This implements the library parts of P1494 as amended by P3641. For GCC the
compiler itself treats stdio operations as equivalent to the observable
checkpoint and thus it does not appear to be necessary to add calls to those
functions (it will not alter the outcome).
This adds the facility for C++26, although there is no reason, in principle,
that it would not work back to C++11 at least.
PR c++/119060
libstdc++-v3/ChangeLog:
* include/bits/version.def: Add observable_checkpoint at present
allowed from C++26.
* include/bits/version.h: Regenerate.
* include/std/utility: Add std::observable_checkpoint().
* src/c++23/std.cc.in: Add obervable_checkpoint () to utility.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
P1494 provides a mechanism that serves to demarc epochs within the code
preventing UB-based optimisations from 'time traveling' across such
boundaries. The additional paper, P3641, alters the name of the function
to 'observable_checkpoint' which is the name used here.
This implementation maintains the observable function call through to
expand, where it produces no code.
PR c++/119060
gcc/ChangeLog:
* builtins.cc (expand_builtin): Handle BUILT_IN_OBSERVABLE_CHKPT.
* builtins.def (BUILT_IN_OBSERVABLE_CHKPT): New.
* tree.cc (build_common_builtin_nodes): Build observable
checkpoint builtin.
gcc/cp/ChangeLog:
* cxxapi-data.csv: Add observable_checkpoint to <utility>.
* std-name-hint.gperf: Add observable_checkpoint to <utility>.
* std-name-hint.h: Regenerate.
gcc/testsuite/ChangeLog:
* g++.dg/cpp26/observable-checkpoint.C: New test.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
[namespace.qual] p1 says that a namespace nominated by a using-directive
is searched if the using-directive precedes that point.
[basic.lookup.general] p2 says that a declaration in a different TU
within a module purview is visible if either the declaration is
exported, or the other TU is part of the same module as the point of
lookup. This patch implements the second half of that.
PR c++/122279
gcc/cp/ChangeLog:
* module.cc (depset:#️⃣:add_namespace_entities): Seed any
purview using-decls.
(module_state::write_using_directives): Stream if the udir was
exported or not.
(module_state::read_using_directives): Add the using-directive
if it's either exported or part of this module.
gcc/testsuite/ChangeLog:
* g++.dg/modules/namespace-13_b.C: Adjust expected results.
* g++.dg/modules/namespace-13_c.C: Test non-exported
using-directive is not used.
* g++.dg/modules/namespace-14_a.C: New test.
* g++.dg/modules/namespace-14_b.C: New test.
* g++.dg/modules/namespace-14_c.C: New test.
* g++.dg/modules/namespace-14_d.C: New test.
Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Patrick Palka <ppalka@redhat.com>
Reviewed-by: Jason Merrill <jason@redhat.com>
SVE2p1 adds 2-way dotproduct which we can use when we have to do a single step
widening addition. This is useful for instance when the value to be widened
does not come from a load. For example for
int foo2_int(unsigned short *x, unsigned short * restrict y) {
int sum = 0;
for (int i = 0; i < 8000; i++)
{
x[i] = x[i] + y[i];
sum += x[i];
}
return sum;
}
we used to generate
.L12:
ld1h z30.h, p7/z, [x0, x2, lsl 1]
ld1h z29.h, p7/z, [x1, x2, lsl 1]
add z30.h, z30.h, z29.h
uaddwb z31.s, z31.s, z30.h
uaddwt z31.s, z31.s, z30.h
st1h z30.h, p7, [x0, x2, lsl 1]
mov x3, x2
inch x2
cmp w2, w4
bls .L12
inch x3
uaddv d31, p7, z31.s
but with +sve2p1
.L12:
ld1h z31.h, p7/z, [x0, x2, lsl 1]
ld1h z29.h, p7/z, [x1, x2, lsl 1]
add z31.h, z31.h, z29.h
udot z30.s, z31.h, z28.h
st1h z31.h, p7, [x0, x2, lsl 1]
mov x3, x2
inch x2
cmp w2, w4
bls .L12
inch x3
uaddv d30, p7, z30.s
gcc/ChangeLog:
PR middle-end/122069
* config/aarch64/aarch64-sve2.md
(widen_ssum<mode><Vnarrow>3): Update.
(widen_usum<mode><Vnarrow>3): Update.
gcc/testsuite/ChangeLog:
PR middle-end/122069
* gcc.target/aarch64/sve2/pr122069_3.c: New test.
* gcc.target/aarch64/sve2/pr122069_4.c: New test.
SVE2 adds [US]ADDW[TB] which we can use when we have to do a single step
widening addition. This is useful for instance when the value to be widened
does not come from a load. For example for
int foo2_int(unsigned short *x, unsigned short * restrict y) {
int sum = 0;
for (int i = 0; i < 8000; i++)
{
x[i] = x[i] + y[i];
sum += x[i];
}
return sum;
}
we used to generate
.L6:
ld1h z1.h, p7/z, [x0, x2, lsl 1]
ld1h z29.h, p7/z, [x1, x2, lsl 1]
add z29.h, z29.h, z1.h
punpklo p6.h, p7.b
uunpklo z0.s, z29.h
add z31.s, p6/m, z31.s, z0.s
punpkhi p6.h, p7.b
uunpkhi z30.s, z29.h
add z31.s, p6/m, z31.s, z30.s
st1h z29.h, p7, [x0, x2, lsl 1]
add x2, x2, x4
whilelo p7.h, w2, w3
b.any .L6
ptrue p7.b, all
uaddv d31, p7, z31.s
but with +sve2
.L12:
ld1h z30.h, p7/z, [x0, x2, lsl 1]
ld1h z29.h, p7/z, [x1, x2, lsl 1]
add z30.h, z30.h, z29.h
uaddwb z31.s, z31.s, z30.h
uaddwt z31.s, z31.s, z30.h
st1h z30.h, p7, [x0, x2, lsl 1]
mov x3, x2
inch x2
cmp w2, w4
bls .L12
inch x3
uaddv d31, p7, z31.s
gcc/ChangeLog:
PR middle-end/122069
* config/aarch64/aarch64-sve2.md: (widen_ssum<mode><Vnarrow>3): New.
(widen_usum<mode><Vnarrow>3): New.
* config/aarch64/iterators.md (Vnarrow): New, to match VNARROW.
gcc/testsuite/ChangeLog:
PR middle-end/122069
* gcc.target/aarch64/sve2/pr122069_1.c: New test.
* gcc.target/aarch64/sve2/pr122069_2.c: New test.
This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab. The result of which requires the output mode to be added to
the existing patterns.
No change in functionality is expected.
gcc/ChangeLog:
PR middle-end/122069
* config/rs6000/altivec.md (widen_usum<mode>3): Rename ...
(widen_usumv4si<mode>3): ... to this.
(widen_ssumv16qi3): Rename ...
