This refactors instantiate_decl, breaking out the actual instantiation
work to instantiate_body. That'll allow me to address the OMP UDR
issue, but it also means we have slightly neater code in
instantiate_decl anyway.
gcc/cp/
* pt.c (instantiate_body): New, broken out of ..
(instantiate_decl): ... here. Call it.
gcc/ChangeLog
2020-09-09 Andrea Corallo <andrea.corallo@arm.com>
* tree-vect-loop.c (vect_need_peeling_or_partial_vectors_p): New
function.
(vect_analyze_loop_2): Make use of it not to select partial
vectors if no peel is required.
(determine_peel_for_niter): Move out some logic into
'vect_need_peeling_or_partial_vectors_p'.
gcc/testsuite/ChangeLog
2020-09-09 Andrea Corallo <andrea.corallo@arm.com>
* gcc.target/aarch64/sve/cost_model_10.c: New test.
* gcc.target/aarch64/sve/clastb_8.c: Update test for new
vectorization strategy.
* gcc.target/aarch64/sve/cost_model_5.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_14.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_15.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_16.c: Likewise.
* gcc.target/aarch64/sve/struct_vect_17.c: Likewise.
Add sp_is_clobbered_by_asm to rtl_data to inform backends that the stack
pointer is clobbered by asm statement.
gcc/
PR target/97032
* cfgexpand.c (asm_clobber_reg_kind): Set sp_is_clobbered_by_asm
to true if the stack pointer is clobbered by asm statement.
* emit-rtl.h (rtl_data): Add sp_is_clobbered_by_asm.
* config/i386/i386.c (ix86_get_drap_rtx): Set need_drap to true
if the stack pointer is clobbered by asm statement.
gcc/testsuite/
PR target/97032
* gcc.target/i386/pr97032.c: New test.
The following testcases will be simplified by the new rule
(T)(A) +- (T)(B) -> (T)(A +- B), so could not keep code pattern
expected by test-check. Adjust test code to suppress simplification.
2020-09-16 Feng Xue <fxue@os.amperecomputing.com>
gcc/testsuite/
PR testsuite/97066
* gcc.dg/ifcvt-3.c: Modified to suppress simplification.
* gcc.dg/tree-ssa/20030807-10.c: Likewise.
Certain alternatives of *vec_tf_to_v1tf use "v" constraint for its
TFmode source operand. Therefore it is assigned to VEC_REGS class,
and when it is reloaded using *movtf_64, whose relevant alternatives
need FP_REGS, LRA loops and ICE happens. The reason is that register
class mismatch causes LRA to emit another reload, which triggers this
issue again.
Fix by using "f" constraint, which is more appropriate for FP register
pairs anyway.
gcc/ChangeLog:
2020-09-02 Ilya Leoshkevich <iii@linux.ibm.com>
* config/s390/vector.md(*vec_tf_to_v1tf): Use "f" instead of "v"
for the source operand.
This removes STMT_VINFO_NUM_SLP_USES by pushing the setting of
the shared stmt_vec_info vector type to where we actually need it
which is alignment analysis and vectorizable_* analysis (where
we could eventually elide it for non-load/store operations).
In particular "uses" in the cache and in disqualified SLP
subgraphs should no longer provide conflicting vector types
this way.
2020-09-16 Richard Biener <rguenther@suse.de>
* tree-vectorizer.h (_stmt_vec_info::num_slp_uses): Remove.
(STMT_VINFO_NUM_SLP_USES): Likewise.
(vect_free_slp_instance): Adjust.
(vect_update_shared_vectype): Declare.
* tree-vectorizer.c (vec_info::~vec_info): Adjust.
* tree-vect-loop.c (vect_analyze_loop_2): Likewise.
(vectorizable_live_operation): Use vector type from
SLP_TREE_REPRESENTATIVE.
(vect_transform_loop): Adjust.
* tree-vect-data-refs.c (vect_slp_analyze_node_alignment):
Set the shared vector type.
* tree-vect-slp.c (vect_free_slp_tree): Remove final_p
parameter, remove STMT_VINFO_NUM_SLP_USES updating.
(vect_free_slp_instance): Adjust.
