In r208350 I improved the diagnostic location of the initializer-list
pedwarn in C++98 mode on crash90.C, but didn't adjust the testcase to verify
the location, so reverting that change didn't break regression testing.
gcc/testsuite/ChangeLog:
* g++.dg/template/crash90.C: Check location of pedwarn.
This fixes handling of the return value of vect_supportable_dr_alignment
in multiple places. We should use the enum type and not int for
storage and not auto-convert the enum return value to bool. It also
commonizes the read/write path in vect_supportable_dr_alignment.
2021-10-18 Richard Biener <rguenther@suse.de>
* tree-vect-data-refs.c (vect_peeling_hash_insert): Do
not auto-convert dr_alignment_support to bool.
(vect_peeling_supportable): Likewise.
(vect_enhance_data_refs_alignment): Likewise.
(vect_supportable_dr_alignment): Commonize read/write case.
* tree-vect-stmts.c (vect_get_store_cost): Use
dr_alignment_support, not int, for the vect_supportable_dr_alignment
result.
(vect_get_load_cost): Likewise.
This is a minor cleanup to bail out early if the result of
__builtin_object_size is not assigned to anything and avoid initializing
the object size arrays.
gcc/ChangeLog:
* tree-object-size.c (object_sizes_execute): Consolidate LHS
null check and do it early.
Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>
This uses the computed alignment scheme in vectorizable_store
much like vectorizable_load does instead of re-querying
it via aligned_access_p.
2021-10-18 Richard Biener <rguenther@suse.de>
* tree-vect-stmts.c (vectorizable_store): Use the
computed alignment scheme instead of querying
aligned_access_p.
The following avoids the recomputation of the alignment scheme
which is already fully determined by get_load_store_type.
2021-10-18 Richard Biener <rguenther@suse.de>
* tree-vect-stmts.c (vectorizable_store): Do not recompute
alignment scheme already determined by get_load_store_type.
If numa-domains is used with num-places count, sometimes the function
could create more places than requested and crash. This depended on the
content of /sys/devices/system/node/online file, e.g. if the file
contains
0-1,16-17
and all NUMA nodes contain at least one CPU in the cpuset of the program,
then numa_domains(2) or numa_domains(4) (or 5+) work fine while
numa_domains(1) or numa_domains(3) misbehave. I.e. the function was able
to stop after reaching limit on the , separators (or trivially at the end),
but not within in the ranges.
2021-10-18 Jakub Jelinek <jakub@redhat.com>
* config/linux/affinity.c (gomp_affinity_init_numa_domains): Add
&& gomp_places_list_len < count after nfirst <= nlast loop condition.
On x86-64,
$ make check RUNTESTFLAGS="--target_board='unix{-m32,}'"
can be used to test both 64-bit and 32-bit targets. Require ia32 target
instead of explicit -m32 for 32-bit only test.
* gcc.target/i386/387-12.c (dg-do compile): Require ia32.
(dg-options): Remove -m32.
My recent attempts to come up with a testcase for my patch to evaluate
ss_plus in simplify-rtx.c, identified a missed optimization opportunity
(that's potentially a long-time regression): The RTL optimizers no longer
place constants in the constant pool.
The motivating x86_64 example is the simple program:
typedef char v8qi __attribute__ ((vector_size (8)));
v8qi foo()
{
v8qi tx = { 1, 0, 0, 0, 0, 0, 0, 0 };
v8qi ty = { 2, 0, 0, 0, 0, 0, 0, 0 };
v8qi t = __builtin_ia32_paddsb(tx, ty);
return t;
}
which (with my previous patch) currently results in:
foo: movq .LC0(%rip), %xmm0
movq .LC1(%rip), %xmm1
paddsb %xmm1, %xmm0
ret
even though the RTL contains the result in a REG_EQUAL note:
(insn 7 6 12 2 (set (reg:V8QI 83)
(ss_plus:V8QI (reg:V8QI 84)
(reg:V8QI 85))) "ssaddqi3.c":7:12 1419 {*mmx_ssaddv8qi3}
(expr_list:REG_DEAD (reg:V8QI 85)
(expr_list:REG_DEAD (reg:V8QI 84)
(expr_list:REG_EQUAL (const_vector:V8QI [
(const_int 3 [0x3])
(const_int 0 [0]) repeated x7
])
(nil)))))
Together with the patch below, GCC will now generate the much
more sensible:
foo: movq .LC2(%rip), %xmm0
ret
My first approach was to look in cse.c (where the REG_EQUAL note gets
added) and notice that the constant pool handling functionality has been
unreachable for a while. A quick search for constant_pool_entries_cost
shows that it's initialized to zero, but never set to a non-zero value,
meaning that force_const_mem is never called. This functionality used
to work way back in 2003, but has been lost over time:
https://gcc.gnu.org/pipermail/gcc-patches/2003-October/116435.html
The changes to cse.c below restore this functionality (placing suitable
constants in the constant pool) with two significant refinements;
(i) it only attempts to do this if the function already uses a constant
pool (thanks to the availability of crtl->uses_constant_pool since 2003).
(ii) it allows different constants (i.e. modes) to have different costs,
so that floating point "doubles" and 64-bit, 128-bit, 256-bit and 512-bit
vectors don't all have the share the same cost. Back in 2003, the
assumption was that everything in a constant pool had the same
cost, hence the global variable constant_pool_entries_cost.
