range_from_dom makes a recursive call to resolve the immediate dominator
when there are multiple incoming edges to a block. This is not necessary
when the dominator already has an on-entry cache value.
PR tree-optimization/106234
* gimple-range-cache.cc (ranger_cache::range_from_dom): Check dominator
cache value before recursively resolving it.
This patch upgrades x86_64's scalar-to-vector (STV) pass to more
aggressively transform 128-bit scalar TImode operations into vector
V1TImode operations performed on SSE registers. TImode functionality
already exists in STV, but only for move operations. This change
brings support for logical operations (AND, IOR, XOR, NOT and ANDN)
and comparisons.
The effect of these changes are conveniently demonstrated by the new
sse4_1-stv-5.c test case:
__int128 a[16];
__int128 b[16];
__int128 c[16];
void foo()
{
for (unsigned int i=0; i<16; i++)
a[i] = b[i] & ~c[i];
}
which when currently compiled on mainline wtih -O2 -msse4 produces:
foo: xorl %eax, %eax
.L2: movq c(%rax), %rsi
movq c+8(%rax), %rdi
addq $16, %rax
notq %rsi
notq %rdi
andq b-16(%rax), %rsi
andq b-8(%rax), %rdi
movq %rsi, a-16(%rax)
movq %rdi, a-8(%rax)
cmpq $256, %rax
jne .L2
ret
but with this patch now produces:
foo: xorl %eax, %eax
.L2: movdqa c(%rax), %xmm0
pandn b(%rax), %xmm0
addq $16, %rax
movaps %xmm0, a-16(%rax)
cmpq $256, %rax
jne .L2
ret
Technically, the STV pass is implemented by three C++ classes, a common
abstract base class "scalar_chain" that contains common functionality,
and two derived classes: general_scalar_chain (which handles SI and
DI modes) and timode_scalar_chain (which handles TI modes). As
mentioned previously, because only TI mode moves were handled the
two worker classes behaved significantly differently. These changes
bring the functionality of these two classes closer together, which
is reflected by refactoring more shared code from general_scalar_chain
to the parent scalar_chain and reusing it from timode_scalar_chain.
There still remain significant differences (and simplifications) so
the existing division of classes (as specializations) continues to
make sense.
2022-07-11 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-features.h (scalar_chain): Add fields
insns_conv, n_sse_to_integer and n_integer_to_sse to this
parent class, moved from general_scalar_chain.
(scalar_chain::convert_compare): Protected method moved
from general_scalar_chain.
(mark_dual_mode_def): Make protected, not private virtual.
(scalar_chain:convert_op): New private virtual method.
(general_scalar_chain::general_scalar_chain): Simplify constructor.
(general_scalar_chain::~general_scalar_chain): Delete destructor.
(general_scalar_chain): Move insns_conv, n_sse_to_integer and
n_integer_to_sse fields to parent class, scalar_chain.
(general_scalar_chain::mark_dual_mode_def): Delete prototype.
(general_scalar_chain::convert_compare): Delete prototype.
(timode_scalar_chain::compute_convert_gain): Remove simplistic
implementation, convert to a method prototype.
(timode_scalar_chain::mark_dual_mode_def): Delete prototype.
(timode_scalar_chain::convert_op): Prototype new virtual method.
* config/i386/i386-features.cc (scalar_chain::scalar_chain):
Allocate insns_conv and initialize n_sse_to_integer and
n_integer_to_sse fields in constructor.
(scalar_chain::scalar_chain): Free insns_conv in destructor.
(general_scalar_chain::general_scalar_chain): Delete
constructor, now defined in the class declaration.
(general_scalar_chain::~general_scalar_chain): Delete destructor.
(scalar_chain::mark_dual_mode_def): Renamed from
general_scalar_chain::mark_dual_mode_def.
(timode_scalar_chain::mark_dual_mode_def): Delete.
(scalar_chain::convert_compare): Renamed from
general_scalar_chain::convert_compare.
(timode_scalar_chain::compute_convert_gain): New method to
determine the gain from converting a TImode chain to V1TImode.
(timode_scalar_chain::convert_op): New method to convert an
operand from TImode to V1TImode.
(timode_scalar_chain::convert_insn) <case REG>: Only PUT_MODE
on REG_EQUAL notes that were originally TImode (not CONST_INT).
Handle AND, ANDN, XOR, IOR, NOT and COMPARE.
(timode_mem_p): Helper predicate to check where operand is
memory reference with sufficient alignment for TImode STV.
(timode_scalar_to_vector_candidate_p): Use convertible_comparison_p
to check whether COMPARE is convertible. Handle SET_DESTs that
that are REG_P or MEM_P and SET_SRCs that are REG, CONST_INT,
CONST_WIDE_INT, MEM, AND, ANDN, IOR, XOR or NOT.
gcc/testsuite/ChangeLog
* gcc.target/i386/sse4_1-stv-2.c: New test case, pand.
* gcc.target/i386/sse4_1-stv-3.c: New test case, por.