(vect_create_new_slp_node): Remove STMT_VINFO_NUM_SLP_USES
updating.
(vect_update_shared_vectype): Always compare with the
present vector type, update if NULL.
(vect_build_slp_tree_1): Do not update the shared vector
type here.
(vect_build_slp_tree_2): Adjust.
(slp_copy_subtree): Likewise.
(vect_attempt_slp_rearrange_stmts): Likewise.
(vect_analyze_slp_instance): Likewise.
(vect_analyze_slp): Likewise.
(vect_slp_analyze_node_operations_1): Update the shared
vector type.
(vect_slp_analyze_operations): Adjust.
(vect_slp_analyze_bb_1): Likewise.
2020-09-16 Jakub Jelinek <jakub@redhat.com>
* config/arm/arm.c (arm_option_restore): Comment out opts argument
name to avoid unused parameter warnings.
When working on the previous patch, I've noticed that all cl_optimization
fields appart from strings are streamed with bp_pack_value (..., 64); so we
waste quite a lot of space, given that many of the options are just booleans
or char options and there are 450-ish of them.
Fixed by streaming the number of bits the corresponding fields have.
While for char fields we have also range information, except for 3
it is either -128, 127 or 0, 255, so it didn't seem worth it to bother
with using range-ish packing.
2020-09-16 Jakub Jelinek <jakub@redhat.com>
* optc-save-gen.awk: In cl_optimization_stream_out use
bp_pack_var_len_{int,unsigned} instead of bp_pack_value. In
cl_optimization_stream_in use bp_unpack_var_len_{int,unsigned}
instead of bp_unpack_value. Formatting fix.
As the testcases show, if we have something like:
MEM <char[12]> [&b + 8B] = {};
MEM[(short *) &b] = 5;
_5 = *x_4(D);
MEM <long long unsigned int> [&b + 2B] = _5;
MEM[(char *)&b + 16B] = 88;
MEM[(int *)&b + 20B] = 1;
then in sort_by_bitpos the stores are almost like in the given order,
except the first store is after the = _5; store.
We can't coalesce the = 5; store with = _5;, because the latter is MEM_REF,
while the former INTEGER_CST, and we can't coalesce the = _5 store with
the = {} store because the former is MEM_REF, the latter INTEGER_CST.
But we happily coalesce the remaining 3 stores, which is wrong, because the
= _5; store overlaps those and is in between them in the program order.
We already have code to deal with similar cases in check_no_overlap, but we
deal only with the following stores in sort_by_bitpos order, not the earlier
ones.
The following patch checks also the earlier ones. In coalesce_immediate_stores
it computes the first one that needs to be checked (all the ones whose
bitpos + bitsize is smaller or equal to merged_store->start don't need to be
checked and don't need to be checked even for any following attempts because
of the sort_by_bitpos sorting) and the end of that (that is the first store
in the merged_store).
2020-09-16 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/97053
* gimple-ssa-store-merging.c (check_no_overlap): Add FIRST_ORDER,
START, FIRST_EARLIER and LAST_EARLIER arguments. Return false if
any stores between FIRST_EARLIER inclusive and LAST_EARLIER exclusive
has order in between FIRST_ORDER and LAST_ORDER and overlaps the to
be merged store.
(imm_store_chain_info::try_coalesce_bswap): Add FIRST_EARLIER argument.
Adjust check_no_overlap caller.
(imm_store_chain_info::coalesce_immediate_stores): Add first_earlier
and last_earlier variables, adjust them during iterations. Adjust
check_no_overlap callers, call check_no_overlap even when extending
overlapping stores by extra INTEGER_CST stores.
* gcc.dg/store_merging_31.c: New test.
* gcc.dg/store_merging_32.c: New test.
This patch is to extend the existing function find_alignment_op to
check all defintions of base_reg are AND operations with mask -16B
to force the alignment. If all are satifised, it passes all AND
operations and instructions to function recombine_lvx_pattern
and recombine_stvx_pattern, they can remove all useless ANDs
further.
Bootstrapped/regtested on powerpc64le-linux-gnu P8.
gcc/ChangeLog:
PR target/97019
* config/rs6000/rs6000-p8swap.c (find_alignment_op): Adjust to
support multiple defintions which are all AND operations with
the mask -16B.
