| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
While testing a later patch, I found that create_degenerate_phi
had an inverted test for bitmap_set_bit. It was assuming that
the return value was the previous bit value, rather than a
"something changed" value. :(
Also, the call to add_live_out_use shouldn't be conditional
on the DF_LR_OUT operation, since the register could be live-out
because of uses later in the same EBB (which do not require a
live-out use to be added to the rtl-ssa instruction). Instead,
add_live_out should itself check whether a live-out use already exists.
gcc/
* rtl-ssa/blocks.cc (function_info::create_degenerate_phi): Fix
inverted test of bitmap_set_bit. Call add_live_out_use even
if the register was previously live-out from the predecessor block.
Instead...
(function_info::add_live_out_use): ...check here whether a live-out
use already exists.
|
|
rtl-ssa already has a find_def function for finding the definition
of a particular resource (register or memory) at a particular point
in the program. This patch adds a similar function for looking
up uses. Both functions have amortised logarithmic complexity.
gcc/
* rtl-ssa/accesses.h (use_lookup): New class.
* rtl-ssa/functions.h (function_info::find_def): Expand comment.
(function_info::find_use): Declare.
* rtl-ssa/member-fns.inl (use_lookup::prev_use, use_lookup::next_use)
(use_lookup::matching_use, use_lookup::matching_or_prev_use)
(use_lookup::matching_or_next_use): New member functions.
* rtl-ssa/accesses.cc (function_info::find_use): Likewise.
|
|
The testcase in the PR shows that it's worth splitting the processing
of the initial workset, which is def_blocks from the main iteration.
This reduces SSA incremental update time from 44.7s to 32.9s. Further
changing the workset bitmap of the main iteration to a vector
speeds up things further to 23.5s for an overall nearly halving of
the SSA incremental update compile-time and an overall 12% compile-time
saving at -O1.
Using bitmap_ior in the first loop or avoiding (immediate) re-processing
of blocks in def_blocks does not make a measurable difference for the
testcase so I left this as-is.
PR tree-optimization/114480
* cfganal.cc (compute_idf): Split processing of the initial
workset from the main iteration. Use a vector for the
workset of the main iteration.
|
|
The linker rejects --relax in relocatable links (-r), hence only
add --relax when -r is not specified.
gcc/
PR target/121608
* config/avr/specs.h (LINK_RELAX_SPEC): Wrap in %{!r...}.
|
|
vect_analyze_slp_instance still handles stores and reduction chains.
The following threads the special handling of those two kinds,
duplicating vect_build_slp_instance into two specialized entries.
* tree-vect-slp.cc (vect_analyze_slp_reduc_chain): New,
copied from vect_analyze_slp_instance and only handle
slp_inst_kind_reduc_chain. Inline vect_build_slp_instance.
(vect_analyze_slp_instance): Only handle slp_inst_kind_store.
Inline vect_build_slp_instance.
(vect_build_slp_instance): Remove now unused stmt_info parameter,
remove special code for store groups and reduction chains.
(vect_analyze_slp): Call vect_analyze_slp_reduc_chain
for reduction chain SLP build and adjust.
|
|
The restriction no longer applies, so remove it.
* tree-vect-data-refs.cc (vect_check_gather_scatter):
Remove restriction on epilogue of epilogue vectorization.
|
|
The following removes the fixup we apply to pattern stmt operands
before code generating vector epilogues. This isn't necessary anymore
since the SLP graph now exclusively records the data flow. Similarly
fixing up of SSA references inside DR_REF of gather/scatter isn't
necessary since we now record the analysis result and avoid re-doing
it during transform.
What we still need to keep is the adjustment of the actual pointers
to gimple stmts from stmt_vec_info and the back-reference from the DRs.
* tree-vect-loop.cc (update_epilogue_loop_vinfo): Remove
fixing up pattern stmt operands and gather/scatter DR_REFs.
(find_in_mapping): Remove.
|
|
The following is a patch to make us record the get_load_store_info
results from load/store analysis and re-use them during transform.
In particular this moves where SLP_TREE_MEMORY_ACCESS_TYPE is stored.
A major hassle was (and still is, to some extent), gather/scatter
handling with it's accompaning gather_scatter_info. As
get_load_store_info no longer fully re-analyzes them but parts of
the information is recorded in the SLP tree during SLP build the
following goes and eliminates the use of this data in
vectorizable_load/store, instead recording the other relevant
part in the load-store info (namely the IFN or decl chosen).
