platform/upstream/gcc.git
21 months agoFix PR 108582: ICE due to PHI-OPT removing a still in use ssa_name.
Andrew Pinski [Sat, 28 Jan 2023 18:27:08 +0000 (18:27 +0000)]
Fix PR 108582: ICE due to PHI-OPT removing a still in use ssa_name.

This patch adds a check in match_simplify_replacement to make sure the middlebb
does not have any phi-nodes as we don't currently move those.
This was just a thinko from before.

Ok? Bootstrapped and tested on x86_64-linux-gnu with no regressions?

PR tree-optimization/108582

gcc/ChangeLog:

* tree-ssa-phiopt.cc (match_simplify_replacement): Add check
for middlebb to have no phi nodes.

gcc/testsuite/ChangeLog:

* gcc.dg/pr108582-1.c: New test.

21 months agotree-optimization/108574 - wrong-code with PRE PHI node processing
Richard Biener [Mon, 30 Jan 2023 08:25:23 +0000 (09:25 +0100)]
tree-optimization/108574 - wrong-code with PRE PHI node processing

The PR108523 was too optimistic in replacing the same value with
an equivalence from a possibly not taken edge.  The following
rectifies this and instead refrains from using the equivalence in
the problematic cases.

PR tree-optimization/108574
* tree-ssa-sccvn.cc (visit_phi): Instead of swapping
sameval and def, ignore the equivalence if there's the
danger of oscillating between two values.

* gcc.dg/torture/pr108574-1.c: New testcase.
* gcc.dg/torture/pr108574-2.c: Likewise.
* gcc.dg/torture/pr108574-3.c: Likewise.

21 months agoriscv: Enable -fasynchronous-unwind-tables by default on Linux
Andreas Schwab [Wed, 25 Jan 2023 11:00:09 +0000 (12:00 +0100)]
riscv: Enable -fasynchronous-unwind-tables by default on Linux

This follows the example of aarch64.

gcc/:
* common/config/riscv/riscv-common.cc
(riscv_option_optimization_table)
[TARGET_DEFAULT_ASYNC_UNWIND_TABLES]: Enable
-fasynchronous-unwind-tables and -funwind-tables.
* config.gcc (riscv*-*-linux*): Define
TARGET_DEFAULT_ASYNC_UNWIND_TABLES.

21 months agoSet CROSS_SYSTEM_HEADER_DIR according includedir
YunQiang Su [Fri, 6 Jan 2023 10:15:28 +0000 (18:15 +0800)]
Set CROSS_SYSTEM_HEADER_DIR according includedir

For cross building with option:
   --sysroot=/ --prefix=/usr --includedir=/usr/<triple>
just like Debian does, fixinc.sh will use the wrong header files
from /usr/include.

gcc/
* Makefile.in (CROSS_SYSTEM_HEADER_DIR): set according the
value of includedir.

21 months agoipa/108511 - relax assert for undefined local statics
Richard Biener [Mon, 30 Jan 2023 07:20:03 +0000 (08:20 +0100)]
ipa/108511 - relax assert for undefined local statics

Since we no longer promote undefined local statics extern the
assert in possibly_call_in_translation_unit_p triggers.  The
following relaxes it according to Honzas advice.

PR ipa/108511
* cgraph.cc (possibly_call_in_translation_unit_p): Relax
assert.

21 months agoChange AVX512FP16 to AVX512-FP16 in the document.
liuhongt [Sat, 28 Jan 2023 14:27:17 +0000 (22:27 +0800)]
Change AVX512FP16 to AVX512-FP16 in the document.

The official name is AVX512-FP16.

gcc/ChangeLog:

* config/i386/i386.opt: Change AVX512FP16 to AVX512-FP16.
* doc/invoke.texi: Ditto.

21 months agoDaily bump.
GCC Administrator [Mon, 30 Jan 2023 00:16:41 +0000 (00:16 +0000)]
Daily bump.

21 months agofortran: Set name for *LOC default BACK argument [PR108450]
Mikael Morin [Sun, 29 Jan 2023 20:57:24 +0000 (21:57 +0100)]
fortran: Set name for *LOC default BACK argument [PR108450]

This change fixes an ICE caused by the double resolution of MINLOC,
MAXLOC and FINDLOC expressions which get a default value for the BACK
argument at resolution time.  That argument  is added without name,
and argument reordering code is not prepared to handle unnamed arguments
coming after named ones, so the second resolution causes a NULL pointer
dereference.
The problem is fixed by explicitly setting the argument name.

PR fortran/108450

gcc/fortran/ChangeLog:

* check.cc (gfc_check_minloc_maxloc): Explicitly set argument name.
(gfc_check_findloc): Ditto.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/minmaxloc_1.f90: New test.

21 months agoICE in gfc_free_namespace. ice-on-invalid.
Jerry DeLisle [Sun, 29 Jan 2023 04:00:34 +0000 (20:00 -0800)]
ICE in gfc_free_namespace. ice-on-invalid.

PR fortran/103506

gcc/fortran/ChangeLog:

* parse.cc (parse_module): Remove use of a bool error value
that prevented proper setting of the namespace pointer.

gcc/testsuite/ChangeLog:

* gfortran.dg/pr103506_1.f90: New test.

21 months agoFix find_always_executed_bbs handling of infinite loops
Jan Hubicka [Sun, 29 Jan 2023 00:27:26 +0000 (01:27 +0100)]
Fix find_always_executed_bbs handling of infinite loops

This patch fixes the stupid typo info find_always_executed_bbs which made
all edges to be considered as ones closing infinite loops.  This fix has
somewhat snowballed as, comparing it to finite_function_p, I noticed a
divergence in handling of volatile asms (find_always_executed_bbs is conservative
and thinks that volatile statement may return or do EH, but it is really
impossible and elsewhere we expects that we dont) and I also noticed
a bug in handling early returns which made some loops to be missed.

I also noticed that function assumes that irreducible loops are marked in CFG
which is not necessarily true and it is easy to detect them during the walk.

Bootstrapped/regtested x86_64-linux, comitted.

gcc/ChangeLog:

2023-01-29  Jan Hubicka  <hubicka@ucw.cz>

* ipa-utils.cc: Include calls.h, cfgloop.h and cfganal.h
(stmt_may_terminate_function_p): If assuming return or EH
volatile asm is safe.
(find_always_executed_bbs): Fix handling of terminating BBS and
infinite loops; add debug output.
* tree-ssa-alias.cc (stmt_kills_ref_p): Fix debug output

gcc/testsuite/ChangeLog:

2023-01-29  Jan Hubicka  <hubicka@ucw.cz>

* gcc.dg/ipa/ipa-sra-30.c: New test.
* gcc.dg/ipa/ipa-sra-31.c: New test.
* gcc.dg/tree-ssa/modref-dse-7.c: New test.

21 months agoDaily bump.
GCC Administrator [Sun, 29 Jan 2023 00:17:14 +0000 (00:17 +0000)]
Daily bump.

21 months agoaarch64: Correct the maximum shift amount for shifted operands
Philipp Tomsich [Sat, 28 Jan 2023 23:07:09 +0000 (00:07 +0100)]
aarch64: Correct the maximum shift amount for shifted operands

The aarch64 ISA specification allows a left shift amount to be applied
after extension in the range of 0 to 4 (encoded in the imm3 field).

This is true for at least the following instructions:

 * ADD (extend register)
 * ADDS (extended register)
 * SUB (extended register)

The result of this patch can be seen, when compiling the following code:

uint64_t myadd(uint64_t a, uint64_t b)
{
    return a+(((uint8_t)b)<<4);
}

Without the patch the following sequence will be generated:

0000000000000000 <myadd>:
   0: d37c1c21  ubfiz x1, x1, #4, #8
   4: 8b000020  add x0, x1, x0
   8: d65f03c0  ret

With the patch the ubfiz will be merged into the add instruction:

0000000000000000 <myadd>:
   0: 8b211000  add x0, x0, w1, uxtb #4
   4: d65f03c0  ret

gcc/ChangeLog:

* config/aarch64/aarch64.cc (aarch64_uxt_size): fix an
off-by-one in checking the permissible shift-amount.

21 months agodoc: Update link to the AVR-Libc manual
Gerald Pfeifer [Sat, 28 Jan 2023 22:19:46 +0000 (23:19 +0100)]
doc: Update link to the AVR-Libc manual

gcc/ChangeLog:

* doc/extend.texi (Named Address Spaces): Update link to the
AVR-Libc manual.

21 months agolibstdc++: Update links in the "Contributing" manual
Gerald Pfeifer [Sat, 28 Jan 2023 22:07:01 +0000 (23:07 +0100)]
libstdc++: Update links in the "Contributing" manual

libstdc++-v3/ChangeLog:

* doc/xml/manual/appendix_contributing.xml: Adjust link to
ISO C++ standard at ANSI.
Move link to www.open-std.org to https.
* doc/html/manual/appendix_contributing.html: Regenerate.

21 months agoFortran: fix ICE in compare_bound_int [PR108527]
Harald Anlauf [Tue, 24 Jan 2023 21:36:25 +0000 (22:36 +0100)]
Fortran: fix ICE in compare_bound_int [PR108527]

gcc/fortran/ChangeLog:

PR fortran/108527
* resolve.cc (compare_bound_int): Expression to compare must be of
type INTEGER.
(compare_bound_mpz_t): Likewise.
(check_dimension): Fix comment on checks applied to array section
and clean up associated logic.

gcc/testsuite/ChangeLog:

PR fortran/108527
* gfortran.dg/pr108527.f90: New test.

