Loop distribution can release SSA names used in nb_iterations, make
sure to release those.
2022-02-24 Richard Biener <rguenther@suse.de>
PR tree-optimization/104676
* tree-loop-distribution.cc (loop_distribution::execute):
Do a full scev_reset.
* gcc.dg/torture/pr104676.c: New testcase.
The following testcase is miscompiled, because -fipa-pure-const discovers
that bar is const, but when sccvn during fre3 sees
# .MEM_140 = VDEF <.MEM_96>
*__pred$__d_43 = _50 (_49);
where _50 value numbers to &bar, it value numbers .MEM_140 to
vuse_ssa_val (gimple_vuse (stmt)). For const/pure calls that return
a SSA_NAME (or don't have lhs) that is fine, those calls don't store
anything, but if the lhs is present and not an SSA_NAME, value numbering
the vdef to anything but itself means that e.g. walk_non_aliased_vuses
won't consider the call, but the call acts as a store to its lhs.
When it is ignored, sccvn will return whatever has been stored to the
lhs earlier.
I've bootstrapped/regtested an earlier version of this patch, which did the
if (!lhs && gimple_call_lhs (stmt))
changed |= set_ssa_val_to (vdef, vdef);
part before else if (vnresult->result_vdef), and that regressed
+FAIL: gcc.dg/pr51879-16.c scan-tree-dump-times pre "foo \\\\(" 1
+FAIL: gcc.dg/pr51879-16.c scan-tree-dump-times pre "foo2 \\\\(" 1
so this updated patch uses result_vdef there as before and only otherwise
(which I think must be the const/pure case) decides based on whether the
lhs is non-SSA_NAME.
2022-02-24 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/104601
* tree-ssa-sccvn.cc (visit_reference_op_call): For calls with
non-SSA_NAME lhs value number vdef to itself instead of e.g. the
vuse value number.
* g++.dg/torture/pr104601.C: New test.
Add openmp test-cases that test the omp declare variant construct:
...
#pragma omp declare variant (f30) match (device={isa("sm_30")})
...
using the available nvptx isas.
Only the one for sm_30 is a dg-do run test-case, the other ones are dg-do
link.
Tested on x86_64 with nvptx accelerator.
libgomp/ChangeLog:
2022-02-24 Tom de Vries <tdevries@suse.de>
* testsuite/libgomp.c/declare-variant-3-sm30.c: New test.
* testsuite/libgomp.c/declare-variant-3-sm35.c: New test.
* testsuite/libgomp.c/declare-variant-3-sm53.c: New test.
* testsuite/libgomp.c/declare-variant-3-sm70.c: New test.
* testsuite/libgomp.c/declare-variant-3-sm75.c: New test.
* testsuite/libgomp.c/declare-variant-3-sm80.c: New test.
* testsuite/libgomp.c/declare-variant-3.h: New header file.
In t-omp-device we list isas that can be used in omp declare variant like so:
...
#pragma omp declare variant (f30) match (device={isa("sm_30")})
...
and in nvptx_omp_device_kind_arch_isa we handle them.
Update both to reflect the current list of isas.
Tested on x86_64-linux with nvptx accelerator.
gcc/ChangeLog:
2022-02-23 Tom de Vries <tdevries@suse.de>
* config/nvptx/nvptx.cc (nvptx_omp_device_kind_arch_isa): Handle
sm_70, sm_75 and sm_80.
* config/nvptx/t-omp-device: Add sm_53, sm_70, sm_75 and sm_80.
Co-Authored-By: Tobias Burnus <tobias@codesourcery.com>
Ptx contains funnel shift operations shf.l.wrap and shf.r.wrap that can be
used to implement 32-bit left or right rotate.
Add define_insns rotlsi3 and rotrsi3.
Tested on nvptx.
gcc/ChangeLog:
2022-02-23 Tom de Vries <tdevries@suse.de>
* config/nvptx/nvptx.md (define_insn "rotlsi3", define_insn
"rotrsi3"): New define_insn.
gcc/testsuite/ChangeLog:
2022-02-23 Tom de Vries <tdevries@suse.de>
* gcc.target/nvptx/rotate-run.c: New test.
* gcc.target/nvptx/rotate.c: New test.
I committed "[nvptx] Add -mptx-comment", but tested it in combination with the
proposed "[final] Handle compiler-generated asm insn" (
https://gcc.gnu.org/pipermail/gcc-patches/2022-February/590721.html ), so
by itself the commit introduced some regressions:
...
FAIL: gcc.dg/20020426-2.c (internal compiler error: Segmentation fault)
FAIL: gcc.dg/analyzer/zlib-3.c (internal compiler error: Segmentation fault)
FAIL: gcc.dg/pr101223.c (internal compiler error: Segmentation fault)
FAIL: gcc.dg/torture/pr80764.c -O2 (internal compiler error: Segmentation fault)
...
There are due to cfun->function_start_locus == 0.
Fix these by using DECL_SOURCE_LOCATION (cfun->decl) instead.
Tested on nvptx.
gcc/ChangeLog:
2022-02-23 Tom de Vries <tdevries@suse.de>
* config/nvptx/nvptx.cc (gen_comment): Use
DECL_SOURCE_LOCATION (cfun->decl) instead of cfun->function_start_locus.
For evex encoding vp{xor,or,and}, suffix is needed.
Or there would be an error for
vpxor %xmm0, %xmm31, %xmm1
Error: unsupported instruction `vpxor'
gcc/ChangeLog:
* config/i386/sse.md (<code>v1ti3): Add suffix and replace
isa attr of alternative 2 from avx to avx512vl.
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx512vl-logicsuffix-1.c: New test.
When testing -fanalyzer on openblas-0.3, I noticed slightly over 2000
false positives from -Wanalyzer-malloc-leak on code like this:
if( LAPACKE_lsame( vect, 'b' ) || LAPACKE_lsame( vect, 'p' ) ) {
pt_t = (lapack_complex_float*)
LAPACKE_malloc( sizeof(lapack_complex_float) *
ldpt_t * MAX(1,n) );
[...snip...]
}
[...snip lots of code...]
if( LAPACKE_lsame( vect, 'b' ) || LAPACKE_lsame( vect, 'q' ) ) {
LAPACKE_free( pt_t );
}
where LAPACKE_lsame is a char-comparison function implemented in a
different TU.
The analyzer naively considers the execution path where:
LAPACKE_lsame( vect, 'b' ) || LAPACKE_lsame( vect, 'p' )
is true at the malloc guard, but then false at the free guard, which
is thus a memory leak.
This patch makes -fanalyer respect __attribute__((const)), so that the
analyzer treats such functions as returning the same value when given
the same inputs.