(recombine_lvx_pattern): Adjust to handle multiple AND operations
from find_alignment_op.
(recombine_stvx_pattern): Likewise.
gcc/testsuite/ChangeLog:
PR target/97019
* gcc.target/powerpc/pr97019.c: New test.
The description in rs6000-builtin.def provides for a builtin named
__builtin_altivec_xst_len_r. However, it is hand-defined in
altivec_init_builtins as __builtin_xst_len_r, against the usual naming
practice. Fix that.
2020-09-15 Bill Schmidt <wschmidt@linux.ibm.com>
gcc/
* config/rs6000/rs6000-call.c (altivec_init_builtins): Fix name
of __builtin_altivec_xst_len_r.
PR analyzer/96650 reports an assertion failure when merging the
intersection of two sets of constraints, due to the resulting
constraints being infeasible.
It turns out that the two input sets were each infeasible if
transitivity were considered, but -fanalyzer-transitivity was off.
However for this case, the merging code was "discovering" the
transitive infeasibility of the intersection of the constraints even
when -fanalyzer-transitivity is off, triggering an assertion failure.
I attempted various fixes for this, but each of them would have
introduced O(N^2) logic into the constraint-handling code into the
-fno-analyzer-transitivity case (with N == the number of constraints).
This patch fixes the ICE by tweaking the assertion, so that we
silently drop such constraints if -fanalyzer-transitivity is off.
gcc/analyzer/ChangeLog:
PR analyzer/96650
* constraint-manager.cc (merger_fact_visitor::on_fact): Replace
assertion that add_constraint succeeded with an assertion that
if it fails, -fanalyzer-transitivity is off.
gcc/testsuite/ChangeLog:
PR analyzer/96650
* gcc.dg/analyzer/pr96650-1-notrans.c: New test.
* gcc.dg/analyzer/pr96650-1-trans.c: New test.
* gcc.dg/analyzer/pr96650-2-notrans.c: New test.
* gcc.dg/analyzer/pr96650-2-trans.c: New test.
The following s390 rtx is errneously considered a no-op:
(set (subreg:DF (reg:TF %f0) 8) (subreg:DF (reg:V1TF %f0) 8))
Here, SET_DEST is a second register in a floating-point register pair,
and SET_SRC is the second half of a vector register, so they refer to
different bits.
Fix by treating subregs of registers in different modes conservatively.
gcc/ChangeLog:
2020-09-11 Ilya Leoshkevich <iii@linux.ibm.com>
* rtlanal.c (set_noop_p): Treat subregs of registers in
different modes conservatively.
Since some time we're only using this argument to communicate from
vect_build_slp_tree_1 to vect_get_and_check_slp_defs. This makes
the direction of information flow clear.
2020-09-15 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (vect_get_and_check_slp_defs): Make swap
argument by-value and do not change it.
(vect_build_slp_tree_2): Adjust, set swap to NULL after last
use.
Turns out I didn't get OMP reductions correct. To address those I
need to do some reorganization, so this patch just reverts the
OMP-specific pieces of the local decl changes.
gcc/cp/
* pt.c (push_template_decl_real): OMP reductions retain a template
header.
(tsubst_function_decl): Likewise.
Add a rule (T)(A) +- (T)(B) -> (T)(A +- B), which works only when (A +- B)
could be folded to a simple value. By this rule, a plusminus-mult-with-convert
expression could be handed over to the rule (A * C) +- (B * C) -> (A +- B).
2020-09-15 Feng Xue <fxue@os.amperecomputing.com>
gcc/
PR tree-optimization/94234
* match.pd (T)(A) +- (T)(B) -> (T)(A +- B): New simplification.
gcc/testsuite/
PR tree-optimization/94234
* gcc.dg/pr94234-3.c: New test.
After the previous patch we are left with an unreachable BB. This will
ICE if either we have -fschedule-fusion, or we do not have peephole2.
2020-09-15 Segher Boessenkool <segher@kernel.crashing.org>
PR rtl-optimization/96475
* bb-reorder.c (duplicate_computed_gotos): If we did anything, run
cleanup_cfg.
