The initial implementation of the GNAT aspect/pragma Volatile_Full_Access
made it incompatible with Atomic, because it was not decided whether the
read-modify-write sequences generated by Volatile_Full_Access would need
to be implemented atomically when Atomic was also specified, which would
have required a compare-and-swap primitive from the target architecture.
But Ada 2022 introduced Full_Access_Only and retrofitted it into Atomic
in the process, answering the above question by the negative, so the
incompatibility between Volatile_Full_Access and Atomic was lifted in
Ada 2012 as well, unfortunately without adjusting the implementation.
gcc/ada/
* gcc-interface/trans.cc (get_atomic_access): Deal specifically with
nodes that are both Atomic and Volatile_Full_Access in Ada 2012.
This has historically been done only on platforms requiring the strict
alignment of memory references, but this can arguably be considered as
being mandated by the language on all of them.
gcc/ada/
* gcc-interface/trans.cc (addressable_p) <COMPONENT_REF>: Take into
account the alignment of the field on all platforms.
When updating the size after making a packable type in gnat_to_gnu_field,
we fail to clear it again when it is not constant.
gcc/ada/
* gcc-interface/decl.cc (gnat_to_gnu_field): Clear again gnu_size
after updating it if it is not constant.
The procedure Note_Uplevel_Bound was implemented as a custom expression
tree walk. This change replaces this custom tree traversal by a more
idiomatic use of Traverse_Proc.
gcc/ada/
* exp_unst.adb (Check_Static_Type::Note_Uplevel_Bound): Refactor
to use the generic Traverse_Proc.
(Check_Static_Type): Adjust calls to Note_Uplevel_Bound as the
previous second parameter was unused, so removed.
The non-strict overflow checking code does a better job of eliminating
overflow checks if given an expression consisting only of predefined
operators (including relationals), literals, identifiers, and conditional
expressions. If it is both feasible and useful, rewrite a
Length attribute reference as such an expression. "Feasible" means
"index type is same type as attribute reference type, so we can rewrite without
using type conversions". "Useful" means "Overflow_Mode is something other than
Strict, so there is value in making overflow check elimination easier".
gcc/ada/
* exp_attr.adb (Expand_N_Attribute_Reference): If it makes sense
to do so, then rewrite a Length attribute reference as an
equivalent conditional expression.
The initial implementation of the GNAT aspect/pragma Volatile_Full_Access
made it incompatible with Atomic, because it was not decided whether the
read-modify-write sequences generated by Volatile_Full_Access would need
to be implemented atomically when Atomic was also specified, which would
have required a compare-and-swap primitive from the target architecture.
But Ada 2022 introduced Full_Access_Only and retrofitted it into Atomic
in the process, answering the above question by the negative, so the
incompatibility between Volatile_Full_Access and Atomic was lifted in
Ada 2012 as well, but the implementation was not entirely adjusted.
In Ada 2012, it does not make sense to warn for the partial access to an
Atomic object if the object is also declared Volatile_Full_Access, since
the object will be accessed as a whole in this case (like in Ada 2022).
gcc/ada/
* sem_res.adb (Is_Atomic_Ref_With_Address): Rename into...
(Is_Atomic_Non_VFA_Ref_With_Address): ...this and adjust the
implementation to exclude Volatile_Full_Access objects.
(Resolve_Indexed_Component): Adjust to above renaming.
(Resolve_Selected_Component): Likewise.
Implement the new legality rules of AI22-0106 which (as discussed in the AI)
are needed to disallow constructs whose semantics would otherwise be poorly
defined.
gcc/ada/
* sem_aggr.adb (Resolve_Array_Aggregate): Implement the two new
legality rules of AI11-0106. Add code to avoid cascading error
messages.
Do not pass null for the Collection parameter when
Finalize_Storage_Only is in effect. If the collection
is null in that case, we will blow up later when we
deallocate the object.
gcc/ada/
* exp_ch6.adb (Add_Collection_Actual_To_Build_In_Place_Call):
Remove Finalize_Storage_Only from the code that checks whether to
pass null to the Collection parameter. Having done that, we don't
need to check for Is_Library_Level_Entity, because
No_Heap_Finalization requires that. And if we ever change
No_Heap_Finalization to allow nested access types, we will still
want to pass null. Note that the comment "Such a type lacks a
collection." is incorrect in the case of Finalize_Storage_Only;
such types have a collection.
This patch implements constant folding for svmul by calling
gimple_folder::fold_const_binary with tree_code MULT_EXPR.
Tests were added to check the produced assembly for different
predicates, signed and unsigned integers, and the svmul_n_* case.
The patch was bootstrapped and regtested on aarch64-linux-gnu, no regression.
OK for mainline?
Signed-off-by: Jennifer Schmitz <jschmitz@nvidia.com>
gcc/
* config/aarch64/aarch64-sve-builtins-base.cc (svmul_impl::fold):
Try constant folding.
gcc/testsuite/
* gcc.target/aarch64/sve/const_fold_mul_1.c: New test.
This patch implements constant folding for svdiv:
The new function aarch64_const_binop was created, which - in contrast to
int_const_binop - does not treat operations as overflowing. This function is
passed as callback to vector_const_binop from the new gimple_folder
method fold_const_binary, if the predicate is ptrue or predication is _x.
From svdiv_impl::fold, fold_const_binary is called with TRUNC_DIV_EXPR as
tree_code.
In aarch64_const_binop, a case was added for TRUNC_DIV_EXPR to return 0
for division by 0, as defined in the semantics for svdiv.
Tests were added to check the produced assembly for different
predicates, signed and unsigned integers, and the svdiv_n_* case.
The patch was bootstrapped and regtested on aarch64-linux-gnu, no regression.
OK for mainline?
Signed-off-by: Jennifer Schmitz <jschmitz@nvidia.com>
gcc/
* config/aarch64/aarch64-sve-builtins-base.cc (svdiv_impl::fold):
Try constant folding.