Co-authored-by: Steven G. Kargl <kargl@gcc.gnu.org>
21 months agoFortran: diagnose USE associated symbols in COMMON blocks [PR108453]
Harald Anlauf [Sat, 28 Jan 2023 16:59:23 +0000 (17:59 +0100)]
Fortran: diagnose USE associated symbols in COMMON blocks [PR108453]

gcc/fortran/ChangeLog:

PR fortran/108453
* match.cc (gfc_match_common): A USE associated name shall not appear
in a COMMON block (F2018:C8121).

gcc/testsuite/ChangeLog:

PR fortran/108453
* gfortran.dg/common_27.f90: New test.

21 months agodoc: Fix markup
Gerald Pfeifer [Sat, 28 Jan 2023 18:26:46 +0000 (19:26 +0100)]
doc: Fix markup

gcc/ChangeLog:

* doc/standards.texi (Standards): Fix markup.

21 months agodoc: Update link to Objective-C book
Gerald Pfeifer [Sat, 28 Jan 2023 18:23:35 +0000 (19:23 +0100)]
doc: Update link to Objective-C book

gcc/ChangeLog:

* doc/standards.texi (Standards): Update link to Objective-C book.

21 months agoPR c/108192 - Fix test for mingw
Jonathan Yong [Wed, 11 Jan 2023 09:51:02 +0000 (09:51 +0000)]
PR c/108192 - Fix test for mingw

gcc/testsuite/ChangeLog:

PR c/108192
* g++.dg/cet-notrack-1.C: Use puts instead of printf,
so function call is not mangled by __mingw_printf when
doing assembly symbol inspection.

Signed-off-by: Jonathan Yong <10walls@gmail.com>
21 months agoPR c/108150 - Fix alignment test for Windows targets
Jonathan Yong [Sun, 8 Jan 2023 01:28:34 +0000 (01:28 +0000)]
PR c/108150 - Fix alignment test for Windows targets

gcc/testsuite/ChangeLog:

PR c/108150
* gcc.dg/attr-aligned.c: Make errors emitted on Windows
target same as on Linux.

Signed-off-by: Jonathan Yong <10walls@gmail.com>
21 months agoModula-2: Claim Wreturn-type in lang.opt.
Iain Sandoe [Sat, 28 Jan 2023 09:38:23 +0000 (09:38 +0000)]
Modula-2: Claim Wreturn-type in lang.opt.

Modula-2 handles this warning so that we need to claim it in the
language options.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/m2/ChangeLog:

* lang.opt: Claim Wreturn-type.

21 months agolibstdc++: Move sourceforge.net links to https
Gerald Pfeifer [Sat, 28 Jan 2023 10:49:13 +0000 (11:49 +0100)]
libstdc++: Move sourceforge.net links to https

libstdc++-v3/ChangeLog:

* doc/xml/manual/documentation_hacking.xml: Move sourceforge.net
links to https.
* doc/html/manual/documentation_hacking.html: Regenerate.

21 months agodoc: Update reference to AddressSanitizer
Gerald Pfeifer [Sat, 28 Jan 2023 10:35:40 +0000 (11:35 +0100)]
doc: Update reference to AddressSanitizer

gcc/ChangeLog:

* doc/invoke.texi (Instrumentation Options): Update reference to
AddressSanitizer.

21 months agodoc: Update Go1 link
Gerald Pfeifer [Sat, 28 Jan 2023 10:23:16 +0000 (11:23 +0100)]
doc: Update Go1 link

gcc/ChangeLog:

* doc/standards.texi: Update Go1 link.

21 months agolibstdc++: Switch www.open-std.org to https (ABI manual)
Gerald Pfeifer [Sat, 28 Jan 2023 10:12:57 +0000 (11:12 +0100)]
libstdc++: Switch open-std.org to https (ABI manual)

libstdc++-v3/ChangeLog:

* doc/xml/manual/abi.xml: Update www.open-std.org link to https.
* doc/html/manual/abi.html: Regenerate.

21 months agoRISC-V: Add vlse/vsse C/C++ intrinsic testcases
Ju-Zhe Zhong [Fri, 27 Jan 2023 23:27:57 +0000 (07:27 +0800)]
RISC-V: Add vlse/vsse C/C++ intrinsic testcases

gcc/testsuite/ChangeLog:

* g++.target/riscv/rvv/base/vlse-1.C: New test.
* g++.target/riscv/rvv/base/vlse_tu-1.C: New test.
* g++.target/riscv/rvv/base/vlse_tum-1.C: New test.
* g++.target/riscv/rvv/base/vlse_tumu-1.C: New test.
* g++.target/riscv/rvv/base/vsse-1.C: New test.
* gcc.target/riscv/rvv/base/vlse-1.c: New test.
* gcc.target/riscv/rvv/base/vlse-2.c: New test.
* gcc.target/riscv/rvv/base/vlse-3.c: New test.
* gcc.target/riscv/rvv/base/vlse-vsse-constraint-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_m-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_m-2.c: New test.
* gcc.target/riscv/rvv/base/vlse_m-3.c: New test.
* gcc.target/riscv/rvv/base/vlse_mu-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_mu-2.c: New test.
* gcc.target/riscv/rvv/base/vlse_mu-3.c: New test.
* gcc.target/riscv/rvv/base/vlse_tu-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_tu-2.c: New test.
* gcc.target/riscv/rvv/base/vlse_tu-3.c: New test.
* gcc.target/riscv/rvv/base/vlse_tum-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_tum-2.c: New test.
* gcc.target/riscv/rvv/base/vlse_tum-3.c: New test.
* gcc.target/riscv/rvv/base/vlse_tumu-1.c: New test.
* gcc.target/riscv/rvv/base/vlse_tumu-2.c: New test.
* gcc.target/riscv/rvv/base/vlse_tumu-3.c: New test.
* gcc.target/riscv/rvv/base/vsse-1.c: New test.
* gcc.target/riscv/rvv/base/vsse-2.c: New test.
* gcc.target/riscv/rvv/base/vsse-3.c: New test.
* gcc.target/riscv/rvv/base/vsse_m-1.c: New test.
* gcc.target/riscv/rvv/base/vsse_m-2.c: New test.
* gcc.target/riscv/rvv/base/vsse_m-3.c: New test.

21 months agoRISC-V: Add vlse/vsse intrinsics support
Ju-Zhe Zhong [Fri, 27 Jan 2023 23:26:08 +0000 (07:26 +0800)]
RISC-V: Add vlse/vsse intrinsics support

gcc/ChangeLog:

* config/riscv/predicates.md (pmode_reg_or_0_operand): New predicate.
* config/riscv/riscv-vector-builtins-bases.cc (class loadstore):
Support vlse/vsse.
(BASE): Ditto.
* config/riscv/riscv-vector-builtins-bases.h: Ditto.
* config/riscv/riscv-vector-builtins-functions.def (vlse): New class.
(vsse): New class.
* config/riscv/riscv-vector-builtins.cc
(function_expander::use_contiguous_load_insn): Support vlse/vsse.
* config/riscv/vector.md (@pred_strided_load<mode>): New md pattern.
(@pred_strided_store<mode>): Ditto.

21 months agoRISC-V: Remove redundant attributes [NFC]
Ju-Zhe Zhong [Fri, 27 Jan 2023 22:57:14 +0000 (06:57 +0800)]
RISC-V: Remove redundant attributes [NFC]

gcc/ChangeLog:

* config/riscv/vector.md (tail_policy_op_idx): Remove.
(mask_policy_op_idx): Remove.
(avl_type_op_idx): Remove.

21 months agoFortran tests: Revise line end tests allowing windows testing.
Jerry DeLisle [Sat, 21 Jan 2023 23:47:19 +0000 (15:47 -0800)]
Fortran tests: Revise line end tests allowing windows testing.

gcc/testsuite/ChangeLog:

* gfortran.dg/ISO_Fortran_binding_17.f90: Replace (\n|\r\n|\r)
with (\r*\n+).
* gfortran.dg/array_temporaries_2.f90: Likewise.
* gfortran.dg/bind-c-contiguous-1.f90: Likewise.
* gfortran.dg/bind-c-contiguous-4.f90: Likewise.
* gfortran.dg/bind-c-contiguous-5.f90: Likewise.
* gfortran.dg/fmt_error_4.f90: Likewise.
* gfortran.dg/fmt_error_5.f90: Likewise.
* gfortran.dg/fmt_float.f90: Likewise.
* gfortran.dg/fmt_l.f90: Likewise.
* gfortran.dg/fmt_nonchar_2.f90: Likewise.
* gfortran.dg/fmt_zero_precision.f90: Likewise.
* gfortran.dg/g77/f77-edit-apostrophe-out.f: Likewise.
* gfortran.dg/g77/f77-edit-colon-out.f: Likewise.
* gfortran.dg/g77/f77-edit-h-out.f: Likewise.
* gfortran.dg/g77/f77-edit-i-out.f: Likewise.
* gfortran.dg/g77/f77-edit-s-out.f: Likewise.
* gfortran.dg/g77/f77-edit-slash-out.f: Likewise.
* gfortran.dg/g77/f77-edit-t-out.f: Likewise.
* gfortran.dg/g77/f77-edit-x-out.f: Likewise.
* gfortran.dg/namelist_40.f90: Likewise.
* gfortran.dg/namelist_47.f90: Likewise.
* gfortran.dg/namelist_print_1.f: Likewise.
* gfortran.dg/parameter_array_dummy.f90: Likewise.

21 months agoDaily bump.
GCC Administrator [Sat, 28 Jan 2023 00:16:39 +0000 (00:16 +0000)]
Daily bump.