* gcc.target/i386/sse4_1-stv-4.c: New test case, pxor.
* gcc.target/i386/sse4_1-stv-5.c: New test case, pandn.
* gcc.target/i386/sse4_1-stv-6.c: New test case, ptest.
In g:76c3041b856cb0 I'd removed a "C ? optab_vector : optab_mixed_sign"
argument from a call to directly_supported_p, thinking that the argument
only existed because of the condition (which I was removing). But the
difference between the scalar and vector forms matters for shifts,
so we do still need the argument.
gcc/
PR tree-optimization/106250
* tree-vect-loop.cc (vectorizable_reduction): Reinstate final
argument to directly_supported_p.
In r13-1544, handle_pragma_diagnostic was refactored to support processing
early pragmas. During that process the part looking up option arguments was
inadvertenly moved too early, prior to checking the option was valid, causing
PR106252. Fixed by moving the check back where it goes.
gcc/c-family/ChangeLog:
PR preprocessor/106252
* c-pragma.cc (handle_pragma_diagnostic_impl): Don't look up the
option argument prior to verifying the option was found.
When working on a smaller region like a loop version copy the main
time spent is now dominance fast query recompute which does a full
function DFS walk. The dominance queries within the region of
interest should be O(log n) without fast queries and we should do
on the order of O(n) of them which overall means reasonable
complexity.
For the artificial testcase I'm looking at this shaves off
considerable time again.
* tree-into-ssa.cc (update_ssa): Do not forcefully
re-compute dominance fast queries for TODO_update_ssa_no_phi.
The following fixes the last commit to honor the case we are not
vectorizing a loop.
PR tree-optimization/106228
* tree-vect-data-refs.cc (vect_setup_realignment): Adjust
VUSE compute for the non-loop case.
When we do TODO_update_ssa_no_phi we already avoid computing
dominance frontiers for all blocks - it is worth to also avoid
walking all dominated blocks in the update domwalk and restrict
the walk to the SEME region with the affected blocks. We can
do that by walking the CFG in reverse from blocks_to_update to
the common immediate dominator, marking blocks in the region
and telling the domwalk to STOP when leaving it.
For an artificial testcase with N adjacent loops with one
unswitching opportunity that takes the incremental SSA updating
off the -ftime-report radar:
tree loop unswitching : 11.25 ( 3%) 0.09 ( 5%) 11.53 ( 3%) 36M ( 9%)
`- tree SSA incremental : 35.74 ( 9%) 0.07 ( 4%) 36.65 ( 9%) 2734k ( 1%)
improves to
tree loop unswitching : 10.21 ( 3%) 0.05 ( 3%) 11.50 ( 3%) 36M ( 9%)
`- tree SSA incremental : 0.66 ( 0%) 0.02 ( 1%) 0.49 ( 0%) 2734k ( 1%)
for less localized updates the SEME region isn't likely constrained
enough so I've restricted the extra work to TODO_update_ssa_no_phi
callers.
* tree-into-ssa.cc (rewrite_mode::REWRITE_UPDATE_REGION): New.
(rewrite_update_dom_walker::rewrite_update_dom_walker): Update.
(rewrite_update_dom_walker::m_in_region_flag): New.
(rewrite_update_dom_walker::before_dom_children): If the region
to update is marked, STOP at exits.
(rewrite_blocks): For REWRITE_UPDATE_REGION mark the region
to be updated.
(dump_update_ssa): Use bitmap_empty_p.
(update_ssa): Likewise. Use REWRITE_UPDATE_REGION when
TODO_update_ssa_no_phi.
* tree-cfgcleanup.cc (cleanup_tree_cfg_noloop): Account
pending update_ssa to the caller.
The following avoids the need to massage the target optimization
node at WPA time when we fixup the optimization node, copying
FP related flags from callee to caller. The target is already
set up to fixup, but that only works when not switching between
functions. After fixing that the fixup is then done at LTRANS
time when materializing the function.
2022-07-01 Richard Biener <rguenthert@suse.de>
PR target/105459
* config/i386/i386-options.cc (ix86_set_current_function):
Rebuild the target optimization node whenever necessary,
not only when the optimization node didn't change.
* gcc.dg/lto/pr105459_0.c: New testcase.
The following adds missing assignment of a virtual use operand to a
created load to vect_setup_realignment which shows as bootstrap
failure on powerpc64-linux and extra testsuite fails for targets
when misaligned loads are not supported or not optimal.
PR tree-optimization/106228
* tree-vect-data-refs.cc (vect_setup_realignment): Properly
set a VUSE operand on the emitted load.
Currently SSA_NAME_RANGE_INFO only handles integer ranges, and loses
half the precision in the process because its use of legacy
value_range's. This patch rewrites all the SSA_NAME_RANGE_INFO
(nonzero bits included) to use the recently contributed
vrange_storage. With it, we'll be able to efficiently save any ranges
supported by ranger in GC memory. Presently this will only be
irange's, but shortly we'll add floating ranges and others to the mix.