* config/aarch64/aarch64-sve-builtins.h: Declare
gimple_folder::fold_const_binary.
* config/aarch64/aarch64-sve-builtins.cc (aarch64_const_binop):
New function to fold binary SVE intrinsics without overflow.
(gimple_folder::fold_const_binary): New helper function for
constant folding of SVE intrinsics.
gcc/testsuite/
* gcc.target/aarch64/sve/const_fold_div_1.c: New test.
This patch sets the stage for constant folding of binary operations for SVE
intrinsics:
In fold-const.cc, the code for folding vector constants was moved from
const_binop to a new function vector_const_binop. This function takes a
function pointer as argument specifying how to fold the vector elements.
The intention is to call vector_const_binop from the backend with an
aarch64-specific callback function.
The code in const_binop for folding operations where the first operand is a
vector constant and the second argument is an integer constant was also moved
into vector_const_binop to to allow folding of binary SVE intrinsics where
the second operand is an integer (_n).
To allow calling poly_int_binop from the backend, the latter was made public.
The patch was bootstrapped and regtested on aarch64-linux-gnu, no regression.
OK for mainline?
Signed-off-by: Jennifer Schmitz <jschmitz@nvidia.com>
gcc/
* fold-const.h: Declare vector_const_binop.
* fold-const.cc (const_binop): Remove cases for vector constants.
(vector_const_binop): New function that folds vector constants
element-wise.
(int_const_binop): Remove call to wide_int_binop.
(poly_int_binop): Add call to wide_int_binop.
The following makes sure we handle a SLP load/store group from
a structure with complex and scalar members. This for example
happens in gcc.target/i386/pr106010-9a.c.
* tree-vect-slp.cc (vect_build_slp_tree_1): Handle mixing
all of handled components besides ARRAY_RANGE_REF, drop
handling of INDIRECT_REF.
Currently vect_get_vector_types_for_stmt only special-cases
IFN_MASK_STORE but there are now very many variants and simply
passing analysis without setting *VECTYPE will ICE duing SLP
discovery (noticed with IFN_SCATTER_STORE). The following
properly uses internal_store_fn_p. I also noticed we're
unnecessarily handing those again to determine the scalar type
but there should always be a data reference for them.
* tree-vect-stmts.cc (vect_get_vector_types_for_stmt):
Handle all internal_store_fn_p the same. Remove special-casing
for the scalar_type of IFN_MASK_STORE.
This patch supports sminmax for partial vectorized V2BF/V4BF.
gcc/ChangeLog:
* config/i386/mmx.md (<code><mode>3): New define_expand for V2BF/V4BFsmaxmin
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx10_2-partial-bf-vector-smaxmin-1.c: New test.
This patch introduces new mode iterators and expands for the i386 architecture to support partial vectorization of bf16 operations using AVX10.2 instructions.
gcc/ChangeLog:
* config/i386/mmx.md (VBF_32_64): New mode iterator for partial vectorized V2BF/V4BF.
(<insn><mode>3): New define_expand for plusminusmultdiv.
(sqrt<mode>2): New define_expand for sqrt.
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx10_2-partial-bf-vector-fast-math-1.c: New test.
* gcc.target/i386/avx10_2-partial-bf-vector-operations-1.c: New test.
This patch would like to support the scalar signed ssadd pattern
for the RISC-V backend. Aka
Form 1:
#define DEF_SAT_S_ADD_FMT_1(T, UT, MIN, MAX) \
T __attribute__((noinline)) \
sat_s_add_##T##_fmt_1 (T x, T y) \
{ \
T sum = (UT)x + (UT)y; \
return (x ^ y) < 0 \
? sum \
: (sum ^ x) >= 0 \
? sum \
: x < 0 ? MIN : MAX; \
}
DEF_SAT_S_ADD_FMT_1(int64_t, uint64_t, INT64_MIN, INT64_MAX)
Before this patch:
10 │ sat_s_add_int64_t_fmt_1:
11 │ mv a5,a0
12 │ add a0,a0,a1
13 │ xor a1,a5,a1
14 │ not a1,a1
15 │ xor a4,a5,a0
16 │ and a1,a1,a4
17 │ blt a1,zero,.L5
18 │ ret
19 │ .L5:
20 │ srai a5,a5,63
21 │ li a0,-1
22 │ srli a0,a0,1
23 │ xor a0,a5,a0
24 │ ret
After this patch:
10 │ sat_s_add_int64_t_fmt_1:
11 │ add a2,a0,a1
12 │ xor a1,a0,a1
13 │ xor a5,a0,a2
14 │ srli a5,a5,63
15 │ srli a1,a1,63
16 │ xori a1,a1,1
17 │ and a5,a5,a1
18 │ srai a4,a0,63
19 │ li a3,-1
20 │ srli a3,a3,1
21 │ xor a3,a3,a4
22 │ neg a4,a5
23 │ and a3,a3,a4
24 │ addi a5,a5,-1
25 │ and a0,a2,a5
26 │ or a0,a0,a3
27 │ ret
The below test suites are passed for this patch:
1. The rv64gcv fully regression test.
gcc/ChangeLog:
* config/riscv/riscv-protos.h (riscv_expand_ssadd): Add new func
decl for expanding ssadd.
* config/riscv/riscv.cc (riscv_gen_sign_max_cst): Add new func
impl to gen the max int rtx.
(riscv_expand_ssadd): Add new func impl to expand the ssadd.
* config/riscv/riscv.md (ssadd<mode>3): Add new pattern for
signed integer .SAT_ADD.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/sat_arith.h: Add test helper macros.
* gcc.target/riscv/sat_arith_data.h: Add test data.
* gcc.target/riscv/sat_s_add-1.c: New test.
* gcc.target/riscv/sat_s_add-2.c: New test.
* gcc.target/riscv/sat_s_add-3.c: New test.
* gcc.target/riscv/sat_s_add-4.c: New test.