Strided load handling keeps the re-analysis but populates the
data back to the SLP tree and the load-store info. That's something
for further improvement. This also shows that early classifying
a SLP tree as load/store and allocating the load-store data might
be a way to move back all of the gather/scatter auxiliary data
into one place.
Rather than mass-replacing references to variables I've kept the
locals but made them read-only, only adjusting a few elsval setters
and adding a FIXME to strided SLP handling of alignment (allowing
local override there).
The FIXME shows that while a lot of analysis is done in
get_load_store_type that's far from all of it. There's also
a possibility that splitting up the transform phase into
separate load/store def types, based on VMAT choosen, will make
the code more maintainable.
* tree-vectorizer.h (vect_load_store_data): New.
(_slp_tree::memory_access_type): Remove.
(SLP_TREE_MEMORY_ACCESS_TYPE): Turn into inline function.
* tree-vect-slp.cc (_slp_tree::_slp_tree): Do not
initialize SLP_TREE_MEMORY_ACCESS_TYPE.
* tree-vect-stmts.cc (check_load_store_for_partial_vectors):
Remove gather_scatter_info pointer argument, instead get
info from the SLP node.
(vect_build_one_gather_load_call): Get SLP node and builtin
decl as argument and remove uses of gather_scatter_info.
(vect_build_one_scatter_store_call): Likewise.
(vect_get_gather_scatter_ops): Remove uses of gather_scatter_info.
(vect_get_strided_load_store_ops): Get SLP node and remove
uses of gather_scatter_info.
(get_load_store_type): Take pointer to vect_load_store_data
instead of individual pointers.
(vectorizable_store): Adjust. Re-use get_load_store_type
result from analysis time.
(vectorizable_load): Likewise.
|
|
gcc/cobol/ChangeLog:
* genutil.cc (get_binary_value): Fix a comment.
* parse.y: udf_args_valid(): Fix loc calculation.
* symbols.cc (assert): extend_66_capacity(): Avoid assert(e < e2) in
-O0 build until symbol_table expansion is fixed.
libgcobol/ChangeLog:
* libgcobol.cc (format_for_display_internal): Handle NumericDisplay
properly.
(compare_88): Fix memory access error.
(__gg__unstring): Likewise.
|
|
gcc/fortran/ChangeLog:
* intrinsic.texi: Correct the example given for FRACTION.
Move the TEAM_NUMBER section to after the TANPI to align
with the order gven in the index.
|
|
We can't place a TLS call before a conditional jump in a basic block like
(code_label 13 11 14 4 2 (nil) [1 uses])
(note 14 13 16 4 [bb 4] NOTE_INSN_BASIC_BLOCK)
(jump_insn 16 14 17 4 (set (pc)
(if_then_else (le (reg:CCNO 17 flags)
(const_int 0 [0]))
(label_ref 27)
(pc))) "x.c":10:21 discrim 1 1462 {*jcc}
(expr_list:REG_DEAD (reg:CCNO 17 flags)
(int_list:REG_BR_PROB 628353713 (nil)))
-> 27)
since the TLS call will clobber flags register nor place a TLS call in a
basic block if any live caller-saved registers aren't dead at the end of
the basic block:
;; live in 6 [bp] 7 [sp] 16 [argp] 17 [flags] 19 [frame] 104
;; live gen 0 [ax] 102 106 108 116 117 118 120
;; live kill 5 [di]
Instead, we should place such call before all register setting basic
blocks which dominate the current basic block.
Keep track the replaced GNU and GNU2 TLS instructions. Use these info to
place the __tls_get_addr call and mark FLAGS register as dead.
gcc/
PR target/121572
* config/i386/i386-features.cc (replace_tls_call): Add a bitmap
argument and put the updated TLS instruction in the bitmap.
(ix86_get_dominator_for_reg): New.
(ix86_check_flags_reg): Likewise.
(ix86_emit_tls_call): Likewise.
(ix86_place_single_tls_call): Add 2 bitmap arguments for updated
GNU and GNU2 TLS instructions. Call ix86_emit_tls_call to emit
TLS instruction. Correct debug dump for before instruction.
gcc/testsuite/
PR target/121572
* gcc.target/i386/pr121572-1a.c: New test.
* gcc.target/i386/pr121572-1b.c: Likewise.