The following allows more BB vectorization by generally building leafs
from scalars rather than giving up. Note this is only a first step
towards this and as can be seen with the exception for node splitting
it is generally hard to get this heuristic sound. I've added variants
of the bb-slp-48.c testcase to make sure we still try permuting for
example.
2020-09-15 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (vect_build_slp_tree_2): Also consider
building an operand from scalars when building it did not
fail fatally but avoid messing with the upcall splitting
of groups.
* gcc.dg/vect/bb-slp-48.c: New testcase.
* gcc.dg/vect/bb-slp-7.c: Adjust.
This patch changes the test to use the effective-target machinery disables the
error message "ARMv8-M Security Extensions incompatible with selected FPU" when
-mfloat-abi=soft.
Further changes 'asm' to '__asm__' to avoid failures with '-std=' options.
gcc/ChangeLog:
2020-07-06 Andre Vieira <andre.simoesdiasvieira@arm.com>
* config/arm/arm.c (arm_options_perform_arch_sanity_checks): Do not
check +D32 for CMSE if -mfloat-abi=soft
gcc/testsuite/ChangeLog:
2020-07-06 Andre Vieira <andre.simoesdiasvieira@arm.com>
* gcc.target/arm/pr95646.c: Fix testism.
gcc/ChangeLog:
PR target/96744
* config/i386/x86-tune-costs.h (struct processor_costs):
Increase mask <-> integer cost for non AVX512 target to avoid
spill gpr to mask. Also retune mask <-> integer and
mask_load/store for skylake_cost.
These patterns printed bogus <>s around the {1to16} and similar strings.
2020-09-15 Jakub Jelinek <jakub@redhat.com>
PR target/97028
* config/i386/sse.md (mul<mode>3<mask_name>_bcs,
<avx512>_div<mode>3<mask_name>_bcst): Use <avx512bcst> instead of
<<avx512bcst>>.
* gcc.target/i386/avx512f-pr97028.c: Untested fix.
2020-09-03 Feng Xue <fxue@os.amperecomputing.com>
gcc/
PR tree-optimization/94234
* genmatch.c (dt_simplify::gen_1): Emit check on final simplification
result when "!" is specified on toplevel output expr.
* match.pd ((A * C) +- (B * C) -> (A +- B) * C): Allow folding on expr
with multi-use operands if final result is a simple gimple value.
gcc/testsuite/
PR tree-optimization/94234
* gcc.dg/pr94234-2.c: New test.
The BPF ISA doesn't have a no-operation instruction, but in practice
the Linux kernel verifier performs some optimizations that rely on
these instructions to be encoded in a particular way. As it turns
out, we were using the "wrong" instruction in GCC.
This patch makes GCC to generate the expected instruction for NOP (a
`ja 0') and also adds a test to make sure this is the case.
Tested in bpf-unknown-none targets.
2020-09-14 Jose E. Marchesi <jose.marchesi@oracle.com>
gcc/
* config/bpf/bpf.md ("nop"): Re-define as `ja 0'.
gcc/testsuite/
* gcc.target/bpf/nop-1.c: New test.
The tests assume non-PIC for m32 (which means that they fail on default
PIC targets, like Darwin). There is also a missing space before the
closing '}' on one of the dg-require-effective-target which means that
test fails on machines without avx512f.
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx512f-broadcast-pr87767-1.c: Make the test
run as non-dynamic for m32 Darwin.
* gcc.target/i386/avx512f-broadcast-pr87767-3.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-5.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-7.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-1.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-3.c: Likewise.
* gcc.target/i386/avx512vl-broadcast-pr87767-5.c: Likewise.
* gcc.target/i386/avx512f-broadcast-pr87767-6.c: Adjust dg-requires
clause.
Now we consistently mark local externs with DECL_LOCAL_DECL_P, we can
teach the template machinery not to give them a TEMPLATE_DECL head,
and the instantiation machinery treat them as the local specialiations
they are. (openmp UDRs also fall into this category, and are dealt
with similarly.)
gcc/cp/
* pt.c (push_template_decl_real): Don't attach a template head to
local externs.
(tsubst_function_decl): Add support for headless local extern
decls.