* gcc.target/riscv/sat_s_add-run-1.c: New test.
* gcc.target/riscv/sat_s_add-run-2.c: New test.
* gcc.target/riscv/sat_s_add-run-3.c: New test.
* gcc.target/riscv/sat_s_add-run-4.c: New test.
* gcc.target/riscv/scalar_sat_binary_run_xxx.h: New test.
Signed-off-by: Pan Li <pan2.li@intel.com>
We have SHF.fmt and HADD_S/U.fmt with MSA, which can be used for
vector reduc.
For min/max for U8/S8, we can
SHF.B W1, W0, 0xb1 # swap byte inner every half
MIN.B W1, W1, W0
SHF.H W2, W1, 0xb1 # swap half inner every word
MIN.B W2, W2, W1
SHF.W W3, W2, 0xb1 # swap word inner every doubleword
MIN.B W4, W3, W2
SHF.W W4, W4, 0x4e # swap the two doubleword
MIN.B W4, W4, W3
For plus of S8/U8, we can use HADD
HADD.H W0, W0, W0
HADD.W W0, W0, W0
HADD.D W0, W0, W0
SHF.W W1, W0, 0x4e # swap the two doubleword
ADDV.D W1, W1, W0
COPY_S.B T0, W1 # COPY_U.B for U8
We can do similar for S16/U16/S32/U32/S64/U64/FLOAT/DOUBLE.
gcc
* config/mips/mips-msa.md: (MSA_NO_HADD): we have HADD for
S8/U8/S16/U16/S32/U32 only.
(reduc_smin_scal_<mode>): New define pattern.
(reduc_smax_scal_<mode>): Ditto.
(reduc_umin_scal_<mode>): Ditto.
(reduc_umax_scal_<mode>): Ditto.
(reduc_plus_scal_<mode>): Ditto.
(reduc_plus_scal_v4si): Ditto.
(reduc_plus_scal_v8hi): Ditto.
(reduc_plus_scal_v16qi): Ditto.
(reduc_<optab>_scal_<mode>): Ditto.
* config/mips/mips-protos.h: New function mips_expand_msa_reduc.
* config/mips/mips.cc: New function mips_expand_msa_reduc.
* config/mips/mips.md: Define any_bitwise iterator.
gcc/testsuite:
* gcc.target/mips/msa-reduc.c: New tests.
The test FAILs on i686-linux because -mfpmath=sse is used without
-msse2 being enabled.
2024-09-02 Jakub Jelinek <jakub@redhat.com>
* gcc.target/i386/optimize_one.c: Add -msse2 to dg-options.
30_threads/future/members/poll.cc has calibration code that, on
systems with very low clock resolution, may spuriously fail to run.
Even when it does run, low resolution and reasonable
timeouts limit severely the viability of increasing the loop counts so
as to reduce measurement noise, so we end up with very noisy results.
On various vxworks targets, high iteration count (low-noise)
measurements confirmed that some of the operations that we expected to
be up to 100x slower than the fastest ones can run a little slower
than that and, with significant noise, may seem to be even slower,
comparatively.
Bump the factors up to 200x, so that we have plenty of margin over
measured results.
for libstdc++-v3/ChangeLog
* testsuite/30_threads/future/members/poll.cc: Factor out
calibration, and run it unconditionally. Lower its
strictness. Bump wait_until_*'s slowness factor.
When we get to test_pr91486_wait_until(), we're about 10s past the
float_steady_clock epoch. This is enough for the 1s delta for the
timeout to come out slightly lower when the futex-less wait_until
converts the deadline from float_steady_clock to __clock_t. So we may
wake up a little too early, and end up looping one extra time to sleep
for e.g. another 954ns until we hit the deadline.
Each iteration calls float_steady_clock::now(), bumping the call_count
that we VERIFY() at the end of the subtest. Since we expect at most 3
calls, and we're going to have at the very least 3 on futex-less
targets (one in the test proper, one before wait_until_impl to compute
the deadline, and one after wait_until_impl to check whether the
deadline was hit), any such imprecision that causes an extra iteration
will reach 5 and cause the test to fail.
Initializing the epoch in the beginning of the test makes such
spurious fails due to loss of precision far less likely. I don't
suppose allowing for an extra couple of calls would be desirable.
While at that, I'm annotating unused status variables as such.
for libstdc++-v3/ChangeLog
PR libstdc++/91486
* testsuite/30_threads/async/async.cc
(test_pr91486_wait_for): Mark status as unused.
(test_pr91486_wait_until): Likewise. Initialize epoch later.
The D testsuite shows it was a mistake to assume that
dg-additional-sources are never to be used for compilation tests.
Even if an output file is specified for compilation, extra module
files can be named and used in the compilation without being flagged
as errors.
Introduce a 'linkonly' flag for dg-additional-sources, and use it in
pr95401.cc and other vector tests that default to run, so that its
additional sources get discarded when vector tests downgrade to
compile-only. This reverts previous workarounds for this very
circumstance, that relied on being able to run vector tests anyway,
even after failing to detect runtime or hardware vector support.
for gcc/ChangeLog
PR d/115295
* doc/sourcebuild.texi (dg-additional-sources): Add linkonly.
for gcc/testsuite/ChangeLog
PR d/115295
* g++.dg/vect/pr95401.cc: Add linkonly to dg-additional-sources.
* g++.dg/vect/pr68762-1.cc: Likewise.
* g++.dg/vect/simd-clone-3.cc: Likewise.
* g++.dg/vect/simd-clone-5.cc: Likewise.
* gcc.dg/vect/vect-simd-clone-10.c: Likewise. Drop dg-do run.
* gcc.dg/vect/vect-simd-clone-12.c: Likewise. Likewise.
* lib/gcc-defs.exp (additional_sources_omit_on_compile): New.
(dg-additional-sources): Add to it on linkonly.
(dg-additional-files-options): Omit select sources on compile.