* gcc.target/i386/pr121572-2a.c: Likewise.
* gcc.target/i386/pr121572-2b.c: Likewise.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
|
|
This testcase is testing the difference between functions that are or are
not declared constexpr.
gcc/testsuite/ChangeLog:
* g++.dg/cpp26/expansion-stmt16.C: Add -fno-implicit-constexpr.
|
|
This testcase caused an ICE when mangling the invalid type-constraint in
write_requirement since write_type_constraint expects a TEMPLATE_TYPE_PARM.
Setting the trailing return type to NULL_TREE when a
return-type-requirement is found in place of a type-constraint prevents the
failed assertion in write_requirement. It also allows the invalid
constraint to be satisfied in some contexts to prevent redundant errors,
e.g. in concepts-requires5.C.
Bootstrapped and tested on x86_64-linux-gnu.
PR c++/120618
gcc/cp/ChangeLog:
* parser.cc (cp_parser_compound_requirement): Set type to
NULL_TREE for invalid type-constraint.
gcc/testsuite/ChangeLog:
* g++.dg/cpp2a/concepts-requires5.C: Don't require
redundant diagnostic in static assertion.
* g++.dg/concepts/pr120618.C: New test.
Suggested-by: Jason Merrill <jason@redhat.com>
|
|
When expanding malloc like functions, we copy the return register into a temporary
and then mark that temporary register with a noalias regnote and the alignment.
This works fine unless you are calling the function with a return type of void.
At this point then the valreg will be null and a crash will happen.
A few cleanups are included in this patch because it was easier to do the fix
with the cleanups added.
The start_sequence/end_sequence for ECF_MALLOC is no longer needed; I can't tell
if it was ever needed.
The emit_move_insn function returns the last emitted instruction anyways so
there is no reason to call get_last_insn as we can just use the return value
of emit_move_insn. This has been true since this code was originally added
so I don't understand why it was done that way beforehand.
Bootstrapped and tested on x86_64-linux-gnu.
PR middle-end/120024
gcc/ChangeLog:
* calls.cc (expand_call): Remove start_sequence/end_sequence
for ECF_MALLOC.
Check valreg before deferencing it when it comes to malloc like
functions. Use the return value of emit_move_insn instead of
calling get_last_insn.
gcc/testsuite/ChangeLog:
* gcc.dg/torture/malloc-1.c: New test.
* gcc.dg/torture/malloc-2.c: New test.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
|
|
When comparing constraints during correspondence checking for a using
from a partial specialization, we need to substitute the partial
specialization arguments into the constraints rather than the primary
template arguments. Otherwise we incorrectly reject e.g. the below
testcase as ambiguous since we substitute T=int* instead of T=int
into #1's constraints and don't notice the correspondence.
This patch corrects the recent r16-2771-gb9f1cc4e119da9 fix by using
outer_template_args instead of TI_ARGS of the DECL_CONTEXT, which
should always give the correct outer arguments for substitution.
PR c++/121351
gcc/cp/ChangeLog:
* class.cc (add_method): Use outer_template_args when
substituting outer template arguments into constraints.
gcc/testsuite/ChangeLog:
* g++.dg/cpp2a/concepts-using7.C: New test.
Reviewed-by: Jason Merrill <jason@redhat.com>
|
|
Historically SLP reduction chains were the only multi-stmt reductions
supported. But since we have check_reduction_path more complicated
cases are handled. As parloops doesn't do any specific chain
processing it can solely rely on that functionality instead.
* tree-parloops.cc (parloops_is_slp_reduction): Remove.
(parloops_is_simple_reduction): Do not call it.
|
|
The following fixes another few missed cases to pass a SLP node
instead of a stmt_info.
* tree-vect-loop.cc (vectorizable_reduction): Pass the
appropriate SLP node for costing of single-def-use-cycle
operations.
(vectorizable_live_operation): Pass the SLP node to the
costing hook.
* tree-vect-stmts.cc (vectorizable_bswap): Likewise.
(vectorizable_store): Likewise.
|
|
The testcase in the PR shows that when we have a reduction chain
with a wrapped conversion we fail to properly fall back to a
regular reduction, resulting in wrong-code. The following fixes
this by failing discovery. The testcase has other issues, so
I'm not including it here.
