maintain that invariant
nor a verifier.
Fixing this it became obvious that the current API of speculative_call_info is
not useful because it really builds on fact tht there are precisely three
components (direct call, ref and indirect call) in every speculative call
sequence. I ended up replacing it with iterator API for direct call
(first_speculative_call_target, next_speculative_call_target) and accessors for
the other coponents updating comment in cgraph.h.
Finally I made the work with call site hash more effetive by updating edge
manipulation to keep them in sequence. So first one can be looked up from the
hash and then they can be iterated by callee.
There are other things that can be improved (for example the speculation should
start with most common target first), but I will try to keep that for next
stage1. This patch is mostly about getting rid of ICE and profile corruption
which is a regression from GCC 9.
Honza
gcc/ChangeLog:
2020-01-28 Jan Hubicka <hubicka@ucw.cz>
PR lto/93318
* cgraph.c (cgraph_add_edge_to_call_site_hash): Update call site
hash only when edge is first within the sequence.
(cgraph_edge::set_call_stmt): Update handling of speculative calls.
(symbol_table::create_edge): Do not set target_prob.
(cgraph_edge::remove_caller): Watch for speculative calls when updating
the call site hash.
(cgraph_edge::make_speculative): Drop target_prob parameter.
(cgraph_edge::speculative_call_info): Remove.
(cgraph_edge::first_speculative_call_target): New member function.
(update_call_stmt_hash_for_removing_direct_edge): New function.
(cgraph_edge::resolve_speculation): Rewrite to new API.
(cgraph_edge::speculative_call_for_target): New member function.
(cgraph_edge::make_direct): Rewrite to new API; fix handling of
multiple speculation targets.
(cgraph_edge::redirect_call_stmt_to_callee): Likewise; fix updating
of profile.
(verify_speculative_call): Verify that targets form an interval.
* cgraph.h (cgraph_edge::speculative_call_info): Remove.
(cgraph_edge::first_speculative_call_target): New member function.
(cgraph_edge::next_speculative_call_target): New member function.
(cgraph_edge::speculative_call_target_ref): New member function.
(cgraph_edge;:speculative_call_indirect_edge): New member funtion.
(cgraph_edge): Remove target_prob.
* cgraphclones.c (cgraph_node::set_call_stmt_including_clones):
Fix handling of speculative calls.
* ipa-devirt.c (ipa_devirt): Fix handling of speculative cals.
* ipa-fnsummary.c (analyze_function_body): Likewise.
* ipa-inline.c (speculation_useful_p): Use new speculative call API.
* ipa-profile.c (dump_histogram): Fix formating.
(ipa_profile_generate_summary): Watch for overflows.
(ipa_profile): Do not require probablity to be 1/2; update to new API.
* ipa-prop.c (ipa_make_edge_direct_to_target): Update to new API.
(update_indirect_edges_after_inlining): Update to new API.
* ipa-utils.c (ipa_merge_profiles): Rewrite merging of speculative call
profiles.
* profile-count.h: (profile_probability::adjusted): New.
* tree-inline.c (copy_bb): Update to new speculative call API; fix
updating of profile.
* value-prof.c (gimple_ic_transform): Rename to ...
(dump_ic_profile): ... this one; update dumping.
(stream_in_histogram_value): Fix formating.
(gimple_value_profile_transformations): Update.
gcc/testsuite/ChangeLog:
2020-01-28 Jan Hubicka <hubicka@ucw.cz>
* g++.dg/tree-prof/indir-call-prof.C: Update template.
* gcc.dg/tree-prof/crossmodule-indircall-1.c: Add more targets.
* gcc.dg/tree-prof/crossmodule-indircall-1a.c: Add more targets.
* gcc.dg/tree-prof/indir-call-prof.c: Update template.
When I implemented the [over.match.ref] rule that a reference conversion
function needs to match l/rvalue of the target reference type it changed our
handling of this testcase. It seems to me that our current behavior is what
the standard says, but it doesn't seem desirable, and all the other
compilers have our old behavior. So let's limit the change to non-templates
until there's some clarification from the committee.
PR c++/90546
* call.c (build_user_type_conversion_1): Allow a template conversion
returning an rvalue reference to bind directly to an lvalue.
This patch started as work to resole Richard's comment on quadratic lookups
in resolve_speculation. While doing it I however noticed multiple problems
in the new speuclative call code which made the patch quite big. In
particular:
1) Before applying speculation we consider only targets with at lest
probability 1/2.
If profile is sane at most two targets can have probability greater or
equal to 1/2. So the new multi-target speculation code got enabled only
in very special scenario when there ae precisely two target with precise
probability 1/2 (which is tested by the single testcase).
As a conseuqence the multiple target logic got minimal test coverage and
this made us to miss several ICEs.
2) Profile updating in profile merging, tree-inline and indirect call
expansion was wrong which led to inconsistent profiles (as already seen
on the testcase).
3) Code responsible to turn speculative call to direct call was broken for
anything with more than one target.
4) There were multiple cases where call_site_hash went out of sync which
eventually leads to an ICE..
5) Some code expects that all speculative call targets forms a sequence in
the callee linked list but there is no code to maintain that invariant
nor a verifier.
Fixing this it became obvious that the current API of speculative_call_info is
not useful because it really builds on fact tht there are precisely three
components (direct call, ref and indirect call) in every speculative call
sequence. I ended up replacing it with iterator API for direct call
(first_speculative_call_target, next_speculative_call_target) and accessors for
the other coponents updating comment in cgraph.h.