I've filed https://github.com/xianyi/OpenBLAS/issues/3543 suggesting that
LAPACKE_lsame be annotated with __attribute__((const)); with that, and
with this patch, the false positives seem to be fixed.
gcc/analyzer/ChangeLog:
PR analyzer/104434
* analyzer.h (class const_fn_result_svalue): New decl.
* region-model-impl-calls.cc (call_details::get_manager): New.
* region-model-manager.cc
(region_model_manager::get_or_create_const_fn_result_svalue): New.
(region_model_manager::log_stats): Log
m_const_fn_result_values_map.
* region-model.cc (const_fn_p): New.
(maybe_get_const_fn_result): New.
(region_model::on_call_pre): Handle fndecls with
__attribute__((const)) by calling the above rather than making
a conjured_svalue.
* region-model.h (visitor::visit_const_fn_result_svalue): New.
(region_model_manager::get_or_create_const_fn_result_svalue): New
decl.
(region_model_manager::const_fn_result_values_map_t): New typedef.
(region_model_manager::m_const_fn_result_values_map): New field.
(call_details::get_manager): New decl.
* svalue.cc (svalue::cmp_ptr): Handle SK_CONST_FN_RESULT.
(const_fn_result_svalue::dump_to_pp): New.
(const_fn_result_svalue::dump_input): New.
(const_fn_result_svalue::accept): New.
* svalue.h (enum svalue_kind): Add SK_CONST_FN_RESULT.
(svalue::dyn_cast_const_fn_result_svalue): New.
(class const_fn_result_svalue): New.
(is_a_helper <const const_fn_result_svalue *>::test): New.
(template <> struct default_hash_traits<const_fn_result_svalue::key_t>):
New.
gcc/testsuite/ChangeLog:
PR analyzer/104434
* gcc.dg/analyzer/attr-const-1.c: New test.
* gcc.dg/analyzer/attr-const-2.c: New test.
* gcc.dg/analyzer/attr-const-3.c: New test.
* gcc.dg/analyzer/pr104434-const.c: New test.
* gcc.dg/analyzer/pr104434-nonconst.c: New test.
* gcc.dg/analyzer/pr104434.h: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
A nice side effect of r12-1822 was improving the diagnostic
we emit for the following test.
PR c++/79493
gcc/testsuite/ChangeLog:
* g++.dg/diagnostic/undeclared1.C: New test.
The following patch avoids infinite recursion during generic folding.
The (cmp (bswap @0) INTEGER_CST@1) simplification relies on
(bswap @1) actually being simplified, if it is not simplified, we just
move the bswap from one operand to the other and if @0 is also INTEGER_CST,
we apply the same rule next.
The reason why bswap @1 isn't folded to INTEGER_CST is that the INTEGER_CST
has TREE_OVERFLOW set on it and fold-const-call.cc predicate punts in
such cases:
static inline bool
integer_cst_p (tree t)
{
return TREE_CODE (t) == INTEGER_CST && !TREE_OVERFLOW (t);
}
The patch uses ! modifier to ensure the bswap is simplified and
extends support to GENERIC by means of requiring !EXPR_P which
is not perfect but a conservative approximation.
2022-02-22 Richard Biener <rguenther@suse.de>
PR tree-optimization/104644
* doc/match-and-simplify.texi: Amend ! documentation.
* genmatch.cc (expr::gen_transform): Code-generate ! support
for GENERIC.
(parser::parse_expr): Allow ! for GENERIC.
* match.pd (cmp (bswap @0) INTEGER_CST@1): Use ! modifier on
bswap.
* gcc.dg/pr104644.c: New test.
Co-Authored-by: Jakub Jelinek <jakub@redhat.com>
Currently we fail to parse
int * _3;
as SSA name and instead get a VAR_DECL because of the way the C
frontends declarator specs work. That causes havoc if those
supposed SSA names are used in PHIs or in other places where
VAR_DECLs are not allowed. The following fixes the pointer case
in an ad-hoc way - for more complex type declarators we probably
have to find a way to re-use the C frontend grokdeclarator without
actually creating a VAR_DECL there (or maybe make it create an
SSA name).
Pointers appear too often to be neglected though, thus the following
ad-hoc fix for this. This also adds verification that we do not
end up with SSA names without definitions as can happen when
reducing a GIMPLE testcase. Instead of working through segfaults
one-by-one we emit errors for all of those at once now.
2022-02-23 Richard Biener <rguenther@suse.de>
gcc/c
* gimple-parser.cc (c_parser_parse_gimple_body): Diagnose
SSA names without definition.
(c_parser_gimple_declaration): Handle pointer typed SSA names.
gcc/testsuite/
* gcc.dg/gimplefe-49.c: New testcase.
* gcc.dg/gimplefe-error-13.c: Likewise.
The following fixes an ICE when vectorizing the defs of a CTOR
results in a different vector type than expected. That can happen
with AARCH64 SVE and a fixed vector length as noted in r10-5979
and on x86 with AVX512 mask CTORs and trying to re-vectorize
using SSE as shown in this bug.
The fix is simply to reject the vectorization when it didn't
produce the desired type.
2022-02-23 Richard Biener <rguenther@suse.de>
PR tree-optimization/101636
* tree-vect-slp.cc (vect_print_slp_tree): Dump the
vector type of the node.
(vect_slp_analyze_operations): Make sure the CTOR
is vectorized with an expected type.
(vectorize_slp_instance_root_stmt): Revert r10-5979 fix.
* gcc.target/i386/pr101636.c: New testcase.
* c-c++-common/torture/pr101636.c: Likewise.
The first two testcases show different ways how e.g. the glibc
_FORTIFY_SOURCE wrappers are implemented, and on Winfinite-recursion-3.c
the new -Winfinite-recursion warning emits a false positive warning.
It is a false positive because when a builtin with 2 names is called
through the __builtin_ name (but not all builtins have a name prefixed
exactly like that) from extern inline function with gnu_inline semantics,
it doesn't mean the compiler will ever attempt to use the user inline
wrapper for the call, the __builtin_ just does what the builtin function
is expected to do and either expands into some compiler generated code,
or if the compiler decides to emit a call it will use an actual definition
of the function, but that is not the extern inline gnu_inline function
which is never emitted out of line.
Compared to that, in Winfinite-recursion-5.c the extern inline gnu_inline
wrapper calls the builtin by the same name as the function's name and in
that case it is infinite recursion, we actuall try to inline the recursive
call and also error because the recursion is infinite during inlining;
without always_inline we wouldn't error but it is still infinite recursion,
the user has no control on how many recursive calls we actually inline.
2022-02-22 Jakub Jelinek <jakub@redhat.com>
PR c/104633
* gimple-warn-recursion.cc (pass_warn_recursion::find_function_exit):
Don't warn about calls to corresponding builtin from extern inline
gnu_inline wrappers.
