When testing for Nios II (gcc-testresults shows this for various other
targets as well), tests scanning cselim dumps produce an UNRESOLVED
result because those dumps do not exist.
cselim is enabled conditionally by code in toplev.c:
if (flag_tree_cselim == AUTODETECT_VALUE)
{
if (HAVE_conditional_move)
flag_tree_cselim = 1;
else
flag_tree_cselim = 0;
}
Add explicit -ftree-cselim to dg-options in the affected tests (as
already used by some other tests of cselim dumps) so that this dump
exists on all architectures.
Tested with no regressions with cross to nios2-elf, where this causes
the tests in question to PASS instead of being UNRESOLVED.
* gcc.dg/tree-ssa/pr89430-1.c, gcc.dg/tree-ssa/pr89430-2.c,
gcc.dg/tree-ssa/pr89430-3.c, gcc.dg/tree-ssa/pr89430-4.c,
gcc.dg/tree-ssa/pr89430-5.c, gcc.dg/tree-ssa/pr89430-6.c,
gcc.dg/tree-ssa/pr89430-7-comp-ref.c,
gcc.dg/tree-ssa/pr89430-8-mem-ref-size.c,
gcc.dg/tree-ssa/pr99473-1.c: Use -ftree-cselim.
We cannot use r12 here, it is already in use as the GEP (for sibling
calls).
2021-09-08 Segher Boessenkool <segher@kernel.crashing.org>
PR target/102107
* config/rs6000/rs6000-logue.c (rs6000_emit_epilogue): For ELFv2 use
r11 instead of r12 for restoring CR.
Thinking about it more this morning, while this patch fixes the problems
revealed in the testcase, the recent PR89984 change was buggy too, but
perhaps that can be fixed incrementally. Because for AVX the new code
destructively modifies op1. If that is different from dest, say on:
float
foo (float x, float y)
{
return x * __builtin_copysignf (1.0f, y) + y;
}
then we get after RA:
(insn 8 7 9 2 (set (reg:SF 20 xmm0 [orig:82 _2 ] [82])
(unspec:SF [
(reg:SF 20 xmm0 [88])
(reg:SF 21 xmm1 [89])
(mem/u/c:V4SF (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [0 S16 A128])
] UNSPEC_XORSIGN)) "hohoho.c":4:12 649 {xorsignsf3_1}
(nil))
(insn 9 8 15 2 (set (reg:SF 20 xmm0 [87])
(plus:SF (reg:SF 20 xmm0 [orig:82 _2 ] [82])
(reg:SF 21 xmm1 [89]))) "hohoho.c":4:44 1021 {*fop_sf_comm}
(nil))
but split the xorsign into:
vandps .LC0(%rip), %xmm1, %xmm1
vxorps %xmm0, %xmm1, %xmm0
and then the addition:
vaddss %xmm1, %xmm0, %xmm0
which means we miscompile it - instead of adding y in the end we add
__builtin_copysignf (0.0f, y).
So, wonder if we don't want instead in addition to the &Yv <- Yv, 0
alternative (enabled for both pre-AVX and AVX as in this patch) the
&Yv <- Yv, Yv where destination must be different from inputs and another
Yv <- Yv, Yv where it can be the same but then need a match_scratch
(with X for the other alternatives and =Yv for the last one).
That way we'd always have a safe register we can store the op1 & mask
value into, either the destination (in the first alternative known to
be equal to op1 which is needed for non-AVX but ok for AVX too), in the
second alternative known to be different from both inputs and in the third
which could be used for those
float bar (float x, float y) { return x * __builtin_copysignf (1.0f, y); }
cases where op1 is naturally xmm1 and dest == op0 naturally xmm0 we'd use
some other register like xmm2.
On Wed, Sep 08, 2021 at 05:23:40PM +0800, Hongtao Liu wrote:
> I'm curious why we need the post_reload splitter @xorsign<mode>3_1
> for scalar mode, can't we just expand them into and/xor operations in
> the expander, just like vector modes did.
Following seems to work for all the testcases I've tried (and in some
generates better code than the post-reload splitter).
2021-09-08 Jakub Jelinek <jakub@redhat.com>
liuhongt <hongtao.liu@intel.com>
PR target/89984
* config/i386/i386.md (@xorsign<mode>3_1): Remove.
* config/i386/i386-expand.c (ix86_expand_xorsign): Expand right away
into AND with mask and XOR, using paradoxical subregs.
(ix86_split_xorsign): Remove.
* config/i386/i386-protos.h (ix86_split_xorsign): Remove.