Now that GCN3 support is gone, TARGET_GCN5_PLUS always evaluates to true, so
we can make that code unconditional, and remove all the "else" cases.
The ISA features TARGET_GLOBAL_ADDRSPACE, TARGET_FLAT_OFFSETS,
TARGET_EXPLICIT_CARRY, and TARGET_MULTIPLY_IMMEDIATE, are similarly also
redundant and can be made unconditional.
The naming of the "gcc_version" attribute has been confusing since the "rdna"
attribute was added and this makes it worse, so it has been renamed to "cdna".
The add-with-carry assembler mnemonics no longer have two forms, so '%^' can be
removed.
gcc/ChangeLog:
* config/gcn/gcn-opts.h (TARGET_GCN5_PLUS): Delete.
(TARGET_GLOBAL_ADDRSPACE): Delete.
(TARGET_FLAT_OFFSETS): Delete.
(TARGET_EXPLICIT_CARRY): Delete.
(TARGET_MULTIPLY_IMMEDIATE): Delete.
* config/gcn/gcn-valu.md (*mov<mode>): Rename "gcn_version" to "cdna".
(*mov<mode>_4reg): Likewise.
(@mov<mode>_sgprbase): Likwise.
(gather<mode>_insn_1offset<exec>): Likewise.
(gather<mode>_insn_1offset_ds<exec>): Likewise.
(gather<mode>_insn_2offsets<exec>): Likewise.
(scatter<mode>_insn_1offset<exec_scatter>): Likewise.
(scatter<mode>_insn_1offset_ds<exec_scatter>): Likewise.
(scatter<mode>_insn_2offsets<exec_scatter>): Likewise.
(gather<mode>_insn_1offset<exec>): Remove TARGET_FLAT_OFFSETS
conditionals.
(scatter<mode>_insn_1offset<exec_scatter>): Likewise.
(scatter<mode>_insn_1offset<exec_scatter>): Likewise.
(add<mode>3<exec_clobber>): Use "_co" instead of "%^".
(add<mode>3_dup<exec_clobber>): Likewise.
(add<mode>3_vcc<exec_vcc>): Likewise.
(add<mode>3_vcc_dup<exec_vcc>): Likewise.
(addc<mode>3<exec_vcc>): Likewise.
(sub<mode>3<exec_clobber>): Likewise.
(sub<mode>3_vcc<exec_vcc>): Likewise.
(subc<mode>3<exec_vcc>): Likewise.
(*plus_carry_dpp_shr_<mode>): Likewise.
(*plus_carry_in_dpp_shr_<mode>): Likewise.
* config/gcn/gcn.cc (gcn_flat_address_p): Remove TARGET_FLAT_OFFSETS
conditionals.
(gcn_addr_space_legitimate_address_p): Likewise.
(gcn_addr_space_legitimize_address): Likewise.
(gcn_expand_scalar_to_vector_address): Likewise.
(print_operand_address): Likewise, and TARGET_GLOBAL_ADDRSPACE also.
(print_operand): Remove "%^" operand code.
Remove TARGET_GLOBAL_ADDRSPACE assertion.
* config/gcn/gcn.h (STACK_ADDR_SPACE): Remove GCN5 conditional.
* config/gcn/gcn.md (gcn_version): Rename attribute ...
(cdna): ... to this, and remove the gcn3 and gcn5 values.
(enabled): Replace old "gcn_version" logic with new "cdna" logic.
(*mov<mode>_insn): Rename "gcn_version" to "cdna".
(*movti_insn): Likewise.
(addsi3): Use "_co" instead of "%^".
(addsi3_scalar_carry): Likewise.
(addsi3_scalar_carry_cst): Likewise.
(addcsi3_scalar): Likewise.
(addcsi3_scalar_zero): Likewise.
(addptrdi3): Likewise.
(subsi3): Likewise.
(<su>mulsi3_highpart): Remove TARGET_MULTIPLY_IMMEDIATE conditions.
(<su>mulsi3_highpart_reg): Remove "gcn_version" attribute.
(muldi3): Likewise.
(atomic_fetch_<bare_mnemonic><mode>): Likewise.
(atomic_<bare_mnemonic><mode>): Likewise.
(sync_compare_and_swap<mode>_insn): Likewise.
(atomic_load<mode>): Likewise.
(atomic_store<mode>): Likewise.
(atomic_exchange<mode>): Likewise.
(<su>mulsi3_highpart_imm): Remove both TARGET_MULTIPLY_IMMEDIATE and
"gcn_version".
(<su>mulsidi3): Likewise.
(<su>mulsidi3_imm): Likewise.
The only GCN3 ISA device was remove (Fiji, gfx803) so all the GCN3-specific
code and features can be removed from the back-end.
gcc/ChangeLog:
* config/gcn/gcn-opts.h (enum gcn_isa): Delete ISA_GCN3.
(TARGET_GCN3): Delete.
(TARGET_GCN3_PLUS): Delete.
(TARGET_M0_LDS_LIMIT): Delete.
* config/gcn/gcn-valu.md
(gather<mode>_insn_1offset<exec>): Remove TARGET_GCN3 from conditions.
(*<reduc_op>_dpp_shr_<mode>): Likewise.
* config/gcn/gcn.cc (enum gcn_isa): Change default to ISA_GCN5.
(gcn_expand_prologue): Remove TARGET_M0_LDS_LIMIT feature.
(gcn_expand_reduc_scalar): Remove TARGET_GCN3 conditions.
* config/gcn/gcn.h (TARGET_CPU_CPP_BUILTINS): Remove TARGET_GCN3.
The gfx803 "Fiji" device was deprecated in GCC 14, removed from LLVM 18, and
hasn't worked properly with the drivers since about ROCm 4.
This patch removes the device from GCC options and documentation, and removes
the direct mentions from the internals.
The TARGET_GCN3 support in the back-end is now unused and can be removed (in a
follow-up patch).
gcc/ChangeLog:
* config.gcc (amdgcn-*-*): Remove "fiji" from with_arch checks.