As per the discussion with the trailing_wide_ints adjustments and
vrange_storage, we'll be able to save integer ranges with a maximum of
5 sub-ranges. This could be adjusted later if more sub-ranges are
needed (unlikely).
Since this is a behavior changing patch, I would like to take a few
days for discussion, and commit early next week if all goes well.
A few notes.
First, we get rid of the SSA_NAME_ANTI_RANGE_P bit in the SSA_NAME
since we store full resolution ranges. Perhaps it could be re-used
for something else.
The range_info_def struct is gone in favor of an opaque type handled
by vrange_storage. It currently supports irange, but will support
frange, prange, etc, in due time.
From the looks of it, set_range_info was an update operation despite
its name, as we improved the nonzero bits with each call, even though
we clobbered the ranges. Presumably this was because doing a proper
intersect of ranges lost information with the anti-range hack. We no
longer have this limitation so now we formalize both set_range_info
and set_nonzero_bits to an update operation. After all, we should
never be losing information, but enhancing it whenever possible. This
means, that if folks' finger-memory is not offended, as a follow-up,
I'd like to rename set_nonzero_bits and set_range_info to update_*.
I have kept the same global API we had in tree-ssanames.h, with the
caveat that all set operations are now update as discussed above.
There is a 2% performance penalty for evrp and a 3% penalty for VRP
that is coincidentally in line with a previous improvement of the same
amount in the vrange abstraction patchset. Interestingly, this
penalty is mostly due to the wide int to tree dance we keep doing with
irange and legacy. In a first draft of this patch where I was
streaming trees directly, there was actually a small improvement
instead. I hope to get some of the gain back when we move irange's to
wide-ints, though I'm not in a hurry ;-).
Tested and benchmarked on x86-64 Linux. Tested on ppc64le Linux.
Comments welcome.
gcc/ChangeLog:
* gimple-range.cc (gimple_ranger::export_global_ranges): Remove
verification against legacy value_range.
(gimple_ranger::register_inferred_ranges): Same.
(gimple_ranger::export_global_ranges): Rename update_global_range
to set_range_info.
* tree-core.h (struct range_info_def): Remove.
(struct irange_storage_slot): New.
(struct tree_base): Remove SSA_NAME_ANTI_RANGE_P documentation.
(struct tree_ssa_name): Add vrange_storage support.
* tree-ssanames.cc (range_info_p): New.
(range_info_fits_p): New.
(range_info_alloc): New.
(range_info_free): New.
(range_info_get_range): New.
(range_info_set_range): New.
(set_range_info_raw): Remove.
(set_range_info): Adjust to use vrange_storage.
(set_nonzero_bits): Same.
(get_nonzero_bits): Same.
(duplicate_ssa_name_range_info): Remove overload taking
value_range_kind.
Rewrite tree overload to use vrange_storage.
(duplicate_ssa_name_fn): Adjust to use vrange_storage.
* tree-ssanames.h (struct range_info_def): Remove.
(set_range_info): Adjust prototype to take vrange.
* tree-vrp.cc (vrp_asserts::remove_range_assertions): Call
duplicate_ssa_name_range_info.
* tree.h (SSA_NAME_ANTI_RANGE_P): Remove.
(SSA_NAME_RANGE_TYPE): Remove.
* value-query.cc (get_ssa_name_range_info): Adjust to use
vrange_storage.
(update_global_range): Remove.
(get_range_global): Remove as_a<irange>.
* value-query.h (update_global_range): Remove.
* tree-ssa-dom.cc (set_global_ranges_from_unreachable_edges):
Rename update_global_range to set_range_info.
* value-range-storage.cc (vrange_storage::alloc_slot): Remove
gcc_unreachable.
The handling of #pragma GCC diagnostic uses input_location, which is not always
as precise as needed; in particular the relative location of some tokens and a
_Pragma directive will crucially determine whether a given diagnostic is enabled
or suppressed in the desired way. PR97498 shows how the C frontend ends up with
input_location pointing to the beginning of the line containing a _Pragma()
directive, resulting in the wrong behavior if the diagnostic to be modified
pertains to some tokens found earlier on the same line. This patch fixes that by
addressing two issues:
a) libcpp was not assigning a valid location to the CPP_PRAGMA token
generated by the _Pragma directive.
b) C frontend was not setting input_location to something reasonable.
With this change, the C frontend is able to change input_location to point to
the _Pragma token as needed.
This is just a two-line fix (one for each of a) and b)), the testsuite changes
were needed only because the location on the tested warnings has been somewhat
improved, so the tests need to look for the new locations.
gcc/c/ChangeLog:
PR preprocessor/97498
* c-parser.cc (c_parser_pragma): Set input_location to the
location of the pragma, rather than the start of the line.
libcpp/ChangeLog:
PR preprocessor/97498
* directives.cc (destringize_and_run): Override the location of
the CPP_PRAGMA token from a _Pragma directive to the location of
the expansion point, as is done for the tokens lexed from it.
gcc/testsuite/ChangeLog:
PR preprocessor/97498
* c-c++-common/pr97498.c: New test.