(widen_ssumv4siv16qi3): ... to this.
(widen_ssumv8hi3): Rename ...
(widen_ssumv4siv8hi3): ... to this.
The target does not seem to have a maintainer listed, I've CC'ed a group of
global maintainers instead hoping one of you could approve it.
This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab. The result of which requires the output mode to be added to
the existing patterns.
No change in functionality is expected.
gcc/ChangeLog:
PR middle-end/122069
* config/ia64/vect.md (widen_usumv8qi3): Renamed ...
(widen_usumv4hiv8qi3): ... into this.
(widen_usumv4hi3): Renamed ...
(widen_usumv2siv4hi3): ... into this.
(widen_ssumv8qi3): Renamed ...
(widen_ssumv4hiv8qi3): ... into this.
(widen_ssumv4hi3): Renamed ...
(widen_ssumv2siv4hi3): ... into this.
This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab. The result of which requires the output mode to be added to
the existing patterns.
No change in functionality is expected.
gcc/ChangeLog:
PR middle-end/122069
* config/arm/iterators.md (v_double_width): New, matching
V_double_width.
* config/arm/neon.md (widen_ssum<mode>3): Renamed ...
(widen_ssum<v_double_width><mode>3, widen_ssum<V_widen_l><mode>3): ...
into these.
(widen_usum<mode>3): Renamed ...
(widen_usum<v_double_width><mode>3, widen_usum<V_widen_l><mode>3): ...
into these.
This patch implements support for using dotproduct to do sum reductions by
changing += a into += (a * 1). i.e. we seed the multiplication with 1.
Given the example
int foo_int(unsigned char *x, unsigned char * restrict y) {
int sum = 0;
for (int i = 0; i < 8000; i++)
sum += char_abs(x[i] - y[i]);
return sum;
}
we used to generate
.L2:
ldr q0, [x0, x2]
ldr q28, [x1, x2]
sub v28.16b, v0.16b, v28.16b
zip1 v29.16b, v28.16b, v31.16b
zip2 v28.16b, v28.16b, v31.16b
uaddw v30.4s, v30.4s, v29.4h
uaddw2 v30.4s, v30.4s, v29.8h
uaddw v30.4s, v30.4s, v28.4h
uaddw2 v30.4s, v30.4s, v28.8h
add x2, x2, 16
cmp x2, x3
bne .L2
addv s31, v30.4s
but now generates with +dotprod
.L2:
ldr q29, [x0, x2]
ldr q28, [x1, x2]
sub v28.16b, v29.16b, v28.16b
udot v31.4s, v28.16b, v30.16b
add x2, x2, 16
cmp x2, x3
bne .L2
addv s31, v31.4s
gcc/ChangeLog:
PR middle-end/122069
* config/aarch64/aarch64-simd.md (widen_ssum<mode><vsi2qi>3): New.
(widen_usum<mode><vsi2qi>3): New.
gcc/testsuite/ChangeLog:
PR middle-end/122069
* gcc.target/aarch64/pr122069_3.c: New test.
* gcc.target/aarch64/pr122069_4.c: New test.
This patch is a mechanical rewrite of the widen_[us]sum optabs from a direct to
a conversion optab. The result of which requires the output mode to be added to
the existing patterns.
No change in functionality is expected.
gcc/ChangeLog:
PR middle-end/122069
* config/aarch64/aarch64-simd.md (widen_ssum<mode>3): Change into..
(widen_ssum<Vdblw><mode>3, widen_ssum<Vwide><mode>3): ... these.
(widen_usum<mode>3): Change into ...
(widen_usum<Vdblw><mode>3, widen_usum<Vwide><mode>3): ... these.
* config/aarch64/iterators.md (Vdblw): New.
(Vwide): Extend to match VWIDE.
gcc/testsuite/ChangeLog:
PR middle-end/122069
* gcc.target/aarch64/pr122069_1.c: New test.
* gcc.target/aarch64/pr122069_2.c: New test.
This patch changes the widen_[us]sum optabs into a convert optabs such that
targets and specify more than one conversion.
Following this patch are patches rewriting all targets using this change.
While working on this I noticed that the pattern does miss some cases it
could handle if it tried multiple attempts. e.g. if the promotion is from
qi to si, and the target doesn't have this, it should try hi -> si.
But I'm leaving that for now.
gcc/ChangeLog:
PR middle-end/122069
* doc/md.texi (widen_ssum@var{n}@var{m}3, widen_usum@var{n}@var{m}3):
Update docs.
* optabs.cc (expand_widen_pattern_expr): Add WIDEN_SUM_EXPR as widening.
* optabs.def (ssum_widen_optab, usum_widen_optab): Convert from direct
to a conversion optab.
* tree-vect-patterns.cc (vect_recog_widen_sum_pattern): Change
vect_supportable_direct_optab_p into vect_supportable_conv_optab_p.
This patch allows the use of character types in conditional expressions.
gcc/fortran/ChangeLog:
* resolve.cc (resolve_conditional): Allow character in cond-expr.
* trans-const.cc (gfc_conv_constant): Handle want_pointer.
* trans-expr.cc (gfc_conv_conditional_expr): Fill se->string_length.
(gfc_conv_string_parameter): Handle COND_EXPR tree code.
gcc/testsuite/ChangeLog:
* gfortran.dg/conditional_1.f90: Test character type.
* gfortran.dg/conditional_2.f90: Test print constants.
* gfortran.dg/conditional_4.f90: Test diagnostic message.
* gfortran.dg/conditional_6.f90: Test character cond-arg.
After commit 51b85dfeb1, when the
pointer offset is a variable in the loop, the object size of the
pointer may also need to be reexamined.
Which make gcc_assert in the check_for_plus_in_loops failed.
gcc/ChangeLog:
PR tree-optimization/122012
* tree-object-size.cc (check_for_plus_in_loops): Skip check
for the variable offset
gcc/testsuite/ChangeLog:
PR tree-optimization/122012
* gcc.dg/torture/pr122012.c: New test.
Signed-off-by: Linsen Zhou <i@lin.moe>
The BPF backend expansion of setmem was broken, because it could elect
to use stores of HI, SI or DI modes based on the destination alignment
when the value was QI, but fail to duplicate the byte value across to
those larger sizes. This resulted in not all bytes of the destination
actually being set to the desired value.
Fix bpf_expand_setmem to ensure the desired byte value is really
duplicated as necessary, whether it is constant or a (sub)reg:QI.
PR target/122139
gcc/
* config/bpf/bpf.cc (bpf_expand_setmem): Duplicate byte value
across to new mode when using larger modes for store.
gcc/testsuite/
* gcc.target/bpf/memset-3.c: New.
* gcc.target/bpf/memset-4.c: New.
In PR121604 the problem was noted that currently the SVE intrinsics
infrastructure assumes that for any predicated operation that the GP is at the
first argument position which has a svbool_t or for a unary merging operation
that it's in the second position.