* config/gcn/gcn-hsa.h (ABI_VERSION_SPEC): Remove fiji alternative.
(NO_XNACK): Likewise.
(NO_SRAM_ECC): Likewise.
(ASM_SPEC): Remove "%{}" around ABI_VERSION_SPEC.
* config/gcn/gcn-opts.h (enum processor_type): Remove PROCESSOR_FIJI.
(TARGET_FIJI): Delete.
* config/gcn/gcn.cc (gcn_option_override): Remove Fiji.
(gcn_omp_device_kind_arch_isa): Likewise.
(output_file_start): Likewise.
* config/gcn/gcn.h (TARGET_CPU_CPP_BUILTINS): Likewise.
* config/gcn/gcn.opt (gpu_type): Likewise.
(march, mtune): Change default to PROCESSOR_VEGA10.
* config/gcn/mkoffload.cc (EF_AMDGPU_MACH_AMDGCN_GFX803): Delete.
(copy_early_debug_info): Remove elf_flags_actual.
Use ELFABIVERSION_AMDGPU_HSA_V4 unconditionally.
(get_arch): Remove Fiji.
(main): Remove gfx803.
* config/gcn/t-omp-device
(omp-device-properties-gcn): Remove fiji and gfx803.
* doc/install.texi (amdgcn*-*-*): Remove fiji and special instructions.
* doc/invoke.texi: Remove fiji.
libgomp/ChangeLog:
* libgomp.texi: Remove fiji and gfx803.
* testsuite/libgomp.c/declare-variant-4.h: Remove fiji and gfx803.
* testsuite/libgomp.c/declare-variant-4-fiji.c: Removed.
* testsuite/libgomp.c/declare-variant-4-gfx803.c: Removed.
Since r15-3254-g3f51f0dc88ec21c1ec79df694200f10ef85915f4
added scan-ltrans-rtl* variants to scanltranstree.exp, it no longer
makes sense to have "tree" in the name. This renames the file
accordingly and updates users.
libatomic/ChangeLog:
* testsuite/lib/libatomic.exp: Load scanltrans.exp instead of
scanltranstree.exp.
libgomp/ChangeLog:
* testsuite/lib/libgomp.exp: Load scanltrans.exp instead of
scanltranstree.exp.
libitm/ChangeLog:
* testsuite/lib/libitm.exp: Load scanltrans.exp instead of
scanltranstree.exp.
libphobos/ChangeLog:
* testsuite/lib/libphobos-dg.exp: Load scanltrans.exp instead of
scanltranstree.exp.
libvtv/ChangeLog:
* testsuite/lib/libvtv.exp: Load scanltrans.exp instead of
scanltranstree.exp.
gcc/testsuite/ChangeLog:
* gcc.dg-selftests/dg-final.exp: Load scanltrans.exp instead of
scanltranstree.exp.
* lib/gcc-dg.exp: Likewise.
* lib/scanltranstree.exp: Rename to ...
* lib/scanltrans.exp: ... this.
ASM_INPUT_P is so named because it causes the eventual rtl insn
pattern to be a top-level ASM_INPUT rather than an ASM_OPERANDS.
However, this name has caused confusion, partly due to earlier
documentation. The name also sounds related to ASM_INPUTS but
is for a different piece of state.
This patch renames it to ASM_BASIC_P, with the inverse meaning
an extended asm. ("Basic asm" is the term used in extend.texi.)
gcc/
* doc/generic.texi (ASM_BASIC_P): Document.
* tree.h (ASM_INPUT_P): Rename to...
(ASM_BASIC_P): ...this.
(ASM_VOLATILE_P, ASM_INLINE_P): Reindent.
* gimplify.cc (gimplify_asm_expr): Update after above renaming.
* tree-core.h (tree_base): Likewise.
gcc/c/
* c-typeck.cc (build_asm_expr): Rename ASM_INPUT_P to ASM_BASIC_P.
gcc/cp/
* pt.cc (tsubst_stmt): Rename ASM_INPUT_P to ASM_BASIC_P.
* parser.cc (cp_parser_asm_definition): Likewise.
gcc/d/
* toir.cc (IRVisitor): Rename ASM_INPUT_P to ASM_BASIC_P.
gcc/jit/
* jit-playback.cc (playback::block::add_extended_asm): Rename
ASM_INPUT_P to ASM_BASIC_P.
gcc/m2/
* gm2-gcc/m2block.cc (flush_pending_note): Rename ASM_INPUT_P
to ASM_BASIC_P.
* gm2-gcc/m2statement.cc (m2statement_BuildAsm): Likewise.
The problem is that the size clause changes the floating-point format used
for the type, but it must not when this format is the widest format that is
supported in hardware on the target. Instead a padding type must be built
and the associated warning given.
gcc/ada/
* gcc-interface/decl.cc (gnat_to_gnu_entity): Cap the Esize of a
floating-point type to the size of the widest format supported in
hardware if it is explicity defined.
Before this patch, the gnat command sent to standard error pieces of
information that are a better match for standard output. This patch
makes this information go to standard output.
gcc/ada/
* gnatcmd.adb (GNATCmd): Fix standard output stream.
Before this patch, the documentation of -gnaty0 used 0-based indexing
for column numbers while 1-based indexing is used everywhere else. This
patch makes this documentation use 1-based indexing, and also adds a
missing parenthesis.
gcc/ada/
* doc/gnat_ugn/building_executable_programs_with_gnat.rst: Fix
minor issues.
* gnat_ugn.texi: Regenerate.
...plus minor improvements to existing documentation.
gcc/ada/
* doc/gnat_rm/gnat_language_extensions.rst: I assume "extended set
of extensions" was a typo for "experimental set of extensions",
because "extended extensions" is repetitive and redundant. "in
addition" clarifies that the one subsumes the other. Add a
reminder at the start of each subsection about what switch/pragma
enables what extensions. Add new section about "Inference of
Dependent Types in Generic Instantiations".