* gcc.target/i386/avx-pr102224.c: Fix up PR number.
* gcc.dg/pr89984.c: New test.
* gcc.target/i386/avx-pr89984.c: New test.
If the first predicate value is different and copied, the comparison will then
be between val->result and the copied one. That can cause inserting extra
vn_pvals.
gcc/ChangeLog:
* tree-ssa-sccvn.c (vn_nary_op_insert_into): fix result compare
The following patch exports it for Linux from config/i386/*.ver where it
IMNSHO belongs, aarch64 already exports some of those at GCC_11* and other
targets might add them at completely different gcc versions.
2021-09-08 Jakub Jelinek <jakub@redhat.com>
Iain Sandoe <iain@sandoe.co.uk>
* config/i386/libgcc-glibc.ver: Add %inherit GCC_12.0.0 GCC_7.0.0
and export *hf* and *hc* functions at GCC_12.0.0.
As the testcase shows, we miscompile @xorsign<mode>3_1 if both input
operands are in the same register, because the splitter overwrites op1
before with op1 & mask before using op0.
For dest = xorsign op0, op0 we can actually simplify it from
dest = (op0 & mask) ^ op0 to dest = op0 & ~mask (aka abs).
The expander change is an optimization improvement, if we at expansion
time know it is xorsign op0, op0, we can emit abs right away and get better
code through that.
The @xorsign<mode>3_1 is a fix for the case where xorsign wouldn't be known
to have same operands during expansion, but during RTL optimizations they
would appear. For non-AVX we need to use earlyclobber, we require
dest and op1 to be the same but op0 must be different because we overwrite
op1 first. For AVX the constraints ensure that at most 2 of the 3 operands
may be the same register and if both inputs are the same, handles that case.
This case can be easily tested with the xorsign<mode>3 expander change
reverted.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Thinking about it more this morning, while this patch fixes the problems
revealed in the testcase, the recent PR89984 change was buggy too, but
perhaps that can be fixed incrementally. Because for AVX the new code
destructively modifies op1. If that is different from dest, say on:
float
foo (float x, float y)
{
return x * __builtin_copysignf (1.0f, y) + y;
}
then we get after RA:
(insn 8 7 9 2 (set (reg:SF 20 xmm0 [orig:82 _2 ] [82])
(unspec:SF [
(reg:SF 20 xmm0 [88])
(reg:SF 21 xmm1 [89])
(mem/u/c:V4SF (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [0 S16 A128])
] UNSPEC_XORSIGN)) "hohoho.c":4:12 649 {xorsignsf3_1}
(nil))
(insn 9 8 15 2 (set (reg:SF 20 xmm0 [87])
(plus:SF (reg:SF 20 xmm0 [orig:82 _2 ] [82])
(reg:SF 21 xmm1 [89]))) "hohoho.c":4:44 1021 {*fop_sf_comm}
(nil))
but split the xorsign into:
vandps .LC0(%rip), %xmm1, %xmm1
vxorps %xmm0, %xmm1, %xmm0
and then the addition:
vaddss %xmm1, %xmm0, %xmm0
which means we miscompile it - instead of adding y in the end we add
__builtin_copysignf (0.0f, y).
So, wonder if we don't want instead in addition to the &Yv <- Yv, 0
alternative (enabled for both pre-AVX and AVX as in this patch) the
&Yv <- Yv, Yv where destination must be different from inputs and another
Yv <- Yv, Yv where it can be the same but then need a match_scratch
(with X for the other alternatives and =Yv for the last one).
That way we'd always have a safe register we can store the op1 & mask
value into, either the destination (in the first alternative known to
be equal to op1 which is needed for non-AVX but ok for AVX too), in the
second alternative known to be different from both inputs and in the third
which could be used for those
float bar (float x, float y) { return x * __builtin_copysignf (1.0f, y); }
cases where op1 is naturally xmm1 and dest == op0 naturally xmm0 we'd use
some other register like xmm2.
2021-09-08 Jakub Jelinek <jakub@redhat.com>
PR target/102224
* config/i386/i386.md (xorsign<mode>3): If operands[1] is equal to
operands[2], emit abs<mode>2 instead.
(@xorsign<mode>3_1): Add early-clobbers for output operand, enable
first alternative even for avx, add another alternative with
=&Yv <- 0, Yv, Yvm constraints.
* config/i386/i386-expand.c (ix86_split_xorsign): If op0 is equal
to op1, emit vpandn instead.
* gcc.dg/pr102224.c: New test.
* gcc.target/i386/avx-pr102224.c: New test.