However you have intrinsics like fmov_lane which have an svbool_t but it's not
a GP but instead it's the inactive lanes.
You also have instructions like BRKB which work only on predicates so it
incorrectly determines the first operand to be the GP, while that's also the
inactive lanes.
However during apply_predication we do have the information about where the GP
is. This patch re-organizes the code to record this information into the
function_instance such that folders have access to this information.
For functions that are outliers like pmov_lane we can now override the
availability of the intrinsics having a GP.
gcc/ChangeLog:
PR target/121604
* config/aarch64/aarch64-sve-builtins-shapes.cc (apply_predication):
Store gp_index.
(struct pmov_to_vector_lane_def): Mark instruction as has no GP.
* config/aarch64/aarch64-sve-builtins.h (function_instance::gp_value,
function_instance::inactive_values, function_instance::gp_index,
function_shape::has_gp_argument_p): New.
* config/aarch64/aarch64-sve-builtins.cc (gimple_folder::fold_pfalse):
Simplify code and use GP helpers.
gcc/testsuite/ChangeLog:
PR target/121604
* gcc.target/aarch64/sve/pr121604_brk.c: New test.
* gcc.target/aarch64/sve2/pr121604_pmov.c: New test.
Co-authored-by: Jennifer Schmitz <jschmitz@nvidia.com>
Just like with loop interchange, unroll-and-jam can leave invariant
stmts in the inner loop from outer loop stmts inbetween the two
inner loop copies. Do a per-function invariant motion when we
applied unroll-and-jam. This avoids failed dataref analysis
and fallback to gather/scatter during vectorization.
PR tree-optimization/122308
* gimple-loop-jam.cc (tree_loop_unroll_and_jam): Do LIM
after applying unroll-and-jam.
* gcc.dg/vect/vect-pr122308.c: New testcase.
This patch enables constant propagation to outlined OpenMP kernels.
It does so using a new function attribute called ' callback' (note the
space).
The attribute ' callback' captures the notion of a function calling one
of its arguments with some of its parameters as arguments. An OpenMP
example of such function is GOMP_parallel.
We implement the attribute with new callgraph edges called callback
edges. They are imaginary edges pointing from the caller of the function
with the attribute (e.g. caller of GOMP_parallel) to the body function
itself (e.g. the outlined OpenMP body). They share their call statement
with the edge from which they are derived (direct edge caller -> GOMP_parallel
in this case). These edges allow passes such as ipa-cp to see the hidden
call site to the body function and optimize the function accordingly.
To illustrate on an example, the body GOMP_parallel looks something
like this:
void GOMP_parallel (void (*fn) (void *), void *data, /* ... */)
{
/* ... */
fn (data);
/* ... */
}
If we extend it with the attribute ' callback(1, 2)', we express that the
function calls its first argument and passes it its second argument.
This is represented in the call graph in this manner:
direct indirect
caller -----------------> GOMP_parallel ---------------> fn
|
----------------------> fn
callback
The direct edge is then the callback-carrying edge, all new edges
are the derived callback edges.
While constant propagation is the main focus of this patch, callback
edges can be useful for different passes (for example, they improve icf
for OpenMP kernels), as they allow for address redirection.
If the outlined body function gets optimized and cloned, from body_fn to
body_fn.optimized, the callback edge allows us to replace the
address in the arguments list:
GOMP_parallel (body_fn, &data_struct, /* ... */);
becomes
GOMP_parallel (body_fn.optimized, &data_struct, /* ... */);
This redirection is possible for any function with the attribute.
This callback attribute implementation is partially compatible with
clang's implementation. Its semantics, arguments and argument indexing style are
the same, but we represent an unknown argument position with 0
(precedent set by attributes such as 'format'), while clang uses -1 or '?'.
We use the index 1 for the 'this' pointer in member functions, clang
uses 0. We also allow for multiple callback attributes on the same function,
while clang only allows one.
The attribute is currently for GCC internal use only, thanks to the
space in its name. Originally, it was supposed to be called
'callback' like its clang counterpart, but we cannot use this name, as
clang uses non-standard indexing style, leading to inconsistencies. The
attribute will be introduced into the public API as 'gnu::callback_only'
in a future patch.
The attribute allows us to propagate constants into body functions of
OpenMP constructs. Currently, GCC won't propagate the value 'c' into the
OpenMP body in the following example:
int a[100];
void test(int c) {
#pragma omp parallel for
for (int i = 0; i < c; i++) {
if (!__builtin_constant_p(c)) {
__builtin_abort();
}
a[i] = i;
}
}
int main() {
test(100);
return a[5] - 5;
}
With this patch, the body function will get cloned and the constant 'c'
will get propagated.
Some functions may utilize the attribute's infrastructure without being
declared with it, for example GOMP_task. These functions are special
cases and use the special case functions found in attr-callback.h. Special
cases use the attribute under certain circumstances, for example
GOMP_task uses it when the copy function is not being used required.
gcc/ChangeLog:
* Makefile.in: Add attr-callback.o to OBJS.
* builtin-attrs.def (ATTR_CALLBACK): Callback attr identifier.
(DEF_CALLBACK_ATTRIBUTE): Macro for callback attr creation.
(GOMP): Attr for libgomp functions.
(ATTR_CALLBACK_GOMP_LIST): ATTR_NOTHROW_LIST with GOMP callback
attr added.
* cgraph.cc (cgraph_add_edge_to_call_site_hash): Always hash the
callback-carrying edge.
(cgraph_node::get_edge): Always return the callback-carrying
edge.
(cgraph_edge::set_call_stmt): Add cascade for callback edges.
(symbol_table::create_edge): Allow callback edges to share call
stmts, initialize new flags.
(cgraph_edge::make_callback): New method, derives a new callback
edge.
(cgraph_edge::get_callback_carrying_edge): New method.
(cgraph_edge::first_callback_edge): Likewise.
(cgraph_edge::next_callback_edge): Likewise.
(cgraph_edge::purge_callback_edges): Likewise.
(cgraph_edge::redirect_callee): When redirecting a callback
edge, redirect its ref as well.
(cgraph_edge::redirect_call_stmt_to_callee): Add callback edge
redirection logic, set update_derived_edges to true hwne
redirecting the carrying edge.
(cgraph_node::remove_callers): Add cascade for callback edges.
(cgraph_edge::dump_edge_flags): Print callback flags.
(cgraph_node::verify_node): Add sanity checks for callback
edges.
* cgraph.h: Add new 1 bit flags and 16 bit callback_id to
cgraph_edge class.
* cgraphclones.cc (cgraph_edge::clone): Copy over callback data.
* cif-code.def (CALLBACK_EDGE): Add CIF_CALLBACK_EDGE code.