* c-c++-common/gomp/pragma-3.c: Adapt for improved warning locations.
* c-c++-common/gomp/pragma-5.c: Likewise.
* gcc.dg/pragma-message.c: Likewise.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-c-c++-common/reduction-5.c: Adapt for
improved warning locations.
* testsuite/libgomp.oacc-c-c++-common/vred2d-128.c: Likewise.
In discussions with Andrew we realized varying_p() was returning true
for a range of the entire domain with a non-empty nonzero mask. This
is confusing as varying_p() should only return true when absolutely no
information is available. A nonzero mask that has any cleared bits is
extra information and must return false for varying_p(). This patch
fixes this oversight. Now a range of the entire domain with nonzero
bits, is internally set to VR_RANGE (with the appropriate end points
set). VR_VARYING ranges must have a null nonzero mask.
Also, the union and intersect code were not quite right in the presence of
nonzero masks. Sometimes we would drop masks to -1 unnecessarily. I
was trying to be too smart in avoiding extra work when the mask was
NULL, but there's also an implicit mask in the range that must be
taken into account. For example, [0,0] may have no nonzero bits set
explicitly, but the mask is really 0x0. This will all be simpler when
we drop trees, because the nonzero bits will always be set, even if
-1.
Finally, I've added unit tests to the nonzero mask code. This should
help us maintain sanity going forward.
There should be no visible changes, as the main consumer of this code
is the SSA_NAME_RANGE_INFO patchset which has yet to be committed.
Tested on x86-64 Linux.
gcc/ChangeLog:
* value-range.cc (irange::operator=): Call verify_range.
(irange::irange_set): Normalize kind after everything else has
been set.
(irange::irange_set_anti_range): Same.
(irange::set): Same.
(irange::verify_range): Disallow nonzero masks for VARYING.
(irange::irange_union): Call verify_range.
Handle nonzero masks better.
(irange::irange_intersect): Same.
(irange::set_nonzero_bits): Calculate mask if either range has an
explicit mask.
(irange::intersect_nonzero_bits): Same.
(irange::union_nonzero_bits): Same.
(range_tests_nonzero_bits): New.
(range_tests): Call range_tests_nonzero_bits.
* value-range.h (class irange): Remove set_nonzero_bits method
with trees.
(irange::varying_compatible_p): Set nonzero mask.
Like add.w/sub.w/mul.w, div.w/mod.w/div.wu/mod.wu also sign-extend the
output on LA64. But, LoongArch v1.00 mandates that the inputs of 32-bit
division to be sign-extended so we have to expand 32-bit division into
RTL sequences.
We defined div.w/mod.w/div.wu/mod.wu as a (DI, DI) -> SI instruction.
This definition does not indicate the fact that these instructions will
store the result as sign-extended value in a 64-bit GR. Then the
compiler would emit unnecessary sign-extend operations. For example:
int div(int a, int b) { return a / b; }
was compiled to:
div.w $r4, $r4, $r5
slli.w $r4, $r4, 0 # this is unnecessary
jr $r1
To remove this unnecessary operation, we change the division
instructions to (DI, DI) -> DI and describe the sign-extend behavior
explicitly in the RTL template. In the expander for 32-bit division we
then use simplify_gen_subreg to extract the lower 32 bits.
gcc/ChangeLog:
* config/loongarch/loongarch.md (<any_div>di3_fake): Describe
the sign-extend of result in the RTL template.
(<any_div><mode>3): Adjust for <any_div>di3_fake change.
gcc/testsuite/ChangeLog:
* gcc.target/loongarch/div-4.c: New test.
Currently in the description of LoongArch integer division instructions,
the output is marked as earlyclobbered ('&'). It's necessary when
loongarch_check_zero_div_p() because clobbering operand 2 (divisor) will
make the checking for zero divisor impossible.
But, for -mno-check-zero-division (the default of GCC >= 12.2 for
optimized code), the output is not earlyclobbered at all. And, the
read of operand 1 only occurs before clobbering the output. So we make
three alternatives for an idiv instruction:
* (=r,r,r): For -mno-check-zero-division.
* (=&r,r,r): For -mcheck-zero-division.
* (=&r,0,r): For -mcheck-zero-division, to explicitly allow patterns
like "div.d $a0, $a0, $a1".
gcc/ChangeLog:
* config/loongarch/loongarch.cc (loongarch_check_zero_div_p):
Remove static, for use in the machine description file.
* config/loongarch/loongarch-protos.h:
(loongarch_check_zero_div_p): Add prototype.
* config/loongarch/loongarch.md (enabled): New attr.
(*<optab><mode>3): Add (=r,r,r) and (=&r,0,r) alternatives for
idiv. Conditionally enable the alternatives using
loongarch_check_zero_div_p.
(<optab>di3_fake): Likewise.
gcc/testsuite/ChangeLog:
* gcc.target/loongarch/div-1.c: New test.