Although this is a useful CSE fix, it turns out that it doesn't cure
my motivating problem above. CSE only considers a single instruction,
so determines that it's cheaper to perform the ss_plus (COSTS_N_INSNS(1))
than read the result from the constant pool (COSTS_N_INSNS(2)). It's
only when the other reads from the constant pool are also eliminated,
that this transformation is a win. Hence a better place to perform
this transformation is in combine, where after failing to "recog" the
load of a suitable constant, it can retry after calling force_const_mem.
This achieves the desired transformation and allows the backend insn_cost
call-back to control whether or not using the constant pool is preferrable.
Alas, it's rare to change code generation without affecting something in
GCC's testsuite. On x86_64-pc-linux-gnu there were two families of new
failures (and I'd predict similar benign fallout on other platforms).
One failure was gcc.target/i386/387-12.c (aka PR target/26915), where
the test is missing an explicit -m32 flag. On i686, it's very reasonable
to materialize -1.0 using "fld1; fchs", but on x86_64-pc-linux-gnu we
currently generate the awkward:
testm1: fld1
fchs
fstpl -8(%rsp)
movsd -8(%rsp), %xmm0
ret
which combine now very reasonably simplifies to just:
testm1: movsd .LC3(%rip), %xmm0
ret
The other class of x86_64-pc-linux-gnu failure was from materialization
of vector constants using vpbroadcast (e.g. gcc.target/i386/pr90773-17.c)
where the decision is finely balanced; the load of an integer register
with an immediate constant, followed by a vpbroadcast is deemed to be
COSTS_N_INSNS(2), whereas a load from the constant pool is also reported
as COSTS_N_INSNS(2). My solution is to tweak the i386.c's rtx_costs
so that all other things being equal, an instruction (sequence) that
accesses memory is fractionally more expensive than one that doesn't.
2021-10-18 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* combine.c (recog_for_combine): For an unrecognized move/set of
a constant, try force_const_mem to place it in the constant pool.
* cse.c (constant_pool_entries_cost, constant_pool_entries_regcost):
Delete global variables (that are no longer assigned a cost value).
(cse_insn): Simplify logic for deciding whether to place a folded
constant in the constant pool using force_const_mem.
(cse_main): Remove zero initialization of constant_pool_entries_cost
and constant_pool_entries_regcost.
* config/i386/i386.c (ix86_rtx_costs): Make memory accesses
fractionally more expensive, when optimizing for speed.
gcc/testsuite/ChangeLog
* gcc.target/i386/387-12.c: Add explicit -m32 option.
Blackfin processors support a ONES instruction that implements a
32-bit popcount returning a 16-bit result. This instruction was
previously described by GCC's bfin backend using an UNSPEC, which
this patch changes to use a popcount:SI rtx thats capture its semantics,
allowing it to evaluated and simplified at compile-time. I've decided
to keep the instruction name the same (avoiding any changes to the
__builtin_bfin_ones machinery), but have provided popcountsi2 and
popcounthi2 expanders so that the middle-end can use this instruction
to implement __builtin_popcount (and __builtin_parity).
The new testcase ones.c
short foo ()
{
int t = 5;
short r = __builtin_bfin_ones(t);
return r;
}
previously generated:
_foo: nop;
nop;
R0 = 5 (X);
R0.L = ONES R0;
rts;
with this patch, now generates:
_foo: nop;
nop;
nop;
R0 = 2 (X);
rts;
The new testcase popcount.c
int foo(int x)
{
return __builtin_popcount(x);
}
previously generated:
_foo: [--SP] = RETS;
SP += -12;
call ___popcountsi2;
SP += 12;
RETS = [SP++];
rts;
now generates:
_foo: nop;
nop;
R0.L = ONES R0;
R0 = R0.L (Z);
rts;
And the new testcase parity.c
int foo(int x)
{
return __builtin_parity(x);
}
previously generated:
_foo: [--SP] = RETS;
SP += -12;
call ___paritysi2;
SP += 12;
RETS = [SP++];
rts;
now generates:
_foo: nop;
R1 = 1 (X);
R0.L = ONES R0;
R0 = R1 & R0;
rts;
2021-10-18 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/bfin/bfin.md (define_constants): Remove UNSPEC_ONES.
(define_insn "ones"): Replace UNSPEC_ONES with a truncate of
a popcount, allowing compile-time evaluation/simplification.
(popcountsi2, popcounthi2): New expanders using a "ones" insn.
gcc/testsuite/ChangeLog
* gcc.target/bfin/ones.c: New test case.
* gcc.target/bfin/parity.c: New test case.
* gcc.target/bfin/popcount.c: New test case.
Bool pattern recog is required for correctness since vectorized
compares otherwise produce -1 for true so any context where bool
is used as value and not as condition or mask needs to be replaced
with CMP ? 1 : 0. When we fail to find a vector type for the
result of such use we may not simply elide such transform since
a new bool result can emerge when for example the cast_forwprop
pattern is applied. So the following avoids failing of the
bool pattern recog process and instead not assign a vector type
for the stmt.
2021-10-18 Richard Biener <rguenther@suse.de>
PR tree-optimization/102788
* tree-vect-patterns.c (vect_init_pattern_stmt): Allow
a NULL vectype.
(vect_pattern_recog_1): Likewise.
(vect_recog_bool_pattern): Continue matching the pattern
even if we do not have a vector type for a conversion
result.
* g++.dg/vect/pr102788.cc: New testcase.