21 months agoPR-108557 Stuck compilation for empty file
Gaius Mulley [Fri, 27 Jan 2023 22:18:46 +0000 (22:18 +0000)]
PR-108557 Stuck compilation for empty file

Trying to compile an empty file causes cc1gm2 to hang.
The bug occurs when M2LexBuf.mod calls m2flex.GetToken after
an eof token has been seen which results in m2flex attempting
to read from stdin.  The bug fix detects eof per file and
blocks subsequent calls to m2flex.GetToken.

gcc/m2/ChangeLog:

* gm2-compiler/M2Comp.mod:  Import MetaString0.
(ExamineCompilationUnit): New variable Message.
Create and format error string.
* gm2-compiler/M2LexBuf.mod: New variable SeenEof.
(GetNonEofToken): New procedure.
(Init): Set SeenEof to FALSE.
(GetToken): Use GetNonEofToken instead of calls to
m2flex.GetToken and GetToken.
(AddTok): Detect eoftok and set SeenEof.

gcc/testsuite/ChangeLog:

* gm2/pim/fail/empty.mod: New test.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agoc: Disallow braces around C2x auto initializers
Joseph Myers [Fri, 27 Jan 2023 21:38:57 +0000 (21:38 +0000)]
c: Disallow braces around C2x auto initializers

WG14 agreed at this week's meeting to remove support for braces around
auto scalar initializers, as incompatible with C++ auto handling of
braced initializers; thus remove that support in GCC.

Bootstrapped with no regressions for x86_64-pc-linux-gnu.

gcc/c/
* c-parser.cc (c_parser_declaration_or_fndef): Do not allow braces
around auto initializer.

gcc/testsuite/
* gcc.dg/c2x-auto-1.c, gcc.dg/c2x-auto-3.c: Expect braces around
auto initializers to be disallowed.

21 months agolibstdc++: Fix up FAIL in 17_intro/names.cc on glibc < 2.19 [PR108568]
Jakub Jelinek [Fri, 27 Jan 2023 19:11:20 +0000 (20:11 +0100)]
libstdc++: Fix up FAIL in 17_intro/names.cc on glibc < 2.19 [PR108568]

On gcc112 which has glibc 2.17 I've noticed
FAIL: 17_intro/names.cc (test for excess errors)
FAIL: experimental/names.cc (test for excess errors)
These are because glibc < 2.19 used __unused as field member of various structs,
including mcontext_t in sys/ucontext.h on ppc64le.
This was changed in glibc with
https://gcc.gnu.org/pipermail/libc-alpha/2013-November/045766.html
names.cc even has
 #ifdef __GLIBC_PREREQ
 #if ! __GLIBC_PREREQ(2, 19)
 // Glibc defines this prior to 2.19
 #undef __unused
 #endif
 #endif
for it, but it doesn't work.  The reason is that __GLIBC_PREREQ is defined in
<features.h> but nothing included that header before this spot (it is included later
from bits/stdc++.h).

The following patch on Linux/Hurd conditionally includes features.h to get
the needed macros before deciding if __unused should be undefined or not.
If needed, I could use __GLIBC_PREREQ then but would need to check if it is
defined and between 1996 and 1999 it wasn't.

2023-01-27  Jakub Jelinek  <jakub@redhat.com>

PR libstdc++/108568
* testsuite/17_intro/names.cc (__unused): For linux or GNU hurd
include features.h if present and then check __GLIBC__ and
__GLIBC_MINOR__ macros for glibc prior to 2.19, instead of testing
__GLIBC_PREREQ which isn't defined yet.

21 months agotestsuite: Two adjustments to gcc.dg/vect/complex
Richard Sandiford [Fri, 27 Jan 2023 17:04:28 +0000 (17:04 +0000)]
testsuite: Two adjustments to gcc.dg/vect/complex

fast-math-bb-slp-complex-add-pattern-half-float.c no longer fails.
The scans in (loop test) fast-math-complex-add-half-float.c were
marked UNRESOLVED because they scanned slp1 rather than vect.

gcc/testsuite/
* gcc.dg/vect/complex/fast-math-bb-slp-complex-add-pattern-half-float.c:
Remove XFAIL.
* gcc.dg/vect/complex/fast-math-complex-add-half-float.c: Fix names
of dump files.

21 months agoaarch64: Prevent simd tests from being optimised away
Richard Sandiford [Fri, 27 Jan 2023 17:04:28 +0000 (17:04 +0000)]
aarch64: Prevent simd tests from being optimised away

The vqdml[as]l[hs]_laneq_* tests were folded at compile time, meaning
that we didn't have any Advanced SIMD instructions in the assembly.
Kyrill's preference was to use wrapper functions, so this patch does
that for the failing tests and for others that had scan-assemblers
with inline intrinsics calls.  (There were some tests that already
used wrapper functions, some that used volatile, some that used
inline asm barriers, and some that had no separation.)

Doing that for vqdmulhs_lane_s32.c meant that we generated the scalar
form of the instruction, rather than a vector instruction operating
on lane 0.  That seems fair enough, so the patch keeps that test but
adds a second one for lane 1.

gcc/testsuite/
* gcc.target/aarch64/simd/vfma_f64.c: Use a wrapper function
rather than an asm barrier.
* gcc.target/aarch64/simd/vfms_f64.c: Likewise.
* gcc.target/aarch64/simd/vmul_f64_1.c: Use a wrapper function
rather than volatile.
* gcc.target/aarch64/simd/vmul_n_f64_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmlalh_laneq_s16_1.c: Use a wrapper
function.  Remove -fno-inline.
* gcc.target/aarch64/simd/vqdmlals_laneq_s32_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmlslh_laneq_s16_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmlsls_laneq_s32_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmulhh_lane_s16.c: Likewise.
* gcc.target/aarch64/simd/vqdmulhh_laneq_s16_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmulhs_laneq_s32_1.c: Likewise.
* gcc.target/aarch64/simd/vqrdmulhh_lane_s16.c: Likewise.
* gcc.target/aarch64/simd/vqrdmulhh_laneq_s16_1.c: Likewise.
* gcc.target/aarch64/simd/vqrdmulhs_lane_s32.c: Likewise.
* gcc.target/aarch64/simd/vqrdmulhs_laneq_s32_1.c: Likewise.
* gcc.target/aarch64/simd/vqdmulhs_lane_s32.c: Likewise.
Allow the scalar form to be used when operating on lane 0.
Add a test for lane 1.

21 months agoAdd support for conditional xorsign [PR96373]
Richard Sandiford [Fri, 27 Jan 2023 17:03:51 +0000 (17:03 +0000)]
Add support for conditional xorsign [PR96373]

This patch is an optimisation, but it's also a prerequisite for
fixing PR96373 without regressing vect-xorsign_exec.c.

Currently the vectoriser vectorises:

  for (i = 0; i < N; i++)
    r[i] = a[i] * __builtin_copysignf (1.0f, b[i]);

as two unconditional operations (copysign and mult).
tree-ssa-math-opts.cc later combines them into an "xorsign" function.
This works for both Advanced SIMD and SVE.

However, with the fix for PR96373, the vectoriser will instead
generate a conditional multiplication (IFN_COND_MUL).  Something then
needs to fold copysign & IFN_COND_MUL to the equivalent of a conditional
xorsign.  Three obvious options were:

(1) Extend tree-ssa-math-opts.cc.
(2) Do the fold in match.pd.
(3) Leave it to rtl combine.

I'm against (3), because this isn't a target-specific optimisation.
(1) would be possible, but would involve open-coding a lot of what
match.pd does for us.  And, in contrast to doing the current
tree-ssa-math-opts.cc optimisation in match.pd, there should be
no danger of (2) happening too early.  If we have an IFN_COND_MUL
then we're already past the stage of simplifying the original
source code.

There was also a choice between adding a conditional xorsign ifn
and simply open-coding the xorsign.  The latter seems simpler,
and means less boiler-plate for target-specific code.

The signed_or_unsigned_type_for change is needed to make sure
that we stay in "SVE space" when doing the optimisation on 128-bit
fixed-length SVE.

gcc/
PR tree-optimization/96373
* tree.h (sign_mask_for): Declare.
* tree.cc (sign_mask_for): New function.
(signed_or_unsigned_type_for): For vector types, try to use the
related_int_vector_mode.
* genmatch.cc (commutative_op): Handle conditional internal functions.
* match.pd: Fold an IFN_COND_MUL+copysign into an IFN_COND_XOR+and.

gcc/testsuite/
PR tree-optimization/96373
* gcc.target/aarch64/sve/cond_xorsign_1.c: New test.
* gcc.target/aarch64/sve/cond_xorsign_2.c: Likewise.

21 months agovect/aarch64: Fix various sve/cond*.c failures
Richard Sandiford [Fri, 27 Jan 2023 17:03:50 +0000 (17:03 +0000)]
vect/aarch64: Fix various sve/cond*.c failures

Quite a few gcc.target/aarch64/sve/cond*.c tests started failing
after g:68e0063397ba820e71adc220b2da0581dce29ffa, but it turns out
that we were cheating passes before the patch.

The tests involve comparing the cost of N wide compares, a pack
sequence, and a narrow COND_EXPR with the cost of a single COND_EXPR
on fewer elements.  The costs for the former included all operations,
but the costs for the latter didn't model the comparison embedded in
the COND_EXPR.  The patch made us include the comparison on both sides,
making it apples-for-apples, but that's enough to tip the balance in
favour of using the wider types.

I think the new choice does reflect the current SVE cost model
correctly.  (Whether and how the model should be tweaked is a
different question.)  This patch therefore changes the tuning
vector length to one that makes the choice more obvious.

That in turn needs a tweak to compare_inside_loop_cost.
The function compares body_cost1/vf1 with body_cost2/vf2,
but for fully-amsked loops, it limits vf to the actual number
of iterations.  This is so that (say) an expensive 16-element
vector body doesn't win over a cheaper 8-element vector body
when there are only 7 elements to process.