* ipa-cp.cc (purge_useless_callback_edges): New function,
deletes callback edges when necessary.
(ipcp_decision_stage): Call purge_useless_callback_edges.
* ipa-fnsummary.cc (ipa_call_summary_t::duplicate): Add
an exception for callback edges.
(analyze_function_body): Copy over summary from carrying to
callback edge.
* ipa-inline-analysis.cc (do_estimate_growth_1): Skip callback
edges when estimating growth.
* ipa-inline-transform.cc (inline_transform): Add redirection
cascade for callback edges.
* ipa-param-manipulation.cc
(drop_decl_attribute_if_params_changed_p): New function.
(ipa_param_adjustments::build_new_function_type): Add
args_modified out param.
(ipa_param_adjustments::adjust_decl): Drop callback attrs when
modifying args.
* ipa-param-manipulation.h: Adjust decl of
build_new_function_type.
* ipa-prop.cc (ipa_duplicate_jump_function): Add decl.
(init_callback_edge_summary): New function.
(ipa_compute_jump_functions_for_edge): Add callback edge
creation logic.
* lto-cgraph.cc (lto_output_edge): Stream out callback data.
(input_edge): Input callback data.
* omp-builtins.def (BUILT_IN_GOMP_PARALLEL_LOOP_STATIC): Use new
attr list.
(BUILT_IN_GOMP_PARALLEL_LOOP_GUIDED): Likewise.
(BUILT_IN_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC): Likewise.
(BUILT_IN_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME): Likewise.
(BUILT_IN_GOMP_PARALLEL): Likewise.
(BUILT_IN_GOMP_PARALLEL_SECTIONS): Likewise.
(BUILT_IN_GOMP_TEAMS_REG): Likewise.
* tree-core.h (ECF_CB_1_2): New constant for callback(1,2).
* tree-inline.cc (copy_bb): Copy callback edges when copying the
carrying edge.
(redirect_all_calls): Redirect callback edges.
* tree.cc (set_call_expr_flags): Create callback attr according
to the ECF_CB flag.
* attr-callback.cc: New file.
* attr-callback.h: New file.
gcc/c-family/ChangeLog:
* c-attribs.cc: Define callback attr.
gcc/fortran/ChangeLog:
* f95-lang.cc (ATTR_CALLBACK_GOMP_LIST): New attr list
corresponding to the list in builtin-attrs.def.
gcc/testsuite/ChangeLog:
* gcc.dg/ipa/ipcp-cb-spec1.c: New test.
* gcc.dg/ipa/ipcp-cb-spec2.c: New test.
* gcc.dg/ipa/ipcp-cb1.c: New test.
Signed-off-by: Josef Melcr <jmelcr02@gmail.com>
Unlike for subprogram instantiation, -gnatyr does not report style violation
for package instantiation, more precisely for the generic package's name.
Fixing it uncovered style violations in the sources of the compiler itself!
gcc/ada/
PR ada/122295
* sem_ch12.adb (Analyze_Package_Instantiation): Force Style_Check
to False only after possibly installing the parent.
* aspects.adb (UAD_Pragma_Map): Fix style violation.
* inline.adb (To_Pending_Instantiations): Likewise.
* lib.ads (Unit_Names): Likewise.
* repinfo.adb (Relevant_Entities): Likewise.
* sem_ch7.adb (Subprogram_Table): Likewise.
(Traversed_Table): Likewise.
* sem_util.adb (Interval_Sorting): Likewise.
gcc/testsuite/
* gnat.dg/specs/style1.ads: New test.
The following fixes another issue with updating of reduc_idx in pattern
sequences. But the testcase also shows the pattern in question is
harmful for vectorization since a reduction path may not contain
promotions/demotions. So the already existing but ineffective check
to guard the pattern is fixed.
PR tree-optimization/122301
* tree-vect-patterns.cc (vect_recog_over_widening_pattern):
Fix reduction guard.
(vect_mark_pattern_stmts): Fix reduction def check.
* gcc.dg/vect/vect-pr122301.c: New testcase.
Added a new helper function "add_code_for_floorceilround_divmod" in
tree-vect-patterns.cc for adding compensating code for each of the op
{FLOOR,ROUND,CEIL}_{DIV,MOD}_EXPR. This function checks if target supports all
required operations required to implement these operation and generates
vectorized code for the respective operations. Based on the following logic
FLOOR_{DIV,MOD}
r = x %[fl] y;
r = x % y; if (r && (x ^ y) < 0) r += y;
r = x/[fl] y;
r = x % y; d = x/y; if (r && (x ^ y) < 0) d--;
CEIL_{DIV,MOD} (unsigned)
r = x %[cl] y;
r = x % y; if (r) r -= y;
r = x/[cl] y;
r = x % y; d = x/y; if (r) d++;
CEIL_{DIV,MOD} (signed)
r = x %[cl] y;
r = x % y; if (r && (x ^ y) >= 0) r -= y;
r = x/[cl] y;
r = x % y; d = x/y; if (r && (x ^ y) >= 0) d++;
ROUND_{DIV,MOD} (unsigned)
r = x %[rd] y;
r = x % y; if (r > ((y-1)/2)) r -= y;
r = x/[rd] y;
r = x % y; d = x/y; if (r > ((y-1)/2)) d++;
ROUND_{DIV,MOD} (signed)
r = x %[rd] y;
r = x % y; if (r > ((y-1)/2))
{if ((x ^ y) >= 0) r -= y; else r += y;}
r = x/[rd] y;
r = x % y; d = x/y; if ((r > ((y-1)/2)) && (x ^ y) >= 0)
{if ((x ^ y) >= 0) d++; else d--;}
each of the case is implemented in a vectorized form.
This function is then called in each of the path in vect_recog_divmod_pattern,
which there are 3, based on value of constant operand1,
1. == 2
2. == power of 2
3. otherwise
2025-10-17 Avinash Jayakar <avinashd@linux.ibm.com>
gcc/ChangeLog:
PR tree-optimization/104116
* tree-vect-patterns.cc (add_code_for_floorceilround_divmod): patt recog
for {FLOOR,ROUND,CEIL}_{DIV,MOD}_EXPR.
(vect_recog_divmod_pattern): Call add_code_for_floorceilround_divmod
after computing div/mod for each control path.
gcc/testsuite/ChangeLog:
PR tree-optimization/104116
* gcc.dg/vect/pr104116-ceil-div-2.c: New test.
* gcc.dg/vect/pr104116-ceil-div-pow2.c: New test.
* gcc.dg/vect/pr104116-ceil-div.c: New test.
* gcc.dg/vect/pr104116-ceil-mod-2.c: New test.
* gcc.dg/vect/pr104116-ceil-mod-pow2.c: New test.
* gcc.dg/vect/pr104116-ceil-mod.c: New test.
* gcc.dg/vect/pr104116-ceil-udiv-2.c: New test.