* gcc.target/loongarch/div-2.c: New test.
* gcc.target/loongarch/div-3.c: New test.
(mult (sign_extend:DI rj:SI) (sign_extend:DI rk:SI)) should be
"mulw.d.w", not "mul.d".
gcc/ChangeLog:
* config/loongarch/loongarch.md (mulsidi3_64bit): Use mulw.d.w
instead of mul.d.
gcc/testsuite/ChangeLog:
* gcc.target/loongarch/mulw_d_w.c: New test.
* gcc.c-torture/execute/mul-sext.c: New test.
The fast union operation is sometimes setting a range of the entire
domain, but leaving the kind bit as VR_RANGE instead of downgrading it
to VR_VARYING.
Tested on x86-64 Linux.
gcc/ChangeLog:
* value-range.cc (irange::irange_single_pair_union): Set
VR_VARYING when appropriate.
gcc/
* expr.cc (store_expr): Identify trailing NULs in a STRING_CST
initializer and use clear_storage rather than copying the
NULs to the destination array.
The comment is obsolete because char_traits.h do not include stl_algobase.h
anymore and stl_algobase.h is included directly from <string> a few lines
below.
libstdc++-v3/ChangeLog:
* include/std/string: Remove obsolete comment about char_traits.h including
stl_algobase.h.
The patch tweaks several peephole2s in i386.md that propagate the flags
register, but take its mode from the SET_SRC rather than preserve the
mode of the original SET_DEST. This encounters problems when the
SET_SRC is a VOIDmode CONST_INT. Fixed by using match_operand with a
flags_reg_operand predicate.
2022-07-09 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386.md (define_peephole2): Use match_operand of
flags_reg_operand to capture and preserve the mode of FLAGS_REG.
(define_peephole2): Likewise.
(define_peephole2): Likewise...
This patch fixes the current two FAILs of pr65105-5.c on x86 when
compiled with -m32. These (temporary) breakages were fallout from my
patches to improve/upgrade (scalar) double word comparisons.
On mainline, the i386 backend currently represents a critical comparison
using (compare (and (not reg1) reg2) (const_int 0)) which isn't/wasn't
recognized by the STV pass' convertible_comparison_p. This simple STV
patch adds support for this pattern (*testdi_not_doubleword) and
generates the vector pandn and ptest instructions expected in the
existing (failing) test case.
2022-07-09 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-features.cc (convert_compare): Add support
for *testdi_not_doubleword pattern, "(compare (and (not ...)))"
by generating a pandn followed by ptest.
(convertible_comparison_p): Recognize both *cmpdi_doubleword and
recent *testdi_not_doubleword comparison patterns.
Hi All,
I seem to have broken the s390 bootstrap because I added a new parameter to the
store_bit_field function to indicate whether the value the field of is being set
is currently undefined.
If it's undefined we use a subreg instead. In this case the value of false
restores the old behavior.
Ok for master?
Thanks,
Tamar
gcc/ChangeLog:
* config/s390/s390.cc (s390_expand_atomic): Pass false to store_bit_field to
indicate that the value is not undefined.
The problem here is that when we mark the ssa name that was referenced in the now removed
dead store (to a write only static variable), the inline-asm would also be removed
even though it was defining another ssa name. This fixes the problem by checking
to make sure that the statement was only defining one ssa name.
Committed as approved after a bootstrapped and tested on x86_64 with no regressions.
PR tree-optimization/106087
gcc/ChangeLog:
* tree-ssa-dce.cc (simple_dce_from_worklist): Check
to make sure the statement is only defining one operand.
gcc/testsuite/ChangeLog:
* gcc.c-torture/compile/inline-asm-1.c: New test.
QNX uses sys/link.h rather than link.h for dl_iterate_phdr
Fixes https://github.com/ianlancetaylor/libbacktrace/issues/86
* configure.ac: Check for sys/link.h. Use either link.h or
sys/link.h when checking for dl_iterate_phdr.
* elf.c: Include sys/link.h if available.
* configure, config.h.in: Regenerate.
For gcc.dg/tree-ssa/alias-access-path-13.c to work, SRA must think of
accesses to foo.inn.val and to foo itself as different ones, i.e. they
need to have different offset and size, which on 32bit platforms they
do not. Fixed by replacing a dummy long int field of the union with a
struct of two integers.
Tested by:
make -k check-gcc RUNTESTFLAGS="tree-ssa.exp=alias-access-path-13.c" and
make -k check-gcc RUNTESTFLAGS="--target_board=unix'{-m32}' tree-ssa.exp=alias-access-path-13.c"
on an x86_64-linux, also with patched SRA to verify it still tests the
original intent.
gcc/testsuite/ChangeLog:
2022-07-08 Martin Jambor <mjambor@suse.cz>
PR testsuite/106216
* gcc.dg/tree-ssa/alias-access-path-13.c (union foo): Replace a long
int field with a struct that is larger than an int also on 32bit
platforms.
libcpp recognizes a lone \r as a valid line ending, so the infrastructure
for retrieving source lines to be output in diagnostics needs to do the
same. This patch fixes file_cache_slot::get_next_line() accordingly so that
diagnostics display the correct part of the source when \r line endings are in
use.
gcc/ChangeLog:
PR preprocessor/91733
* input.cc (find_end_of_line): New helper function.