* gcc.dg/Winfinite-recursion-3.c: New test.
* gcc.dg/Winfinite-recursion-4.c: New test.
* gcc.dg/Winfinite-recursion-5.c: New test.
This one line fix/tweak is the back-end specific change for a fix for
PR target/104489, that allows the ISA for GCC's nvptx backend to be bumped
to sm_53. The machine-independent middle-end pieces were posted here:
https://gcc.gnu.org/pipermail/gcc-patches/2022-February/590139.html
2022-02-23 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/104489
* config/nvptx/nvptx.md (*movhf_insn): Add subregs_ok attribute.
I made a last minute renaming of mve_const_bool_vec_to_hi () into
mve_bool_vec_to_const () and forgot to update the call sites in vfp.md
accordingly.
Committed as obvious.
2022-02-23 Christophe Lyon <christophe.lyon@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/vfp.md (thumb2_movhi_vfp, thumb2_movhi_fp16): Fix
typo.
Since the ISA supported by Intel architectures in the documentation
are inconsistent with the actual, modify them all.
gcc/Changelog:
* doc/invoke.texi: Update documents for Intel architectures.
This patch resolves the bootstrap failure on x86_64-pc-linux-gnu.
2022-02-22 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-expand.cc (ix86_expand_cmpxchg_loop): Restore
bootstrap.
Now that I've resolved GCC 'hash_map' issues (a while ago already), we may
further simplify this after commit 049eda8274
"Avoid 'GTY' use for 'gcc/omp-oacc-neuter-broadcast.cc:field_map'": as
'hash_map' Value, directly store 'field_map_t' objects, not pointers to
manually allocated 'field_map_t' objects.
gcc/
* omp-oacc-neuter-broadcast.cc (record_field_map_t): Further
simplify. Adjust all users.
This was a latent problem, and this commit here now resolves a regression that
after recent commit a78b1ab1df
"amdgcn: Tune default OpenMP/OpenACC GPU utilization" we had (only) seen on a
GCN offloading '-march=gfx908' system:
{+WARNING: program timed out.+}
[-PASS:-]{+FAIL:+} libgomp.oacc-fortran/privatized-ref-2.f90 -DACC_DEVICE_TYPE_radeon=1 -DACC_MEM_SHARED=0 -foffload=amdgcn-amdhsa -O0 execution test
Same for other optimization levels.
Make sure that we're not executing non-parallelized code in gang-redundant
mode, by putting these parts into their own 'parallel' constructs, which then
default to 'num_gangs(1)'.
libgomp/
* testsuite/libgomp.oacc-fortran/privatized-ref-2.f90: Fix OpenACC
gang-redundant execution.
In PR88134 it is pointed out that we do not have GTY markup for some
variables we use for atomic. So, let's add that.
2022-02-22 Segher Boessenkool <segher@kernel.crashing.org>
PR target/88134
* config/rs6000/rs6000.cc (atomic_hold_decl, atomic_clear_decl,
atomic_update_decl): Add GTY markup.
VPR_REG should be part of ALL_REGS, this patch fixes this omission.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
* config/arm/arm.h (REG_CLASS_CONTENTS): Add VPR_REG to ALL_REGS.
This patch covers a few non-load/store builtins where we do not use
the <mode> iterator and thus we cannot use <MVE_vpred>.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/arm-builtins.cc (CX_UNARY_UNONE_QUALIFIERS): Use
predicate.
(CX_BINARY_UNONE_QUALIFIERS): Likewise.
(CX_TERNARY_UNONE_QUALIFIERS): Likewise.
(TERNOP_NONE_NONE_NONE_UNONE_QUALIFIERS): Delete.
(QUADOP_NONE_NONE_NONE_NONE_UNONE_QUALIFIERS): Delete.
(QUADOP_UNONE_UNONE_UNONE_UNONE_UNONE_QUALIFIERS): Delete.
* config/arm/arm_mve_builtins.def: Use predicated qualifiers.
* config/arm/mve.md: Use VxBI instead of HI.
This patch covers a few builtins where we do not use the <mode>
iterator and thus we cannot use <MVE_vpred>.
For v2di instructions, we keep the HI mode for predicates.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/arm-builtins.cc (STRSBS_P_QUALIFIERS): Use predicate
qualifier.
(STRSBU_P_QUALIFIERS): Likewise.
(LDRGBS_Z_QUALIFIERS): Likewise.
(LDRGBU_Z_QUALIFIERS): Likewise.
(LDRGBWBXU_Z_QUALIFIERS): Likewise.
(LDRGBWBS_Z_QUALIFIERS): Likewise.
(LDRGBWBU_Z_QUALIFIERS): Likewise.
(STRSBWBS_P_QUALIFIERS): Likewise.
(STRSBWBU_P_QUALIFIERS): Likewise.
* config/arm/mve.md: Use VxBI instead of HI.
This patch covers all builtins that have an HI operand and use the
<mode> iterator, thus we can replace HI whe <MVE_vpred>.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/arm-builtins.cc (TERNOP_UNONE_UNONE_NONE_UNONE_QUALIFIERS): Change to ...
(TERNOP_UNONE_UNONE_NONE_PRED_QUALIFIERS): ... this.
(TERNOP_UNONE_UNONE_IMM_UNONE_QUALIFIERS): Change to ...
(TERNOP_UNONE_UNONE_IMM_PRED_QUALIFIERS): ... this.
(TERNOP_NONE_NONE_IMM_UNONE_QUALIFIERS): Change to ...
(TERNOP_NONE_NONE_IMM_PRED_QUALIFIERS): ... this.
(TERNOP_NONE_NONE_UNONE_UNONE_QUALIFIERS): Change to ...
(TERNOP_NONE_NONE_UNONE_PRED_QUALIFIERS): ... this.
(QUADOP_UNONE_UNONE_NONE_NONE_UNONE_QUALIFIERS): Change to ...
(QUADOP_UNONE_UNONE_NONE_NONE_PRED_QUALIFIERS): ... this.
(QUADOP_NONE_NONE_NONE_NONE_PRED_QUALIFIERS): New.
(QUADOP_NONE_NONE_NONE_IMM_UNONE_QUALIFIERS): Change to ...
(QUADOP_NONE_NONE_NONE_IMM_PRED_QUALIFIERS): ... this.
(QUADOP_UNONE_UNONE_UNONE_UNONE_PRED_QUALIFIERS): New.
(QUADOP_UNONE_UNONE_NONE_IMM_UNONE_QUALIFIERS): Change to ...
(QUADOP_UNONE_UNONE_NONE_IMM_PRED_QUALIFIERS): ... this.
(QUADOP_NONE_NONE_UNONE_IMM_UNONE_QUALIFIERS): Change to ...