This simple patch performs compile-time constant folding of
signed saturating negation and signed saturating absolute value
in the RTL optimizers. Normally in two's complement arithmetic
the lowest representable signed value overflows on negation,
With these saturating operators they "saturate" to the maximum
representable signed value, so SS_NEG:QI -128 is 127, and
SS_ABS:HI -32768 is 32767.
On bfin-elf, the following two short functions:
short foo()
{
short t = -32768;
short r = __builtin_bfin_negate_fr1x16(t);
return r;
}
int bar()
{
int t = -2147483648;
int r = __builtin_bfin_abs_fr1x32(t);
return r;
}
currently compile to:
_foo: nop;
nop;
R0 = -32768 (X);
R0 = -R0 (V);
rts;
_bar: nop;
R0 = -1 (X);
R0 <<= 31;
R0 = abs R0;
rts;
but with this middle-end patch now compile to:
_foo: nop;
nop;
nop;
R0 = 32767 (X);
rts;
_bar: nop;
nop;
R0 = -1 (X);
R0.H = 32767;
rts;
2021-10-18 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* simplify-rtx.c (simplify_const_unary_operation) [SS_NEG, SS_ABS]:
Evalute SS_NEG and SS_ABS of a constant argument.
gcc/testsuite/ChangeLog
* gcc.target/bfin/ssabs.c: New test case.
* gcc.target/bfin/ssneg.c: New test case.
This refactors the strlen pass to avoid passing around as much state.
It is meant to be a start. There's still some more refactoring
that could be done, especially in cleaning up the interface between
the strlen internals and the sprintf pass.
Tested on x86-64 Linux.
gcc/ChangeLog:
* tree-ssa-strlen.c (class strlen_pass): Rename from
strlen_dom_walker.
(handle_builtin_stxncpy_strncat): Move to strlen_pass.
(handle_assign): Same.
(adjust_last_stmt): Same.
(maybe_warn_overflow): Same.
(handle_builtin_strlen): Same.
(handle_builtin_strchr): Same.
(handle_builtin_strcpy): Same.
(handle_builtin_strncat): Same.
(handle_builtin_stxncpy_strncat): Same.
(handle_builtin_memcpy): Same.
(handle_builtin_strcat): Same.
(handle_alloc_call): Same.
(handle_builtin_memset): Same.
(handle_builtin_memcmp): Same.
(get_len_or_size): Same.
(strxcmp_eqz_result): Same.
(handle_builtin_string_cmp): Same.
(handle_pointer_plus): Same.
(count_nonzero_bytes_addr): Same.
(count_nonzero_bytes): Same.
(handle_store): Same.
(strlen_check_and_optimize_call): Same.
(handle_integral_assign): Same.
(check_and_optimize_stmt): Same.
(printf_strlen_execute): Rename strlen_dom_walker to strlen_pass.
gfortran uses internally a different array descriptor ("gfc") as
Fortran 2018 alias TS291113 defines for C interoperability via
ISO_Fortran_binding.h ("CFI"). Hence, when calling a C function
from Fortran, it has to be converted in the callee - and if a
BIND(C) procedure is written in Fortran, the CFI argument has
to be converted to gfc in order work with the rest of the FE
code and the library calls.
Before this patch, part was handled in the FE generated code and
other parts in libgfortran. With this patch, all code is generated
and CFI is defined as proper type - visible in the debugger and to
the middle end - avoiding both alias issues and missed optimization
issues.
This patch also fixes issues like: intent(out) deallocation in
the bind(C) callee, using the CFI descriptor also for allocatable
and pointer scalars and for len=* character strings.
For 'select rank', it also optimizes the code + avoid accessing
uninitialized memory if the dummy argument is allocatable/a pointer.
It additionally rejects passing a descriptorless type(*) to an
assumed-rank dummy argument. [F2018:C711]
PR fortran/102086
PR fortran/92189
PR fortran/92621
PR fortran/101308
PR fortran/101309
PR fortran/101635
PR fortran/92482
gcc/fortran/ChangeLog:
* decl.c (gfc_verify_c_interop_param): Remove 'sorry' for
scalar allocatable/pointer and len=*.
* expr.c (is_CFI_desc): Return true for for those.
* gfortran.h (CFI_type_kind_shift, CFI_type_mask,
CFI_type_from_type_kind, CFI_VERSION, CFI_MAX_RANK,
CFI_attribute_pointer, CFI_attribute_allocatable,
CFI_attribute_other, CFI_type_Integer, CFI_type_Logical,
CFI_type_Real, CFI_type_Complex, CFI_type_Character,
CFI_type_ucs4_char, CFI_type_struct, CFI_type_cptr,
CFI_type_cfunptr, CFI_type_other): New #define.
* trans-array.c (CFI_FIELD_BASE_ADDR, CFI_FIELD_ELEM_LEN,
CFI_FIELD_VERSION, CFI_FIELD_RANK, CFI_FIELD_ATTRIBUTE,
CFI_FIELD_TYPE, CFI_FIELD_DIM, CFI_DIM_FIELD_LOWER_BOUND,
CFI_DIM_FIELD_EXTENT, CFI_DIM_FIELD_SM,
gfc_get_cfi_descriptor_field, gfc_get_cfi_desc_base_addr,
gfc_get_cfi_desc_elem_len, gfc_get_cfi_desc_version,
gfc_get_cfi_desc_rank, gfc_get_cfi_desc_type,
gfc_get_cfi_desc_attribute, gfc_get_cfi_dim_item,
gfc_get_cfi_dim_lbound, gfc_get_cfi_dim_extent, gfc_get_cfi_dim_sm):
New define/functions to access the CFI array descriptor.