(tsubst_decl): Add support for headless local extern decls.
On attempting to run the full test suite with -fanalyzer via
make check RUNTESTFLAGS="-v -v --target_board=unix/-fanalyzer"
I saw it get stuck on:
gcc.c-torture/compile/20001226-1.c
It turns out this was on a debug build, rather than a release build;
but a release build with -fanalyzer took:
real 1m33.689s
user 1m30.239s
sys 0m2.727s
as compared to:
real 0m2.361s
user 0m2.107s
sys 0m0.214s
without -fanalyzer.
This torture test performs 64 * 64 uniqely-coded comparisons between
elements of a pair of arrays until it finds an element that's different,
leading to an accumulation of 4096 constraints along the path where
no difference is found.
"perf" shows most of the time is spent in canonicalizing and copying
constraint_manager instances, presumably as it copies and merges states
with increasingly more complex sets of constraints as it analyzes
further along the "no differences yet" path.
This patch crudely works around this by adding a
-param=analyzer-max-constraints=
limit, defaulting to 20, above which constraints will be silently
dropped. With this -fanalyzer takes:
real 0m6.935s
user 0m6.413s
sys 0m0.396s
on the above case.
gcc/analyzer/ChangeLog:
* analyzer.opt (-param=analyzer-max-constraints=): New param.
* constraint-manager.cc
(constraint_manager::add_constraint_internal): Silently reject
attempts to add constraints when the above limit is reached.
PR analyzer/96653 reports a CPU-time and memory explosion in -fanalyzer
seen in Linux 5.9-rc1:drivers/media/v4l2-core/v4l2-ctrls.c on a switch
statement with many cases.
The issue is some old code in constraint_manager::get_or_add_equiv_class
for ensuring that comparisons between equivalence classes containing
constants work correctly. The old code added constraints for every
pair of ECs containing constants, leading to O(N^2) constraints (for
N constants). Given that get_or_add_equiv_class also involves O(N)
comparisons, this led to at least O(N^3) CPU time, and O(N^2) memory
consumption when handling the "default" case, where N is the number of
other cases in the switch statement.
The state rewrite of r11-2694-g808f4dfeb3a95f50f15e71148e5c1067f90a126d
added checking for comparisons between constants, making these explicit
constraints redundant, but failed to remove the code mentioned above.
This patch removes it, fixing the blow-up of constraints in the default
case.
gcc/analyzer/ChangeLog:
PR analyzer/96653
* constraint-manager.cc
(constraint_manager::get_or_add_equiv_class): Don't accumulate
transitive closure of all constraints on constants.
gcc/testsuite/ChangeLog:
PR analyzer/96653
* gcc.dg/analyzer/pr96653.c: New test.
Downstream bug report:
https://bugzilla.redhat.com/show_bug.cgi?id=1878600
describes a false positive from -Wanalyzer-file-leak seen with
gcc 10.2, but which has been fixed in gcc 11.
This patch adds the reproducer as a regression test.
gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/rhbz1878600.c: New test.
gcc/analyzer/ChangeLog:
PR analyzer/97029
* analyzer.cc (is_setjmp_call_p): Require the initial arg to be a
pointer.
* region-model.cc (region_model::deref_rvalue): Assert that the
svalue is of pointer type.
gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/pr97029.c: New test.
When a thunk cannot be emitted in assembly directly, expand_thunk
generates regular GIMPLE code but unconditionally forces a tail
call to the target of the thunk. That's theoretically OK because
the thunk essentially forwards its parameters to the target, but
in practice the RTL expander can spill parameters passed by reference
on the stack in assign_parm_setup_reg.
gcc/ChangeLog:
* cgraphunit.c (cgraph_node::expand_thunk): Make sure to set
cfun->tail_call_marked when forcing a tail call.
* function.c (assign_parm_setup_reg): Always use a register to
load a parameter passed by reference if cfun->tail_call_marked.
gcc/testsuite/ChangeLog:
* gnat.dg/thunk1.adb: New test.
* gnat.dg/thunk1_pkg1.ads: New helper.
* gnat.dg/thunk1_pkg2.ads: Likewise.
* gnat.dg/thunk1_pkg2.adb: Likewise.