Finally I made the work with call site hash more effetive by updating edge
manipulation to keep them in sequence. So first one can be looked up from the
hash and then they can be iterated by callee.
There are other things that can be improved (for example the speculation should
start with most common target first), but I will try to keep that for next
stage1. This patch is mostly about getting rid of ICE and profile corruption
which is a regression from GCC 9.
gcc/ChangeLog:
PR lto/93318
* cgraph.c (cgraph_add_edge_to_call_site_hash): Update call site
hash only when edge is first within the sequence.
(cgraph_edge::set_call_stmt): Update handling of speculative calls.
(symbol_table::create_edge): Do not set target_prob.
(cgraph_edge::remove_caller): Watch for speculative calls when updating
the call site hash.
(cgraph_edge::make_speculative): Drop target_prob parameter.
(cgraph_edge::speculative_call_info): Remove.
(cgraph_edge::first_speculative_call_target): New member function.
(update_call_stmt_hash_for_removing_direct_edge): New function.
(cgraph_edge::resolve_speculation): Rewrite to new API.
(cgraph_edge::speculative_call_for_target): New member function.
(cgraph_edge::make_direct): Rewrite to new API; fix handling of
multiple speculation targets.
(cgraph_edge::redirect_call_stmt_to_callee): Likewise; fix updating
of profile.
(verify_speculative_call): Verify that targets form an interval.
* cgraph.h (cgraph_edge::speculative_call_info): Remove.
(cgraph_edge::first_speculative_call_target): New member function.
(cgraph_edge::next_speculative_call_target): New member function.
(cgraph_edge::speculative_call_target_ref): New member function.
(cgraph_edge;:speculative_call_indirect_edge): New member funtion.
(cgraph_edge): Remove target_prob.
* cgraphclones.c (cgraph_node::set_call_stmt_including_clones):
Fix handling of speculative calls.
* ipa-devirt.c (ipa_devirt): Fix handling of speculative cals.
* ipa-fnsummary.c (analyze_function_body): Likewise.
* ipa-inline.c (speculation_useful_p): Use new speculative call API.
* ipa-profile.c (dump_histogram): Fix formating.
(ipa_profile_generate_summary): Watch for overflows.
(ipa_profile): Do not require probablity to be 1/2; update to new API.
* ipa-prop.c (ipa_make_edge_direct_to_target): Update to new API.
(update_indirect_edges_after_inlining): Update to new API.
* ipa-utils.c (ipa_merge_profiles): Rewrite merging of speculative call
profiles.
* profile-count.h: (profile_probability::adjusted): New.
* tree-inline.c (copy_bb): Update to new speculative call API; fix
updating of profile.
* value-prof.c (gimple_ic_transform): Rename to ...
(dump_ic_profile): ... this one; update dumping.
(stream_in_histogram_value): Fix formating.
(gimple_value_profile_transformations): Update.
gcc/testsuite/ChangeLog:
* g++.dg/tree-prof/indir-call-prof.C: Update template.
* gcc.dg/tree-prof/crossmodule-indircall-1.c: Add more targets.
* gcc.dg/tree-prof/crossmodule-indircall-1a.c: Add more targets.
* gcc.dg/tree-prof/indir-call-prof.c: Update template.
Changed in v2:
- rename from warning_with_metadata_at to warning_meta
- fix test plugins
While C++ can have overloads, xgettext can't deal with overloads that have
different argument positions, leading to two failures in "make gcc.pot":
emit_diagnostic_valist used incompatibly as both
--keyword=emit_diagnostic_valist:4
--flag=emit_diagnostic_valist:4:gcc-internal-format and
--keyword=emit_diagnostic_valist:5
--flag=emit_diagnostic_valist:5:gcc-internal-format
warning_at used incompatibly as both
--keyword=warning_at:3 --flag=warning_at:3:gcc-internal-format and
--keyword=warning_at:4 --flag=warning_at:4:gcc-internal-format
The emit_diagnostic_valist overload isn't used anywhere (I think it's
a leftover from an earlier iteration of the analyzer patch kit).
The warning_at overload is used throughout the analyzer but nowhere else.
Ideally I'd like to consolidate this argument with something
constructable in various ways:
- from a metadata and an int or
- from an int (or, better an "enum opt_code"),
so that the overload happens when implicitly choosing the ctor, but
that feels like stage 1 material.
In the meantime, fix xgettext by deleting the unused overload and
renaming the used one.
gcc/analyzer/ChangeLog:
* region-model.cc (poisoned_value_diagnostic::emit): Update for
renaming of warning_at overload to warning_meta.
* sm-file.cc (file_leak::emit): Likewise.
* sm-malloc.cc (double_free::emit): Likewise.
(possible_null_deref::emit): Likewise.
(possible_null_arg::emit): Likewise.
(null_deref::emit): Likewise.
(null_arg::emit): Likewise.
(use_after_free::emit): Likewise.
(malloc_leak::emit): Likewise.
(free_of_non_heap::emit): Likewise.
* sm-sensitive.cc (exposure_through_output_file::emit): Likewise.
* sm-signal.cc (signal_unsafe_call::emit): Likewise.
* sm-taint.cc (tainted_array_index::emit): Likewise.
gcc/ChangeLog:
* diagnostic-core.h (warning_at): Rename overload to...
(warning_meta): ...this.
(emit_diagnostic_valist): Delete decl of overload taking
diagnostic_metadata.
* diagnostic.c (emit_diagnostic_valist): Likewise for defn.