(QUADOP_NONE_NONE_UNONE_IMM_PRED_QUALIFIERS): ... this.
(QUADOP_UNONE_UNONE_UNONE_IMM_UNONE_QUALIFIERS): Change to ...
(QUADOP_UNONE_UNONE_UNONE_IMM_PRED_QUALIFIERS): ... this.
(QUADOP_UNONE_UNONE_UNONE_NONE_UNONE_QUALIFIERS): Change to ...
(QUADOP_UNONE_UNONE_UNONE_NONE_PRED_QUALIFIERS): ... this.
(STRS_P_QUALIFIERS): Use predicate qualifier.
(STRU_P_QUALIFIERS): Likewise.
(STRSU_P_QUALIFIERS): Likewise.
(STRSS_P_QUALIFIERS): Likewise.
(LDRGS_Z_QUALIFIERS): Likewise.
(LDRGU_Z_QUALIFIERS): Likewise.
(LDRS_Z_QUALIFIERS): Likewise.
(LDRU_Z_QUALIFIERS): Likewise.
(QUINOP_UNONE_UNONE_UNONE_UNONE_IMM_UNONE_QUALIFIERS): Change to ...
(QUINOP_UNONE_UNONE_UNONE_UNONE_IMM_PRED_QUALIFIERS): ... this.
(BINOP_NONE_NONE_PRED_QUALIFIERS): New.
(BINOP_UNONE_UNONE_PRED_QUALIFIERS): New.
* config/arm/arm_mve_builtins.def: Use new predicated qualifiers.
* config/arm/mve.md: Use MVE_VPRED instead of HI.
This is mostly a mechanical change, only tested by the intrinsics
expansion tests.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/arm-builtins.cc (BINOP_UNONE_NONE_NONE_QUALIFIERS):
Delete.
(TERNOP_UNONE_NONE_NONE_UNONE_QUALIFIERS): Change to ...
(TERNOP_PRED_NONE_NONE_PRED_QUALIFIERS): ... this.
(TERNOP_PRED_UNONE_UNONE_PRED_QUALIFIERS): New.
* config/arm/arm_mve_builtins.def (vcmp*q_n_, vcmp*q_m_f): Use new
predicated qualifiers.
* config/arm/mve.md (mve_vcmp<mve_cmp_op>q_n_<mode>)
(mve_vcmp*q_m_f<mode>): Use MVE_VPRED instead of HI.
The problem in this PR is that we call VPSEL with a mask of vector
type instead of HImode. This happens because operand 3 in vcond_mask
is the pre-computed vector comparison and has vector type.
This patch fixes it by implementing TARGET_VECTORIZE_GET_MASK_MODE,
returning the appropriate VxBI mode when targeting MVE. In turn, this
implies implementing vec_cmp<mode><MVE_vpred>,
vec_cmpu<mode><MVE_vpred> and vcond_mask_<mode><MVE_vpred>, and we can
move vec_cmp<mode><v_cmp_result>, vec_cmpu<mode><mode> and
vcond_mask_<mode><v_cmp_result> back to neon.md since they are not
used by MVE anymore. The new *<MVE_vpred> patterns listed above are
implemented in mve.md since they are only valid for MVE. However this
may make maintenance/comparison more painful than having all of them
in vec-common.md.
In the process, we can get rid of the recently added vcond_mve
parameter of arm_expand_vector_compare.
Compared to neon.md's vcond_mask_<mode><v_cmp_result> before my "arm:
Auto-vectorization for MVE: vcmp" patch (r12-834), it keeps the VDQWH
iterator added in r12-835 (to have V4HF/V8HF support), as well as the
(!<Is_float_mode> || flag_unsafe_math_optimizations) condition which
was not present before r12-834 although SF modes were enabled by VDQW
(I think this was a bug).
Using TARGET_VECTORIZE_GET_MASK_MODE has the advantage that we no
longer need to generate vpsel with vectors of 0 and 1: the masks are
now merged via scalar 'ands' instructions operating on 16-bit masks
after converting the boolean vectors.
In addition, this patch fixes a problem in arm_expand_vcond() where
the result would be a vector of 0 or 1 instead of operand 1 or 2.
Since we want to skip gcc.dg/signbit-2.c for MVE, we also add a new
arm_mve effective target.
Reducing the number of iterations in pr100757-3.c from 32 to 8, we
generate the code below:
float a[32];
float fn1(int d) {
float c = 4.0f;
for (int b = 0; b < 8; b++)
if (a[b] != 2.0f)
c = 5.0f;
return c;
}
fn1:
ldr r3, .L3+48
vldr.64 d4, .L3 // q2=(2.0,2.0,2.0,2.0)
vldr.64 d5, .L3+8
vldrw.32 q0, [r3] // q0=a(0..3)
adds r3, r3, #16
vcmp.f32 eq, q0, q2 // cmp a(0..3) == (2.0,2.0,2.0,2.0)
vldrw.32 q1, [r3] // q1=a(4..7)
vmrs r3, P0
vcmp.f32 eq, q1, q2 // cmp a(4..7) == (2.0,2.0,2.0,2.0)
vmrs r2, P0 @ movhi
ands r3, r3, r2 // r3=select(a(0..3]) & select(a(4..7))
vldr.64 d4, .L3+16 // q2=(5.0,5.0,5.0,5.0)
vldr.64 d5, .L3+24
vmsr P0, r3
vldr.64 d6, .L3+32 // q3=(4.0,4.0,4.0,4.0)
vldr.64 d7, .L3+40
vpsel q3, q3, q2 // q3=vcond_mask(4.0,5.0)
vmov.32 r2, q3[1] // keep the scalar max
vmov.32 r0, q3[3]
vmov.32 r3, q3[2]
vmov.f32 s11, s12
vmov s15, r2
vmov s14, r3
vmaxnm.f32 s15, s11, s15
vmaxnm.f32 s15, s15, s14
vmov s14, r0
vmaxnm.f32 s15, s15, s14
vmov r0, s15
bx lr
.L4:
.align 3
.L3:
.word 1073741824 // 2.0f
.word 1073741824
.word 1073741824
.word 1073741824
.word 1084227584 // 5.0f
.word 1084227584
.word 1084227584
.word 1084227584
.word 1082130432 // 4.0f
.word 1082130432
.word 1082130432
.word 1082130432
This patch adds tests that trigger an ICE without this fix.
The pr100757*.c testcases are derived from
gcc.c-torture/compile/20160205-1.c, forcing the use of MVE, and using
various types and return values different from 0 and 1 to avoid
commonalization with boolean masks. In addition, since we should not
need these masks, the tests make sure they are not present.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
PR target/100757
gcc/
* config/arm/arm-protos.h (arm_get_mask_mode): New prototype.