The following is mostly a mechanical change to rename the mffgpr/mftgpr
insn types to mtvsr/mfvsr to be more clear. It also removes Power6
references to those insn types since we no longer generate those
instructions.
2020-09-14 Pat Haugen <pthaugen@linux.ibm.com>
gcc/
* config/rs6000/power10.md (power10-mffgpr, power10-mftgpr): Rename to
power10-mtvsr/power10-mfvsr.
* config/rs6000/power6.md (X2F_power6, power6-mftgpr, power6-mffgpr):
Remove.
* config/rs6000/power8.md (power8-mffgpr, power8-mftgpr): Rename to
power8-mtvsr/power8-mfvsr.
* config/rs6000/power9.md (power9-mffgpr, power9-mftgpr): Rename to
power9-mtvsr/power9-mfvsr.
* config/rs6000/rs6000.c (rs6000_adjust_cost): Remove Power6
TYPE_MFFGPR cases.
* config/rs6000/rs6000.md (mffgpr, mftgpr, zero_extendsi<mode>2,
extendsi<mode>2, @signbit<mode>2_dm, lfiwax, lfiwzx, *movsi_internal1,
movsi_from_sf, *movdi_from_sf_zero_ext, *mov<mode>_internal,
movsd_hardfloat, movsf_from_si, *mov<mode>_hardfloat64, p8_mtvsrwz,
p8_mtvsrd_df, p8_mtvsrd_sf, p8_mfvsrd_3_<mode>, *movdi_internal64,
unpack<mode>_dm): Rename mffgpr/mftgpr to mtvsr/mfvsr.
* config/rs6000/vsx.md (vsx_mov<mode>_64bit, vsx_extract_<mode>,
vsx_extract_si, *vsx_extract_<mode>_p8): Likewise.
The following patch on top of the
https://gcc.gnu.org/pipermail/gcc-patches/2020-September/553801.html
patch fixes the gcc.target/arm/lto/pr96939_* test in certain ARM
configurations.
As said in the above mentioned patch, the generic code takes care of
saving/restoring TargetVariables or Target Save options, so this just
arranges for the generic code to save those instead of needing the
arm backend to do it manually.
2020-09-14 Jakub Jelinek <jakub@redhat.com>
* config/arm/arm.opt (x_arm_arch_string, x_arm_cpu_string,
x_arm_tune_string): Remove TargetSave entries.
(march=, mcpu=, mtune=): Add Save keyword.
* config/arm/arm.c (arm_option_save): Remove.
(TARGET_OPTION_SAVE): Don't redefine.
(arm_option_restore): Don't restore x_arm_*_string here.
> Seems a latent issue.
> Neither cl_optimization_{save,restore} nor cl_target_option_{save,restore}
> (nor any of the target hooks they call) saves or restores any opts_set
> values, so I think opts_set can be trusted only during option processing (if
> at all), but not later.
> So, short term a fix would be IMHO just stop using opts_set altogether in
> arm_configure_build_target, it doesn't make much sense to me, it should test
> if those strings are non-NULL instead, or at least do that when it is
> invoked from arm_option_restore (e.g. could be done by calling it with
> opts instead of &global_options_set ).
> Longer term, the question is if cl_optimization_{save,restore} and
> cl_target_option_{save,restore} shouldn't be changed not to only
> save/restore the options, but also save the opts_set flags.
> It could be done e.g. by adding a bool array or set of bool members
> to struct cl_optimization and struct cl_target_option , or even more compact
> by using bitmasks, pack each 64 adjacent option flags into a UHWI element
> of an array.
So, I've tried under debugger how it behaves and seems global_options_set
is really an or of whether an option has been ever seen as explicit, either
on the command line or in any of the option pragmas or optimize/target
attributes seen so far, so it isn't something that can be relied on.
The following patch implements the saving/restoring of the opts_set bits
(though only for the options/variables saved by the generic options-save.c
code, for the target specific stuff that isn't handled by the generic code
the opts_set argument is now passed to the hook and the backends can choose
e.g. to use a TargetSave variable to save the flags either individually or
together in some bitmask (or ignore it if they never need opts_set for the
options).
This patch itself doesn't fix the testcase failing on arm, but a follow up
patch will.