However, the limit was applied using known_le, regardless of
the tuning target.  For a heuristic like this, it seems better
to use the likely minimum (which is a concept that was only
added after this code went in).

g:68e0063397ba820e71adc220b2da0581dce29ffa also fixed
vcond_4_costly.c.

gcc/
* tree-vectorizer.cc (vector_costs::compare_inside_loop_cost):
Use the likely minimum VF when bounding the denominators to
the estimated number of iterations.

gcc/testsuite/
* gcc.target/aarch64/sve/cond_asrd_1.c: Tune for a 256-bit
vector length.
* gcc.target/aarch64/sve/cond_cnot_4.c: Likewise.
* gcc.target/aarch64/sve/cond_cnot_6.c: Likewise.
* gcc.target/aarch64/sve/cond_unary_5.c: Likewise.
* gcc.target/aarch64/sve/cond_unary_6.c: Likewise.
* gcc.target/aarch64/sve/cond_uxt_5.c: Likewise.
* gcc.target/aarch64/sve/vcond_4_costly.c: Remove XFAILs.

21 months agoTidy up to declarations allowing files to be built by gm2
Gaius Mulley [Fri, 27 Jan 2023 16:38:29 +0000 (16:38 +0000)]
Tidy up to declarations allowing files to be built by gm2

This patch adds missing declarations in export qualified lists
and fixes comparisons of an address type against NIL.
These changes allow make m2/stage2/cc1gm2 to succeed when in
maintainer mode.

gcc/m2/ChangeLog:

* gm2-compiler/M2Options.def: Export GetMQ, SetMQ.
* gm2-compiler/M2Preprocess.mod: (MakeSaveTempsFileName):
Test NewDir against NIL.  Test Dumpdir against NIL.
Test GetMD () against NIL.  Test GetMMD () against NIL.
Test GetMQ () against NIL.  Test GetObj () against NIL.
Test tempfile against NIL.
* gm2-compiler/P2SymBuild.def: Export
BuildNoReturnAttribute.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agolibstdc++: Use constant for name of tzdata file
Jonathan Wakely [Fri, 27 Jan 2023 11:29:21 +0000 (11:29 +0000)]
libstdc++: Use constant for name of tzdata file

There's a string_view with this filename, which should have been used
instead of a string literal.

libstdc++-v3/ChangeLog:

* src/c++20/tzdb.cc (tzdata_stream): Use constant instead of
string literal.

21 months agolibstdc++: Use dg-bogus in new test [PR108554]
Jonathan Wakely [Fri, 27 Jan 2023 11:28:37 +0000 (11:28 +0000)]
libstdc++: Use dg-bogus in new test [PR108554]

libstdc++-v3/ChangeLog:

PR libstdc++/108554
* testsuite/23_containers/map/modifiers/108554.cc: Use dg-bogus.

21 months agoClarify -shared effect on crtfastmath.o
Richard Biener [Fri, 13 Jan 2023 07:57:12 +0000 (08:57 +0100)]
Clarify -shared effect on crtfastmath.o

This rewords the note to not specifically mention crtfastmath.o
but FP environment altering by -ffast-math or -Ofast.

PR target/55522
* doc/invoke.texi (-shared): Clarify effect on -ffast-math
and -Ofast FP environment side-effects.

21 months agomips: Don't add crtfastmath.o for -shared
Richard Biener [Fri, 13 Jan 2023 07:53:44 +0000 (08:53 +0100)]
mips: Don't add crtfastmath.o for -shared

Don't add crtfastmath.o for -shared to avoid altering the FP
environment when loading a shared library.

PR target/55522
* config/mips/gnu-user.h (GNU_USER_TARGET_MATHFILE_SPEC):
Don't add crtfastmath.o for -shared.

21 months agoia64: Don't add crtfastmath.o for -shared
Richard Biener [Fri, 13 Jan 2023 07:52:07 +0000 (08:52 +0100)]
ia64: Don't add crtfastmath.o for -shared

Don't add crtfastmath.o for -shared to avoid altering the FP
environment when loading a shared library.

PR target/55522
* config/ia64/linux.h (ENDFILE_SPEC): Don't add crtfastmath.o
for -shared.

21 months agoalpha: Don't add crtfastmath.o for -shared
Richard Biener [Fri, 13 Jan 2023 07:50:14 +0000 (08:50 +0100)]
alpha: Don't add crtfastmath.o for -shared

Don't add crtfastmath.o for -shared to avoid altering the FP
environment when loading a shared library.

PR target/55522
* config/alpha/linux.h (ENDFILE_SPEC): Don't add
crtfastmath.o for -shared.

21 months agoCorrectly detect shifts out of range
Andrew MacLeod [Mon, 16 Jan 2023 20:02:51 +0000 (15:02 -0500)]
Correctly detect shifts out of range

get_shift_range was incorrectly communicating that it couldn't calculate
a range when the shift values was always out fo range.  Fix this and
alwasy return [0, 0] when the shift value is always out of range.

PR tree-optimization/108306
gcc/
* range-op.cc (operator_lshift::fold_range): Return [0, 0] not
varying for shifts that are always out of void range.
(operator_rshift::fold_range): Return [0, 0] not
varying for shifts that are always out of void range.

gcc/testsuite/
* gcc.dg/pr108306.c: New.

21 months agoDo not try to logical fold floating point relations.
Andrew MacLeod [Wed, 25 Jan 2023 16:13:23 +0000 (11:13 -0500)]
Do not try to logical fold floating point relations.

relation_fold_and_or looks for relations among common operands feeding
logical ands and ors.  With no knowledge of NANs, it should not attempt
to do this with floating point ssa names.

PR tree-optimization/108447
gcc/
* gimple-range-fold.cc (old_using_range::relation_fold_and_or):
Do not attempt to fold HONOR_NAN types.

gcc/testsuite/
* gcc.dg/pr108447.c: New.

21 months agoRISC-V: Fix testcases check.
Ju-Zhe Zhong [Fri, 27 Jan 2023 12:30:20 +0000 (20:30 +0800)]
RISC-V: Fix testcases check.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/avl_multiple-7.c: Fix testcase check.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-8.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vsetvl-18.c: Ditto.

21 months agoRISC-V: Add vle/vse C++ overloaded API intrinsic testcases
Ju-Zhe Zhong [Fri, 20 Jan 2023 02:24:34 +0000 (10:24 +0800)]
RISC-V: Add vle/vse C++ overloaded API intrinsic testcases

gcc/testsuite/ChangeLog:

* g++.target/riscv/rvv/base/vle-1.C: New test.
* g++.target/riscv/rvv/base/vle_tu-1.C: New test.
* g++.target/riscv/rvv/base/vle_tum-1.C: New test.
* g++.target/riscv/rvv/base/vle_tumu-1.C: New test.
* g++.target/riscv/rvv/base/vse-1.C: New test.
* g++.target/riscv/rvv/base/riscv_vector.h: New.

21 months agoRISC-V: Fix vop_m overloaded C++ API name.
Ju-Zhe Zhong [Fri, 20 Jan 2023 02:20:29 +0000 (10:20 +0800)]
RISC-V: Fix vop_m overloaded C++ API name.

According to https://github.com/riscv-non-isa/rvv-intrinsic-doc/tree/master/
For "vop_m" intrinsics, C++ overloaded API does not have "_m" suffix.

gcc/ChangeLog:

* config/riscv/riscv-vector-builtins-shapes.cc (struct loadstore_def):
Remove _m suffix for "vop_m" C++ overloaded API name.

21 months agoRISC-V: Add vse.v C API intrinsics testcases
Ju-Zhe Zhong [Thu, 19 Jan 2023 14:31:08 +0000 (22:31 +0800)]
RISC-V: Add vse.v C API intrinsics testcases

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/vse-1.c: New test.
* gcc.target/riscv/rvv/base/vse-2.c: New test.
* gcc.target/riscv/rvv/base/vse-3.c: New test.
* gcc.target/riscv/rvv/base/vse_m-1.c: New test.
* gcc.target/riscv/rvv/base/vse_m-2.c: New test.
* gcc.target/riscv/rvv/base/vse_m-3.c: New test.

21 months agoRISC-V: Add vle.v C API intrinsics testcases
Ju-Zhe Zhong [Thu, 19 Jan 2023 14:12:49 +0000 (22:12 +0800)]
RISC-V: Add vle.v C API intrinsics testcases

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/vle-1.c: New test.
* gcc.target/riscv/rvv/base/vle-2.c: New test.
* gcc.target/riscv/rvv/base/vle-3.c: New test.
* gcc.target/riscv/rvv/base/vle_m-1.c: New test.
* gcc.target/riscv/rvv/base/vle_m-2.c: New test.
* gcc.target/riscv/rvv/base/vle_m-3.c: New test.
* gcc.target/riscv/rvv/base/vle_mu-1.c: New test.
* gcc.target/riscv/rvv/base/vle_mu-2.c: New test.
* gcc.target/riscv/rvv/base/vle_mu-3.c: New test.
* gcc.target/riscv/rvv/base/vle_tu-1.c: New test.
* gcc.target/riscv/rvv/base/vle_tu-2.c: New test.
* gcc.target/riscv/rvv/base/vle_tu-3.c: New test.
* gcc.target/riscv/rvv/base/vle_tum-1.c: New test.
* gcc.target/riscv/rvv/base/vle_tum-2.c: New test.
* gcc.target/riscv/rvv/base/vle_tum-3.c: New test.
* gcc.target/riscv/rvv/base/vle_tumu-1.c: New test.
* gcc.target/riscv/rvv/base/vle_tumu-2.c: New test.
* gcc.target/riscv/rvv/base/vle_tumu-3.c: New test.