(warning_at): Rename overload taking diagnostic_metadata to...
(warning_meta): ...this.
gcc/testsuite/ChangeLog:
* gcc.dg/plugin/diagnostic_plugin_test_metadata.c: Update for
renaming of warning_at overload to warning_meta.
* gcc.dg/plugin/diagnostic_plugin_test_paths.c: Likewise.
PR fortran/93461
* trans.h: Increase GFC_MAX_MANGLED_SYMBOL_LEN to
GFC_MAX_SYMBOL_LEN*3+5 to allow for inclusion of submodule name,
plus the "." between module and submodule names.
* gfortran.dg/pr93461.f90: New test.
Increase length of char variables "parent1" and "parent2" in
set_syms_host_assoc() to allow them to hold concatenated module +
submodule names.
PR fortran/93473
* parse.c: Increase length of char variables to allow them to hold
a concatenated module + submodule name.
* gfortran.dg/pr93473.f90: New test.
The clever hack of '#define __has_include __has_include' breaks -dD
and -fdirectives-only, because that emits definitions. This turns
__has_include into a proper builtin macro. Thus it's never emitted
via -dD, and because use outside of directive processing is undefined,
we can just expand it anywhere.
PR preprocessor/93452
* internal.h (struct spec_nodes): Drop n__has_include{,_next}.
* directives.c (lex_macro_node): Don't check __has_include redef.
* expr.c (eval_token): Drop __has_include eval.
(parse_has_include): Move to ...
* macro.c (builtin_has_include): ... here.
(_cpp_builtin_macro_text): Eval __has_include{,_next}.
* include/cpplib.h (enum cpp_builtin_type): Add BT_HAS_INCLUDE{,_NEXT}.
* init.c (builtin_array): Add them.
(cpp_init_builtins): Drop __has_include{,_next} init here ...
* pch.c (cpp_read_state): ... and here.
* traditional.c (enum ls): Drop has_include states ...
(_cpp_scan_out_logical_line): ... and here.
We just need to handle the exception specification like other properties of
a function typedef.
PR c++/90731
* decl.c (grokdeclarator): Propagate eh spec from typedef.
gcc/fortran/
* gfortran.h (gfc_symbol): Add comp_mark bitfield.
* openmp.c (resolve_omp_clauses): Disallow mixed component and
full-derived-type accesses to the same variable within a single
directive.
libgomp/
* testsuite/libgomp.oacc-fortran/deep-copy-2.f90: Remove test from here.
* testsuite/libgomp.oacc-fortran/deep-copy-3.f90: Don't use mixed
component/non-component variable refs in a single directive.
* testsuite/libgomp.oacc-fortran/classtypes-1.f95: Likewise.
gcc/testsuite/
* gfortran.dg/goacc/deep-copy-2.f90: Move test here (from libgomp
testsuite). Make a compilation test, and expect rejection of mixed
component/non-component accesses.
* gfortran.dg/goacc/mapping-tests-1.f90: New test.
It's wrong to assume that clock_gettime is unavailable on any *-*-linux*
target that doesn't have glibc 2.17 or later. Use a generic test instead
of using __GLIBC_PREREQ. Only do that test when is_hosted=yes so that we
don't get an error for cross targets without a working linker.
This ensures that C library's clock_gettime will be used on non-glibc
targets, instead of an incorrect syscall to SYS_clock_gettime.
PR libstdc++/93325
* acinclude.m4 (GLIBCXX_ENABLE_LIBSTDCXX_TIME): Use AC_SEARCH_LIBS for
clock_gettime instead of explicit glibc version check.
* configure: Regenerate.
Autopar was doing clique bookkeeping too early when creating destination
functions but then later introducing new cliques via versioning loops.
The following moves the bookkeeping to the actual outlining process.
2020-01-28 Richard Biener <rguenther@suse.de>
PR tree-optimization/93439
* tree-parloops.c (create_loop_fn): Move clique bookkeeping...
* tree-cfg.c (move_sese_region_to_fn): ... here.
(verify_types_in_gimple_reference): Verify used cliques are
tracked.
* gfortran.dg/graphite/pr93439.f90: New testcase.
There are
static void
parse_mtune_ctrl_str (bool dump)
{
if (!ix86_tune_ctrl_string)
return;
parse_mtune_ctrl_str is only called from set_ix86_tune_features, which
is only called from ix86_function_specific_restore and
ix86_option_override_internal. parse_mtune_ctrl_str shouldn't use
ix86_tune_ctrl_string which is defined with global_options. Instead,
opts should be passed to parse_mtune_ctrl_str.
PR target/91399
* config/i386/i386-options.c (set_ix86_tune_features): Add an
argument of a pointer to struct gcc_options and pass it to
parse_mtune_ctrl_str.
(ix86_function_specific_restore): Pass opts to
set_ix86_tune_features.
(ix86_option_override_internal): Likewise.
(parse_mtune_ctrl_str): Add an argument of a pointer to struct
gcc_options and use it for x_ix86_tune_ctrl_string.
This change makes sure that if the driver is invoked with
"-mcode-density" flag, then the assembler will receive it too.
Note Claudiu Zissulescu:
This is an old patch of which I forgot to add the test.
testsuite/
2019-09-03 Sahahb Vahedi <shahab@synopsys.com>
* gcc.target/arc/code-density-flag.c: New test.