2020-09-14 Jakub Jelinek <jakub@redhat.com>
gcc/
* opt-read.awk: Also initialize extra_target_var_types array.
* opth-gen.awk: Emit explicit_mask arrays to struct cl_optimization
and cl_target_option. Adjust cl_optimization_save,
cl_optimization_restore, cl_target_option_save and
cl_target_option_restore declarations.
* optc-save-gen.awk: Add opts_set argument to cl_optimization_save,
cl_optimization_restore, cl_target_option_save and
cl_target_option_restore functions and save or restore opts_set
next to the opts values into or from explicit_mask arrays.
In cl_target_option_eq and cl_optimization_option_eq compare
explicit_mask arrays, in cl_target_option_hash and cl_optimization_hash
hash them and in cl_target_option_stream_out,
cl_target_option_stream_in, cl_optimization_stream_out and
cl_optimization_stream_in stream them.
* tree.h (build_optimization_node, build_target_option_node): Add
opts_set argument.
* tree.c (build_optimization_node): Add opts_set argument, pass it
to cl_optimization_save.
(build_target_option_node): Add opts_set argument, pass it to
cl_target_option_save.
* function.c (invoke_set_current_function_hook): Adjust
cl_optimization_restore caller.
* ipa-inline-transform.c (inline_call): Adjust cl_optimization_restore
and build_optimization_node callers.
* target.def (TARGET_OPTION_SAVE, TARGET_OPTION_RESTORE): Add opts_set
argument.
* target-globals.c (save_target_globals_default_opts): Adjust
cl_optimization_restore callers.
* toplev.c (process_options): Adjust build_optimization_node and
cl_optimization_restore callers.
(target_reinit): Adjust cl_optimization_restore caller.
* tree-streamer-in.c (lto_input_ts_function_decl_tree_pointers):
Adjust build_optimization_node and cl_optimization_restore callers.
* doc/tm.texi: Updated.
* config/aarch64/aarch64.c (aarch64_override_options): Adjust
build_target_option_node caller.
(aarch64_option_save, aarch64_option_restore): Add opts_set argument.
(aarch64_set_current_function): Adjust cl_target_option_restore
caller.
(aarch64_option_valid_attribute_p): Adjust cl_target_option_save,
cl_target_option_restore, cl_optimization_restore,
build_optimization_node and build_target_option_node callers.
* config/aarch64/aarch64-c.c (aarch64_pragma_target_parse): Adjust
cl_target_option_restore and build_target_option_node callers.
* config/arm/arm.c (arm_option_save, arm_option_restore): Add
opts_set argument.
(arm_option_override): Adjust cl_target_option_save,
build_optimization_node and build_target_option_node callers.
(arm_set_current_function): Adjust cl_target_option_restore caller.
(arm_valid_target_attribute_tree): Adjust build_target_option_node
caller.
(add_attribute): Formatting fix.
(arm_valid_target_attribute_p): Adjust cl_optimization_restore,
cl_target_option_restore, arm_valid_target_attribute_tree and
build_optimization_node callers.
* config/arm/arm-c.c (arm_pragma_target_parse): Adjust
cl_target_option_restore callers.
* config/csky/csky.c (csky_option_override): Adjust
build_target_option_node and cl_target_option_save callers.
* config/gcn/gcn.c (gcn_fixup_accel_lto_options): Adjust
build_optimization_node and cl_optimization_restore callers.
* config/i386/i386-builtins.c (get_builtin_code_for_version):
Adjust cl_target_option_save and cl_target_option_restore
callers.
* config/i386/i386-c.c (ix86_pragma_target_parse): Adjust
build_target_option_node and cl_target_option_restore callers.
* config/i386/i386-options.c (ix86_function_specific_save,
ix86_function_specific_restore): Add opts_set arguments.
(ix86_valid_target_attribute_tree): Adjust build_target_option_node
caller.
(ix86_valid_target_attribute_p): Adjust build_optimization_node,
cl_optimization_restore, cl_target_option_restore,
ix86_valid_target_attribute_tree and build_optimization_node callers.
(ix86_option_override_internal): Adjust build_target_option_node
caller.