* gcc.dg/vect/pr104116-ceil-udiv-pow2.c: New test.
* gcc.dg/vect/pr104116-ceil-udiv.c: New test.
* gcc.dg/vect/pr104116-ceil-umod-2.c: New test.
* gcc.dg/vect/pr104116-ceil-umod-pow2.c: New test.
* gcc.dg/vect/pr104116-ceil-umod.c: New test.
* gcc.dg/vect/pr104116-floor-div-2.c: New test.
* gcc.dg/vect/pr104116-floor-div-pow2.c: New test.
* gcc.dg/vect/pr104116-floor-div.c: New test.
* gcc.dg/vect/pr104116-floor-mod-2.c: New test.
* gcc.dg/vect/pr104116-floor-mod-pow2.c: New test.
* gcc.dg/vect/pr104116-floor-mod.c: New test.
* gcc.dg/vect/pr104116-round-div-2.c: New test.
* gcc.dg/vect/pr104116-round-div-pow2.c: New test.
* gcc.dg/vect/pr104116-round-div.c: New test.
* gcc.dg/vect/pr104116-round-mod-2.c: New test.
* gcc.dg/vect/pr104116-round-mod-pow2.c: New test.
* gcc.dg/vect/pr104116-round-mod.c: New test.
* gcc.dg/vect/pr104116-round-udiv-2.c: New test.
* gcc.dg/vect/pr104116-round-udiv-pow2.c: New test.
* gcc.dg/vect/pr104116-round-udiv.c: New test.
* gcc.dg/vect/pr104116-round-umod-2.c: New test.
* gcc.dg/vect/pr104116-round-umod-pow2.c: New test.
* gcc.dg/vect/pr104116-round-umod.c: New test.
* gcc.dg/vect/pr104116.h: New test.
There are 2 fixes for these 2 patterns.
1) Reuse the (a|b) expression instead of recreating it
Fixed by capturing the bit_ior expression and using that instead
of a new expression.
2) Use the correct 0. Fixed by capturing the integer_zerop and using that
instead of integer_zero_node.
2) could be fuxed by using `build_cst_zero (TREE_TYPE (@0))` But since
we already have the correct 0, capturing it would be faster.
Pushed as obvious after a bootstrap/test on x86_64-linux-gnu.
PR tree-optimization/122296
gcc/ChangeLog:
* match.pd (`(a != b) | ((a|b) != 0)`): Reuse both
the ior and zero instead of recreating them.
(`(a == b) & ((a|b) == 0)`): Likewise
gcc/testsuite/ChangeLog:
* gcc.dg/torture/int-bwise-opt-1.c: New test.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
The pattern `(a == b) | ((a|b) != 0)` uses build_one_cst to build boolean true
but boolean can be a signed multi-bit type. So this changes the result to
use constant_boolean_node isntead.
`(a != b) & ((a|b) == 0)` has a similar issue but in that case it is less likely
to be an issue as false is almost always just 0 but this changes it to be consistent.
Pushed as obvious after a bootstrap/test on x86_64-linux-gnu.
PR tree-optimization/122296
gcc/ChangeLog:
* match.pd (`(a == b) | ((a|b) != 0)`): Fix true value.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/int-bwise-opt-vect01.c: New test.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
On 64-bit windows, long can't be used, because it is 32 bits. Use
__PTRDIFF_TYPE__ instead of long.
gcc/ChangeLog:
PR target/122119
* config/i386/amxmovrsintrin.h
(_tile_loaddrs_internal): Use __PTRDIFF_TYPE__ instead of long.
(_tile_loaddrst1_internal): Ditto.
In r16-1631-g2334d30cd8feac I added support for capturing state
information from -fanalyzer in the form of embedded XML strings
in SARIF output.
In r16-2211-ga5d9debedd2f46 I rewrote this so the state was captured in
the form of a SARIF directed graph, using various custom types.
I want to add the ability to capture other kinds of graph in our SARIF
output (e.g. inheritance hierarchies, CFGs, etc), so the following patch
reworks the state graph handling code to minimize the use of custom types.
Instead, the patch introduces various json::property types, and
describes the state graph serialization in terms of instances of these
properties, rather than hardcoding string attribute names in readers and
writers. The custom SARIF properties live in a new
"gcc/custom-sarif-properties/" directory.
The "experimental-html" scheme keys "show-state-diagrams-dot-src" and
"show-state-diagrams-sarif" become "show-graph-dot-src" and
"show-graph-dot-src" in preparation for new kinds of graph in the output.
This is an updated version of the patch, tested to build with GCC 5
(which the previous version didn't leading to PR bootstrap/122151)
contrib/ChangeLog:
* gcc.doxy (INPUT): Add gcc/custom-sarif-properties
gcc/ChangeLog:
* Makefile.in (OBJS-libcommon): Add
custom-sarif-properties/digraphs.o and
custom-sarif-properties/state-graphs.o. Remove
diagnostics/state-graphs.o.
* configure: Regenerate.
* configure.ac: Add custom-sarif-properties to subdir iteration.
* custom-sarif-properties/digraphs.cc: New file.
* custom-sarif-properties/digraphs.h: New file.
* custom-sarif-properties/state-graphs.cc: New file.
* custom-sarif-properties/state-graphs.h: New file.
* diagnostics/diagnostics-selftests.cc
(run_diagnostics_selftests): Drop call of state_graphs_cc_tests.
* diagnostics/diagnostics-selftests.h (state_graphs_cc_tests):
Delete decl.
* diagnostics/digraphs.cc: Include
"custom-sarif-properties/digraphs.h". Move include of
"selftest.h" to within CHECKING_P section.
(using digraph_object): New.
(namespace properties): New.
(diagnostics::digraphs::object::get_attr): Delete.
(diagnostics::digraphs::object::set_attr): Delete.
(diagnostics::digraphs::object::set_json_attr): Delete.
(digraph_object::get_property): New definitions, for various
property types.
(digraph_object::set_property): Likewise.
(digraph_object::maybe_get_property): New.
(digraph_object::get_property_as_tristate): New.
(digraph_object::ensure_property_bag): New.
(digraph::get_graph_kind): New.
(digraph::set_graph_kind): New.
Add include of "custom-sarif-properties/state-graphs.h".
(selftest::test_simple_graph): Rewrite to use json::property
instances rather than string attribute names.
(selftest::test_property_objects): New test.
(selftest::digraphs_cc_tests): Call it.
* diagnostics/digraphs.h: Include "tristate.h".
(object::get_attr): Delete.
(object::set_attr): Delete.
(object::get_property): New decls.
(object::set_property): New decls.
(object::maybe_get_property): New.
(object::get_property_as_tristate): New.
(object::set_json_attr): Delete.
(object::ensure_property_bag): New.
(graph::get_graph_kind): New.
(graph::set_graph_kind): New.