In the gcc.target/aarch64/lsl_asr_sbfiz.c part of this PR, we have:
Failed to match this instruction:
(set (reg:SI 95)
(ashift:SI (subreg:SI (sign_extract:DI (subreg:DI (reg:SI 97) 0)
(const_int 3 [0x3])
(const_int 0 [0])) 0)
(const_int 19 [0x13])))
If we perform the natural simplification to:
(set (reg:SI 95)
(ashift:SI (sign_extract:SI (reg:SI 97)
(const_int 3 [0x3])
(const_int 0 [0])) 0)
(const_int 19 [0x13])))
then the pattern matches. And it turns out that we do have a
simplification like that already, but it would only kick in for
extractions from a reg, not a subreg. E.g.:
(set (reg:SI 95)
(ashift:SI (subreg:SI (sign_extract:DI (reg:DI X)
(const_int 3 [0x3])
(const_int 0 [0])) 0)
(const_int 19 [0x13])))
would simplify to:
(set (reg:SI 95)
(ashift:SI (sign_extract:SI (subreg:SI (reg:DI X) 0)
(const_int 3 [0x3])
(const_int 0 [0])) 0)
(const_int 19 [0x13])))
IMO the subreg case is even more obviously a simplification
than the bare reg case, since the net effect is to remove
either one or two subregs, rather than simply change the
position of a subreg/truncation.
However, doing that regressed gcc.dg/tree-ssa/pr64910-2.c
for -m32 on x86_64-linux-gnu, because we could then simplify
a :HI zero_extract to a :QI one. The associated *testqi_ext_3
pattern did already seem to want to handle QImode extractions:
"ix86_match_ccmode (insn, CCNOmode)
&& ((TARGET_64BIT && GET_MODE (operands[2]) == DImode)
|| GET_MODE (operands[2]) == SImode
|| GET_MODE (operands[2]) == HImode
|| GET_MODE (operands[2]) == QImode)
but I'm not sure how often the QI case would trigger in practice,
since the zero_extract mode was restricted to HI and above. I checked
the other x86 patterns and couldn't see any other instances of this.
2020-01-28 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR rtl-optimization/87763
* simplify-rtx.c (simplify_truncation): Extend sign/zero_extract
simplification to handle subregs as well as bare regs.
* config/i386/i386.md (*testqi_ext_3): Match QI extracts too.
gcc.dg/pr56350.c started ICEing for SVE in GCC 10 because we
pattern-matched a division reduction:
a /= 8;
into a signed shift with division semantics:
... = IFN_SDIV_POW2 (..., 3);
whereas the reduction code expected it still to be a gassign.
One fix would be to check for a reduction in the pattern matcher
(but current patterns don't generally do that). Another would be
to fail gracefully for reductions involving calls. Since we can't
vectorise the reduction either way, and probably have a better shot
with the shift form, this patch goes for the "fail gracefully" approach.
2020-01-28 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-loop.c (vectorizable_reduction): Fail gracefully
for reduction chains that (now) include a call.
PR fortran/93464
* openmp.c (gfc_omp_check_optional_argument): Avoid ICE when
DECL_LANG_SPECIFIC and GFC_DESCRIPTOR_TYPE_P but not
GFC_DECL_SAVED_DESCRIPTOR as for local allocatable character vars.
PR fortran/93464
* gfortran.dg/goacc/pr93464.f90: New.
For the 2s failures in the PR, we have a V4SF VEC_PERM_EXPR in
which the first two elements are duplicates of one element and
the other two are don't-care:
v4sf_b = VEC_PERM_EXPR <v4sf_a, v4sf_a, { 1, 1, ?, ? }>;
The heuristic was to extend this with a blend:
v4sf_b = VEC_PERM_EXPR <v4sf_a, v4sf_a, { 1, 1, 2, 3 }>;
but it seems better to extend a partial duplicate to a full duplicate:
v4sf_b = VEC_PERM_EXPR <v4sf_a, v4sf_a, { 1, 1, 1, 1 }>;
Obviously this is still just a heuristic though.
I wondered whether to restrict this to two elements or more
but couldn't find any examples in which it made a difference.
Either way should be fine for the purposes of fixing this PR.
2020-01-28 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR tree-optimization/92822
* tree-ssa-forwprop.c (simplify_vector_constructor): When filling
out the don't-care elements of a vector whose significant elements
are duplicates, make the don't-care elements duplicates too.
predcom has the following code to stop one rogue load from
interfering with other store-load opportunities:
/* If A is read and B write or vice versa and there is unsuitable
dependence, instead of merging both components into a component
that will certainly not pass suitable_component_p, just put the
read into bad component, perhaps at least the write together with
all the other data refs in it's component will be optimizable. */
But when store-store commoning was added later, this had the effect
of ignoring loads that occur between two candidate stores.
There is code further up to handle loads and stores with unknown
dependences:
/* Don't do store elimination if there is any unknown dependence for
any store data reference. */
if ((DR_IS_WRITE (dra) || DR_IS_WRITE (drb))
&& (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know
|| DDR_NUM_DIST_VECTS (ddr) == 0))
eliminate_store_p = false;
But the store-load code above skips loads for *known* dependences
if (a) the load has already been marked "bad" or (b) the data-ref
machinery knows the dependence distance, but determine_offsets
can't handle the combination.
(a) happens to be the problem in the testcase, but a different
sequence could have given (b) instead. We have writes to individual
fields of a structure and reads from the whole structure. Since
determine_offsets requires the types to be the same, it returns false
for each such read/write combination.
This patch records which components have had loads removed and
prevents store-store commoning for them. It's a bit too pessimistic,
since there shouldn't be a problem if a "bad" load dominates all stores
in a component. But (a) we can't AFAIK use pcom_stmt_dominates_stmt_p
here and (b) the handling for that case would probably need to be
removed again if we handled more exotic cases in future.