PR tree-optimization/121592
* tree-vect-slp.cc (vect_analyze_slp): When SLP reduction chain
discovery fails, fail overall when the tail of the chain
isn't also the entry for the non-SLP reduction.
|
|
Building riscv no longer works with python2:
> python ./config/riscv/arch-canonicalize -misa-spec=20191213 rv64gc
File "./config/riscv/arch-canonicalize", line 229
print(f"ERROR: Unhandled conditional dependency: '{ext_name}' with condition:", file=sys.stderr)
^
SyntaxError: invalid syntax
On systems that have python aliased to python2 we chose that, even
when python3 is available. Don't.
* config.gcc (riscv*-*-*): Look for python3, then fall back
to python. Never use python2.
|
|
SRA handles outermost VIEW_CONVERT_EXPRs but it wrongly ignores
those when building an access which leads to the wrong size
used when the VIEW_CONVERT_EXPR does not have the same size as
its operand which is valid GENERIC and is used by Ada upcasting.
PR tree-optimization/121527
* tree-sra.cc (build_access_from_expr_1): Do not strip an
outer VIEW_CONVERT_EXPR as it's relevant for the size of
the access.
(get_access_for_expr): Likewise.
|
|
commit g:1786be14e94bf1a7806b9dc09186f021737f0227 stops storing in
STMT_VINFO_VECTYPE the vectype of the current stmt being vectorized and instead
requires the use of SLP_TREE_VECTYPE for everything but data-refs.
This means that STMT_VINFO_VECTYPE (stmt_info) will always be NULL and so
aarch64_bool_compound_p will never properly cost predicate AND operations
anymore resulting in less vectorization.
This patch changes it to use SLP_TREE_VECTYPE and pass the slp_node to
aarch64_bool_compound_p.
gcc/ChangeLog:
PR target/121536
* config/aarch64/aarch64.cc (aarch64_bool_compound_p): Use
SLP_TREE_VECTYPE instead of STMT_VINFO_VECTYPE.
(aarch64_adjust_stmt_cost, aarch64_vector_costs::count_ops): Pass SLP
node to aarch64_bool_compound_p.
gcc/testsuite/ChangeLog:
PR target/121536
* g++.target/aarch64/sve/pr121536.cc: New test.
|
|
commit g:1786be14e94bf1a7806b9dc09186f021737f0227 stops storing in
STMT_VINFO_VECTYPE the vectype of the current stmt being vectorized and instead
requires the use of SLP_TREE_VECTYPE for everything but data-refs.
However contrary to what the commit says not all usages of STMT_VINFO_VECTYPE
have been purged from vectorizable_* as the costing hooks which don't pass the
SLP tree as an agrument will extract vectype using STMT_VINFO_VECTYPE.
This results in no vector type being passed to the backends and results in a few
costing test failures in AArch64.
This commit replaces the last few cases I could find, all except for in
vectorizable_reduction when single_defuse_cycle where the stmt being costed is
not the representative of the PHI in the SLP tree but rather the out of tree
reduction statement. So I've left that alone, but it does mean vectype is NULL.
Most likely this needs to use the overload where we pass an explicit vectype but
I wasn't sure so left it for now.
gcc/ChangeLog:
PR target/121536
* tree-vect-loop.cc (vectorizable_phi, vectorizable_recurr,
vectorizable_nonlinear_induction, vectorizable_induction): Pass slp_node
instead of stmt_info to record_stmt_cost.
|
|
commit g:fb59c5719c17a04ecfd58b5e566eccd6d2ac583a stops passing the scalar type
(confusingly named vectype) to the costing hook when doing scalar costing.
As a result, we could no longer distinguish between FPR and GPR scalar stmts.
A later commit also removed STMT_VINFO_VECTYPE from stmt_info.
This leaves the only remaining option to get the type of the original stmt in
the stmt_info. This patch does this when we're performing scalar costing.
Ideally I'd refactor this a bit because a lot of the hooks just need to know if
it's FP or not, but this seems pointless with the ongoing costing churn. So for
now this restores our costing.
gcc/ChangeLog:
PR target/121536
* config/aarch64/aarch64.cc (aarch64_vector_costs::add_stmt_cost): Set
vectype from type of lhs of gimple stmt.
|
|
This testcase (added in r16-3233-g7921bb4afcb7a3) mistakenly only
required C++14, but auto template paramaters are a C++17 feature.
PR c++/121578
gcc/testsuite/ChangeLog:
* g++.dg/abi/mangle83.C: Requires C++17.
Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
|
|
[PR120195]
We have logic to adjust a function decl if it gets re-declared as a
using-decl with different purviewness, but we also need to do the same
if it gets redeclared with different exportedness.
PR c++/120195
gcc/cp/ChangeLog:
* name-lookup.cc (do_nonmember_using_decl): Also handle change
in exportedness of a function.
gcc/testsuite/ChangeLog:
* g++.dg/modules/using-32_a.C: New test.
* g++.dg/modules/using-32_b.C: New test.
Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
|
|
I added a testcase for the (temporary) warning that we don't currently
support the 'gnu::optimize' or 'gnu::target' attributes in r15-10183;
however, some targets produce target nodes even when only an optimize
attribute is present. This adjusts the warning.
PR c++/108080
PR c++/121396
gcc/testsuite/ChangeLog:
* g++.dg/modules/pr108080.H: Also allow target warnings.
Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
|
|
|
|
This example used to work (with C) in GCC 14 before the
warning for different pointer types without a cast was changed
to an error.
The fix is to make the q variable `int*` rather than the current `char*`.
This also fixes the example for C++ too.
Pushed as obvious after doing a `make html`.
PR middle-end/121581
gcc/ChangeLog:
* doc/extend.texi (__builtin_object_size): Fix example.
Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>
|
|
Currently, the data type of sanitizer flags is unsigned int, with
SANITIZE_SHADOW_CALL_STACK (1UL << 31) being highest individual
enumerator for enum sanitize_code. Use 'sanitize_code_type' data type
to allow for more distinct instrumentation modes be added when needed.
gcc/ChangeLog:
* flag-types.h (sanitize_code_type): Define.
* asan.h (sanitize_flags_p): Use 'sanitize_code_type' instead of
'unsigned int'.
* common.opt: Likewise.
* dwarf2asm.cc (dw2_output_indirect_constant_1): Likewise.
* opts.cc (find_sanitizer_argument): Likewise.
(report_conflicting_sanitizer_options): Likewise.
(parse_sanitizer_options): Likewise.
(parse_no_sanitize_attribute): Likewise.
* opts.h (parse_sanitizer_options): Likewise.
(parse_no_sanitize_attribute): Likewise.
* tree-cfg.cc (print_no_sanitize_attr_value): Likewise.
* tree.cc (tree_fits_sanitize_code_type_p): Define.
(tree_to_sanitize_code_type): Likewise.
* tree.h (tree_fits_sanitize_code_type_p): Declare.
(tree_to_sanitize_code_type): Likewise.
gcc/c-family/ChangeLog:
* c-attribs.cc (add_no_sanitize_value): Use 'sanitize_code_type'
instead of 'unsigned int'.
(handle_no_sanitize_attribute): Likewise.
(handle_no_sanitize_address_attribute): Likewise.
(handle_no_sanitize_thread_attribute): Likewise.
(handle_no_address_safety_analysis_attribute): Likewise.
* c-common.h (add_no_sanitize_value): Likewise.
gcc/c/ChangeLog:
* c-parser.cc (c_parser_declaration_or_fndef): Use
'sanitize_code_type' instead of 'unsigned int'.
gcc/cp/ChangeLog:
* typeck.cc (get_member_function_from_ptrfunc): Use
'sanitize_code_type' instead of 'unsigned int'.
gcc/d/ChangeLog:
* d-attribs.cc (d_handle_no_sanitize_attribute): Use
'sanitize_code_type' instead of 'unsigned int'.
Signed-off-by: Claudiu Zissulescu <claudiu.zissulescu-ianculescu@oracle.com>
|
|
Define new constants to be used by the MTE pattern definitions.
gcc/
* config/aarch64/aarch64.md (MEMTAG_TAG_MASK): New define
constant.
(MEMTAG_ADDR_MASK): Likewise.
(irg, subp, ldg): Use new constants.
Signed-off-by: Claudiu Zissulescu <claudiu.zissulescu-ianculescu@oracle.com>
|
|
This patch fixes an internal disagreement in gcse about how to
handle partial clobbers. Like many passes, gcse doesn't track
the modes of live values, so if a call clobbers only part of
a register, the pass has to make conservative assumptions.
As the comment in the patch says, this means:
(1) ignoring partial clobbers when computing liveness and reaching
definitions
(2) treating partial clobbers as full clobbers when computing
availability
DF is mostly concerned with (1), so ignores partial clobbers.
compute_hash_table_work did (2) when calculating kill sets,
but compute_transp didn't do (2) when computing transparency.