* diagnostics/html-sink.cc
(html_generation_options::html_generation_options): Update for
field renamings.
(html_generation_options::dump): Likewise.
(html_builder::maybe_make_state_diagram): Likewise.
(html_builder::add_graph): Show SARIF and .dot src inline, if
requested.
* diagnostics/html-sink.h
(html_generation_options::m_show_state_diagrams_sarif): Rename
to...
(html_generation_options::m_show_graph_sarif): ...this.
(html_generation_options::m_show_state_diagrams_dot_src): Rename
to...
(html_generation_options::m_show_graph_dot_src0): ...this.
* diagnostics/output-spec.cc
(html_scheme_handler::maybe_handle_kv): Rename keys.
(html_scheme_handler::get_keys): Likewise.
* diagnostics/state-graphs-to-dot.cc: : Reimplement throughout to
use json::property instances found within custom_sarif_properties
throughout, rather than types in diagnostics::state_graphs.
* diagnostics/state-graphs.cc: Deleted file.
* diagnostics/state-graphs.h: Delete almost all, except decl of
diagnostics::state_graphs::make_dot_graph.
* doc/invoke.texi: Update for changes to "experimental-html" sink
keys.
* json.cc (json::object::set_string): New.
(json::object::set_integer): New.
(json::object::set_bool): New.
(json::object::set_array_of_string): New.
* json.h: Include "label-text.h".
(struct json::property): New template.
(json::string_property): New.
(json::integer_property): New.
(json::bool_property): New.
(json::json_property): New.
(using json::array_of_string_property): New.
(struct json::enum_traits): New.
(enum_json::property): New.
(json::value::dyn_cast_array): New vfunc.
(json::value::dyn_cast_integer_number): New vfunc.
(json::value::set_string): New.
(json::value::set_integer): New.
(json::value::set_bool): New.
(json::value::set_array_of_string): New.
(json::value::maybe_get_enum): New.
(json::value::set_enum): New.
(json::array::dyn_cast_array): New.
(json::integer_number::dyn_cast_integer_number): New.
(object::maybe_get_enum): New.
(object::set_enum): New.
gcc/analyzer/ChangeLog:
* ana-state-to-diagnostic-state.cc: Reimplement throughout to use
json::property instances found within custom_sarif_properties
throughout, rather than types in diagnostics::state_graphs.
* ana-state-to-diagnostic-state.h: Likewise.
* checker-event.cc: Likewise.
* sm-malloc.cc: Likewise.
gcc/testsuite/ChangeLog:
* gcc.dg/plugin/diagnostic_plugin_test_graphs.cc
(report_diag_with_graphs): Port from set_attr to set_property.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
With this change stacktrace entries always output the frame address, and
source file information no longer results in " at :0", e.g.
16# myfunc(int) at /tmp/bt.cc:48 [0x4008b7]
17# main at /tmp/bt.cc:61 [0x40091a]
18# __libc_start_call_main [0x7efc3d6d3574]
19# __libc_start_main@GLIBC_2.2.5 [0x7efc3d6d3627]
20# _start [0x400684]
This replaces the previous output:
16# myfunc(int) at /tmp/bt.cc:48
17# main at /tmp/bt.cc:61
18# __libc_start_call_main at :0
19# __libc_start_main@GLIBC_2.2.5 at :0
20# _start at :0
A change that is not visible in the examples above is that for a
non-empty stacktrace_entry, we now print "<unknown>" for the function
name if description() returns an empty string. For an empty (e.g.
default constructed) stacktrace_entry the entire string representation
is now "<unknown>" instead of an empty string.
Instead of printing "<unknown>" for the function name, we could set that
string in the stacktrace_entry::_Info object, so that description()
returns "<unknown>" and then operator<< wouldn't need to handle an empty
description() string. However, returning an empty string from that
function seems simpler for users to detect, rather than having to parse
"<unknown>".
We could also choose a different string for an empty stacktrace_entry,
maybe "<none>" or "<invalid>", but "<unknown>" seems good.
libstdc++-v3/ChangeLog:
* include/std/stacktrace
(operator<<(ostream&, const stacktrace_entry&)): Improve output
when description() or source_file() returns an empty string,
or the stacktrace_entry is invalid. Append frame address to
output.
(operator<<(ostream&, const basic_stacktrace<A>&)): Use the
size_type of the correct specialization.
Reviewed-by: Tomasz Kamiński <tkaminsk@redhat.com>
Reviewed-by: Nathan Myers <nmyers@redhat.com>
stack-protector is not supported in GCC on AIX. This patch is to fail the
compilation if -fstack-protector option is passed.
gcc/ChangeLog:
* config/rs6000/aix.h (SUBTARGET_DRIVER_SELF_SPECS):
Error out when stack-protector option is used in AIX
as it is not supported on AIX
Approved By: Segher Boessenkool <segher@kernel.crashing.org>
Add missing tests for gfx* context selectors; mark all but the
default-arch declare-variant-4.c with 'dg-excess-errors' to
silence libgomp not-found errors (still passing the
scan-offload-tree-dump check) - or at least causing just
UNRESOLVED errors if the error is
"built without library support ... consider compiling for
the associated generic architecture".
In case the multilib is configured, the result will be
an XPASS.
libgomp/ChangeLog:
* testsuite/libgomp.c/declare-variant-4-gfx10-3-generic.c: Add
dg-excess-errors to handle possible missing libgomp multi lib.
* testsuite/libgomp.c/declare-variant-4-gfx1030.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1036.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx11-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1100.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1103.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx9-4-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx9-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx900.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx906.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx908.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx90a.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx90c.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx942.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1031.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1032.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1033.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1034.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1035.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1101.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1102.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1150.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1151.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1152.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1153.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx902.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx904.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx909.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx950.c: New test.
The following fixes a mixup of vector types checked when looking
at a conditional reduction operation. We want the actual data
vector input type, so look at the SLP trees type instead and
special-case lane-reducing ops like the original code did.
PR tree-optimization/122292
* tree-vect-loop.cc (vect_transform_reduction): Compute the
input vector type the same way the analysis phase does.
Range bounds adjustments based on a bitmask were lazily set. This lead
to some inconsitencies which were causing problems. Improve the bounds,
and do it every time the bitmask is adjusted.
PR tree-optimization/121468
PR tree-optimization/121206
PR tree-optimization/122200
gcc/
* value-range.cc (irange_bitmask::range_from_mask): New.
(irange::snap): Add explicit overflow flag.
(irange::snap_subranges): Use overflow flag.
(irange::set_range_from_bitmask): Use range_from_mask.
(test_irange_snap_bounds): Adjust for improved ranges.
* value-range.h (irange::range_from_mask): Add prototype.
(irange::snap): Adjust prototype.
gcc/testsuite/
* gcc.dg/pr121468.c: New.