(file_cache_slot::get_next_line): Recognize \r as a line ending.
* diagnostic-show-locus.cc (test_escaping_bytes_1): Adapt selftest
since \r will now be interpreted as a line-ending.
gcc/testsuite/ChangeLog:
PR preprocessor/91733
* c-c++-common/pr91733.c: New test.
Split report_conflicting_sanitizer_options(..., SANITIZE_ADDRESS | SANITIZE_HWADDRESS)
call into 2 calls as we don't have any option that would be
address+hwaddress (that conflicts) as well.
PR sanitizer/106132
gcc/ChangeLog:
* opts.cc (finish_options): Use 2 calls to
report_conflicting_sanitizer_options.
gcc/testsuite/ChangeLog:
* c-c++-common/hwasan/arguments-3.c: Cover new ICE.
When we knowingly have broken virtual SSA form we need to update
it before we eventually perform slpeel manual updating which will
call delete_update_ssa. Currently that's done on-demand but
communicating whether it's a known unavoidable case is broken
there. The following makes that a synchronous operation but
instead of actually performing the update we instead recod the
need, clear the update SSA sub-state and force virtual renaming
at the very end of the vectorization pass.
PR tree-optimization/106226
* tree-vect-loop-manip.cc (vect_do_peeling): Assert that
no SSA update is needed. Move virtual SSA update ...
* tree-vectorizer.cc (pass_vectorize::execute): ... here,
via forced virtual renaming when TODO_update_ssa_only_virtuals
is queued.
(vect_transform_loops): Return TODO_update_ssa_only_virtuals
when virtual SSA update is required.
(try_vectorize_loop_1): Adjust.
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Allow
virtual renaming if the ABI forces an aggregate return
but the original call did not have a virtual definition.
* gfortran.dg/pr106226.f: New testcase.
Right now the following is printed:
lto-dump
.file "<artificial>"
.ident "GCC: (GNU) 13.0.0 20220707 (experimental)"
.section .note.GNU-stack,"",@progbits
After the patch we print -help and do not emit any assembly output:
lto-dump
Usage: lto-dump [OPTION]... SUB_COMMAND [OPTION]...
LTO dump tool command line options.
-list [options] Dump the symbol list.
-demangle Dump the demangled output.
-defined-only Dump only the defined symbols.
...
gcc/lto/ChangeLog:
* lto-dump.cc (lto_main): Exit in the function
as we don't want any LTO bytecode processing.
gcc/ChangeLog:
* toplev.cc (init_asm_output): Do not init asm_out_file.
Hi All,
My previous patch can cause a problem if the pattern matches after veclower
as it may replace the construct with a vector sequence which the target may not
directly support.
As such don't perform the rewriting if after veclower unless the target supports
the operation. If before veclower do the rewriting as well if the target didn't
support the original operation either.
gcc/ChangeLog:
PR tree-optimization/106063
* match.pd: Do not apply pattern after veclower is not supported.
gcc/testsuite/ChangeLog:
PR tree-optimization/106063
* gcc.dg/pr106063.c: New test.
When lowering COMPLEX_EXPR we currently emit two VEC_EXTRACTs. One for the
lowpart and one for the highpart.
The problem with this is that in RTL the lvalue of the RTX is the only thing
tying the two instructions together.
This means that e.g. combine is unable to try to combine the two instructions
for setting the lowpart and highpart.
For ISAs that have bit extract instructions we can eliminate one of the extracts
if, and only if we're setting the entire complex number.
This change changes the expand code when we're setting the entire complex number
to generate a subreg for the lowpart instead of a vec_extract.
This allows us to optimize sequences such as:
_Complex int f(int a, int b) {
_Complex int t = a + b * 1i;
return t;
}
from:
f:
bfi x2, x0, 0, 32
bfi x2, x1, 32, 32
mov x0, x2
ret
into:
f:
bfi x0, x1, 32, 32
ret
I have also confirmed the codegen for x86_64 did not change.
gcc/ChangeLog:
* expmed.cc (store_bit_field_1): Add parameter that indicates if value is
still undefined and if so emit a subreg move instead.
(store_integral_bit_field): Likewise.
(store_bit_field): Likewise.
* expr.h (write_complex_part): Likewise.
* expmed.h (store_bit_field): Add new parameter.
* builtins.cc (expand_ifn_atomic_compare_exchange_into_call): Use new
parameter.
(expand_ifn_atomic_compare_exchange): Likewise.
* calls.cc (store_unaligned_arguments_into_pseudos): Likewise.
* emit-rtl.cc (validate_subreg): Likewise.
* expr.cc (emit_group_store): Likewise.