(gfc_conv_descriptor_type): New function for the GFC descriptor.
(gfc_get_array_span): Handle expr of CFI descriptors and
assumed-type descriptors.
(gfc_trans_array_bounds): Remove 'static'.
(gfc_conv_expr_descriptor): For assumed type, use the dtype of
the actual argument.
(structure_alloc_comps): Remove ' ' inside tabs.
* trans-array.h (gfc_trans_array_bounds, gfc_conv_descriptor_type,
gfc_get_cfi_desc_base_addr, gfc_get_cfi_desc_elem_len,
gfc_get_cfi_desc_version, gfc_get_cfi_desc_rank,
gfc_get_cfi_desc_type, gfc_get_cfi_desc_attribute,
gfc_get_cfi_dim_lbound, gfc_get_cfi_dim_extent, gfc_get_cfi_dim_sm):
New prototypes.
* trans-decl.c (gfor_fndecl_cfi_to_gfc, gfor_fndecl_gfc_to_cfi):
Remove global vars.
(gfc_build_builtin_function_decls): Remove their initialization.
(gfc_get_symbol_decl, create_function_arglist,
gfc_trans_deferred_vars): Update for CFI.
(convert_CFI_desc): Remove and replace by ...
(gfc_conv_cfi_to_gfc): ... this function
(gfc_generate_function_code): Call it; create local GFC var for CFI.
* trans-expr.c (gfc_maybe_dereference_var): Handle CFI.
(gfc_conv_subref_array_arg): Handle the if-noncontigous-only copy in
when the result should be a descriptor.
(gfc_conv_gfc_desc_to_cfi_desc): Completely rewritten.
(gfc_conv_procedure_call): CFI fixes.
* trans-openmp.c (gfc_omp_is_optional_argument,
gfc_omp_check_optional_argument): Handle optional
CFI.
* trans-stmt.c (gfc_trans_select_rank_cases): Cleanup, avoid invalid
code for allocatable/pointer dummies, which cannot be assumed size.
* trans-types.c (gfc_cfi_descriptor_base): New global var.
(gfc_get_dtype_rank_type): Skip rank init for rank < 0.
(gfc_sym_type): Handle CFI dummies.
(gfc_get_function_type): Update call.
(gfc_get_cfi_dim_type, gfc_get_cfi_type): New.
* trans-types.h (gfc_sym_type): Update prototype.
(gfc_get_cfi_type): New prototype.
* trans.c (gfc_trans_runtime_check): Make conditions more consistent
to avoid '<logical> AND_THEN <long int>' in conditions.
* trans.h (gfor_fndecl_cfi_to_gfc, gfor_fndecl_gfc_to_cfi): Remove
global-var declaration.
libgfortran/ChangeLog:
* ISO_Fortran_binding.h (CFI_type_cfunptr): Make unique type again.
* runtime/ISO_Fortran_binding.c (cfi_desc_to_gfc_desc,
gfc_desc_to_cfi_desc): Add comment that those are no longer called
by new code.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/optional-bind-c.f90: New test.
gcc/testsuite/ChangeLog:
* gfortran.dg/ISO_Fortran_binding_4.f90: Extend testcase.
* gfortran.dg/PR100914.f90: Remove xfail.
* gfortran.dg/PR100915.c: Expect CFI_type_cfunptr.
* gfortran.dg/PR100915.f90: Handle CFI_type_cfunptr != CFI_type_cptr.
* gfortran.dg/PR93963.f90: Extend select-rank tests.
* gfortran.dg/bind-c-intent-out.f90: Change to dg-do run,
update scan-dump.
* gfortran.dg/bind_c_array_params_2.f90: Update/extend scan-dump.
* gfortran.dg/bind_c_char_10.f90: Update scan-dump.
* gfortran.dg/bind_c_char_8.f90: Remove dg-error "sorry".
* gfortran.dg/c-interop/allocatable-dummy.f90: Remove xfail.
* gfortran.dg/c-interop/c1255-1.f90: Likewise.
* gfortran.dg/c-interop/c407c-1.f90: Update dg-error.
* gfortran.dg/c-interop/cf-descriptor-5.f90: Remove xfail.
* gfortran.dg/c-interop/cf-out-descriptor-3.f90: Likewise.
* gfortran.dg/c-interop/cf-out-descriptor-4.f90: Likewise.
* gfortran.dg/c-interop/cf-out-descriptor-5.f90: Likewise.
* gfortran.dg/c-interop/contiguous-2.f90: Likewise.
* gfortran.dg/c-interop/contiguous-3.f90: Likewise.
* gfortran.dg/c-interop/deferred-character-1.f90: Likewise.
* gfortran.dg/c-interop/deferred-character-2.f90: Likewise.
* gfortran.dg/c-interop/fc-descriptor-3.f90: Likewise.
* gfortran.dg/c-interop/fc-descriptor-5.f90: Likewise.
* gfortran.dg/c-interop/fc-descriptor-6.f90: Likewise.
* gfortran.dg/c-interop/fc-out-descriptor-3.f90: Likewise.
* gfortran.dg/c-interop/fc-out-descriptor-4.f90: Likewise.
* gfortran.dg/c-interop/fc-out-descriptor-5.f90: Likewise.
* gfortran.dg/c-interop/fc-out-descriptor-6.f90: Likewise.