(arm_expand_vector_compare): Update prototype.
* config/arm/arm.cc (TARGET_VECTORIZE_GET_MASK_MODE): New.
(arm_vector_mode_supported_p): Add support for VxBI modes.
(arm_expand_vector_compare): Remove useless generation of vpsel.
(arm_expand_vcond): Fix select operands.
(arm_get_mask_mode): New.
* config/arm/mve.md (vec_cmp<mode><MVE_vpred>): New.
(vec_cmpu<mode><MVE_vpred>): New.
(vcond_mask_<mode><MVE_vpred>): New.
* config/arm/vec-common.md (vec_cmp<mode><v_cmp_result>)
(vec_cmpu<mode><mode, vcond_mask_<mode><v_cmp_result>): Move to ...
* config/arm/neon.md (vec_cmp<mode><v_cmp_result>)
(vec_cmpu<mode><mode, vcond_mask_<mode><v_cmp_result>): ... here
and disable for MVE.
* doc/sourcebuild.texi (arm_mve): Document new effective-target.
gcc/testsuite/
PR target/100757
* gcc.target/arm/simd/pr100757-2.c: New.
* gcc.target/arm/simd/pr100757-3.c: New.
* gcc.target/arm/simd/pr100757-4.c: New.
* gcc.target/arm/simd/pr100757.c: New.
* gcc.dg/signbit-2.c: Skip when targeting ARM/MVE.
* lib/target-supports.exp (check_effective_target_arm_mve): New.
We make use of qualifier_predicate to describe MVE builtins
prototypes, restricting to auto-vectorizable vcmp* and vpsel builtins,
as they are exercised by the tests added earlier in the series.
Special handling is needed for mve_vpselq because it has a v2di
variant, which has no natural VPR.P0 representation: we keep HImode
for it.
The vector_compare expansion code is updated to use the right VxBI
mode instead of HI for the result.
We extend the existing thumb2_movhi_vfp and thumb2_movhi_fp16 patterns
to use the new MVE_7_HI iterator which covers HI and the new VxBI
modes, in conjunction with the new DB constraint for a constant vector
of booleans.
This patch also adds tests derived from the one provided in PR
target/101325: there is a compile-only test because I did not have
access to anything that could execute MVE code until recently. I have
been able to add an executable test since QEMU supports MVE.
Instead of adding arm_v8_1m_mve_hw, I update arm_mve_hw so that it
uses add_options_for_arm_v8_1m_mve_fp, like arm_neon_hw does. This
ensures arm_mve_hw passes even if the toolchain does not generate MVE
code by default.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR target/100757
PR target/101325
* config/arm/arm-builtins.cc (BINOP_PRED_UNONE_UNONE_QUALIFIERS)
(BINOP_PRED_NONE_NONE_QUALIFIERS)
(TERNOP_NONE_NONE_NONE_PRED_QUALIFIERS)
(TERNOP_UNONE_UNONE_UNONE_PRED_QUALIFIERS): New.
* config/arm/arm-protos.h (mve_bool_vec_to_const): New.
* config/arm/arm.cc (arm_hard_regno_mode_ok): Handle new VxBI
modes.
(arm_mode_to_pred_mode): New.
(arm_expand_vector_compare): Use the right VxBI mode instead of
HI.
(arm_expand_vcond): Likewise.
(simd_valid_immediate): Handle MODE_VECTOR_BOOL.
(mve_bool_vec_to_const): New.
(neon_make_constant): Call mve_bool_vec_to_const when needed.
* config/arm/arm_mve_builtins.def (vcmpneq_, vcmphiq_, vcmpcsq_)
(vcmpltq_, vcmpleq_, vcmpgtq_, vcmpgeq_, vcmpeqq_, vcmpneq_f)
(vcmpltq_f, vcmpleq_f, vcmpgtq_f, vcmpgeq_f, vcmpeqq_f, vpselq_u)
(vpselq_s, vpselq_f): Use new predicated qualifiers.
* config/arm/constraints.md (DB): New.
* config/arm/iterators.md (MVE_7, MVE_7_HI): New mode iterators.
(MVE_VPRED, MVE_vpred): New attribute iterators.
* config/arm/mve.md (@mve_vcmp<mve_cmp_op>q_<mode>)
(@mve_vcmp<mve_cmp_op>q_f<mode>, @mve_vpselq_<supf><mode>)
(@mve_vpselq_f<mode>): Use MVE_VPRED instead of HI.
(@mve_vpselq_<supf>v2di): Define separately.
(mov<mode>): New expander for VxBI modes.
* config/arm/vfp.md (thumb2_movhi_vfp, thumb2_movhi_fp16): Use
MVE_7_HI iterator and add support for DB constraint.
gcc/testsuite/
PR target/100757
PR target/101325
* gcc.dg/rtl/arm/mve-vxbi.c: New test.
* gcc.target/arm/simd/pr101325.c: New.
* gcc.target/arm/simd/pr101325-2.c: New.
* lib/target-supports.exp (check_effective_target_arm_mve_hw): Use
add_options_for_arm_v8_1m_mve_fp.
This patch implements support for vectors of booleans to support MVE
predicates, instead of HImode. Since the ABI mandates pred16_t (aka
uint16_t) to represent predicates in intrinsics prototypes, we
introduce a new "predicate" type qualifier so that we can map relevant
builtins HImode arguments and return value to the appropriate vector
of booleans (VxBI).
We have to update test_vector_ops_duplicate, because it iterates using
an offset in bytes, where we would need to iterate in bits: we stop
iterating when we reach the end of the vector of booleans.
In addition, we have to fix the underlying definition of vectors of
booleans because ARM/MVE needs a different representation than
AArch64/SVE. With ARM/MVE the 'true' bit is duplicated over the
element size, so that a true element of V4BI is represented by
'0b1111'. This patch updates the aarch64 definition of VNx*BI as
needed.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR target/100757
PR target/101325
* config/aarch64/aarch64-modes.def (VNx16BI, VNx8BI, VNx4BI,
VNx2BI): Update definition.
* config/arm/arm-builtins.cc (arm_init_simd_builtin_types): Add new
simd types.
(arm_init_builtin): Map predicate vectors arguments to HImode.
(arm_expand_builtin_args): Move HImode predicate arguments to VxBI
rtx. Move return value to HImode rtx.
* config/arm/arm-builtins.h (arm_type_qualifiers): Add qualifier_predicate.
* config/arm/arm-modes.def (B2I, B4I, V16BI, V8BI, V4BI): New modes.
* config/arm/arm-simd-builtin-types.def (Pred1x16_t,
Pred2x8_t,Pred4x4_t): New.
* emit-rtl.cc (init_emit_once): Handle all boolean modes.