2020-01-28 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR tree-optimization/93434
* tree-predcom.c (split_data_refs_to_components): Record which
components have had aliasing loads removed. Prevent store-store
commoning for all such components.
gcc/testsuite/
PR tree-optimization/93434
* gcc.c-torture/execute/pr93434.c: New test.
PR c++/92440
* pt.c (redeclare_class_template): Use inform
for the second location.
PR c++/92440
* g++.dg/template/pr92440.C: New test.
* g++.dg/cpp0x/vt-34314.C: Update error to note.
* g++.dg/template/pr59930-2.C: Likewise.
* g++.old-deja/g++.pt/redecl1.C: Likewise.
The following testcase is miscompiled, because the variable shift left
operand, { -1, -1, -1, -1 } is represented as a VECTOR_CST with
VECTOR_CST_NPATTERNS 1 and VECTOR_CST_NELTS_PER_PATTERN 1, so when
we call builder.new_unary_operation, builder.encoded_nelts () will be just 1
and thus we encode the resulting vector as if all the elements were the
same.
For non-masked is_vshift, we could perhaps call builder.new_binary_operation
(TREE_TYPE (args[0]), args[0], args[1], false), but then there are masked
shifts, for non-is_vshift we could perhaps call it too but with args[2]
instead of args[1], but there is no builder.new_ternary_operation.
All this stuff is primarily for aarch64 anyway, on x86 we don't have any
variable length vectors, and it is not a big deal to compute all elements
and just let builder.finalize () find the most efficient VECTOR_CST
representation of the vector. So, instead of doing too much, this just
keeps using new_unary_operation only if only one VECTOR_CST is involved
(i.e. non-masked shift by constant) and for the rest just compute all elts.
2020-01-28 Jakub Jelinek <jakub@redhat.com>
PR target/93418
* config/i386/i386.c (ix86_fold_builtin) <do_shift>: If mask is not
-1 or is_vshift is true, use new_vector with number of elts npatterns
rather than new_unary_operation.
* gcc.target/i386/avx2-pr93418.c: New test.
libgcrypt FAILs to build on aarch64-linux with
*** stack smashing detected ***: terminated
when gcc is compiled with -D_FORTIFY_SOURCE=2. The problem is if
fold_array_ctor_reference is called with size equal to or very close to
MAX_BITSIZE_MODE_ANY_MODE bits and non-zero inner_offset.
The first native_encode_expr is called with that inner_offset and bufoff 0,
the subsequent ones with offset of 0, and bufoff elt_size - inner_offset,
2 * elt_size - inner_offset etc. So, e.g. on the testcase where we start
with inner_offset 1 and size is e.g. 256 bytes and elt_size 4 bytes
we then call native_encode_expr at bufoff 251 and then 255, but that one
overwrites 3 bytes beyond the buf array.
The following patch fixes that. In addition, it avoids calling
elt_size.to_uhwi () all the time, and punts if elt_sz would be too large.
2020-01-28 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/93454
* gimple-fold.c (fold_array_ctor_reference): Perform
elt_size.to_uhwi () just once, instead of calling it in every
iteration. Punt if that value is above size of the temporary
buffer. Decrease third native_encode_expr argument when
bufoff + elt_sz is above size of buf.
* gcc.dg/pr93454.c: New test.
PR analyzer/93451 reports an ICE when canonicalizing the constants
in a region_model, with a failed qsort_chk when attempting to sort
the constants within the region_model.
The svalues in the model were:
sv0: {poisoned: uninit}
sv1: {type: ‘double’, ‘0.0’}
sv2: {type: ‘double’, ‘1.0e+0’}
sv3: {type: ‘double’, ‘ Nan’}
The qsort_chk of the 3 constants fails due to tree_cmp using the
LT_EXPR ordering of the REAL_CSTs, which doesn't work for NaN.
This patch adjusts tree_cmp to impose an arbitrary ordering during
canonicalization for UNORDERED_EXPR cases w/o relying on the LT_EXPR
ordering, fixing the ICE.
gcc/analyzer/ChangeLog:
PR analyzer/93451
* region-model.cc (tree_cmp): For the REAL_CST case, impose an
arbitrary order on NaNs relative to other NaNs and to non-NaNs;
const-correctness tweak.
(ana::selftests::build_real_cst_from_string): New function.
(ana::selftests::append_interesting_constants): New function.
(ana::selftests::test_tree_cmp_on_constants): New test.
(ana::selftests::test_canonicalization_4): New test.
(ana::selftests::analyzer_region_model_cc_tests): Call the new
tests.
gcc/testsuite/ChangeLog:
PR analyzer/93451
* gcc.dg/analyzer/torture/pr93451.c: New test.
The motivation behind this change is to make it easier for a user to
link against static libraries on a target where dynamic libraries are
the default library type (for example GNU/Linux).
Further, my motivation is really for linking libraries into GDB,
however, the binutils-gdb/config/ directory is a copy of gcc/config/
so changes for GDB need to be approved by the GCC project first.
After making this change in the gcc/config/ directory I've run
autoreconf on all of the configure scripts in the GCC tree and a
couple have been updated, so I'll use one of these to describe what my
change does.
Consider libcpp, this library links against libiconv. Currently if
the user builds on a system with both static and dynamic libiconv
installed then autotools will pick up the dynamic libiconv by
default. This is almost certainly the right thing to do.