* gfortran.dg/c-interop/ff-descriptor-5.f90: Likewise.
* gfortran.dg/c-interop/ff-descriptor-6.f90: Likewise.
* gfortran.dg/c-interop/fc-descriptor-7.f90: Remove xfail + extend.
* gfortran.dg/c-interop/fc-descriptor-7-c.c: Update for changes.
* gfortran.dg/c-interop/shape.f90: Add implicit none.
* gfortran.dg/c-interop/typecodes-array-char-c.c: Add kind=4 char.
* gfortran.dg/c-interop/typecodes-array-char.f90: Likewise.
* gfortran.dg/c-interop/typecodes-array-float128.f90: Remove xfail.
* gfortran.dg/c-interop/typecodes-scalar-basic.f90: Likewise.
* gfortran.dg/c-interop/typecodes-scalar-float128.f90: Likewise.
* gfortran.dg/c-interop/typecodes-scalar-int128.f90: Likewise.
* gfortran.dg/c-interop/typecodes-scalar-longdouble.f90: Likewise.
* gfortran.dg/iso_c_binding_char_1.f90: Remove dg-error "sorry".
* gfortran.dg/pr93792.f90: Turn XFAIL into PASS.
* gfortran.dg/ISO_Fortran_binding_19.f90: New test.
* gfortran.dg/assumed_type_12.f90: New test.
* gfortran.dg/assumed_type_13.c: New test.
* gfortran.dg/assumed_type_13.f90: New test.
* gfortran.dg/bind-c-char-descr.f90: New test.
* gfortran.dg/bind-c-contiguous-1.c: New test.
* gfortran.dg/bind-c-contiguous-1.f90: New test.
* gfortran.dg/bind-c-contiguous-2.f90: New test.
* gfortran.dg/bind-c-contiguous-3.c: New test.
* gfortran.dg/bind-c-contiguous-3.f90: New test.
* gfortran.dg/bind-c-contiguous-4.c: New test.
* gfortran.dg/bind-c-contiguous-4.f90: New test.
* gfortran.dg/bind-c-contiguous-5.c: New test.
* gfortran.dg/bind-c-contiguous-5.f90: New test.
The vectorizer duplicates pointer-info to created pointer bases
but it has to avoid changing points-to info on existing SSA names
because there's now flow-sensitive info in there (pt->pt_null as
set from VRP).
2021-10-18 Richard Biener <rguenther@suse.de>
PR tree-optimization/102798
* tree-vect-data-refs.c (vect_create_addr_base_for_vector_ref):
Only copy points-to info to newly generated SSA names.
* gcc.dg/pr102798.c: New testcase.
PR fortran/102745
gcc/fortran/ChangeLog
* intrinsic.c (gfc_convert_type_warn): Fix checks by checking CLASS
and do typcheck in correct order for type extension.
* misc.c (gfc_typename): Print proper not internal CLASS type name.
gcc/testsuite/ChangeLog
* gfortran.dg/class_72.f90: New.
This refactors the backward threader pass so that it can be called in
either fully resolving mode, or in classic mode where any unknowns
default to VARYING. Doing so opens the door for
"pass_thread_jumps_full" which has the resolving bits set.
This pass has not been added to the pipeline, but with it in place, we
can now experiment with it to see how to reduce the number of
jump threaders. The first suspect will probably be enabling fully
resolving in the backward threader pass immediately preceeding VRP2,
and removing the VRP2 threader pass. Now that VRP and the backward
threader are sharing a solver, and most of the threads get handcuffed
by cancel_threads(), we should have a variety of scenarios to try.
In the process, I have cleaned up things to make it trivial to see
what the difference between the 3 variants are (early jump
threading, quick jump threading without resolving SSAs, and fully
resolving jump threading). Since I moved stuff around, it's probably
easier to just look at the last section in tree-ssa-threadbackward to
see how it's all laid out.
No functional changes as the new pass hasn't been added to the
pipeline.
gcc/ChangeLog:
* tree-pass.h (make_pass_thread_jumps_full): New.
* tree-ssa-threadbackward.c (pass_thread_jumps::gate): Inline.
(try_thread_blocks): Add resolve and speed arguments.
(pass_thread_jumps::execute): Inline.
(do_early_thread_jumps): New.
(do_thread_jumps): New.
(make_pass_thread_jumps): Move.
(pass_early_thread_jumps::gate): Inline.
(pass_early_thread_jumps::execute): Inline.
(class pass_thread_jumps_full): New.
Only powerpc64-unknown-freebsd was checked for.
Signed-off-by: Piotr Kubaj <pkubaj@FreeBSD.org>
gcc/
* configure.ac: Treat powerpc64*-*-freebsd* the same as
powerpc64-*-freebsd*.
* configure: Regenerate.
THis fixes teh following error seen with Clang:
error: function '_S_convert<std::basic_string_view<char8_t>>' with deduced
return type cannot be used before it is defined
return string_type(_S_convert(std::u8string_view(__str)));
^
libstdc++-v3/ChangeLog:
* include/bits/fs_path.h (path::_S_convert(T)): Avoid recursive
call to function with deduced return type.
A recently approved change for the C++23 working draft.
libstdc++-v3/ChangeLog:
* include/bits/basic_string.h (__cpp_lib_string_resize_and_overwrite):
Define for C++23.
(basic_string::resize_and_overwrite): Declare.
* include/bits/basic_string.tcc (basic_string::resize_and_overwrite):
Define.