21 months agoRISC-V: Add vlm/vsm C/C++ API intrinsics support
Ju-Zhe Zhong [Thu, 19 Jan 2023 06:07:49 +0000 (14:07 +0800)]
RISC-V: Add vlm/vsm C/C++ API intrinsics support

gcc/ChangeLog:

* config/riscv/riscv-vector-builtins-bases.cc (BASE): Add vlm/vsm support.
* config/riscv/riscv-vector-builtins-bases.h: Ditto.
* config/riscv/riscv-vector-builtins-functions.def (vlm): New define.
(vsm): Ditto.
* config/riscv/riscv-vector-builtins-shapes.cc (struct loadstore_def): Add vlm/vsm support.
* config/riscv/riscv-vector-builtins-types.def (DEF_RVV_B_OPS): Ditto.
(vbool64_t): Ditto.
(vbool32_t): Ditto.
(vbool16_t): Ditto.
(vbool8_t): Ditto.
(vbool4_t): Ditto.
(vbool2_t): Ditto.
(vbool1_t): Ditto.
* config/riscv/riscv-vector-builtins.cc (DEF_RVV_B_OPS): Ditto.
(rvv_arg_type_info::get_tree_type): Ditto.
(function_expander::use_contiguous_load_insn): Ditto.
* config/riscv/vector.md (@pred_store<mode>): Ditto.

gcc/testsuite/ChangeLog:

* g++.target/riscv/rvv/base/vsm-1.C: New test.
* g++.target/riscv/rvv/rvv.exp: New test.
* gcc.target/riscv/rvv/base/vlm_vsm-1.c: New test.
* gcc.target/riscv/rvv/base/vlm_vsm-2.c: New test.
* gcc.target/riscv/rvv/base/vlm_vsm-3.c: New test.

21 months agoRISC-V: Finalize testcases for final version VSETVL PASS.
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:29:15 +0000 (11:29 +0800)]
RISC-V: Finalize testcases for final version VSETVL PASS.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/avl_single-14.c: Adjust for final implementation.
* gcc.target/riscv/rvv/vsetvl/avl_single-23.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-30.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-44.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-47.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-50.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-51.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-6.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-65.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-66.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-67.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-68.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-70.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-71.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_single-9.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-2.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-3.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-4.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-4.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-5.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-13.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-13.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-14.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-15.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-16.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-21.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-22.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-23.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-24.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-5.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-6.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-7.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-8.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_conflict-1.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_conflict-6.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-1.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-10.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-11.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-12.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-13.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-14.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-2.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-3.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-4.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-5.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-6.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-7.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-8.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_miss_default-9.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_switch_vtype-10.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-14.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-15.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-16.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_multiple-9.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-73.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-74.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-75.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-14.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-15.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-16.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-17.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-18.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-19.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-20.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvl-9.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-14.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-15.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-16.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-17.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-18.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-19.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-20.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/vsetvlmax-9.c: New test.

21 months agoRISC-V: Finalize VSETVL PASS implementation
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:24:34 +0000 (11:24 +0800)]
RISC-V: Finalize VSETVL PASS implementation

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (vsetvl_insn_p): Add condition to avoid ICE.
(vsetvl_discard_result_insn_p): New function.
(reg_killed_by_bb_p): rename to find_reg_killed_by.
(find_reg_killed_by): New name.
(get_vl): allow it to be called by more functions.
(has_vsetvl_killed_avl_p): Add condition.
(get_avl): allow it to be called by more functions.
(insn_should_be_added_p): New function.
(get_all_nonphi_defs): Refine function.
(get_all_sets): Ditto.
(get_same_bb_set): New function.
(any_insn_in_bb_p): Ditto.
(any_set_in_bb_p): Ditto.
(get_vl_vtype_info): Add VLMAX forward optimization.
(source_equal_p): Fix issues.
(extract_single_source): Refine.
(avl_info::multiple_source_equal_p): New function.
(avl_info::operator==): Adjust for final version.
(vl_vtype_info::operator==): Ditto.
(vl_vtype_info::same_avl_p): Ditto.
(vector_insn_info::parse_insn): Ditto.
(vector_insn_info::available_p): New function.
(vector_insn_info::merge): Adjust for final version.
(vector_insn_info::dump): Add hard_empty.
(pass_vsetvl::hard_empty_block_p): New function.
(pass_vsetvl::backward_demand_fusion): Adjust for final version.
(pass_vsetvl::forward_demand_fusion): Ditto.
(pass_vsetvl::demand_fusion): Ditto.
(pass_vsetvl::cleanup_illegal_dirty_blocks): New function.
(pass_vsetvl::compute_local_properties): Adjust for final version.
(pass_vsetvl::can_refine_vsetvl_p): Ditto.
(pass_vsetvl::refine_vsetvls): Ditto.
(pass_vsetvl::commit_vsetvls): Ditto.
(pass_vsetvl::propagate_avl): New function.
(pass_vsetvl::lazy_vsetvl): Adjust for new version.
* config/riscv/riscv-vsetvl.h (enum def_type): New enum.

21 months agotestsuite: Use noipa attribute for pr95115 test
Xi Ruoyao [Fri, 27 Jan 2023 11:03:05 +0000 (19:03 +0800)]
testsuite: Use noipa attribute for pr95115 test

This prevent the compiler from deeming the NaN result "unused" and
remove the calculation raising INVALID exception. See the discussion
in PR107608 for details.

gcc/testsuite/ChangeLog:

* gcc.dg/pr95115.c (x): Add noipa attribute.

21 months agoOpenMP/Fortran: Fix has_device_addr clause splitting [PR108558]
Tobias Burnus [Fri, 27 Jan 2023 10:32:19 +0000 (11:32 +0100)]
OpenMP/Fortran: Fix has_device_addr clause splitting [PR108558]

gcc/fortran/ChangeLog:

PR fortran/108558
* trans-openmp.cc (gfc_split_omp_clauses): Handle has_device_addr.

libgomp/ChangeLog:

PR fortran/108558
* testsuite/libgomp.fortran/has_device_addr.f90: New test.

21 months agodoc: Fix up return type of __builtin_va_arg_pack_len [PR108560]
Jakub Jelinek [Fri, 27 Jan 2023 10:17:35 +0000 (11:17 +0100)]
doc: Fix up return type of __builtin_va_arg_pack_len [PR108560]

__builtin_va_arg_pack_len as implemented returned int since its introduction
in 2007.  The initial documentation didn't mention any return type,
which changed in 2010 in r0-103077-gab940b73bfabe2cec4 during some
documentation formatting cleanups
https://gcc.gnu.org/legacy-ml/gcc-patches/2010-09/msg01632.html
I can understand that for formatting some type was needed there
but what exactly hasn't been really discussed.

So, I think we should change documentation to match the implementation,
rather than change implementation to match the documentation.
Most people don't use more than 2147483647 arguments to inline functions,
and on poor targets with 16-bit ints I bet even having more than 65535
arguments to inline functions would be highly unexpected.

2023-01-27  Jakub Jelinek  <jakub@redhat.com>

PR other/108560
* doc/extend.texi: Fix up return type of __builtin_va_arg_pack_len
from size_t to int.

21 months agocgraph: Adjust verify_corresponds_to_fndecl [PR106061]
Jakub Jelinek [Fri, 27 Jan 2023 10:16:43 +0000 (11:16 +0100)]
cgraph: Adjust verify_corresponds_to_fndecl [PR106061]

IPA passes redirect some calls in what it determines to be unreachable code
to builtin_decl_unreachable.  But that function returns sometimes
builtin_decl_explicit (BUILT_IN_UNREACHABLE) (which was what GCC 12
and earlier did always), or builtin_decl_explicit (BUILT_IN_TRAP)
(e.g. for -funreachable-traps, -O0, -Og).
Now the cgraph verification code has a code to verify cgraph edges
and has there an exception for these redirections to BUILT_IN_UNREACHABLE,
but doesn't have for BUILT_IN_TRAP, so e.g. the following testcase
ICEs during that verification.

The following patch just adds BUILT_IN_TRAP to those exceptions.

2023-01-27  Jakub Jelinek  <jakub@redhat.com>

PR ipa/106061
* cgraph.cc (cgraph_edge::verify_corresponds_to_fndecl): Allow
redirection of calls to __builtin_trap in addition to redirection
to __builtin_unreachable.

* gcc.dg/pr106061.c: New test.

21 months agoRISC-V: Fix bug of before_p function
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:09:21 +0000 (11:09 +0800)]
RISC-V: Fix bug of before_p function

compare_with will return other than -1, so it should check less than 0
rather than check exactly with -1.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (before_p): Fix bug.

21 months agoRISC-V: Refine function args of some functions.
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:13:05 +0000 (11:13 +0800)]
RISC-V: Refine function args of some functions.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (gen_vsetvl_pat): Refine function args.
(emit_vsetvl_insn): Ditto.

21 months agoRISC-V: Fix pred_mov constraint for vle.v
Ju-Zhe Zhong [Thu, 19 Jan 2023 07:02:58 +0000 (15:02 +0800)]
RISC-V: Fix pred_mov constraint for vle.v

The original constraint is incorrect in pred_mov pattern.
Take a look at Alternative 2, the operands[0] is "vr",
operands[1] which is mask operand can be "vm".
Such alternative matching will give the wrong codegen (vle.v v0,0(a5),v0.t)
This is illegal according to RVV ISA.

To fix this issue and not destroy the RA performance, fix this pattern in
this patch.

gcc/ChangeLog:

* config/riscv/vector.md: Fix constraints.