* gnat_rm.texi: Regenerate.
Size of pthread data types now need to be defined for FreeBSD ports.
Traceback support for AArch64 FreeBSD is now defined.
gcc/ada/
* s-oscons-tmplt.c: Define sizes of pthread data types on FreeBSD.
* tracebak.c: Use GCC unwinder and adjust PC appropriately on
aarch64-freebsd.
When changing the scope for entities found in the entry body that is
mutated into a procedure, the compiler needs to look deeper than only
the top level entities as expansion may produce object declarations
which scopes are also the entry. For example, the tree after expansion
may look like:
procedure This_Is_An_Entry_Proc is
...
O1 : Typ := do
TMP1 : OTyp := ...;
...
in TMP1;
O1's scope needs to be reset to This_Is_An_Entry_Proc, but so does
TMP1's scope.
This change also fix a small oversight where
N_Implicit_Label_Declaration scope must be reset and its content
skipped.
gcc/ada/
* exp_ch9.adb (Reset_Scopes_To): Adjust comment.
(Reset_Scopes_To.Reset_Scope): Adjust the scope reset for object
declaration. In particular, visit the children nodes if any. Also
extend the handling of other declarations to
N_Implicit_Label_Declaration.
Replace repeated calls to Sloc with uses of local constant Loc.
Code cleanup; behavior is unaffected.
gcc/ada/
* exp_ch3.adb (Expand_N_Object_Declaration): Replace calls to Sloc
with uses of Loc; turn variable Prag into constant.
Routine Insert_Valid_Check only applies checks when Expr_Known_Valid
query returns False; there is no need to call this query before
inserting checks.
Code cleanup; behavior is unaffected.
gcc/ada/
* exp_imgv.adb (Expand_User_Defined_Enumeration_Image)
(Expand_Image_Attribute): Remove redundant guards.
The initial CLZ gimple-range-op.cc implementation handled just the
case where second argument to .CLZ is equal to prec, but in
r15-1014 I've added also handling of the -1 case. As the following
testcase shows, incorrectly though for the case where the first argument
has [0,0] range. If the second argument is prec, then the result should
be [prec,prec] and that was handled correctly, but when the second argument
is -1, the result should be [-1,-1] but instead it was incorrectly computed
as [prec-1,prec-1] (when second argument is prec, mini is 0 and maxi is
prec, while when second argument is -1, mini is -1 and maxi is prec-1).
Fixed thusly (the actual handling is then similar to the CTZ [0,0] case).
2024-09-02 Jakub Jelinek <jakub@redhat.com>
PR middle-end/116486
* gimple-range-op.cc (cfn_clz::fold_range): If lh is [0,0]
and mini is -1, return [-1,-1] range rather than [prec-1,prec-1].
* gcc.dg/bitint-109.c: New test.
The following is a prototype for how to represent load/store-lanes
within SLP. I've for now settled with having a single load node
with multiple permute nodes acting as selection, one for each loaded lane
and a single store node fed from all stored lanes. For
for (int i = 0; i < 1024; ++i)
{
a[2*i] = b[2*i] + 7;
a[2*i+1] = b[2*i+1] * 3;
}
you have the following SLP graph where I explain how things are set
up and code-generated:
t.c:23:21: note: SLP graph after lowering permutations:
t.c:23:21: note: node 0x50dc8b0 (max_nunits=1, refcnt=1) vector(4) int
t.c:23:21: note: op template: *_6 = _7;
t.c:23:21: note: stmt 0 *_6 = _7;
t.c:23:21: note: stmt 1 *_12 = _13;
t.c:23:21: note: children 0x50dc488 0x50dc6e8
This is the store node, it's marked with ldst_lanes = true during
SLP discovery. This node code-generates
vect_array.65[0] = vect__7.61_29;
vect_array.65[1] = vect__13.62_28;
MEM <int[8]> [(int *)vectp_a.63_27] = .STORE_LANES (vect_array.65);
...
t.c:23:21: note: node 0x50dc520 (max_nunits=4, refcnt=2) vector(4) int
t.c:23:21: note: op: VEC_PERM_EXPR
t.c:23:21: note: stmt 0 _5 = *_4;
t.c:23:21: note: lane permutation { 0[0] }
t.c:23:21: note: children 0x50dc948
t.c:23:21: note: node 0x50dc780 (max_nunits=4, refcnt=2) vector(4) int
t.c:23:21: note: op: VEC_PERM_EXPR
t.c:23:21: note: stmt 0 _11 = *_10;
t.c:23:21: note: lane permutation { 0[1] }
t.c:23:21: note: children 0x50dc948
These are the selection nodes, marked with ldst_lanes = true.
They code generate nothing.
t.c:23:21: note: node 0x50dc948 (max_nunits=4, refcnt=3) vector(4) int
t.c:23:21: note: op template: _5 = *_4;
t.c:23:21: note: stmt 0 _5 = *_4;
t.c:23:21: note: stmt 1 _11 = *_10;
t.c:23:21: note: load permutation { 0 1 }
This is the load node, marked with ldst_lanes = true (the load
permutation is only accurate when taking into account the lane permute
in the selection nodes). It code generates
vect_array.58 = .LOAD_LANES (MEM <int[8]> [(int *)vectp_b.56_33]);
vect__5.59_31 = vect_array.58[0];
vect__5.60_30 = vect_array.58[1];
This scheme allows to leave code generation in vectorizable_load/store
mostly as-is.
While this should support both load-lanes and (masked) store-lanes
the decision to do either is done during SLP discovery time and
cannot be reversed without altering the SLP tree - as-is the SLP
tree is not usable for non-store-lanes on the store side, the
load side is OK representation-wise but will very likely fail
permute handling as the lowering to deal with the two input vector
restriction isn't done - but of course since the permute node is
marked as to be ignored that doesn't work out. So I've put
restrictions in place that fail vectorization if a load/store-lane
SLP tree is later classified differently by get_load_store_type.