* genmodes.cc (mode_data): Add boolean field.
(blank_mode): Initialize it.
(make_complex_modes): Fix handling of boolean modes.
(make_vector_modes): Likewise.
(VECTOR_BOOL_MODE): Use new COMPONENT parameter.
(make_vector_bool_mode): Likewise.
(BOOL_MODE): New.
(make_bool_mode): New.
(emit_insn_modes_h): Fix generation of boolean modes.
(emit_class_narrowest_mode): Likewise.
* machmode.def: (VECTOR_BOOL_MODE): Document new COMPONENT
parameter. Use new BOOL_MODE instead of FRACTIONAL_INT_MODE to
define BImode.
* rtx-vector-builder.cc (rtx_vector_builder::find_cached_value):
Fix handling of constm1_rtx for VECTOR_BOOL.
* simplify-rtx.cc (native_encode_rtx): Fix support for VECTOR_BOOL.
(native_decode_vector_rtx): Likewise.
(test_vector_ops_duplicate): Skip vec_merge test
with vectors of booleans.
* varasm.cc (output_constant_pool_2): Likewise.
The vmvnq_n* intrinsics and have [u]int[16|32]_t arguments, so use
<V_elem> iterator instead of HI in mve_vmvnq_n_<supf><mode>.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
* config/arm/mve.md (mve_vmvnq_n_<supf><mode>): Use V_elem mode
for operand 1.
VPR_REG is the only register in its class, so it should be handled by
TARGET_CLASS_LIKELY_SPILLED_P, which is achieved by calling
default_class_likely_spilled_p. No test fails without this patch, but
it seems it should be implemented.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
* config/arm/arm.cc (arm_class_likely_spilled_p): Handle VPR_REG.
At some point during the development of this patch series, it appeared
that in some cases the register allocator wants “VPR or general”
rather than “VPR or general or FP” (which is the same thing as
ALL_REGS). The series does not seem to require this anymore, but it
seems to be a good thing to do anyway, to give the register allocator
more freedom.
CLASS_MAX_NREGS and arm_hard_regno_nregs need adjustment to avoid a
regression in gcc.dg/stack-usage-1.c when compiled with -mthumb
-mfloat-abi=hard -march=armv8.1-m.main+mve.fp+fp.dp.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/
* config/arm/arm.h (reg_class): Add GENERAL_AND_VPR_REGS.
(REG_CLASS_NAMES): Likewise.
(REG_CLASS_CONTENTS): Likewise.
(CLASS_MAX_NREGS): Handle VPR.
* config/arm/arm.cc (arm_hard_regno_nregs): Handle VPR.
This patch mainly adds Neon tests similar to existing MVE ones,
to make sure we do not break Neon when fixing MVE.
mve-vcmp-f32-2.c is similar to mve-vcmp-f32.c but uses a conditional
with 2.0f and 3.0f constants to help scan-assembler-times.
Most of the work of this patch series was carried out while I was
working at STMicroelectronics as a Linaro assignee.
2022-02-22 Christophe Lyon <christophe.lyon@arm.com>
gcc/testsuite/
* gcc.target/arm/simd/mve-vcmp-f32-2.c: New.
* gcc.target/arm/simd/neon-compare-1.c: New.
* gcc.target/arm/simd/neon-compare-2.c: New.
* gcc.target/arm/simd/neon-compare-3.c: New.
* gcc.target/arm/simd/neon-compare-scalar-1.c: New.
* gcc.target/arm/simd/neon-vcmp-f16.c: New.
* gcc.target/arm/simd/neon-vcmp-f32-2.c: New.
* gcc.target/arm/simd/neon-vcmp-f32-3.c: New.
* gcc.target/arm/simd/neon-vcmp-f32.c: New.
* gcc.target/arm/simd/neon-vcmp.c: New.
Consider the following omp fragment.
...
#pragma omp target
#pragma omp parallel num_threads (2)
#pragma omp task
;
...
This hangs at -O0 for nvptx.
Investigating the behaviour gives us the following trace of events:
- both threads execute GOMP_task, where they:
- deposit a task, and
- execute gomp_team_barrier_wake
- thread 1 executes gomp_team_barrier_wait_end and, not being the last thread,
proceeds to wait at the team barrier
- thread 0 executes gomp_team_barrier_wait_end and, being the last thread, it
calls gomp_barrier_handle_tasks, where it:
- executes both tasks and marks the team barrier done
- executes a gomp_team_barrier_wake which wakes up thread 1
- thread 1 exits the team barrier
- thread 0 returns from gomp_barrier_handle_tasks and goes to wait at
the team barrier.
- thread 0 hangs.
To understand why there is a hang here, it's good to understand how things
are setup for nvptx. The libgomp/config/nvptx/bar.c implementation is
a copy of the libgomp/config/linux/bar.c implementation, with uses of both
futex_wake and do_wait replaced with uses of ptx insn bar.sync:
...
if (bar->total > 1)
asm ("bar.sync 1, %0;" : : "r" (32 * bar->total));
...
The point where thread 0 goes to wait at the team barrier, corresponds in
the linux implementation with a do_wait. In the linux case, the call to
do_wait doesn't hang, because it's waiting for bar->generation to become
a certain value, and if bar->generation already has that value, it just
proceeds, without any need for coordination with other threads.
In the nvtpx case, the bar.sync waits until thread 1 joins it in the same
logical barrier, which never happens: thread 1 is lingering in the
thread pool at the thread pool barrier (using a different logical barrier),
waiting to join a new team.
The easiest way to fix this is to revert to the posix implementation for
bar.{c,h}. That however falls back on a busy-waiting approach, and
does not take advantage of the ptx bar.sync insn.
Instead, we revert to the linux implementation for bar.c,
and implement bar.c local functions futex_wait and futex_wake using the
bar.sync insn.
The bar.sync insn takes an argument specifying how many threads are
participating, and that doesn't play well with the futex syntax where it's
not clear in advance how many threads will be woken up.
This is solved by waking up all waiting threads each time a futex_wait or
futex_wake happens, and possibly going back to sleep with an updated thread
count.
Tested libgomp on x86_64 with nvptx accelerator.
libgomp/ChangeLog:
2021-04-20 Tom de Vries <tdevries@suse.de>
PR target/99555
* config/nvptx/bar.c (generation_to_barrier): New function, copied
from config/rtems/bar.c.
(futex_wait, futex_wake): New function.
(do_spin, do_wait): New function, copied from config/linux/wait.h.
(gomp_barrier_wait_end, gomp_barrier_wait_last)
(gomp_team_barrier_wake, gomp_team_barrier_wait_end):
(gomp_team_barrier_wait_cancel_end, gomp_team_barrier_cancel): Remove
and replace with include of config/linux/bar.c.