2021-09-07 Takayuki 'January June' Suwa <jjsuwa_sys3175@yahoo.co.jp>
gcc/
PR target/102115
* config/xtensa/xtensa.c (xtensa_emit_move_sequence): Add
'CONST_INT_P (src)' to the condition of the block that tries to
eliminate literal when loading integer contant.
Document the new command line options (-mco-re and -mno-co-re), the new
BPF target builtin (__builtin_preserve_access_index), and the new BPF
target attribute (preserve_access_index) introduced with BPF CO-RE.
gcc/ChangeLog:
* doc/extend.texi (BPF Type Attributes) New node.
Document new preserve_access_index attribute.
Document new preserve_access_index builtin.
* doc/invoke.texi: Document -mco-re and -mno-co-re options.
Expose the function get_btf_id, so that it may be used by the BPF
backend. This enables the BPF CO-RE machinery in the BPF backend to
lookup BTF type IDs, in order to create CO-RE relocation records.
A prototype is added in ctfc.h
gcc/ChangeLog:
* btfout.c (get_btf_id): Function is no longer static.
* ctfc.h: Expose it here.
Add a new function, ctf_lookup_tree_type, to return the CTF type ID
associated with a type via its is TREE node. The function is exposed via
a prototype in ctfc.h.
gcc/ChangeLog:
* ctfc.c (ctf_lookup_tree_type): New function.
* ctfc.h: Likewise.
Expose the function ctf_dtd_lookup, so that it can be used by the BPF
CO-RE machinery. The function is no longer static, and an extern
prototype is added in ctfc.h.
gcc/ChangeLog:
* ctfc.c (ctf_dtd_lookup): Function is no longer static.
* ctfc.h: Analogous change.
Expose the function lookup_type_die in dwarf2out, so that it can be used
by CTF/BTF when adding BPF CO-RE information. The function is now
non-static, and an extern prototype is added in dwarf2out.h.
gcc/ChangeLog:
* dwarf2out.c (lookup_type_die): Function is no longer static.
* dwarf2out.h: Expose it here.
Dejagnu is unfortunately brittle: a syntax error in a
directive can abort the test-run for the current "tool"
(gcc, g++, gfortran), and if you don't check for this
condition or actually read the stdout log yourself, your
tools may make you believe the test was successful without
regressions. At the very least, always grep for ^ERROR: in
the stdout log!
With r12-3379, the testsuite got such a fatal syntax error,
causing the gcc test-run to abort at (e.g.):
...
FAIL: gcc.dg/memchr.c (test for excess errors)
FAIL: gcc.dg/memcmp-3.c (test for excess errors)
ERROR: (DejaGnu) proc "scan-tree-dump-not\" = foo {\(\)"} optimized" does not exist.
The error code is TCL LOOKUP COMMAND scan-tree-dump-not\"
The info on the error is:
invalid command name "scan-tree-dump-not""
while executing
"::tcl_unknown scan-tree-dump-not\" = foo {\(\)"} optimized"
("uplevel" body line 1)
invoked from within
"uplevel 1 ::tcl_unknown $args"
=== gcc Summary ===
# of expected passes 63740
# of unexpected failures 38
# of unexpected successes 2
# of expected failures 351
# of unresolved testcases 3
# of unsupported tests 662
x/cris-elf/gccobj/gcc/xgcc version 12.0.0 20210907 (experimental)\
[master r12-3391-g849d5f5929fc] (GCC)
testsuite:
* gcc.dg/no_profile_instrument_function-attr-2.c: Fix
typo in last change.
gcc/fortran/ChangeLog:
PR fortran/101327
* expr.c (find_array_element): When bounds cannot be determined as
constant, return error instead of aborting.
gcc/testsuite/ChangeLog:
PR fortran/101327
* gfortran.dg/pr101327.f90: New test.
DWARF generation is split between early and late phases when LTO is in effect.
This poses challenges for CTF/BTF generation especially if late debug info
generation is desirable, as turns out to be the case for BPF CO-RE.
The approach taken here in this patch is:
1. LTO is disabled for BPF CO-RE
The reason to disable LTO for BPF CO-RE is that if LTO is in effect, BPF CO-RE
relocations need to be generated in the LTO link phase _after_ the optimizations
are done. This means we need to devise way to combine early and late BTF. At
this time, in absence of linker support for BTF sections, it makes sense to
steer clear of LTO for BPF CO-RE and bypass the issue.