I'll note that for example gcc.target/aarch64/sve/mask_struct_store_3.c
will not get SLP store-lanes used because the full store SLPs just
fine though we then fail to handle the "splat" load-permutation
t2.c:5:21: note: node 0x4db2630 (max_nunits=4, refcnt=2) vector([4,4]) int
t2.c:5:21: note: op template: _6 = *_5;
t2.c:5:21: note: stmt 0 _6 = *_5;
t2.c:5:21: note: stmt 1 _6 = *_5;
t2.c:5:21: note: stmt 2 _6 = *_5;
t2.c:5:21: note: stmt 3 _6 = *_5;
t2.c:5:21: note: load permutation { 0 0 0 0 }
the load permute lowering code currently doesn't consider it worth
lowering single loads from a group (or in this case not grouped loads).
The expectation is the target can handle this by two interleaves with
itself.
So what we see here is that while the explicit SLP representation is
helpful in some cases, in cases like this it would require changing
it when we make decisions how to vectorize. My idea is that this
all will change a lot when we re-do SLP discovery (for loops) and
when we get rid of non-SLP as I think vectorizable_* should be
allowed to alter the SLP graph during analysis.
The patch also removes the code cancelling SLP if we can use
load/store-lanes from the main loop vector analysis code and
re-implements it as re-discovering the SLP instance with
forced single-lane splits so SLP load/store-lanes scheme can be
used.
This is now done after SLP discovery and SLP pattern recog are
complete to not disturb the latter but per SLP instance instead
of being a global decision on the whole loop.
This is a behavioral change that for example shows in
gcc.dg/vect/slp-perm-6.c on ARM where we formerly used SLP permutes
but now a mix of SLP without permutes and load/store lanes. The
previous flaky heuristic is now flaky in a different way.
Testing on RISC-V and aarch64 reveal several testcases that require
adjustment as to now expect SLP even when load/store lanes are being
used. If in doubt I've adjusted them to the final expectation which
will lead to one or two new FAILs where we still do the SLP cancelling.
I have a followup that implements that while remaining in SLP that's
in final testing.
Note that gcc.dg/vect/slp-42.c and gcc.dg/vect/pr68445.c will FAIL
on aarch64 with SVE because for some odd reason vect_stridedN
is true for any N for check_effective_target_vect_fully_masked
targets but SVE cannot do ld8 while risc-v can.
I have not bothered to adjust target tests that now fail assembly-scan.
* tree-vectorizer.h (_slp_tree::ldst_lanes): New flag to mark
load, store and permute nodes.
* tree-vect-slp.cc (_slp_tree::_slp_tree): Initialize ldst_lanes.
(vect_build_slp_instance): For stores iff the target prefers
store-lanes discover single-lane sub-groups, do not perform
interleaving lowering but mark the node with ldst_lanes.
Also allow i == 0 - fatal failure - for splitting up a store group
when we're not doing single-lane discovery already.
(vect_lower_load_permutations): When the target supports
load lanes and the loads all fit the pattern split out
a single level of permutes only and mark the load and
permute nodes with ldst_lanes.
(vectorizable_slp_permutation_1): Handle the load-lane permute
forwarding of vector defs.
(vect_analyze_slp): After SLP pattern recog is finished see if
there are any SLP instances that would benefit from using
load/store-lanes and re-discover those with forced single lanes.
* tree-vect-stmts.cc (get_group_load_store_type): Support
load/store-lanes for SLP.
(vectorizable_store): Support SLP code generation for store-lanes.
(vectorizable_load): Support SLP code generation for load-lanes.
* tree-vect-loop.cc (vect_analyze_loop_2): Do not cancel SLP
when store-lanes can be used.
* gcc.dg/vect/slp-55.c: New testcase.
* gcc.dg/vect/slp-56.c: Likewise.
* gcc.dg/vect/slp-11c.c: Adjust.
* gcc.dg/vect/slp-53.c: Likewise.
* gcc.dg/vect/slp-cond-1.c: Likewise.
* gcc.dg/vect/vect-complex-5.c: Likewise.
* gcc.dg/vect/slp-1.c: Likewise.
* gcc.dg/vect/slp-54.c: Remove riscv XFAIL.
* gcc.dg/vect/slp-perm-5.c: Adjust.
* gcc.dg/vect/slp-perm-7.c: Likewise.
* gcc.dg/vect/slp-perm-8.c: Likewise.
* gcc.dg/vect/slp-multitypes-11.c: Likewise.
* gcc.dg/vect/slp-multitypes-11-big-array.c: Likewise.
* gcc.dg/vect/slp-perm-9.c: Remove expected SLP fail due to
three-vector permute.
* gcc.dg/vect/slp-perm-6.c: Remove XFAIL.
* gcc.dg/vect/slp-perm-1.c: Adjust.
* gcc.dg/vect/slp-perm-2.c: Likewise.
* gcc.dg/vect/slp-perm-3.c: Likewise.
* gcc.dg/vect/slp-perm-4.c: Likewise.
* gcc.dg/vect/pr68445.c: Likewise.
* gcc.dg/vect/slp-11b.c: Likewise.
* gcc.dg/vect/slp-2.c: Likewise.
* gcc.dg/vect/slp-23.c: Likewise.
* gcc.dg/vect/slp-33.c: Likewise.
* gcc.dg/vect/slp-42.c: Likewise.
* gcc.dg/vect/slp-46.c: Likewise.
* gcc.dg/vect/slp-perm-10.c: Likewise.