* config/nvptx/bar.h (gomp_barrier_t): Add fields waiters and lock.
(gomp_barrier_init): Init new fields.
* testsuite/libgomp.c-c++-common/task-detach-6.c: Remove nvptx-specific
workarounds.
* testsuite/libgomp.c/pr99555-1.c: Same.
* testsuite/libgomp.fortran/task-detach-6.f90: Same.
Add -misa=sm_70, and use it to specify the misa value in test-case
gcc.target/nvptx/atomic-store-2.c.
Tested on nvptx.
gcc/ChangeLog:
* config/nvptx/nvptx-c.cc (nvptx_cpu_cpp_builtins): Handle SM70.
* config/nvptx/nvptx.cc (first_ptx_version_supporting_sm):
Likewise.
* config/nvptx/nvptx.opt (misa): Add sm_70 alias PTX_ISA_SM70.
gcc/testsuite/ChangeLog:
2022-02-22 Tom de Vries <tdevries@suse.de>
* gcc.target/nvptx/atomic-store-2.c: Use -misa=sm_70.
* gcc.target/nvptx/uniform-simt-3.c: Same.
Co-Authored-By: Tom de Vries <tdevries@suse.de>
This implements the wording changes in P2415R2 "What is a view?", which
is a DR for C++20.
libstdc++-v3/ChangeLog:
* include/bits/ranges_base.h (__detail::__is_initializer_list):
Define.
(viewable_range): Adjust as per P2415R2.
* include/bits/ranges_cmp.h (__cpp_lib_ranges): Adjust value.
* include/std/ranges (owning_view): Define as per P2415R2.
(enable_borrowed_range<owning_view>): Likewise.
(views::__detail::__can_subrange): Replace with ...
(views::__detail::__can_owning_view): ... this.
(views::_All::_S_noexcept): Sync with operator().
(views::_All::operator()): Use owning_view instead of subrange
as per P2415R2.
* include/std/version (__cpp_lib_ranges): Adjust value.
* testsuite/std/ranges/adaptors/all.cc (test06): Adjust now that
views::all uses owning_view instead of subrange.
(test08): New test.
* testsuite/std/ranges/adaptors/lazy_split.cc (test09): Adjust
now that rvalue non-view non-borrowed ranges are viewable.
* testsuite/std/ranges/adaptors/split.cc (test06): Likewise.
Currently supported internally are 3.1, 6.0, 6.3 and 7.0.
However, -mptx= supports 3.1, 6.3, 7.0 – but not the internal default 6.0.
Add -mptx=6.0 for consistency.
Tested on nvptx.
gcc/ChangeLog:
* config/nvptx/nvptx.opt (mptx): Add 6.0 alias PTX_VERSION_6_0.
* doc/invoke.texi (-mptx): Update for new values and defaults.
Co-Authored-By: Tom de Vries <tdevries@suse.de>
Add functionality that indicates which insns are added by -minit-regs, such
that for instance we have for pr53465.s:
...
// #APP
// 9 "gcc/testsuite/gcc.c-torture/execute/pr53465.c" 1
// Start: Added by -minit-regs=3:
// #NO_APP
mov.u32 %r26, 0;
// #APP
// 9 "gcc/testsuite/gcc.c-torture/execute/pr53465.c" 1
// End: Added by -minit-regs=3:
// #NO_APP
...
Can be switched off using -mno-ptx-comment.
Tested on nvptx.
gcc/ChangeLog:
2022-02-21 Tom de Vries <tdevries@suse.de>
* config/nvptx/nvptx.cc (gen_comment): New function.
(workaround_uninit_method_1, workaround_uninit_method_2)
(workaround_uninit_method_3): : Use gen_comment.
* config/nvptx/nvptx.opt (mptx-comment): New option.
This makes the SLP vectorizer dump the def we use for a splat to
aid debugging.
2022-02-22 Richard Biener <rguenther@suse.de>
* tree-vect-slp.cc (vect_build_slp_tree_2): Dump the def used
for a splat.
This patch addresses a code quality regression in GCC 12 by implementing
some constant folding/simplification transformations for REDUC_PLUS_EXPR
in match.pd. The motivating example is gcc.dg/vect/pr89440.c which with
-O2 -ffast-math (with vectorization now enabled) gets optimized to:
float f (float x)
{
vector(4) float vect_x_14.11;
vector(4) float _2;
float _32;
_2 = {x_9(D), 0.0, 0.0, 0.0};
vect_x_14.11_29 = _2 + { 1.0e+1, 2.6e+1, 4.2e+1, 5.8e+1 };
_32 = .REDUC_PLUS (vect_x_14.11_29); [tail call]
return _32;
}
With these proposed new transformations, we can simplify the
above code even further.
float f (float x)
{
float _32;
_32 = x_9(D) + 1.36e+2;
return _32;
}
[which happens to match what we'd produce with -fno-tree-vectorize,
and with GCC 11].
2022-02-22 Roger Sayle <roger@nextmovesoftware.com>
Richard Biener <rguenther@suse.de>
gcc/ChangeLog
* fold-const.cc (ctor_single_nonzero_element): New function to
return the single non-zero element of a (vector) constructor.
* fold-const.h (ctor_single_nonzero_element): Prototype here.
* match.pd (reduc (constructor@0)): Simplify reductions of a
constructor containing a single non-zero element.
(reduc (@0 op VECTOR_CST) -> (reduc @0) op CONST): Simplify
reductions of vector operations of the same operator with
constant vector operands.
gcc/testsuite/ChangeLog
* gcc.dg/fold-reduc-1.c: New test case.
On
#define A(n) int foo1##n(void) { return 1##n; }
#define B(n) A(n##0) A(n##1) A(n##2) A(n##3) A(n##4) A(n##5) A(n##6) A(n##7) A(n##8) A(n##9)
#define C(n) B(n##0) B(n##1) B(n##2) B(n##3) B(n##4) B(n##5) B(n##6) B(n##7) B(n##8) B(n##9)
#define D(n) C(n##0) C(n##1) C(n##2) C(n##3) C(n##4) C(n##5) C(n##6) C(n##7) C(n##8) C(n##9)
#define E(n) D(n##0) D(n##1) D(n##2) D(n##3) D(n##4) D(n##5) D(n##6) D(n##7) D(n##8) D(n##9)
E(0) E(1) E(2) D(30) D(31) C(320) C(321) C(322) C(323) C(324) C(325)
B(3260) B(3261) B(3262) B(3263) A(32640) A(32641) A(32642)
testcase with
./xgcc -B ./ -c -g -fpic -ffat-lto-objects -flto -O0 -o foo1.o foo1.c -ffunction-sections
./xgcc -B ./ -shared -g -fpic -flto -O0 -o foo1.so foo1.o
/tmp/ccTW8mBm.debug.temp.o: file not recognized: file format not recognized
(testcase too slow to be included into testsuite).