(copy_blkmode_from_reg): Likewise.
(copy_blkmode_to_reg): Likewise.
(clear_storage_hints): Likewise.
(write_complex_part): Likewise.
(emit_move_complex_parts): Likewise.
(expand_assignment): Likewise.
(store_expr): Likewise.
(store_field): Likewise.
(expand_expr_real_2): Likewise.
* ifcvt.cc (noce_emit_move_insn): Likewise.
* internal-fn.cc (expand_arith_set_overflow): Likewise.
(expand_arith_overflow_result_store): Likewise.
(expand_addsub_overflow): Likewise.
(expand_neg_overflow): Likewise.
(expand_mul_overflow): Likewise.
(expand_arith_overflow): Likewise.
gcc/testsuite/ChangeLog:
* g++.target/aarch64/complex-init.C: New test.
Under the LA architecture, when the stack is dropped too far, the process
of dropping the stack is divided into two steps.
step1: After dropping the stack, save callee saved registers on the stack.
step2: The rest of it.
The stack drop operation is optimized when frame->total_size minus
frame->sp_fp_offset is an integer multiple of 4096, can reduce the number
of instructions required to drop the stack. However, this optimization is
not effective because of the original calculation method
The following case:
int main()
{
char buf[1024 * 12];
printf ("%p\n", buf);
return 0;
}
As you can see from the generated assembler, the old GCC has two more
instructions than the new GCC, lines 14 and line 24.
new old
10 main: | 11 main:
11 addi.d $r3,$r3,-16 | 12 lu12i.w $r13,-12288>>12
12 lu12i.w $r13,-12288>>12 | 13 addi.d $r3,$r3,-2032
13 lu12i.w $r5,-12288>>12 | 14 ori $r13,$r13,2016
14 lu12i.w $r12,12288>>12 | 15 lu12i.w $r5,-12288>>12
15 st.d $r1,$r3,8 | 16 lu12i.w $r12,12288>>12
16 add.d $r12,$r12,$r5 | 17 st.d $r1,$r3,2024
17 add.d $r3,$r3,$r13 | 18 add.d $r12,$r12,$r5
18 add.d $r5,$r12,$r3 | 19 add.d $r3,$r3,$r13
19 la.local $r4,.LC0 | 20 add.d $r5,$r12,$r3
20 bl %plt(printf) | 21 la.local $r4,.LC0
21 lu12i.w $r13,12288>>12 | 22 bl %plt(printf)
22 add.d $r3,$r3,$r13 | 23 lu12i.w $r13,8192>>12
23 ld.d $r1,$r3,8 | 24 ori $r13,$r13,2080
24 or $r4,$r0,$r0 | 25 add.d $r3,$r3,$r13
25 addi.d $r3,$r3,16 | 26 ld.d $r1,$r3,2024
26 jr $r1 | 27 or $r4,$r0,$r0
| 28 addi.d $r3,$r3,2032
| 29 jr $r1
gcc/ChangeLog:
* config/loongarch/loongarch.cc (loongarch_compute_frame_info):
Modify fp_sp_offset and gp_sp_offset's calculation method,
when frame->mask or frame->fmask is zero, don't minus UNITS_PER_WORD
or UNITS_PER_FP_REG.
gcc/testsuite/ChangeLog:
* gcc.target/loongarch/prolog-opt.c: New test.
In r13-1045-gcb7fd1ea85feea I assumed that substitution into generic
DECL_TI_ARGS corresponds to an identity mapping of the given arguments,
and hence its safe to always elide such substitution. But this PR
demonstrates that such a substitution isn't always the identity mapping,
in particular when there's an ARGUMENT_PACK_SELECT argument, which gets
handled specially during substitution:
* when substituting an APS into a template parameter, we strip the
APS to its underlying argument;
* and when substituting an APS into a pack expansion, we strip the
APS to its underlying argument pack.
In this testcase, when expanding the pack expansion pattern (idx + Ns)...
with Ns={0,1}, we specialize idx twice, first with Ns=APS<0,{0,1}> and
then Ns=APS<1,{0,1}>. The DECL_TI_ARGS of idx are the generic template
arguments of the enclosing class template impl, so before r13-1045,
we'd substitute into its DECL_TI_ARGS which gave Ns={0,1} as desired.
But after r13-1045, we elide this substitution and end up attempting to
hash the original Ns argument, an APS, which ICEs.
So this patch reverts that part of r13-1045. I considered using
preserve_args in this case instead, but that'd break the static_assert
in the testcase because preserve_args always strips APS to its
underlying argument, but here we want to strip it to its underlying
argument pack, so we'd incorrectly end up forming the specializations
impl<0>::idx and impl<1>::idx instead of impl<0,1>::idx.
Although we can't elide the substitution into DECL_TI_ARGS in light of
ARGUMENT_PACK_SELECT, it should still be safe to elide template argument
coercion in the case of a non-template decl, which this patch preserves.