21 months agoRISC-V: Add TARGET_MIN_VLEN > 32 into iterators of EEW = 64 vector modes
Ju-Zhe Zhong [Fri, 20 Jan 2023 09:33:09 +0000 (17:33 +0800)]
RISC-V: Add TARGET_MIN_VLEN > 32 into iterators of EEW = 64 vector modes

According to RVV ISA, RVV doesn't support EEW == 64 vector type for zve32x
and zve32f. So it makes sense add predicate in the iterators of EEW = 64
vector modes.

gcc/ChangeLog:

* config/riscv/vector-iterators.md: Add TARGET_MIN_VLEN > 32 predicates.

21 months agotree: Fix up tree_code_{length,type}
Jakub Jelinek [Fri, 27 Jan 2023 09:51:35 +0000 (10:51 +0100)]
tree: Fix up tree_code_{length,type}

On Thu, Jan 26, 2023 at 09:45:35AM -0500, Patrick Palka via Gcc-patches wrote:
> > +#define DEFTREECODE(SYM, NAME, TYPE, LENGTH) TYPE,
> > +#define END_OF_BASE_TREE_CODES tcc_exceptional,
> > +
> > +
> >  /* Class of tree given its code.  */
> > -extern const enum tree_code_class tree_code_type[];
> > +constexpr enum tree_code_class tree_code_type[] = {
> > +#include "all-tree.def"
> > +};
> > +
> > +#undef DEFTREECODE
> > +#undef END_OF_BASE_TREE_CODES
> >
> >  /* Each tree code class has an associated string representation.
> >     These must correspond to the tree_code_class entries.  */
> >  extern const char *const tree_code_class_strings[];
> >
> >  /* Number of argument-words in each kind of tree-node.  */
> > -extern const unsigned char tree_code_length[];
> > +
> > +#define DEFTREECODE(SYM, NAME, TYPE, LENGTH) LENGTH,
> > +#define END_OF_BASE_TREE_CODES 0,
> > +constexpr unsigned char tree_code_length[] = {
> > +#include "all-tree.def"
> > +};
> > +
> > +#undef DEFTREECODE
> > +#undef END_OF_BASE_TREE_CODES
>
> IIUC defining these globals as non-inline constexpr gives them internal
> linkage, and so each TU contains its own unique copy of these globals.
> This bloats cc1plus by a tiny bit and is technically an ODR violation
> because some inline functions such as tree_class_check also ODR-use
> these variables and so each defn of tree_class_check will refer to a
> "different" tree_code_class.  Since inline variables are a C++17
> feature, I guess we could fix this by defining the globals the old way
> before C++17 and as inline constexpr otherwise?

And I'd argue with the tiny bit.
In my x86_64-linux cc1plus from today, I see 193 _ZL16tree_code_length vars,
374 bytes each, and 324 _ZL14tree_code_type vars, 1496 bytes each.
So, that means waste of 555016 .rodata bytes, plus being highly non-cache
friendly.

The following patch does that.

Tested on x86_64-linux in my -O0 working tree (system gcc 12
compiler) where .rodata shrunk with the patch by 928896 bytes, in last
stage of a bootstrapped tree (built by today's prev-gcc) where .rodata
shrunk by 561728 bytes (in neither case .text or most other sections
changed sizes) and on powerpc64le-linux --disable-bootstrap
(system gcc 4.8.5) to test also the non-C++17 case plus with
fully x86_64-linux, i686-linux and powerpc64le-linux bootstraps/regtests.

BTW, wonder if tree_code_type couldn't be an array of unsigned char
elements rather than enum tree_code_class and we'd then cast it
to the enum in the macro, that would shrink that array from 1496 bytes
to 374.  Of course, that sounds like stage1 material.

2023-01-27  Patrick Palka  <ppalka@redhat.com>
    Jakub Jelinek  <jakub@redhat.com>

* tree-core.h (tree_code_type, tree_code_length): For
C++17 and later, add inline keyword, otherwise don't define
the arrays, but declare extern arrays.
* tree.cc (tree_code_type, tree_code_length): Define these
arrays for C++14 and older.

21 months agoRISC-V: Change parse_insn into public for future use.
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:06:59 +0000 (11:06 +0800)]
RISC-V: Change parse_insn into public for future use.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.h: Change it into public.

21 months agoRISC-V: Reorder VSETVL pass location
Ju-Zhe Zhong [Wed, 18 Jan 2023 03:03:47 +0000 (11:03 +0800)]
RISC-V: Reorder VSETVL pass location

Insert before dce means we don't invoke DCE by this pass itself, and
also we can leverage the effort of BB reorder.

gcc/ChangeLog:

* config/riscv/riscv-passes.def (INSERT_PASS_BEFORE): Reorder VSETVL
pass.

21 months agoRISC-V: Change VSETVL PASS always call split_all_insns
Ju-Zhe Zhong [Wed, 18 Jan 2023 02:50:14 +0000 (10:50 +0800)]
RISC-V: Change VSETVL PASS always call split_all_insns

Since LCM will destroy CFG, we are going to reorder the location of VSETVL PASS
at least before bbro (block-reorder PASS) which is before split3 PASS. We need
to call it in VSETVL PASS to get final RVV instructions patterns.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (pass_vsetvl::execute): Always call split_all_insns.

21 months agoRISC-V: Fix incorrect attributes of vsetvl instructions pattern
Ju-Zhe Zhong [Wed, 18 Jan 2023 02:44:15 +0000 (10:44 +0800)]
RISC-V: Fix incorrect attributes of vsetvl instructions pattern

gcc/ChangeLog:

* config/riscv/vector.md: Fix incorrect attributes.

21 months agoModula-2: Add claimed command line options to lang.opt [PR108555].
Iain Sandoe [Thu, 26 Jan 2023 13:49:11 +0000 (13:49 +0000)]
Modula-2: Add claimed command line options to lang.opt [PR108555].

This is a partial reversion of the changes in r13-5373-g80cf2c5e8f496b.

As observed in the PR, handling the C and Driver options in the Modula-2
lang-specific code could be difficult to emulate; This reverts to adding
the required options to the language-specific .opt file.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
PR modula2/108555
PR modula2/108182
PR modula2/102343

gcc/m2/ChangeLog:

* gm2-lang.cc (gm2_langhook_option_lang_mask): Do not claim CL_C
or CL_DRIVER.
(gm2_langhook_init_options): Handle options that we want to pass
to the preprocessor.
* lang-specs.h: Pass -B and -save-temps to regular compile lines.
* lang.opt: Add C and Driver options that Modula-2 intercepts for
internal use. Reorder options into two sections and to collate.

21 months agogomp/declare-variant-1*.f90: Update for Windows
Tobias Burnus [Fri, 27 Jan 2023 08:13:16 +0000 (09:13 +0100)]
gomp/declare-variant-1*.f90: Update for Windows

Replace target selector 'lp64' by '! ilp32' to handle
Windows which uses 32bit long (and vice versa for '! lp64').

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/declare-variant-10.f90: Update scan-tree's
target selector to handle Windows.
* gfortran.dg/gomp/declare-variant-11.f90: Likewise.
* gfortran.dg/gomp/declare-variant-12.f90: Likewise.

21 months agoLoongArch: Don't add crtfastmath.o for -shared
Richard Biener [Fri, 13 Jan 2023 08:01:12 +0000 (09:01 +0100)]
LoongArch: Don't add crtfastmath.o for -shared

Don't add crtfastmath.o for -shared to avoid altering the FP
environment when loading a shared library.

PR target/55522
* config/loongarch/gnu-user.h (GNU_USER_TARGET_MATHFILE_SPEC):
Don't add crtfastmath.o for -shared.

21 months ago[docs] note that -g opts are implicitly negatable too
Alexandre Oliva [Fri, 27 Jan 2023 01:52:07 +0000 (22:52 -0300)]
[docs] note that -g opts are implicitly negatable too

Back in 2017, I made -g* options implicitly negatable, without
realizing there was documentation that required updating.  Oops.
Fixed, at last!

for gcc/ChangeLog

* doc/options.texi (option, RejectNegative): Mention that
-g-started options are also implicitly negatable.

21 months agoDaily bump.
GCC Administrator [Fri, 27 Jan 2023 00:17:53 +0000 (00:17 +0000)]
Daily bump.

21 months agoPR-108551 gcc/m2/gm2-libs-pim/Termbase.mod:128:1 error end of non-void
Gaius Mulley [Thu, 26 Jan 2023 21:43:22 +0000 (21:43 +0000)]
PR-108551 gcc/m2/gm2-libs-pim/Termbase.mod:128:1 error end of non-void

cc1gm2 generates an error: control reaches end of non-void function when
compiling Termbase.mod if -Werror=return-type is present.