The following emulates classical interleaving for SLP load permutes
that we are unlikely handling natively. This is to handle cases
where interleaving (or load/store-lanes) is the optimal choice for
vectorizing even when we are doing that within SLP. An example
would be
void foo (int * __restrict a, int * b)
{
for (int i = 0; i < 16; ++i)
{
a[4*i + 0] = b[4*i + 0] * 3;
a[4*i + 1] = b[4*i + 1] + 3;
a[4*i + 2] = (b[4*i + 2] * 3 + 3);
a[4*i + 3] = b[4*i + 3] * 3;
}
}
where currently the SLP store is merging four single-lane SLP
sub-graphs but none of the loads in it can be code-generated
with V4SImode vectors and a VF of four as the permutes would need
three vectors.
The patch introduces a lowering phase after SLP discovery but
before SLP pattern recognition or permute optimization that
analyzes all loads from the same dataref group and creates an
interleaving scheme starting from an unpermuted load.
What can be handled is power-of-two group size and a group size of
three. The possibility for doing the interleaving with a load-lanes
like instruction is done as followup.
For a group-size of three this is done by using
the non-interleaving fallback code which then creates at VF == 4 from
{ { a0, b0, c0 }, { a1, b1, c1 }, { a2, b2, c2 }, { a3, b3, c3 } }
the intermediate vectors { c0, c0, c1, c1 } and { c2, c2, c3, c3 }
to produce { c0, c1, c2, c3 }. This turns out to be more effective
than the scheme implemented for non-SLP for SSE and only slightly
worse for AVX512 and a bit more worse for AVX2. It seems to me that
this would extend to other non-power-of-two group-sizes though (but
the patch does not). Optimal schemes are likely difficult to lay out
in VF agnostic form.
I'll note that while the lowering assumes even/odd extract is
generally available for all vector element sizes (which is probably
a good assumption), it doesn't in any way constrain the other
permutes it generates based on target availability. Again difficult
to do in a VF agnostic way (but at least currently the vector type
is fixed).
I'll also note that the SLP store side merges lanes in a way
producing three-vector permutes for store group-size of three, so
the testcase uses a store group-size of four.
The patch has a fallback for when there are multi-lane groups
and the resulting permutes to not fit interleaving. Code
generation is not optimal when this triggers and might be
worse than doing single-lane group interleaving.
The patch handles gaps by representing them with NULL
entries in SLP_TREE_SCALAR_STMTS for the unpermuted load node.
The SLP discovery changes could be elided if we manually build the
load node instead.
SLP load nodes covering enough lanes to not need intermediate
permutes are retained as having a load-permutation and do not
use the single SLP load node for each dataref group. That's
something we might want to change, making load-permutation
something purely local to SLP discovery (but then SLP discovery
could do part of the lowering).
The patch misses CSEing intermediate generated permutes and
registering them with the bst_map which is possibly required
for SLP pattern detection in some cases - this re-spin of the
patch moves the lowering after SLP pattern detection.
* tree-vect-slp.cc (vect_build_slp_tree_1): Handle NULL stmt.
(vect_build_slp_tree_2): Likewise. Release load permutation
when there's a NULL in SLP_TREE_SCALAR_STMTS and assert there's
no actual permutation in that case.
(vllp_cmp): New function.
(vect_lower_load_permutations): Likewise.
(vect_analyze_slp): Call it.
* gcc.dg/vect/slp-11a.c: Expect SLP.
* gcc.dg/vect/slp-12a.c: Likewise.
* gcc.dg/vect/slp-51.c: New testcase.
* gcc.dg/vect/slp-52.c: New testcase.
Currently, in RV32, even with the D extension enabled, the cost of DFmode
register moves is still set to 'COSTS_N_INSNS (2)'. This results in the
'lower-subreg' pass splitting DFmode register moves into two SImode SUBREG
register moves, leading to the generation of many redundant instructions.
As an example, consider the following test case:
double foo (int t, double a, double b)
{
if (t > 0)
return a;
else
return b;
}
When compiling with -march=rv32imafdc -mabi=ilp32d, the following code is generated:
.cfi_startproc
addi sp,sp,-32
.cfi_def_cfa_offset 32
fsd fa0,8(sp)
fsd fa1,16(sp)
lw a4,8(sp)
lw a5,12(sp)
lw a2,16(sp)
lw a3,20(sp)
bgt a0,zero,.L1
mv a4,a2
mv a5,a3
.L1:
sw a4,24(sp)
sw a5,28(sp)
fld fa0,24(sp)
addi sp,sp,32
.cfi_def_cfa_offset 0
jr ra
.cfi_endproc
After adjust the DFmode register move's cost to 'COSTS_N_INSNS (1)', the
generated code is as follows, with a significant reduction in the number
of instructions.
.cfi_startproc
ble a0,zero,.L5
ret
.L5:
fmv.d fa0,fa1
ret
.cfi_endproc
gcc/
* config/riscv/riscv.cc (riscv_rtx_costs): Optimize the cost of the
DFmode register move for RV32.
gcc/testsuite/
* gcc.target/riscv/rv32-movdf-cost.c: New test.
This is a small bug in the ext-dce code's handling of promoted subregs.
Essentially when we see a promoted subreg we need to make additional bit groups
live as various parts of the RTL path know that an extension of a suitably
promoted subreg can be trivially eliminated.
When I added support for dealing with this quirk I failed to account for the
larger modes properly and it ignored the case when the size of the inner object
was > 32 bits. Opps.
This does _not_ fix the outstanding x86 issue. That's caused by something
completely different and more concerning ;(
Bootstrapped and regression tested on x86. Obviously fixes the testcase on
riscv as well.
Pushing to the trunk.
PR rtl-optimization/116544
gcc/
* ext-dce.cc (ext_dce_process_uses): Fix thinko in promoted subreg
handling.
gcc/testsuite/
* gcc.dg/torture/pr116544.c: New test.
gcc/ChangeLog:
* config/i386/sse.md
(<code><mode>3): New define expand pattern for BF smaxmin.
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx10_2-512-bf-vector-smaxmin-1.c: New test.
* gcc.target/i386/avx10_2-bf-vector-smaxmin-1.c: New test.