However, if the user wants to link against static libiconv then things
are a little harder, they could remove the dynamic libiconv from their
system, but this is probably a bad idea (other things might depend on
that library), or the user can build their own version of libiconv,
install it into a unique prefix, and then configure gcc using the
--with-libiconv-prefix=DIR flag. This works fine, but is somewhat
annoying, the static library available, I just can't get autotools to
use it.
My change then adds a new flag --with-libiconv-type=TYPE, where type
is either auto, static, or shared. The default auto, ensures we keep
the existing behaviour unchanged.
If the user configures with --with-libiconv-type=static then the
configure script will ignore any dynamic libiconv it finds, and will
only look for a static libiconv, if no static libiconv is found then
the configure will continue as though there is no libiconv at all
available.
Similarly a user can specify --with-libiconv-type=shared and force the
use of shared libiconv, any static libiconv will be ignored.
As I've implemented this change within the AC_LIB_LINKFLAGS_BODY macro
then only libraries configured using the AC_LIB_LINKFLAGS or
AC_LIB_HAVE_LINKFLAGS macros will gain the new configure flag.
If this is accepted into GCC then there will be follow on patches for
binutils and GDB to regenerate some configure scripts in those
projects.
For GCC only two configure scripts needed updated after this commit,
libcpp and libstdc++-v3, both of which link against libiconv.
config/ChangeLog:
* lib-link.m4 (AC_LIB_LINKFLAGS_BODY): Add new
--with-libXXX-type=... option. Use this to guide the selection of
either a shared library or a static library.
libcpp/ChangeLog:
* configure: Regenerate.
libstdc++-v3/ChangeLog:
* configure: Regenerate.
mips_declare_object_name is missing the support for declaring symbols
as gnu_unique_object that is present in the generic
ASM_DECLARE_OBJECT_NAME in elfos.h. I'm not aware of any
MIPS-specific reason for that support to be absent;
mips_declare_object_name predates the addition of gnu_unique_object
support and as far as I can tell this was simply an oversight when
that support was added. This patch adds the missing support,
following the code in elfos.h.
Tested with no regressions with cross to mips-linux-gnu. In
particular, this fixes the failure of the recently-added glibc test
elf/tst-dlopen-nodelete-reloc, which relies on the compiler generating
such symbols, for MIPS.
* config/mips/mips.c (mips_declare_object_name)
[USE_GNU_UNIQUE_OBJECT]: Support use of gnu_unique_object.
Since Martin Sebor's patch for PR 71625 to change braced array initializers
to STRING_CST in some cases, we need to be ready for STRING_CST with types
that are changed by tsubst. fold_convert doesn't know how to deal with
STRING_CST, which is reasonable; we really shouldn't expect it to here. So
let's handle STRING_CST separately.
PR c++/90966
* pt.c (tsubst_copy) [STRING_CST]: Don't use fold_convert.
As the PR shows, there is a pathway through the code where the
no_warning value is not set, which corresponds to a missing check
of the ramp return when it was constructed from the 'get return
object' Fixed by ensuring that the check of the return value is
carried out for both return cases.
gcc/cp/ChangeLog:
2020-01-27 Iain Sandoe <iain@sandoe.co.uk>
PR c++/93443
* coroutines.cc (morph_fn_to_coro): Check the ramp return
value when it is constructed from the 'get return object'.
This was committed following offline approval by Kyryl.
One minor intended optimisation introduced by :
https://gcc.gnu.org/ml/gcc-patches/2020-01/msg01237.html
was to set a preference for both __fp16 types and __bf16 types to be
loaded/stored directly into/from the FP/NEON registers (if they are available
and if the vld1.16 is compatible), rather than be passed through the regular
r-registers.
This would convert many observed instances of:
** ldrh r3, [r3] @ __fp16
** vmov.f16 s15, r3 @ __fp16
Into a single:
** vld1.16 {d7[2]}, [r3]
This resulted in a regression of a dg-scan-assembler in a __fp16 test.
This patch updates the test to the same testing standard used by the BFloat
tests (use check-function-bodies to explicitly check for correct assembler
generated by each function) and updates it for the latest optimisation.
Cheers,
Stam
gcc/testsuite/ChangeLog:
2020-01-27 Stam Markianos-Wright <stam.markianos-wright@arm.com>
* gcc.target/arm/armv8_2-fp16-move-1.c: Update following load/store
optimisation.
PR analyzer/93349 reports an ICE in IPA pass: simdclone for
some input files when -fanalyzer is supplied, with:
error: location references block not in block tree
The root cause is that the analyzer touches input_location in some
places (to make it easier to track down which source construct the
analyzer can't handle in the case of an analyzer ICE) and fails to reset
it. For the ICE in question, this sets input_location to a location_t
that references some arbitrary block (specifically, that of the last
statement to be analyzed, within the original CFG of whichever is the
last such function to be analyzed).
Later, within omp-simd-clone.c, input_location is used by gimplify_expr
(called via gimplify_and_add), which has:
14492 if (!gimple_seq_empty_p (*pre_p))
14493 annotate_all_with_location_after (*pre_p, pre_last_gsi, input_location);
thus using whatever the value of input_location is, leading
to statements that reference some arbitrary block in the original CFG.
For the reproducer, this happens to be a block in the CFG for the
original function, rather than that of the clone, but in general it
could be some arbitrary other function in the TU.
This code appears to assume that input_location has some arbitrary
value *not* in the block tree, which is potentially violated by the
analyzer's changes to input_location.