* gcc.dg/pr122200.c: New.
libstdc++-v3/ChangeLog:
* python/libstdcxx/v6/printers.py (StdStacktraceEntryPrinter):
New printer for std::stacktrace_entry.
(StdStacktracePrinter): New printer for std::basic_stacktrace.
The backtrace_simple function seems to consistently invoke the callback
with an invalid -1UL value as the last entry, which seems to come from
_Unwind_Backtrace. The glibc backtrace(3) function has a special case to
not include that final invalid address, but libbacktrace doesn't seem to
handle it. Do so in std::stacktrace::current() instead.
libstdc++-v3/ChangeLog:
* include/std/stacktrace (basic_stacktrace::current): Call
_M_trim before returning.
(basic_stacktrace::_M_trim): New member function.
The preprocessor condition for defining the new __to_timeout_timespec
function templates did not match all the conditions under which it's
needed.
std::this_thread::sleep_for is defined #if ! defined _GLIBCXX_NO_SLEEP
but it relies on __to_timeout_timespec which was only being defined for
targets that use nanosleep, or clock_gettime, or use gthreads.
For a non-gthreads target that uses POSIX sleep to implement
std::this_thread::sleep_for, the build fails with:
include/bits/this_thread_sleep.h:71:40: error: '__to_timeout_timespec' is not a member of 'std::chrono' [-Wtemplate-body]
71 | struct timespec __ts = chrono::__to_timeout_timespec(__rtime);
| ^~~~~~~~~~~~~~~~~~~~~
Presumably the same would happen for mingw-w64 if configured with
--disable-threads (as that would be a non-gthreads target that doesn't
use nanosleep or clock_gettime).
libstdc++-v3/ChangeLog:
PR libstdc++/122293
* include/bits/chrono.h (__to_timeout_timespec): Fix
preprocessor condition to match the conditions under which
callers of this function are defined.
* include/bits/this_thread_sleep.h: Remove unused include.
The extra \; parameter in the find command causes it to fail immediately and
not clean any config.cache:
$ find . -name config.cache -exec rm -f {} \; \;
find: paths must precede expression: `;'
This is benign in most cases but the binutils is also using this Makefile.tpl and
as the result its 'make distclean' can leave config.cache files around, which
fails subsequent attempts to configure and build it.
I have modified the Makefile.tpl and regenerated Makefile.in from it. For testing
I ran a config/make/make distclean loop.
* Makefile.tpl (distclean): Remove extraenous semicolon.
* Makefile.in: Rebuilt.
The definition for gfx942 and gfx950 missed the GFX9_4_GENERIC
family flag.
For OpenMP context selectors: The t-omp-device file missed the
generic selectors.
Additionally, there is now a note in the OpenMP documentation that
there is a one-to-one match for ISA names, ignoring any compatibility.
For instance, for Nvidia GPUs 'isa("sm_70")' is only true when compiling
for 'sm_70', even though sm < 7.0 code also runs on sm_70 hardware.
And, for AMD GPUs, gfx9-4-generic neither matches 'gfx942'
(even though such generic code runs on gfx942) - nor the reverse
(although all gfx9-4-generic code runs on gfx942).
gcc/ChangeLog:
* config/gcn/gcn-devices.def (gfx942, gfx950): Set generic name
to GFX9_4_GENERIC.
* config/gcn/t-omp-device: Include generic names for OpenMP's
ISA trait.
libgomp/ChangeLog:
* libgomp.texi (OpenMP Context Selectors): Add note that there is
currently an exact match between ISA and compilation, ignoring
compatibilities in both ways.
* testsuite/libgomp.c/declare-variant-4.h: Add missing variant
functions for specific and generic AMD GPUs.
* testsuite/libgomp.c/declare-variant-4-gfx10-3-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx11-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx9-4-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx9-generic.c: New test.
While debugging PR 122273, I noticed that print_node was not
printing out the clique/base for MEM_REF/TARGET_MEM_REF. This
made harder to understand why operand_equal_p (without looking
into the code) would be rejecting two looking the same MEM_REFs.
Changes since v1:
* v2: Don't print out clique/base if clique is 0.
Bootstrapped and tested on x86_64-linux-gnu.
gcc/ChangeLog:
* print-tree.cc (print_node): Print out clique/base
for MEM_REF and TARGET_MEM_REF.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
When compiling for m-profile with the floating-point extension we have
a vsel instruction that takes a limited set of comparisons. In most
cases we can use this with careful selection of the operand order, but
we need to expand things in the right way. This patch is in two parts:
1) We validate that the expansion will produce correct RTL;
2) We canonicalize the comparison to increase the chances that the
above check will pass.
gcc:
PR target/118460
* config/arm/arm.cc (arm_canonicalize_comparison): For floating-
point comparisons, swap the operand order if that will be more
likely to produce a comparison that can be used with VSEL.
(arm_validize_comparison): Make sure that HFmode comparisons
are compatible with VSEL.
gcc/testsuite:
PR target/118460
* gcc.target/arm/armv8_2-fp16-move-1.c: Adjust expected output.
* gcc.target/arm/armv8_2-fp16-move-2.c: Likewise.
__builtin_stack_save can be removed when the lhs becomes unused
as it is just recording the current StackPointer into another register.
Bootstrapped and tested on x86_64-linux-gnu.
PR tree-optimization/122037
gcc/ChangeLog:
* tree-ssa-dce.cc (eliminate_unnecessary_stmts): Remove
__builtin_stack_save when the lhs is unused.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/vla-1.c: New test.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
This patch incorporates changes to this file in Binutils since March
2024 (excluding one patch that was already cherry-picked by
Ezra in July 2025). It includes:
- New system registers in the 2024 and 2025 architecture extensions.
- Updated feature requirements for most system register accessors.
- Removal of registers that were dropped from the architecture.
- Removal of the unnecessary F_ARCHEXT flag.
- Fixed encoding for pmsdsfr_el1.
The updated architecture feature requirements are only relevant when the
new `-menable-sysreg-checking' option is enabled.
gcc/ChangeLog:
* config/aarch64/aarch64-sys-regs.def: Copy from Binutils.
* config/aarch64/aarch64.cc (F_ARCHEXT): Delete flag.
* config/aarch64/aarch64.h
(AARCH64_FL_AMU): Delete unused macro.
(AARCH64_FL_SCXTNUM): Ditto.
(AARCH64_FL_ID_PFR2): Ditto.
(AARCH64_FL_AIE): Ditto.
(AARCH64_FL_DEBUGv8p9): Ditto.
(AARCH64_FL_FGT2): Ditto.
(AARCH64_FL_PFAR): Ditto.
(AARCH64_FL_PMUv3_ICNTR): Ditto.
(AARCH64_FL_PMUv3_SS): Ditto.
(AARCH64_FL_PMUv3p9): Ditto.
(AARCH64_FL_S1PIE): Ditto.