../gcc/m2/gm2-libs-pim/Termbase.mod: In function 'Termbase_KeyPressed':
../gcc/m2/gm2-libs-pim/Termbase.mod:128:1: error: control reaches end
of non-void function [-Werror=return-type]
   128 | END KeyPressed ;
       | ^~~

This occurs as cc1gm2 does skips over the <* noreturn *> attribute.  This
patch records the <* noreturn *> attribute in the m2 symbol table and
later on sets TREE_THIS_VOLATILE when creating the function decl.
The patch also contains a fix for the main scaffold which also omitted
a return 0 after the exception handler code.

gcc/m2/ChangeLog:

* gm2-compiler/M2GCCDeclare.mod: Import IsProcedureNoReturn.
(DeclareProcedureToGccWholeProgram): New variable declared and set
returnType.  Pass returnType to BuildEndFunctionDeclaration.
Extra parameter IsProcedureNoReturn passed to
BuildEndFunctionDeclaration.
* gm2-compiler/M2Quads.mod (BuildM2MainFunction): Correct
scaffold comment and add extra return 0.
* gm2-compiler/P2Build.bnf: Import BuildNoReturnAttribute.
(ProcedureHeading): Process EndBuildFormalParameters before
parsing AttributeNoReturn.
(DefProcedureHeading): Process EndBuildFormalParameters before
parsing AttributeNoReturn.
(AttributeNoReturn): Call BuildNoReturnAttribute.
* gm2-compiler/P2SymBuild.def (BuildNoReturnAttribute): New
procedure.
* gm2-compiler/P2SymBuild.mod (BuildNoReturnAttribute): New
procedure.
* gm2-compiler/SymbolTable.def (PutProcedureInline): Corrected
comment.
(PutProcedureNoReturn): New procedure.
(IsProcedureNoReturn): New procedure function.
* gm2-compiler/SymbolTable.mod (SymProcedure): IsNoReturn
new field.
(MakeProcedure): Initialize IsNoReturn to FALSE.
(PutProcedureNoReturn): New procedure.
(IsProcedureNoReturn): New procedure function.
* gm2-gcc/m2decl.cc (m2decl_BuildEndFunctionDeclaration):
Add extra parameter isnoreturn.  Set TREE_THIS_VOLATILE
to isnoreturn.
* gm2-gcc/m2decl.def (BuildEndFunctionDeclaration): Add
extra parameter isnoreturn.
* gm2-gcc/m2decl.h (m2decl_BuildEndFunctionDeclaration): Add
extra parameter isnoreturn.
* gm2-gcc/m2except.cc (m2except_InitExceptions): Change all
function decl to pass an extra parameter isnoreturn.

gcc/testsuite/ChangeLog:

* gm2/warnings/returntype/fail/badreturn.mod: New test.
* gm2/warnings/returntype/fail/warnings-returntype-fail.exp:
New test.
* gm2/warnings/returntype/pass/Termbase.mod: New test.
* gm2/warnings/returntype/pass/goodreturn.mod: New test.
* gm2/warnings/returntype/pass/keypressedsimple.mod: New test.
* gm2/warnings/returntype/pass/warnings-returntype-pass.exp:
New test.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agoFix comment so that /* does not appear inside a comment.
Gaius Mulley [Thu, 26 Jan 2023 21:41:53 +0000 (21:41 +0000)]
Fix comment so that /* does not appear inside a comment.

The block comment in m2.flex associated with splitSlashStar
contains /* which causes a warning.  This patch obfuscates
the comment /* symbols.

* m2.flex (splitSlashStar): Fix comment so that /* does not
appear inside the comment.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agoRISC-V: Use get_typenode_from_name to get fixed-width integer type nodes
Kito Cheng [Tue, 17 Jan 2023 16:14:57 +0000 (00:14 +0800)]
RISC-V: Use get_typenode_from_name to get fixed-width integer type nodes

[u]int32_t are using different type between glibc and newlib, so getting
those node by int or long type isn't portable way, I also update all
other fixed-width integer types to prevent this happened again in future :P

gcc/ChangeLog:

* config/riscv/riscv-vector-builtins.cc (register_builtin_types):
Use get_typenode_from_name to get fixed-width integer type
nodes.
* config/riscv/riscv-vector-builtins.def: Update define with
fixed-width integer type nodes.

21 months agoRISC-V: Add testcases for AVL=REG support
Ju-Zhe Zhong [Mon, 9 Jan 2023 23:38:38 +0000 (07:38 +0800)]
RISC-V: Add testcases for AVL=REG support

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/avl_single-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-20.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-21.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-22.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-23.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-24.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-25.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-26.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-27.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-28.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-29.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-30.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-31.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-32.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-33.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-34.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-35.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-36.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-37.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-38.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-39.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-40.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-41.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-42.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-43.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-44.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-45.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-46.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-47.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-48.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-49.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-50.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-51.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-52.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-53.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-54.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-55.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-56.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-57.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-58.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-59.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-60.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-61.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-62.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-63.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-64.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-65.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-66.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-67.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-68.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-69.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-14.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-15.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-16.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-17.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-18.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-19.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-70.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-71.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-9.c: New test.

21 months agoRISC-V: Add testcases for IMM (0 ~ 31) AVL
Ju-Zhe Zhong [Wed, 4 Jan 2023 13:48:48 +0000 (21:48 +0800)]
RISC-V: Add testcases for IMM (0 ~ 31) AVL

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-9.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_conflict-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-10.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-11.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-12.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-13.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-14.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-15.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-16.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-17.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_loop_invariant-9.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-1.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-2.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-4.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-5.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-6.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-7.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-8.c: New test.
* gcc.target/riscv/rvv/vsetvl/imm_switch-9.c: New test.

21 months agoRISC-V: Fix bugs of supporting AVL=REG (single-real-def) in VSETVL PASS
Ju-Zhe Zhong [Mon, 9 Jan 2023 23:29:11 +0000 (07:29 +0800)]
RISC-V: Fix bugs of supporting AVL=REG (single-real-def) in VSETVL PASS

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (same_bb_and_before_p): Remove it.
(real_insn_and_same_bb_p): New function.
(same_bb_and_after_or_equal_p): Remove it.
(before_p): New function.
(reg_killed_by_bb_p): Ditto.
(has_vsetvl_killed_avl_p): Ditto.
(get_vl): Move location so that we can call it.
(anticipatable_occurrence_p): Fix issue of AVL=REG support.
(available_occurrence_p): Ditto.
(dominate_probability_p): Remove it.
(can_backward_propagate_p): Remove it.
(get_all_nonphi_defs): New function.
(get_all_predecessors): Ditto.
(any_insn_in_bb_p): Ditto.
(insert_vsetvl): Adjust AVL REG.
(source_equal_p): New function.
(extract_single_source): Ditto.
(avl_info::single_source_equal_p): Ditto.
(avl_info::operator==): Adjust for AVL=REG.
(vl_vtype_info::same_avl_p): Ditto.
(vector_insn_info::set_demand_info): Remove it.
(vector_insn_info::compatible_p): Adjust for AVL=REG.
(vector_insn_info::compatible_avl_p): New function.
(vector_insn_info::merge): Adjust AVL=REG.
(vector_insn_info::dump): Ditto.
(pass_vsetvl::merge_successors): Remove it.
(enum fusion_type): New enum.
(pass_vsetvl::get_backward_fusion_type): New function.
(pass_vsetvl::backward_demand_fusion): Adjust for AVL=REG.
(pass_vsetvl::forward_demand_fusion): Ditto.
(pass_vsetvl::demand_fusion): Ditto.
(pass_vsetvl::prune_expressions): Ditto.
(pass_vsetvl::compute_local_properties): Ditto.
(pass_vsetvl::cleanup_vsetvls): Ditto.
(pass_vsetvl::commit_vsetvls): Ditto.
(pass_vsetvl::init): Ditto.
* config/riscv/riscv-vsetvl.h (enum fusion_type): New enum.
(enum merge_type): New enum.

21 months agoRISC-V: Add probability model of each block to prevent endless loop of Phase 3
Ju-Zhe Zhong [Mon, 9 Jan 2023 23:17:20 +0000 (07:17 +0800)]
RISC-V: Add probability model of each block to prevent endless loop of Phase 3

Notice that the PASS is just simpily pick the probability >= 50%
to do the backward fusion which will create endless loop on Phase 3.

Adding this probability to fix this bug.
gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc
(vector_infos_manager::vector_infos_manager): Add probability.
(vector_infos_manager::dump): Ditto.
(pass_vsetvl::compute_probabilities): Ditto.
* config/riscv/riscv-vsetvl.h (struct vector_block_info): Ditto.

21 months agoRISC-V: Remove dirty_pat since it is redundant
Ju-Zhe Zhong [Mon, 9 Jan 2023 23:10:59 +0000 (07:10 +0800)]
RISC-V: Remove dirty_pat since it is redundant

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (vector_insn_info::operator==): Remove dirty_pat.
(vector_insn_info::merge): Ditto.
(vector_insn_info::dump): Ditto.
(pass_vsetvl::merge_successors): Ditto.
(pass_vsetvl::backward_demand_fusion): Ditto.
(pass_vsetvl::forward_demand_fusion): Ditto.
(pass_vsetvl::commit_vsetvls): Ditto.
* config/riscv/riscv-vsetvl.h: Ditto.

21 months agoRISC-V: Rename insn into rinsn for rtx_insn * [NFC]
Ju-Zhe Zhong [Mon, 9 Jan 2023 22:56:43 +0000 (06:56 +0800)]
RISC-V: Rename insn into rinsn for rtx_insn * [NFC]

Since the PASS is implemented base on RTL_SSA framework.
According to rtl_ssa, they name insn_info * as insn and
name rtx_insn * rinsn. I follow this rule in this pass but I missed
this function. So rename it to make codes be consistent to RTL_SSA
framework.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (add_label_notes): Rename insn to
rinsn.

21 months agoRISC-V: Refine codes in backward fusion [NFC]
Ju-Zhe Zhong [Mon, 9 Jan 2023 22:47:26 +0000 (06:47 +0800)]
RISC-V: Refine codes in backward fusion [NFC]

This NFC patch is preparing for the following patches.
gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (pass_vsetvl::backward_demand_fusion): Refine codes.

21 months agoRISC-V: Avoid redundant flow in forward fusion
Ju-Zhe Zhong [Mon, 9 Jan 2023 22:40:07 +0000 (06:40 +0800)]
RISC-V: Avoid redundant flow in forward fusion

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (pass_vsetvl::forward_demand_fusion):
Add pre-check for redundant flow.

21 months agoRISC-V: Cleanup the codes of bitmap create and free [NFC]
Ju-Zhe Zhong [Mon, 9 Jan 2023 22:33:07 +0000 (06:33 +0800)]
RISC-V: Cleanup the codes of bitmap create and free [NFC]

This patch is a NFC patch to move the codes into a wrapper function so that
they can be reused. I will reuse them in the following patches.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (vector_infos_manager::create_bitmap_vectors): New function.
(vector_infos_manager::free_bitmap_vectors): Ditto.
(pass_vsetvl::pre_vsetvl): Adjust codes.
* config/riscv/riscv-vsetvl.h: New function declaration.