* include/std/version (__cpp_lib_resize_and_overwrite): Define
for C++23.
* testsuite/21_strings/basic_string/capacity/char/resize_and_overwrite.cc:
New test.
PRs 101402, 102033, etc. demonstrated that the fix for PR92010 wasn't
handling all cases of the CWG1001/1322 issue with parameter type qual
stripping and arrays with templates. The problem turned out to be in
determine_specialization, which did an extra substitution without the 92010
fix and then complained that the result didn't match.
But just removing that wrong/redundant code meant that we were accepting
specializations with different numbers of parameters, because the code in
fn_type_unification that compares types in this case wasn't checking for
length mismatch.
After fixing that, I realized that fn_type_unification couldn't tell the
difference between variadic and non-variadic function types, because the
args array doesn't include the terminal void we use to indicate non-variadic
function type. So I added it, and made the necessary adjustments.
Thanks to qingzhe "nick" huang <nickhuang99@hotmail.com> for the patch that
led me to dig more into this, and the extensive testcases.
PR c++/51851
PR c++/101402
PR c++/102033
PR c++/102034
PR c++/102039
PR c++/102044
gcc/cp/ChangeLog:
* pt.c (determine_specialization): Remove redundant code.
(fn_type_unification): Check for mismatched length.
(type_unification_real): Ignore terminal void.
(get_bindings): Don't stop at void_list_node.
* class.c (resolve_address_of_overloaded_function): Likewise.
gcc/testsuite/ChangeLog:
* g++.dg/template/fnspec2.C: New test.
* g++.dg/template/parm-cv1.C: New test.
* g++.dg/template/parm-cv2.C: New test.
* g++.dg/template/parm-cv3.C: New test.
We were not wrapping all the default libraries in checks for whether
they should be used. We were also wasting a process launch calling
dsymutil for 'r' link lines (a NOP in practice). Order the checks
that exclude linking from most likely to occur, downwards.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* config/darwin.h (LINK_COMMAND_SPEC_A): Update 'r' handling to
skip gomp and itm when r or nodefaultlibs is given.
(DSYMUTIL_SPEC): Do not call dsymutil for '-r' link lines.
Update ordering of exclusions, remove duplicate 'v' addition
(collect2 will add this from the main command line).
Darwin has a user convenience feature where some linker options are exposed
at the driver level (so one can type '-all_load' instead of '-Wl,-all_load'
or '-Xlinker -all_load'). We retain this feature, but now these options are
all marked as 'Driver' and we process them as early as possible so that they
get allocated to the right toolchain command. There are a couple of special
cases where these driver opts are used multiple times, or to control
operations on more than one command (e.g. dynamiclib). These are handled
specially and we then add %<xxxx specs for the commands that _do not_ need
them. NOTE: the ordering of 'shared' and 'dynamiclib' is significant, hence
they are placed out of alphabetical order at the start. Likewise, we keep
a couple of cases where a negative option originally appeared after the
positive alternate, potentially overriding it.
When we report an error with %e, it seems necessary to strip the option
before doing so, otherwise it survives to the cc1 command line (%e does not
appear to abort the program before this).
Right now there is no mechanism to split up the 'variable portion' (%*) of
the matched spec string, so where we have some driver specs that take 2 or
3 arguments, these cannot be processed here, but are deferred until the
LINK_SPEC, where they are copied verbatim.
We have a 'safe' version of the macOS version string, that has been sanity-
checked and truncated to minor version. If the 'tiny' (3rd) portion of the
value is not significant, it is better to use the safe one in version-compare().
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* config/darwin-driver.c (darwin_driver_init): Revise comments, handle
filelist and framework options in specs instead of code.
* config/darwin.h (SUBTARGET_DRIVER_SELF_SPECS): Update to handle link
specs that are really driver ones.
(DARWIN_CC1_SPEC): Likewise.
(CPP_SPEC): Likewise.
(SYSROOT_SPEC): Append space.
(LINK_SYSROOT_SPEC): Remove most driver link specs.
(STANDARD_STARTFILE_PREFIX_2): Update link-related specs.
(STARTFILE_SPEC): Likewise.
(ASM_MMACOSX_VERSION_MIN_SPEC): Fix line wrap.
(ASM_SPEC): Update driver-related specs.
(ASM_FINAL_SPEC): Likewise.
* config/darwin.opt: Remove now unused option aliases.
* config/i386/darwin.h (EXTRA_ASM_OPTS): Ensure space after opt.
(ASM_SPEC): Update driver-related specs.
My previous patch, which was intended to reduce the differences seen by
the combination of -march=cascadelake and -m32, has additionally found
some more instances where this combination behaves differently to regular
x86_64-pc-linux-gnu. The middle-end always, and backends usually, use
emit_move_insn to emit/expand move instructions allowing the backend
control over placing things in constant pools, adding REG_EQUAL notes,
and so on. Several of the AVX512 built-in expanders bypass this logic,
and instead generate moves directly using emit_insn(gen_rtx_SET (dst,src)).
For example, i386-expand.c line 12004 contains:
for (i = 0; i < 8; i++)
emit_insn (gen_rtx_SET (xmm_regs[i], const0_rtx));
I suspect that in this case, loading of standard_sse_constant_p, my
change to require loading of likely spilled hard registers via a
pseudo is perhaps overly strict, so this patch/fix reallows these
immediate constants values to be loaded directly prior to reload.