(AARCH64_FL_S1POE): Ditto.
(AARCH64_FL_S2PIE): Ditto.
(AARCH64_FL_S2POE): Ditto.
(AARCH64_FL_SCTLR2): Ditto.
(AARCH64_FL_SEBEP): Ditto.
(AARCH64_FL_SPE_FDS): Ditto.
(AARCH64_FL_TCR2): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/acle/rwsr-armv8p9.c: Fix incorrect encoding.
This patch adds runtime thread count detection to auto-parallelization.
-ftree-parallelize-loops option generates parallelized loops without
specifying a fixed thread count, deferring this decision to program execution
time where it is controlled by the OMP_NUM_THREADS environment variable.
Bootstrap and regression tested on aarch64-linux. Compiled SPEC HPC pot3d
https://www.spec.org/hpc2021/docs/benchmarks/628.pot3d_s.html with
-ftree-parallelize-loops and tested without having OMP_NUM_THREADS set in the
environment and with OMP_NUM_THREADS set to different values.
gcc/ChangeLog:
* doc/invoke.texi (ftree-parallelize-loops): Update.
* common.opt (ftree-parallelize-loops): Add alias that maps to
special value INT_MAX for runtime thread detection.
* tree-parloops.cc (create_parallel_loop): Use INT_MAX for runtime
detection. Call gimple_build_omp_parallel without building a
OMP_CLAUSE_NUM_THREADS clause.
(gen_parallel_loop): For auto-detection, use a conservative
estimate of 2 threads.
(parallelize_loops): Same.
gcc/testsuite/ChangeLog:
* gcc.dg/autopar/runtime-auto.c: New test.
Signed-off-by: Sebastian Pop <spop@nvidia.com>
This patch resolves PR rtl-optimization/122266 by changing the types
of the last_set_sign_bit_copies and sign_bit_copies fields in combine.cc's
reg_stat_type struct to be "unsigned short". This makes both types
consistent, and fixes the issue that on platforms where char is by
default signed, combine.cc can overflow when handling TImode values,
where sign_bit_copies can be 128 bits.
Conveniently, there are holes (caused by field alignment/padding) in the
reg_stat_type struct that allows us to upgrade to "unsigned short" without
increasing the total size of the struct. This should help reduce problems
in future handling OImode or XImode values, or possible issues with 256-bit
and 512-bit vector modes. Note that it's important to take care when
reordering the fields of this struct, as the (partial) ordering of fields
is significant: See the use of offsetof in combine.cc's init_reg_last.
Before:
(gdb) ptype /o reg_stat_type
/* offset | size */ type = struct reg_stat_type {
/* 0 | 8 */ rtx_insn *last_death;
/* 8 | 8 */ rtx_insn *last_set;
/* 16 | 8 */ rtx last_set_value;
/* 24 | 4 */ int last_set_table_tick;
/* 28 | 4 */ int last_set_label;
/* 32 | 8 */ unsigned long last_set_nonzero_bits;
/* 40 | 1 */ char last_set_sign_bit_copies;
/* 41: 0 | 4 */ machine_mode last_set_mode : 16;
/* 43 | 1 */ bool last_set_invalid;
/* 44 | 1 */ unsigned char sign_bit_copies;
/* XXX 3-byte hole */
/* 48 | 8 */ unsigned long nonzero_bits;
/* 56 | 4 */ int truncation_label;
/* 60: 0 | 4 */ machine_mode truncated_to_mode : 16;
/* XXX 2-byte padding */
/* total size (bytes): 64 */
}
After:
/* offset | size */ type = struct reg_stat_type {
/* 0 | 8 */ rtx_insn *last_death;
/* 8 | 8 */ rtx_insn *last_set;
/* 16 | 8 */ rtx last_set_value;
/* 24 | 4 */ int last_set_table_tick;
/* 28 | 4 */ int last_set_label;
/* 32 | 8 */ unsigned long last_set_nonzero_bits;
/* 40 | 2 */ unsigned short last_set_sign_bit_copies;
/* 42: 0 | 4 */ machine_mode last_set_mode : 16;
/* 44 | 1 */ bool last_set_invalid;
/* XXX 1-byte hole */
/* 46 | 2 */ unsigned short sign_bit_copies;
/* 48 | 8 */ unsigned long nonzero_bits;
/* 56 | 4 */ int truncation_label;
/* 60: 0 | 4 */ machine_mode truncated_to_mode : 16;
/* XXX 2-byte padding */
/* total size (bytes): 64 */
}
2025-10-15 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR rtl-optimization/122266
* combine.cc (struct reg_stat_type): Change types of sign_bit_copies
and last_set_sign_bit_copies to unsigned short, to avoid overflows
on TImode (and wider) values.
gcc/testsuite/ChangeLog
PR rtl-optimization/122266
* gcc.target/i386/pr122266.c: New test case.
profile_count::max is not implemented same way as other arithmetics on
profile counts which generally require counts to be compatible and
returns minimum of qualities of input counts. Reason is that originally
it was used to compute statistics of whole callgraph profile so inliner
weights can be scaled to reasonable integers interprocedurally. It also
combines qulities weird way so the same counter could be used to
determine what quality of profile is available. That code had roundoff
error issues and was replaced by sreals.
Now max is mostly used to determine cfg->max_count which is used to
scale counts to reasonable integers intraprocedurally and is still being
used i.e. by IRA. There are also few places where max is used for
normal arithmetics when updating profile.
For computing max_count we need max to still be a bit special so max
(uninitialized, initialized) returns initialized rather then
uninitialized. Partial profiles are later handled specially.
This patch renames max to max_prefer_initialized to make it clear and updates
implementation to require compatible profiles. I checked this behaviour is good
for other places using it as well.
I also turned function to static, since a = a->max (b) looks odd.
gcc/ChangeLog:
* auto-profile.cc (scale_bb_profile): Use
profile_count::max_prefer_initialized.
(afdo_adjust_guessed_profile): Likewise.
* bb-reorder.cc (edge_order): Do not use max.
* cfghooks.cc (merge_blocks): Likewise.
* ipa-fnsummary.cc (param_change_prob): Likewise.
* ipa-inline-transform.cc (inline_transform): Likewise.
* predict.cc (update_max_bb_count): Likewise.
(estimate_bb_frequencies): Likewise.
(rebuild_frequencies): Likewise.
* tree-ssa-loop-unswitch.cc (struct unswitch_predicate): Likewise.
* profile-count.h (profile_count::max): Rename to
(profile_count::max_prefer_initialized): this; update handling
of qualities.
Add Wildcat Lake support according to ISE.
gcc/ChangeLog:
* common/config/i386/cpuinfo.h
(get_intel_cpu): Handle Wildcat Lake.
* common/config/i386/i386-common.cc (processor_name):
Add Wildcat Lake.
* doc/invoke.texi: Ditto.
Iain Sandoe