This patch adds a save and restore of input_location at the top-level
function of the analyzer, fixing the ICE.
gcc/analyzer/ChangeLog:
PR analyzer/93349
* engine.cc (run_checkers): Save and restore input_location.
gcc/testsuite/ChangeLog:
PR analyzer/93349
* gcc.dg/analyzer/torture/pr93349.c: New test.
Amongst the inputs to the analyzer state machines that can lead to state
transitions are conditions on CFG edges, such as a test for a pointer
being non-NULL.
These conditionals can be non-trivial to determine in the face of
optimization. For example, at -O2:
if (p == NULL || q == NULL)
is optimized on some targets (e.g. x86_64) to a bitwise-or:
_1 = p_5(D) == 0B;
_2 = q_6(D) == 0B;
_3 = _1 | _2;
if (_3 != 0)
goto <bb 4>; [51.12%]
else
goto <bb 3>; [48.88%]
but on other targets (e.g. powerpc64le) as control flow:
if (p_2(D) == 0B)
goto <bb 5>; [18.09%]
else
goto <bb 3>; [81.91%]
<bb 3> [local count: 879501929]:
if (q_3(D) == 0B)
goto <bb 5>; [30.95%]
else
goto <bb 4>; [69.05%]
region_model::add_any_constraints_from_ssa_def_stmt attempts to walk
SSA def chains to reconstruct the conditions that hold, so that
e.g. in the above case of bitwise-or, the state machine for
"p" can transition to the "known-null" state along the edge leading
to bb 3.
In gcc.dg/analyzer/pattern-test-2.c I attempted to write test coverage
for this, but the test fails on those targets for which the || is
expressed via control flow.
This patch rewrites the test to make explicit use of bitwise-or, and
adds coverage for bitwise-and for good measure.
gcc/testsuite/ChangeLog:
PR analyzer/93291
* gcc.dg/analyzer/pattern-test-2.c: Remove include of stdlib.h.
(test_2): Rewrite to explicitly perform a bitwise-or of two
boolean conditions.
(test_3): New function, to test bitwise-and.
This test started failing after the switch to -fno-common because we can
now force the array to be aligned to 16 bytes, which in turn lets us use
SIMD accesses. Locally restoring -fcommon seems the most faithful to
the original PR.
2020-01-27 Richard Sandiford <richard.sandiford@arm.com>
gcc/testsuite/
PR testsuite/71727
* gcc.target/aarch64/pr71727.c: Add -fcommon.
region_model.cc's tree_cmp attempted to verify that the ordering
is symmetric by asserting that
tree_cmp (x, y) == -tree_cmp (y, x)
This condition is too strong: it's only required for a comparator that
sign (tree_cmp (x, y)) == -sign (tree_cmp (y, x))
and the incorrect form of the assertion doesn't hold e.g. on s390x where
for certain inputs x, y, tree_cmp (x, y) == 1 and tree_cmp (y, x) == -2,
breaking the build in "make selftest" in stage1.
In any case, these checks are redundant, since qsort_chk performs them.
Additionally, there is a potential lack of transitivity in
worklist::key_t::cmp where hashval_t values are compared by subtraction,
which could fail to be transitive if overflows occur.
This patch eliminates the redundant checks and reimplements the hashval_t
comparisons in terms of < and >, fixing these issues.
gcc/analyzer/ChangeLog:
* call-string.cc (call_string::cmp_1): Delete, moving body to...
(call_string::cmp): ...here.
* call-string.h (call_string::cmp_1): Delete decl.
* engine.cc (worklist::key_t::cmp_1): Delete, moving body to...
(worklist::key_t::cmp): ...here. Implement hash comparisons
via comparison rather than subtraction to avoid overflow issues.
* exploded-graph.h (worklist::key_t::cmp_1): Delete decl.
* region-model.cc (tree_cmp): Eliminate buggy checking for
symmetry.
Part of the problem in this PR is that we don't provide patterns
to extract a 64-bit vector from one half of a 128-bit vector.
Adding them fixes:
FAIL: gcc.target/aarch64/fmul_intrinsic_1.c scan-assembler-times fmul\\td[0-9]+, d[0-9]+, d[0-9]+ 1
FAIL: gcc.target/aarch64/fmul_intrinsic_1.c scan-assembler-times fmul\\tv[0-9]+.2d, v[0-9]+.2d, v[0-9]+.d\\[[0-9]+\\] 3
The 2s failures need target-independent changes, after which they rely
on these patterns too.
2020-01-27 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR target/92822
* config/aarch64/aarch64-simd.md (aarch64_get_half<mode>): New
expander.
(@aarch64_split_simd_mov<mode>): Use it.
(aarch64_simd_mov_from_<mode>low): Add a GPR alternative.
Leave the vec_extract patterns to handle 2-element vectors.
(aarch64_simd_mov_from_<mode>high): Likewise.
(vec_extract<VQMOV_NO2E:mode><Vhalf>): New expander.
(vec_extractv2dfv1df): Likewise.
gcc.target/aarch64/cmpimm_branch_1.c started failing after Bernd's
fix to make combine take the costs of jumps into account
(g:391500af1932e696a007). This is because the rtx costs
of *compare_condjump<GPI:mode> were higher than the costs
of the instructions it combines.
2020-01-27 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* config/aarch64/aarch64.c (aarch64_if_then_else_costs): Match
jump conditions for *compare_condjump<GPI:mode>.