The problem is clearly reported by readelf:
readelf: foo1.o.debug.temp.o: Warning: Section 2 has an out of range sh_link value of 65321
readelf: foo1.o.debug.temp.o: Warning: Section 5 has an out of range sh_link value of 65321
readelf: foo1.o.debug.temp.o: Warning: Section 10 has an out of range sh_link value of 65323
readelf: foo1.o.debug.temp.o: Warning: [ 2]: Link field (65321) should index a symtab section.
readelf: foo1.o.debug.temp.o: Warning: [ 5]: Link field (65321) should index a symtab section.
readelf: foo1.o.debug.temp.o: Warning: [10]: Link field (65323) should index a string section.
because simple_object_elf_copy_lto_debug_sections doesn't adjust sh_info and
sh_link fields in ElfNN_Shdr if they are in between SHN_{LO,HI}RESERVE
inclusive. Not adjusting those is incorrect though, SHN_{LO,HI}RESERVE
range is only relevant to the 16-bit fields, mainly st_shndx in ElfNN_Sym
where if one needs >= SHN_LORESERVE section number, SHN_XINDEX should be
used instead and .symtab_shndx section should contain the real section
index, and in ElfNN_Ehdr e_shnum and e_shstrndx fields, where if >=
SHN_LORESERVE value is needed it should put those into
Shdr[0].sh_{size,link}. But, sh_{link,info} are 32-bit fields which can
contain any section index.
Note, as simple-object-elf.c mentions, binutils from 2.12 to 2.18 (so before
2011) used to mishandle the > 63.75K sections case and assumed there is a
hole in between the sections, but what
simple_object_elf_copy_lto_debug_sections does wouldn't help in that case
for the debug temp object creation, we'd need to detect the case also in
that routine and take it into account in the remapping etc. I think
it is not worth it given that it is over 10 years, if somebody needs
63.75K or more sections, better use more recent binutils.
2022-02-22 Jakub Jelinek <jakub@redhat.com>
PR lto/104617
* simple-object-elf.c (simple_object_elf_match): Fix up URL
in comment.
(simple_object_elf_copy_lto_debug_sections): Remap sh_info and
sh_link even if they are in the SHN_LORESERVE .. SHN_HIRESERVE
range (inclusive).
The following testcase is miscompiled since r12-3328.
That change assumed that if rhs1 of a GIMPLE_ASSIGN is COMPLEX_CST, then
that is the value of the lhs of the stmt, but that is not the case always,
only if it is a GIMPLE_SINGLE_RHS stmt. If it is e.g.
GIMPLE_UNARY_RHS or GIMPLE_BINARY_RHS (the latter happens in the testcase),
then it can be e.g.
__complex__ (3, 0) / var
and the REALPART_EXPR of that isn't 3, but the realpart of the division.
I assume once the ranger can do complex numbers adjust_*part_expr will just
fetch one or the other range from a underlying complex range, but until
then, we should limit this to what r12-3328 meant to do.
2022-02-22 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/104604
* gimple-range-fold.cc (adjust_imagpart_expr, adjust_realpart_expr):
Only check if gimple_assign_rhs1 is COMPLEX_CST if
gimple_assign_rhs_code is COMPLEX_CST.
* gcc.c-torture/execute/pr104604.c: New test.
We ICE on the following testcase for -m32 since r12-3435. because
operands[2] is (subreg:SF (reg:DI ...) 0) and
lowpart_subreg (V4SFmode, operands[2], SFmode)
returns NULL, and that is what we use in AND etc. insns we emit.
My earlier version of the patch fixes that by calling force_reg for the
input operands, to make sure they are really REGs and so lowpart_subreg
will succeed on them - even for theoretical MEMs using REGs there seems
desirable, we don't want to read following memory slots for the paradoxical
subreg. For the outputs, I thought we'd get better code by always computing
result into a new pseudo and them move lowpart of that pseudo into dest.
Unfortunately it regressed
FAIL: gcc.target/i386/pr89984-2.c scan-assembler-not vmovaps
on which the patch changes:
vandps .LC0(%rip), %xmm1, %xmm1
- vxorps %xmm0, %xmm1, %xmm0
+ vxorps %xmm0, %xmm1, %xmm1
+ vmovaps %xmm1, %xmm0
ret
The RA sees:
(insn 8 4 9 2 (set (reg:V4SF 85)
(and:V4SF (subreg:V4SF (reg:SF 90) 0)
(mem/u/c:V4SF (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [0 S16 A128]))) "pr89984-2.c":7:12 2838 {*andv4sf3}
(expr_list:REG_DEAD (reg:SF 90)
(nil)))
(insn 9 8 10 2 (set (reg:V4SF 87)
(xor:V4SF (reg:V4SF 85)
(subreg:V4SF (reg:SF 89) 0))) "pr89984-2.c":7:12 2842 {*xorv4sf3}
(expr_list:REG_DEAD (reg:SF 89)
(expr_list:REG_DEAD (reg:V4SF 85)
(nil))))
(insn 10 9 14 2 (set (reg:SF 82 [ <retval> ])
(subreg:SF (reg:V4SF 87) 0)) "pr89984-2.c":7:12 142 {*movsf_internal}
(expr_list:REG_DEAD (reg:V4SF 87)
(nil)))
(insn 14 10 15 2 (set (reg/i:SF 20 xmm0)
(reg:SF 82 [ <retval> ])) "pr89984-2.c":8:1 142 {*movsf_internal}
(expr_list:REG_DEAD (reg:SF 82 [ <retval> ])
(nil)))
(insn 15 14 0 2 (use (reg/i:SF 20 xmm0)) "pr89984-2.c":8:1 -1
(nil))
and doesn't know that if it would use xmm0 not just for pseudo 82
but also for pseudo 87, it could create a noop move in insn 10 and
so could avoid an extra register copy and nothing later on is able
to figure that out either. I don't know how the RA should know
that though.
So that we don't regress, this version of the patch
will do this stuff (i.e. use fresh vector pseudo as destination and
then move lowpart of that to dest) over what it used before (i.e.
use paradoxical subreg of the dest) only if lowpart_subreg returns NULL.
2022-02-22 Jakub Jelinek <jakub@redhat.com>
PR target/104612
* config/i386/i386-expand.cc (ix86_expand_copysign): Call force_reg
on input operands before calling lowpart_subreg on it. For output
operand, use a vmode pseudo as destination and then move its lowpart
subreg into operands[0] if lowpart_subreg fails on dest.
(ix86_expand_xorsign): Likewise.
* gcc.dg/pr104612.c: New test.