This led to a nonsensical situation of a register being in both
the transparency and kill sets.
gcc/
PR rtl-optimization/97497
* function-abi.h (predefined_function_abi::only_partial_reg_clobbers)
(function_abi::only_partial_reg_clobbers): New member functions.
* gcse-common.cc: Include regs.h and function-abi.h.
(compute_transp): Check for partially call-clobbered registers
and treat them as not being transparent in blocks with calls.
|
|
gcc/ChangeLog:
* config/loongarch/sync.md (UNSPEC_TI_FETCH_ADD): New unspec.
(UNSPEC_TI_FETCH_SUB): Likewise.
(UNSPEC_TI_FETCH_AND): Likewise.
(UNSPEC_TI_FETCH_XOR): Likewise.
(UNSPEC_TI_FETCH_OR): Likewise.
(UNSPEC_TI_FETCH_NAND_MASK_INVERTED): Likewise.
(ALL_SC): New define_mode_iterator.
(_scq): New define_mode_attr.
(atomic_fetch_nand<mode>): Accept ALL_SC instead of only GPR.
(UNSPEC_TI_FETCH_DIRECT): New define_int_iterator.
(UNSPEC_TI_FETCH): New define_int_iterator.
(amop_ti_fetch): New define_int_attr.
(size_ti_fetch): New define_int_attr.
(atomic_fetch_<amop_ti_fetch>ti_scq): New define_insn.
(atomic_fetch_<amop_ti_fetch>ti): New define_expand.
|
|
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_exchangeti_scq): New
define_insn.
(atomic_exchangeti): New define_expand.
|
|
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_compare_and_swapti_scq): New
define_insn.
(atomic_compare_and_swapti): New define_expand.
|
|
When LSX is not available but sc.q is (for example on LA664 where the
SIMD unit is not enabled), we can use a LL-SC loop for 16-byte atomic
store.
gcc/ChangeLog:
* config/loongarch/loongarch.cc (loongarch_print_operand_reloc):
Accept "%t" for printing the number of the 64-bit machine
register holding the upper half of a TImode.
* config/loongarch/sync.md (atomic_storeti_scq): New
define_insn.
(atomic_storeti): expand to atomic_storeti_scq if !ISA_HAS_LSX.
|
|
We'll use the sc.q instruction for some 16-byte atomic operations, but
it's only added in LoongArch 1.1 evolution so we need to gate it with
an option.
gcc/ChangeLog:
* config/loongarch/genopts/isa-evolution.in (scq): New evolution
feature.
* config/loongarch/loongarch-evolution.cc: Regenerate.
* config/loongarch/loongarch-evolution.h: Regenerate.
* config/loongarch/loongarch-str.h: Regenerate.
* config/loongarch/loongarch.opt: Regenerate.
* config/loongarch/loongarch.opt.urls: Regenerate.
* config/loongarch/loongarch-def.cc: Make -mscq the default for
-march=la664 and -march=la64v1.1.
* doc/invoke.texi (LoongArch Options): Document -m[no-]scq.
|
|
If the vector is naturally aligned, it cannot cross cache lines so the
LSX store is guaranteed to be atomic. Thus we can use LSX to do the
lock-free atomic store, instead of using a lock.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_storeti_lsx): New
define_insn.
(atomic_storeti): New define_expand.
|
|
If the vector is naturally aligned, it cannot cross cache lines so the
LSX load is guaranteed to be atomic. Thus we can use LSX to do the
lock-free atomic load, instead of using a lock.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_loadti_lsx): New define_insn.
(atomic_loadti): New define_expand.
|
|
Without atomic_fetch_nandsi and atomic_fetch_nanddi, __atomic_fetch_nand
is expanded to a loop containing a CAS in the body, and CAS itself is a
LL-SC loop so we have a nested loop. This is obviously not a good idea
as we just need one LL-SC loop in fact.
As ~(atom & mask) is (~mask) | (~atom), we can just invert the mask
first and the body of the LL-SC loop would be just one orn instruction.
gcc/ChangeLog:
* config/loongarch/sync.md
(atomic_fetch_nand_mask_inverted<GPR:mode>): New define_insn.