This patch removes the hack in is_setjmp_call_p of looking for
"setjmp" and "_setjmp", replacing it with some logic adapted from
special_function_p in calls.c, ignoring up to 2 leading underscores from
the fndecl's name when checking for a function by name.
It also requires that such functions are "extern" and at file scope
for them to be matched.
The patch also generalizes the setjmp/longjmp handling in the analyzer
to also work with sigsetjmp/siglongjmp. Doing so requires generalizing
some hardcoded functions in diagnostics (which were hardcoded to avoid
user-facing messages referring to "_setjmp", which is an implementation
detail) - the patch adds a new function, get_user_facing_name for this,
for use on calls that matched is_named_call_p and
is_specical_named_call_p.
gcc/analyzer/ChangeLog:
* analyzer.cc (is_named_call_p): Check that fndecl is "extern"
and at file scope. Potentially disregard prefix _ or __ in
fndecl's name. Bail if the identifier is NULL.
(is_setjmp_call_p): Expect a gcall rather than plain gimple.
Remove special-case check for leading prefix, and also check for
sigsetjmp.
(is_longjmp_call_p): Also check for siglongjmp.
(get_user_facing_name): New function.
* analyzer.h (is_setjmp_call_p): Expect a gcall rather than plain
gimple.
(get_user_facing_name): New decl.
* checker-path.cc (setjmp_event::get_desc): Use
get_user_facing_name to avoid hardcoding the function name.
(rewind_event::rewind_event): Add rewind_info param, using it to
initialize new m_rewind_info field, and strengthen the assertion.
(rewind_from_longjmp_event::get_desc): Use get_user_facing_name to
avoid hardcoding the function name.
(rewind_to_setjmp_event::get_desc): Likewise.
* checker-path.h (setjmp_event::setjmp_event): Add setjmp_call
param and use it to initialize...
(setjmp_event::m_setjmp_call): New field.
(rewind_event::rewind_event): Add rewind_info param.
(rewind_event::m_rewind_info): New protected field.
(rewind_from_longjmp_event::rewind_from_longjmp_event): Add
rewind_info param.
(class rewind_to_setjmp_event): Move rewind_info field to parent
class.
* diagnostic-manager.cc (diagnostic_manager::add_events_for_eedge):
Update setjmp-handling for is_setjmp_call_p requiring a gcall;
pass the call to the new setjmp_event.
* engine.cc (exploded_node::on_stmt): Update for is_setjmp_call_p
requiring a gcall.
(stale_jmp_buf::emit): Use get_user_facing_name to avoid
hardcoding the function names.
(exploded_node::on_longjmp): Pass the longjmp_call when
constructing rewind_info.
(rewind_info_t::add_events_to_path): Pass the rewind_info_t to the
rewind_from_longjmp_event's ctor.
* exploded-graph.h (rewind_info_t::rewind_info_t): Add
longjmp_call param.
(rewind_info_t::get_longjmp_call): New.
(rewind_info_t::m_longjmp_call): New.
* region-model.cc (region_model::on_setjmp): Update comment to
indicate this is also for sigsetjmp.
* region-model.h (struct setjmp_record): Likewise.
(class setjmp_svalue): Likewise.
gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/sigsetjmp-5.c: New test.
* gcc.dg/analyzer/sigsetjmp-6.c: New test.
This patch fixes various build failures seen with gcc 4.4
gcc prior to 4.6 complains about:
error: #pragma GCC diagnostic not allowed inside functions
for various uses of PUSH_IGNORE_WFORMAT and POP_IGNORE_WFORMAT.
This patch makes them a no-op with such compilers.
The patch also fixes various errors with template base initializers
and redundant uses of "typename" that older g++ implementations
can't cope with.
gcc/analyzer/ChangeLog:
PR analyzer/93276
* analyzer.h (PUSH_IGNORE_WFORMAT, POP_IGNORE_WFORMAT): Guard these
macros with GCC_VERSION >= 4006, making them no-op otherwise.
* engine.cc (exploded_edge::exploded_edge): Specify template for
base class initializer.
(exploded_graph::add_edge): Specify template when chaining up to
base class add_edge implementation.
(viz_callgraph_node::dump_dot): Drop redundant "typename".
(viz_callgraph_edge::viz_callgraph_edge): Specify template for
base class initializer.
* program-state.cc (sm_state_map::clone_with_remapping): Drop
redundant "typename".
(sm_state_map::print): Likewise.
(sm_state_map::hash): Likewise.
(sm_state_map::operator==): Likewise.
(sm_state_map::remap_svalue_ids): Likewise.
(sm_state_map::on_svalue_purge): Likewise.
(sm_state_map::validate): Likewise.
* program-state.h (sm_state_map::iterator_t): Likewise.
* supergraph.h (superedge::superedge): Specify template for base
class initializer.
gcc/ChangeLog:
PR analyzer/93276
* digraph.cc (test_edge::test_edge): Specify template for base
class initializer.
My changes to is_nested_namespace broke is_ancestor's use where a namespace
alias might be passed in. This changes is_ancestor to look through the alias.
PR c++/91826
* name-lookup.c (is_ancestor): Allow CHILD to be a namespace alias.
Jan Hubicka
Jason Merrill
H.J. Lu
David Malcolm
Andrew Benson
Nathan Sidwell
Jonathan Wakely
Julian Brown
Richard Biener
Claudiu Zissulescu
Richard Sandiford
Tobias Burnus
Martin Liska
Jakub Jelinek
GCC Administrator
Andrew Burgess
Joseph Myers
Ian Lance Taylor
Iain Sandoe
Stam Markianos-Wright