It's unfortunate that we need to remove this optimization just because
it doesn't hold for one special tree code. So this patch implements a
heuristic in tsubst_template_args to avoid allocating a new TREE_VEC if
the substituted elements are identical to those of a level from ARGS, as
well as a similar heuristic for tsubst_argument_pack. It turns out that
about 40% of all calls to tsubst_template_args benefit from this, and it
reduces memory usage by about 4% for e.g. range-v3's zip.cpp (relative to
r13-1045) which more than makes up for the reversion.
PR c++/105956
gcc/cp/ChangeLog:
* pt.cc (template_arg_to_parm): Define.
(tsubst_argument_pack): Try to reuse the corresponding
ARGUMENT_PACK from 'args' when substituting into a generic
ARGUMENT_PACK for a variadic template parameter.
(tsubst_template_args): Move variable declarations closer to
their first use. Replace 'orig_t' with 'r'. Rename 'need_new'
to 'const_subst_p'. Heuristically detect if the substituted
elements are identical to that of a level from 'args' and avoid
allocating a new TREE_VEC if so. Add sanity check for the
length of the new TREE_VEC, and remove dead ARGUMENT_PACK_P test.
(tsubst_decl) <case TYPE_DECL, case VAR_DECL>: Revert
r13-1045-gcb7fd1ea85feea change for avoiding substitution into
DECL_TI_ARGS, but still avoid coercion in this case.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/variadic183.C: New test.
libcpp's class label_text stores a char * for a string and a flag saying
whether it owns the buffer. I added this class before we could use
C++11, and so to avoid lots of copying it required an explicit call
to label_text::maybe_free to potentially free the buffer.
Now that we can use C++11, this patch removes label_text::maybe_free in
favor of doing the cleanup in the destructor, and using C++ move
semantics to avoid any copying. This allows lots of messy cleanup code
to be eliminated in favor of implicit destruction (mostly in the
analyzer).
No functional change intended.
gcc/analyzer/ChangeLog:
* call-info.cc (call_info::print): Update for removal of
label_text::maybe_free in favor of automatic memory management.
* checker-path.cc (checker_event::dump): Likewise.
(checker_event::prepare_for_emission): Likewise.
(state_change_event::get_desc): Likewise.
(superedge_event::should_filter_p): Likewise.
(start_cfg_edge_event::get_desc): Likewise.
(warning_event::get_desc): Likewise.
(checker_path::dump): Likewise.
(checker_path::debug): Likewise.
* diagnostic-manager.cc
(diagnostic_manager::prune_for_sm_diagnostic): Likewise.
(diagnostic_manager::prune_interproc_events): Likewise.
* program-state.cc (sm_state_map::to_json): Likewise.
* region.cc (region::to_json): Likewise.
* sm-malloc.cc (inform_nonnull_attribute): Likewise.
* store.cc (binding_map::to_json): Likewise.
(store::to_json): Likewise.
* svalue.cc (svalue::to_json): Likewise.
gcc/c-family/ChangeLog:
* c-format.cc (range_label_for_format_type_mismatch::get_text):
Update for removal of label_text::maybe_free in favor of automatic
memory management.
gcc/ChangeLog:
* diagnostic-format-json.cc (json_from_location_range): Update for
removal of label_text::maybe_free in favor of automatic memory
management.
* diagnostic-format-sarif.cc
(sarif_builder::make_location_object): Likewise.
* diagnostic-show-locus.cc (struct pod_label_text): New.
(class line_label): Convert m_text from label_text to pod_label_text.
(layout::print_any_labels): Move "text" to the line_label.
* tree-diagnostic-path.cc (path_label::get_text): Update for
removal of label_text::maybe_free in favor of automatic memory
management.
(event_range::print): Likewise.
(default_tree_diagnostic_path_printer): Likewise.
(default_tree_make_json_for_path): Likewise.
libcpp/ChangeLog:
* include/line-map.h: Include <utility>.
(class label_text): Delete maybe_free method in favor of a
destructor. Add move ctor and assignment operator. Add deletion
of the copy ctor and copy-assignment operator. Rename field
m_caller_owned to m_owned. Add std::move where necessary; add
moved_from member function.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
gcc/analyzer/ChangeLog:
PR analyzer/106225
* sm-taint.cc (taint_state_machine::on_stmt): Move handling of
assignments from division to...
(taint_state_machine::check_for_tainted_divisor): ...this new
function. Reject warning when the divisor is known to be non-zero.
* sm.cc: Include "analyzer/program-state.h".
(sm_context::get_old_region_model): New.
* sm.h (sm_context::get_old_region_model): New decl.
gcc/testsuite/ChangeLog:
PR analyzer/106225
* gcc.dg/analyzer/taint-divisor-1.c: Add test coverage for various
correct and incorrect checks against zero.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
The front-end bug that prevented this constexpr loop from working has
been fixed since GCC 12.1 so we can remove the workaround.
libstdc++-v3/ChangeLog:
PR c++/89074
* include/bits/char_traits.h (__gnu_cxx::char_traits::move):
Remove workaround for front-end bug.