(atomic_fetch_nand<GPR:mode>): New define_expand.
|
|
With -mlam-bh, we should negate the addend first, and use an amadd
instruction. Disabling the expander makes the compiler do it correctly.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_fetch_sub<SHORT:mode>):
Disable if ISA_HAS_LAM_BH.
|
|
We can just shift the mask and fill the other bits with 0 (for ior/xor)
or 1 (for and), and use an am*.w instruction to perform the atomic
operation, instead of using a LL-SC loop.
gcc/ChangeLog:
* config/loongarch/sync.md (UNSPEC_COMPARE_AND_SWAP_AND):
Remove.
(UNSPEC_COMPARE_AND_SWAP_XOR): Remove.
(UNSPEC_COMPARE_AND_SWAP_OR): Remove.
(atomic_test_and_set): Rename to ...
(atomic_fetch_<any_bitwise:amop><SHORT:mode>): ... this, and
adapt the expansion to use it for any bitwise operations and any
val, instead of just ior 1.
(atomic_test_and_set): New define_expand.
|
|
atomic_test_and_set
On LoongArch sll.w and srl.w instructions only take the [4:0] bits of
rk (shift amount) into account, and we've already defined
SHIFT_COUNT_TRUNCATED to 1 so the compiler knows this fact, thus we
don't need this instruction.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_test_and_set): Remove
unneeded andi instruction from the expansion.
|
|
This instruction is used to skip an redundant barrier if -mno-ld-seq-sa
or the memory model requires a barrier on failure. But with -mld-seq-sa
and other memory models the barrier may be nonexisting at all, and we
should remove the "b 3f" instruction as well.
The implementation uses a new operand modifier "%T" to output a comment
marker if the operand is a memory order for which the barrier won't be
generated. "%T", and also "%t", are not really used before and the code
for them in loongarch_print_operand_reloc is just some MIPS legacy.
gcc/ChangeLog:
* config/loongarch/loongarch.cc (loongarch_print_operand_reloc):
Make "%T" output a comment marker if the operand is a memory
order for which the barrier won't be generated; remove "%t".
* config/loongarch/sync.md (atomic_cas_value_strong<mode>): Add
%T before "b 3f".
(atomic_cas_value_cmp_and_7_<mode>): Likewise.
|
|
For LL-SC loops, if the atomic operation has succeeded, the SC
instruction always imply a full barrier, so the barrier we manually
inserted only needs to take the account for the failure memorder, not
the success memorder (the barrier is skipped with "b 3f" on success
anyway).
Note that if we use the AMCAS instructions, we indeed need to consider
both the success memorder an the failure memorder deciding if "_db"
suffix is needed. Thus the semantics of atomic_cas_value_strong<mode>
and atomic_cas_value_strong<mode>_amcas start to be different. To
prevent the compiler from being too clever, use a different unspec code
for AMCAS instructions.
gcc/ChangeLog:
* config/loongarch/sync.md (UNSPEC_COMPARE_AND_SWAP_AMCAS): New
UNSPEC code.
(atomic_cas_value_strong<mode>): NFC, update the comment to note
we only need to consider failure memory order.
(atomic_cas_value_strong<mode>_amcas): Use
UNSPEC_COMPARE_AND_SWAP_AMCAS instead of
UNSPEC_COMPARE_AND_SWAP.
(atomic_compare_and_swap<mode:GPR>): Pass failure memorder to
gen_atomic_cas_value_strong<mode>.
(atomic_compare_and_swap<mode:SHORT>): Pass failure memorder to
gen_atomic_cas_value_cmp_and_7_si.
|
|
We can use bstrins for masking the address here. As people are already
working on LA32R (which lacks bstrins instructions), for future-proofing
we check whether (const_int -4) is an and_operand and force it into an
register if not.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_test_and_set): Use bstrins
for masking the address if possible.
|
|
Atomic load does not modify the memory. Atomic store does not read the
memory, thus we can use "=" instead.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_load<mode>): Remove "+" for
the memory operand.
(atomic_store<mode>): Use "=" instead of "+" for the memory
operand.
|
|
They are the same.
gcc/ChangeLog:
* config/loongarch/sync.md: Use <size> instead of <amo>.
(amo): Remove.
|
|
They are the same.
gcc/ChangeLog:
* config/loongarch/sync.md (atomic_optab): Remove.
(atomic_<atomic_optab><mode>): Change atomic_optab to amop.
(atomic_fetch_<atomic_optab><mode>): Likewise.
|
|
|