2021-10-15 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386.c (ix86_hardreg_mov_ok): For vector modes,
allow standard_sse_constant_p immediate constants.
This implements the changes in P2231R1 which make std::variant fully
constexpr in C++20.
We need to replace placement new with std::construct_at, but that isn't
defined for C++17. Use std::_Construct instead, which forwards to
std::construct_at in C++20 mode (since the related changes to make
std::optional fully constexpr, in r12-4389).
We also need to replace the untyped char buffer in _Uninitialized with a
union, which can be accessed in constexpr functions. But the union needs
to have a non-trivial destructor if its variant type is non-trivial,
which means that the _Variadic_union also needs a non-trivial
destructor. This adds a constrained partial specialization of
_Variadic_union for the C++20-only case where a non-trivial destructor
is needed.
We can't use concepts to constrain the specialization (or the primary
template's destructor) in C++17, so retain the untyped char buffer
solution for C++17 mode.
libstdc++-v3/ChangeLog:
* include/std/variant (__cpp_lib_variant): Update value for
C++20.
(__variant_cast, __variant_construct): Add constexpr for C++20.
(__variant_construct_single, __construct_by_index) Likewise. Use
std::_Construct instead of placement new.
(_Uninitialized<T, false>) [__cplusplus >= 202002]: Replace
buffer with a union and define a destructor.
(_Variadic_union) [__cplusplus >= 202002]: Add a specialization
for non-trivial destruction.
(_Variant_storage::__index_of): New helper variable template.
(_Variant_storage::~_Variant_storage()): Add constexpr.
(_Variant_storage::_M_reset()): Likewise.
(_Copy_ctor_base, _Move_ctor_base): Likewise.
(_Copy_assign_base, _Move_assign_base): Likewise.
(variant, swap): Likewise.
* include/std/version (__cpp_lib_variant): Update value for
C++20.
* testsuite/20_util/optional/version.cc: Check for exact value
in C++17.
* testsuite/20_util/variant/87619.cc: Increase timeout for
C++20 mode.
* testsuite/20_util/variant/constexpr.cc: New test.
* testsuite/20_util/variant/version.cc: New test.
The __variant_construct_by_index helper function sets the new index
before constructing the new object. This means that if the construction
throws then the exception needs to be caught, so the index can be reset
to variant_npos, and then the exception rethrown. This means callers are
responsible for restoring the variant's invariants and they need the
overhead of a catch handler and a rethrow.
If we don't set the index until after construction completes then the
invariant is never broken, and callers can ignore the exception and let
it propagate. The callers all call _M_reset() first, which sets index to
variant_npos as required while the variant is valueless.
We need to be slightly careful here, because changing the order of
operations in __variant_construct_by_index and removing the try-block
from variant::emplace<I> changes an implicit ABI contract between those
two functions. If the linker were to create an executable containing an
instantiation of the old __variant_construct_by_index and an
instantiation of the new variant::emplace<I> code then we would have a
combination that breaks the invariant and doesn't have the exception
handling to restore it. To avoid this problem, we can rename the
__variant_construct_by_index function so that the new emplace<I> code
calls a new symbol, and is unaffected by the behaviour of the old
symbol.
libstdc++-v3/ChangeLog:
* include/std/variant (__detail::__variant::__get_storage):
Remove unused function.
(__variant_construct_by_index): Set index after construction is
complete. Rename to ...
(__detail::__variant::__construct_by_index): ... this.
(variant): Use new name for __variant_construct_by_index friend
declaration. Remove __get_storage friend declaration.
(variant::emplace): Use new name and remove try-blocks.
These functions aren't used, and accessing the storage as a void* isn't
compatible with C++20 constexpr requirements anyway, so we're unlikely
to ever start using them in future.
libstdc++-v3/ChangeLog:
* include/std/variant (_Variant_storage::_M_storage()): Remove.
(__detail::__variant::__get_storage): Remove.
(variant): Remove friend declaration of __get_storage.
After r12-4432-g7bfe7d634f60b0a9 Darwin fails to bootstrap with D
enabled since there is no definition of either DWARF2_DEBUG_INFO or
PREFERRED_DEBUGGING_TYPE.
Fixed here by adding the tm-dwarf2.h file to tm_d_file for Darwin.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* config.gcc: Add tm-dwarf2.h to tm_d-file.
* gimple-range-fold.h (gimple_range_ssa_p): Don't process names
that occur in abnormal phis.
* gimple-range.cc (gimple_ranger::range_on_edge): Return false for
abnormal and EH edges.
* gimple-ssa-evrp.c (rvrp_folder::value_of_expr): Ditto.
(rvrp_folder::value_on_edge): Ditto.
(rvrp_folder::value_of_stmt): Ditto.
(hybrid_folder::value_of_expr): Ditto for ranger queries.
(hybrid_folder::value_on_edge): Ditto.
(hybrid_folder::value_of_stmt): Ditto.
* value-query.cc (gimple_range_global): Always return a range if
the type is supported.
I've noticed that while I have added hopefully sufficient test coverage
for the case where one uses simple number or !number as p-interval,
I haven't added any coverage for number:len:stride or number:len.
This patch adds that.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.c/affinity-1.c (struct places): Change name field
type from char [50] to const char *.
(places_array): Add a testcase for simplified syntax place followed
by length or length and stride.
2021-10-15 John David Anglin <danglin@gcc.gnu.org>
gcc/ChangeLog:
* config/pa/pa.md: Consistently use "rG" constraint for copy
instruction in move patterns.