(ix86_reset_previous_fndecl, ix86_set_current_function): Adjust
cl_target_option_restore callers.
* config/i386/i386-options.h (ix86_function_specific_save,
ix86_function_specific_restore): Add opts_set argument.
* config/nios2/nios2.c (nios2_option_override): Adjust
build_target_option_node caller.
(nios2_option_save, nios2_option_restore): Add opts_set argument.
(nios2_valid_target_attribute_tree): Adjust build_target_option_node
caller.
(nios2_valid_target_attribute_p): Adjust build_optimization_node,
cl_optimization_restore, cl_target_option_save and
cl_target_option_restore callers.
(nios2_set_current_function, nios2_pragma_target_parse): Adjust
cl_target_option_restore callers.
* config/pru/pru.c (pru_option_override): Adjust
build_target_option_node caller.
(pru_set_current_function): Adjust cl_target_option_restore
callers.
* config/rs6000/rs6000.c (rs6000_debug_reg_global): Adjust
cl_target_option_save caller.
(rs6000_option_override_internal): Adjust build_target_option_node
caller.
(rs6000_valid_attribute_p): Adjust build_optimization_node,
cl_optimization_restore, cl_target_option_save,
cl_target_option_restore and build_target_option_node callers.
(rs6000_pragma_target_parse): Adjust cl_target_option_restore and
build_target_option_node callers.
(rs6000_activate_target_options): Adjust cl_target_option_restore
callers.
(rs6000_function_specific_save, rs6000_function_specific_restore):
Add opts_set argument.
* config/s390/s390.c (s390_function_specific_restore): Likewise.
(s390_option_override_internal): Adjust s390_function_specific_restore
caller.
(s390_option_override, s390_valid_target_attribute_tree): Adjust
build_target_option_node caller.
(s390_valid_target_attribute_p): Adjust build_optimization_node,
cl_optimization_restore and cl_target_option_restore callers.
(s390_activate_target_options): Adjust cl_target_option_restore
caller.
* config/s390/s390-c.c (s390_cpu_cpp_builtins): Adjust
cl_target_option_save caller.
(s390_pragma_target_parse): Adjust build_target_option_node and
cl_target_option_restore callers.
gcc/c-family/
* c-attribs.c (handle_optimize_attribute): Adjust
cl_optimization_save, cl_optimization_restore and
build_optimization_node callers.
* c-pragma.c (handle_pragma_optimize): Adjust
build_optimization_node caller.
(handle_pragma_push_options): Adjust
build_optimization_node and build_target_option_node callers.
(handle_pragma_pop_options, handle_pragma_reset_options):
Adjust cl_optimization_restore callers.
gcc/go/
* go-gcc.cc (Gcc_backend::function): Adjust
cl_optimization_save, cl_optimization_restore and
build_optimization_node callers.
gcc/ada/
* gcc-interface/trans.c (gigi): Adjust build_optimization_node
caller.
As reported here
( https://gcc.gnu.org/pipermail/gcc-patches/2020-September/553070.html ),
when running test-case libgomp.c-c++-common/reduction-16.c for powerpc host
with nvptx accelerator, we run into:
...
unresolved symbol __sync_val_compare_and_swap_16
...
I can reproduce the problem on x86_64 with a trigger patch that:
- initializes ix86_isa_flags2 to TARGET_ISA2_CX16
- enables define_expand "atomic_load<mode>" in gcc/config/i386/sync.md
for TImode
The problem is that omp-expand.c generates atomic builtin calls based on
checks whether those are supported on the host, which forces the target to
support these, even though those checks fail for the accelerator target.
Fix this by:
- adding a __sync_val_compare_and_swap_16 in libgomp for nvptx,
which falls back onto libatomic's __atomic_compare_and_swap_16
- adding -foffload=-latomic in the test-case
Tested libgomp on x86_64-linux with nvptx accelerator.
Tested libgomp with trigger patch on x86_64-linux with nvptx accelerator.
libgomp/ChangeLog:
* config/nvptx/atomic.c: New file. Add
__sync_val_compare_and_swap_16.
* testsuite/libgomp.c-c++-common/reduction-16.c: Add -latomic for
target offload_target_nvptx.