2. The BPF backend updates the write_symbols with BPF_WITH_CORE_DEBUG to convey
the case that BTF with CO-RE support needs to be generated. This information
is used by the debug info emission routines to defer the emission of BTF/CO-RE
until dwarf2out_finish.
So, in other words,
dwarf2out_early_finish
- Always emit CTF here.
- if (BTF && !BTF_WITH_CORE), emit BTF now.
dwarf2out_finish
- if (BTF_WITH_CORE) emit BTF now.
gcc/ChangeLog:
* dwarf2ctf.c (ctf_debug_finalize): Make it static.
(ctf_debug_early_finish): New definition.
(ctf_debug_finish): Likewise.
* dwarf2ctf.h (ctf_debug_finalize): Remove declaration.
(ctf_debug_early_finish): New declaration.
(ctf_debug_finish): Likewise.
* dwarf2out.c (dwarf2out_finish): Invoke ctf_debug_finish.
(dwarf2out_early_finish): Invoke ctf_debug_early_finish.
-mco-re in the BPF backend enables code generation for the CO-RE usecase. LTO is
disabled for CO-RE compilations.
gcc/ChangeLog:
* config/bpf/bpf.c (bpf_option_override): For BPF backend, disable LTO
support when compiling for CO-RE.
* config/bpf/bpf.opt: Add new command line option -mco-re.
gcc/testsuite/ChangeLog:
* gcc.target/bpf/core-lto-1.c: New test.
To best handle BTF/CO-RE in GCC, a distinct BTF_WITH_CORE_DEBUG debug format is
being added. This helps the compiler detect whether BTF with CO-RE relocations
needs to be emitted.
gcc/ChangeLog:
* flag-types.h (enum debug_info_type): Add new enum
DINFO_TYPE_BTF_WITH_CORE.
(BTF_WITH_CORE_DEBUG): New bitmask.
* flags.h (btf_with_core_debuginfo_p): New declaration.
* opts.c (btf_with_core_debuginfo_p): New definition.
I've thought for a while that many of the macros in tree.h and such should
become inline functions. This one in particular was confusing Coverity; the
null check in the macro made it think that all code guarded by
error_operand_p would also need null checks.
gcc/ChangeLog:
* tree.h (error_operand_p): Change to inline function.
We do not save bodies of constexpr clones and instead evaluate the bodies
of the constexpr functions they were cloned from.
I believe that is just fine for constructors because complete vs. base
ctors differ only in classes that have virtual bases and such constructors
aren't constexpr, similarly complete/base destructors.
But as the testcase below shows, for deleting destructors it is not fine,
deleting dtors while marked as clones in fact are just artificial functions
with synthetized body which calls the user destructor and deallocation.
So, either we'd need to evaluate the destructor and afterwards synthetize
and evaluate the deallocation, or we can just save and use the deleting
dtors bodies. The latter seems much easier to me.
2021-09-07 Jakub Jelinek <jakub@redhat.com>
PR c++/100495
* constexpr.c (maybe_save_constexpr_fundef): Save body even for
constexpr deleting dtors.
(cxx_eval_call_expression): Don't use DECL_CLONED_FUNCTION for
deleting dtors.
* g++.dg/cpp2a/constexpr-new21.C: New test.
libgfortran/ChangeLog:
* Makefile.am (gfor_built_src): Depend on
include/ISO_Fortran_binding.h not on ISO_Fortran_binding.h.
(ISO_Fortran_binding.h): Rename make target to ...
(include/ISO_Fortran_binding.h): ... this.
* Makefile.in: Regenerate.
This is the recent LTO bootstrap failure with Ada enabled. The compiler now
generates DW_OP_deref_type for a unit of the Ada front-end, which means that
the offset of base types in the CU must be computed during early DWARF too.
gcc/
PR debug/101947
* dwarf2out.c (mark_base_types): New overloaded function.
(dwarf2out_early_finish): Invoke it on the COMDAT type list as well
as the compilation unit, and call move_marked_base_types afterward.
This fixes re-use of the reduction value in epilogue vectorization
when a conversion from/to variable lenght vectors is required.
2021-09-07 Richard Biener <rguenther@suse.de>
PR tree-optimization/102226
* tree-vect-loop.c (vect_transform_cycle_phi): Record
the converted value for the epilogue PHI use.
* g++.dg/vect/pr102226.cc: New testcase.