In addition to adding ll_caches and numa_domain abstract names
to OMP_PLACES syntax, OpenMP 5.1 also added one syntax simplification:
https://github.com/OpenMP/spec/issues/2080https://github.com/OpenMP/spec/pull/2081
in particular that in the grammar place non-terminal is now
not only { res-list } but also res (i.e. a non-negative integer),
which stands as a shortcut for { res }
So, one can specify OMP_PLACES=0,4,8,12 with the meaning
OMP_PLACES={0},{4},{8},{12} or OMP_PLACES=0:4 instead of OMP_PLACES={0}:4
or OMP_PLACES={0},{1},{2},{3} etc.
This patch implements that.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_one_place): Handle non-negative-number the same
as { non-negative-number }. Reject even !number:1 and
!number:1:stride or !place:1 or !place:1:stride instead of just
length other than 1.
* libgomp.texi (OpenMP 5.1): Document OMP_PLACES syntax extensions
and OMP_NUM_TEAMS/OMP_TEAMS_THREAD_LIMIT and
omp_{set_num,get_max}_teams/omp_{s,g}et_teams_thread_limit features
as implemented.
* testsuite/libgomp.c/affinity-1.c: Add a test for the 5.1 place
simplified syntax.
Yesterday when working on numa_domains, I've noticed because of a bug
in my patch a hang on a large NUMA machine. I've fixed the bug, but
also discovered that the hang was a result of making wrong assumptions
about strtoul/strtoull. All the uses were for portability setting
errno = 0 before the calls and treating non-zero errno after the call
as invalid input, but for the case where there are no valid digits at
all strtoul may set errno to EINVAL, but doesn't have to and with
glibc doesn't do that. So, this patch goes through all the strtoul calls
and next to errno != 0 checks adds also endptr == startptr check.
Haven't done it in places where we immediately reject strtoul returning 0
the same as we reject errno != 0, because strtoul must return 0 in the
case where it sets endptr to the start pointer. In some spots the code
was using errno = 0; x = strtoul (p, &p, 10); if (errno) { /*invalid*/ }
and those spots had to be changed to
errno = 0; x = strtoul (p, &end, 10); if (errno || end == p) { /*invalid*/ }
p = end;
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_schedule): For strtoul or strtoull calls which don't
clearly reject return value 0 as invalid handle the case where end
pointer is the same as first argument as invalid.
(parse_unsigned_long_1): Likewise.
(parse_one_place): Likewise.
(parse_places_var): Likewise.
(parse_stacksize): Likewise.
(parse_spincount): Likewise.
(parse_affinity): Likewise.
(parse_gomp_openacc_dim): Likewise. Avoid strict aliasing violation.
Make code valid C89.
* config/linux/affinity.c (gomp_affinity_find_last_cache_level):
For strtoul calls which don't clearly reject return value 0 as
invalid handle the case where end pointer is the same as first
argument as invalid.
(gomp_affinity_init_level_1): Likewise.
(gomp_affinity_init_numa_domains): Likewise.
* config/rtems/proc.c (parse_thread_pools): Likewise.
When writing the places-*.c tests, I've noticed that we mishandle threads
abstract name with specified num-places if num-places isn't a multiple of
number of hw threads in a core. It then happily ignores the maximum count
and overwrites for the remaining hw threads in a core further places that
haven't been allocated.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* config/linux/affinity.c (gomp_affinity_init_level_1): For level 1
after creating count places clean up and return immediately.
* testsuite/libgomp.c/places-6.c: New test.
* testsuite/libgomp.c/places-7.c: New test.
* testsuite/libgomp.c/places-8.c: New test.
* testsuite/libgomp.c/places-9.c: New test.
* testsuite/libgomp.c/places-10.c: New test.
Between LLVM 9 and LLVM 13 the attribute works differently in several ways,
and this needs to be allowed for in GCC and mkoffload independently.
This patch fixes up mkoffload when debug info is enabled, which is made more
complicated because the configure tests checks whether the attribute option
is accepted silently, but does not check if the assembler actually sets the
ELF flags for that attribute, and mkoffload needs to mimick that behaviour
exactly. The patch therefore removes some of the conditionals.
gcc/ChangeLog:
* config/gcn/gcn-hsa.h (S_FIJI): Set unconditionally.
(S_900): Likewise.
(S_906): Likewise.
* config/gcn/gcn.c: Hard code SRAM ECC settings for old architectures.
* config/gcn/mkoffload.c (ELFABIVERSION_AMDGPU_HSA): Rename to ...
(ELFABIVERSION_AMDGPU_HSA_V3): ... this.
(ELFABIVERSION_AMDGPU_HSA_V4): New.
(SET_SRAM_ECC_UNSUPPORTED): New.
(copy_early_debug_info): Create elf flags to match the other objects.
(main): Just let the attribute flags pass through.
While determining the precission of reduction_var an SSA_NAME instead of
its TREE_TYPE is used. Streamlined with other TREE_TYPE (reduction_var)
uses.
gcc/ChangeLog:
* tree-loop-distribution.c (reduction_var_overflows_first):
Pass the type of reduction_var as first argument as it is also
done for the load type.
(loop_distribution::transform_reduction_loop): Add missing
TREE_TYPE while determining precission of reduction_var.
This test is failing on ppc64* due to different default signness for
chars.
Tested on x86-64 Linux and ppc64le Linux.
PR testsuite/pr102751
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/pr102736.c: Make sign explicit.