This patch adds support for the 'seq_cst' memory order clause on the 'flush'
directive which was introduced in OpenMP 5.1.
gcc/c-family/ChangeLog:
* c-omp.c (c_finish_omp_flush): Handle MEMMODEL_SEQ_CST.
gcc/c/ChangeLog:
* c-parser.c (c_parser_omp_flush): Parse 'seq_cst' clause on 'flush'
directive.
gcc/cp/ChangeLog:
* parser.c (cp_parser_omp_flush): Parse 'seq_cst' clause on 'flush'
directive.
* semantics.c (finish_omp_flush): Handle MEMMODEL_SEQ_CST.
gcc/fortran/ChangeLog:
* openmp.c (gfc_match_omp_flush): Parse 'seq_cst' clause on 'flush'
directive.
* trans-openmp.c (gfc_trans_omp_flush): Handle OMP_MEMORDER_SEQ_CST.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/flush-1.c: Add test case for 'seq_cst'.
* c-c++-common/gomp/flush-2.c: Add test case for 'seq_cst'.
* g++.dg/gomp/attrs-1.C: Adapt test to handle all flush clauses.
* g++.dg/gomp/attrs-2.C: Adapt test to handle all flush clauses.
* gfortran.dg/gomp/flush-1.f90: Add test case for 'seq_cst'.
* gfortran.dg/gomp/flush-2.f90: Add test case for 'seq_cst'.
PR gcov-profile/80223
gcc/ChangeLog:
* ipa-inline.c (can_inline_edge_p): Similarly to sanitizer
options, do not inline when no_profile_instrument_function
attributes are different in early inliner. It's fine to inline
it after PGO instrumentation.
gcc/testsuite/ChangeLog:
* gcc.dg/no_profile_instrument_function-attr-2.c: New test.
This avoids doing redundant work during PHI translation to invalidate
mems when translating their corresponding VUSE through the blocks
virtual PHI node. All the invalidation work is already done by
prune_clobbered_mems.
This speeds up the compile of the testcase from 275s with PRE
taking 91% of the compile-time down to 43s with PRE taking 16%
of the compile-time.
2021-09-07 Richard Biener <rguenther@suse.de>
PR tree-optimization/101555
* tree-ssa-pre.c (translate_vuse_through_block): Do not
perform an alias walk to determine the validity of the
mem at the start of the block which is already guaranteed
by means of prune_clobbered_mems.
(phi_translate_1): Pass edge to translate_vuse_through_block.
Commit fef67987cf changed the
libgfortran build process to generate multilib-specific versions of
ISO_Fortran_binding.h from a template, by running gfortran to identify
the values of the Fortran kind constants C_LONG_DOUBLE, C_FLOAT128,
and C_INT128_T. This caused multiple problems with search paths, both
for build-tree testing and installed-tree use, not all of which have
been fixed.
This patch reverts to a non-multilib-specific .h file that uses GCC's
predefined preprocessor symbols to detect the supported types and map
them to kind values in the same way as the Fortran front end.
2021-09-06 Sandra Loosemore <sandra@codesourcery.com>
libgfortran/
* ISO_Fortran_binding-1-tmpl.h: Deleted.
* ISO_Fortran_binding-2-tmpl.h: Deleted.
* ISO_Fortran_binding-3-tmpl.h: Deleted.
* ISO_Fortran_binding.h: New file to replace the above.
* Makefile.am (gfor_cdir): Remove MULTISUBDIR.
(ISO_Fortran_binding.h): Simplify to just copy the file.
* Makefile.in: Regenerated.
* mk-kinds-h.sh: Revert pieces no longer needed for
ISO_Fortran_binding.h.
Currently, the asm output file for MIPS has no rev info.
It can make some trouble, for example:
assembler is mips1 by default,
gcc is fpxx by default.
To assemble the output of gcc -S, we have to pass -mips2
to assembler.
The same situation is for some CPU has extension insn.
Octeon is an example.
So we can just add ".set arch=octeon".
If an ASE is enabled, .module ase will also be used.
gcc/ChangeLog:
* config/mips/mips.c (mips_file_start): add .module for
arch and ase.
As diagnosed with Jakub and Richard in the analysis of PR 102134, the
current implementation of wi::clz has incorrect/inconsistent behaviour.
As mentioned by Richard in comment #7, clz should (always) return zero
for negative values, but the current implementation can only return 0
when precision is a multiple of HOST_BITS_PER_WIDE_INT. The fix is
simply to reorder/shuffle the existing tests.
2021-09-06 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* wide-int.cc (wi::clz): Reorder tests to ensure the result
is zero for all negative values.
contrib/Changelog:
* gcc_update: Derive human readable name for HEAD using git describe
like "git gcc-descr" with short commit hash. Drop "revision" from
gcc/REVISION.