21 months agoRISC-V: Refine Phase 3 of VSETVL PASS
Ju-Zhe Zhong [Wed, 4 Jan 2023 13:45:26 +0000 (21:45 +0800)]
RISC-V: Refine Phase 3 of VSETVL PASS

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (can_backward_propagate_p): Fix for null iter_bb.
(vector_insn_info::set_demand_info): New function.
(pass_vsetvl::emit_local_forward_vsetvls): Adjust for refinement of Phase 3.
(pass_vsetvl::merge_successors): Ditto.
(pass_vsetvl::compute_global_backward_infos): Ditto.
(pass_vsetvl::backward_demand_fusion): Ditto.
(pass_vsetvl::forward_demand_fusion): Ditto.
(pass_vsetvl::demand_fusion): New function.
(pass_vsetvl::lazy_vsetvl): Adjust for refinement of phase 3.
* config/riscv/riscv-vsetvl.h: New function declaration.

gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-27.c: Update
testcase.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-28.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-45.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-25.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-26.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-27.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-28.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_bb_prop-3.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_conflict-7.c: Ditto.
* gcc.target/riscv/rvv/vsetvl/vlmax_switch_vtype-12.c: Ditto.

21 months agoRISC-V: Fix bugs of available condition.
Ju-Zhe Zhong [Tue, 3 Jan 2023 07:30:30 +0000 (15:30 +0800)]
RISC-V: Fix bugs of available condition.

Suppose there are 2 demand infos:

Demand 1: demand TAIL.
Demand 2: not demand TAIL.

If a block is demand 1, we should adjust this block is available both for demand 1 && 2.
However, if a block is demand 2, we should only adjust this block is available for demand 2 only.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (vector_insn_info::operator>=): Fix available condition.

21 months agoRISC-V: Simplify codes of changing vsetvl instruction
Ju-Zhe Zhong [Tue, 3 Jan 2023 07:24:36 +0000 (15:24 +0800)]
RISC-V: Simplify codes of changing vsetvl instruction

This patch is NFC patch. I move these code as a function since we will
reuse it in the following patch (Refine phase 3 of VSETVL PASS)

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (change_vsetvl_insn): New function.
(pass_vsetvl::compute_global_backward_infos): Simplify codes.

21 months agoRISC-V: Fix backward_propagate_worthwhile_p
Ju-Zhe Zhong [Tue, 3 Jan 2023 07:16:41 +0000 (15:16 +0800)]
RISC-V: Fix backward_propagate_worthwhile_p

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (loop_basic_block_p): Adjust function.
(backward_propagate_worthwhile_p): Fix non-worthwhile.

21 months agoRISC-V: Fix wrong in_group flag in validate_change call function
Ju-Zhe Zhong [Tue, 3 Jan 2023 07:11:59 +0000 (15:11 +0800)]
RISC-V: Fix wrong in_group flag in validate_change call function

Since we only change insn which is not in group. The flag currently is not correct.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (change_insn): Adjust in_group in validate_change.

21 months agoRISC-V: Fix bugs for refine vsetvl a5, zero into vsetvl zero, zero incorrectly
Ju-Zhe Zhong [Tue, 3 Jan 2023 06:55:30 +0000 (14:55 +0800)]
RISC-V: Fix bugs for refine vsetvl a5, zero into vsetvl zero, zero incorrectly

Currently we support this optimization:

bb 0:
 vsetvli a5,zero,e32,mf2
bb 1:
 vsetvli a5,zero,e64,m1 --> vsetvli zero,zero,e64,m1

According RVV ISA, we can do this optimization only if both RATIO and AVL are equal.
However, current VSETVL PASS missed the check of AVL. This patch add this condition
check to fix bugs.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (vector_infos_manager::all_same_avl_p): New function.
(pass_vsetvl::can_refine_vsetvl_p): Add AVL check.
(pass_vsetvl::commit_vsetvls): Ditto.
* config/riscv/riscv-vsetvl.h: New function declaration.

21 months agoRISC-V: Fix vsetivli instruction asm for IMM AVL
Ju-Zhe Zhong [Tue, 3 Jan 2023 01:39:57 +0000 (09:39 +0800)]
RISC-V: Fix vsetivli instruction asm for IMM AVL

Notice that we should used vsetivli zero,4 instead of vsetvli zero,4
for IMM AVL (0 ~ 31) according to RVV ISA.

This patch fix vsetivli instruction asm bug.

gcc/ChangeLog:

* config/riscv/vector.md:

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/vle-constraint-1.c:

21 months agoRISC-V: Fix inferior codegen for vse intrinsics.
Ju-Zhe Zhong [Thu, 29 Dec 2022 15:34:02 +0000 (23:34 +0800)]
RISC-V: Fix inferior codegen for vse intrinsics.

Currently we use pred_mov to to do the codegen for vse intrinsics. However, it
generates inferior codegen when I am testing AVL model of VSETVL PASS using vse
intrinsics.

Consider this following code:
void f2 (int * restrict in, int * restrict out, void * restrict mask_in, int n)
{
  vfloat32mf2_t v = __riscv_vle32_v_f32mf2 ((float *)(in + 10000), 19);
  __riscv_vse32_v_f32mf2 ((float *)(out + 10000), v, 19);
  vbool64_t mask = *(vbool64_t*)mask_in;
  for (int i = 0; i < n; i++)
    {
      vint16mf2_t v1 = __riscv_vle16_v_i16mf2 ((int16_t *)(in + i + 1), 19);
      __riscv_vse16_v_i16mf2 ((int16_t *)(out + i + 1), v1, 19);

      vint32mf2_t v2 = __riscv_vle32_v_i32mf2 ((int32_t *)(in + i + 2), 19);
      __riscv_vse32_v_i32mf2 ((int32_t *)(out + i + 2), v2, 19);

      vint32mf2_t v3 = __riscv_vle32_v_i32mf2_tumu (mask, v2, (int32_t *)(in + i + 200), 13);
      __riscv_vse32_v_i32mf2 ((int32_t *)(out + i + 200), v2, 13);

      vfloat64m1_t v4 = __riscv_vle64_v_f64m1_m (mask, (double *)(in + i + 300), 11);
      __riscv_vse64_v_f64m1 ((double *)(out + i + 300), v4, 11);

      vfloat64m1_t v5 = __riscv_vle64_v_f64m1_tum (mask, v4, (double *)(in + i + 500), 11);
      __riscv_vse64_v_f64m1 ((double *)(out + i + 500), v5, 11);

      vfloat64m1_t v6 = __riscv_vle64_v_f64m1_mu (mask, v5, (double *)(in + i + 600), 11);
      __riscv_vse64_v_f64m1_m (mask, (double *)(out + i + 600), v6, 11);

      vuint8mf4_t v7 = __riscv_vle8_v_u8mf4 ((uint8_t *)(in + i + 700), 11);
      __riscv_vse8_v_u8mf4 ((uint8_t *)(out + i + 700), v7, 11);
    }
}

Before this patch:
csrr t2,vlenb
srli t2,t2,1
slli s0,t2,2
vsetvli zero,19,e16,mf2,ta,ma
sub s0,s0,t2
csrr t2,vlenb
vle16.v v24,0(a3)
mv a4,a3
vse16.v v24,0(a1)
srli t2,t2,1
add a2,a3,t6
add s0,s0,sp
vsetvli zero,19,e32,mf2,ta,ma
addi a3,a3,4
vle32.v v24,0(a3)
vsetvli zero,t0,e32,mf2,ta,ma
vse32.v v24,0(s0)
slli s0,t2,2
sub s0,s0,t2
add s0,s0,sp
vsetvli t0,zero,e32,mf2,ta,ma
vle32.v v24,0(s0)
mv s0,t2
slli t2,t2,2
mv a5,a1
vsetvli zero,19,e32,mf2,ta,ma
addi a1,a1,4
sub t2,t2,s0
vse32.v v24,0(a1)
add t2,t2,sp
vsetvli t0,zero,e32,mf2,ta,ma
addi t1,a5,796
vle32.v v24,0(t2)
addi t5,a4,1196
addi a7,a5,1196
addi t4,a4,1996
addi a6,a5,1996
vsetvli zero,13,e32,mf2,ta,ma
add a4,a4,t3
vse32.v v24,0(t1)
add a5,a5,t3
vsetvli zero,11,e64,m1,tu,mu
vle64.v v24,0(t5),v0.t
vse64.v v24,0(a7)
vle64.v v24,0(t4),v0.t
vse64.v v24,0(a6)
vle64.v v24,0(a4),v0.t
vse64.v v24,0(a5),v0.t
vsetvli zero,11,e8,mf4,ta,ma
vle8.v v24,0(a2)
vse8.v v24,0(a2)
bne a0,a3,.L8
csrr t0,vlenb
slli t1,t0,1
add sp,sp,t1
lw s0,12(sp)
addi sp,sp,16
jr ra

We are generating redundant spilling codes.
Here we introduce a dedicated pred_store pattern for vse intrinsics like
maskstore in ARM SVE.

After this patch:
vsetvli zero,19,e16,mf2,ta,ma
mv a5,a4
vle16.v v24,0(a0)
mv a3,a0
vse16.v 19,0(a4)
addi t1,a4,796
vsetvli zero,19,e32,mf2,ta,ma
addi a0,a0,4
addi a4,a4,4
vle32.v v24,0(a0)
addi t0,a3,1196
vse32.v 19,0(a4)
addi a7,a5,1196
addi t6,a3,1996
addi a6,a5,1996
add t5,a3,t4
vsetvli zero,13,e32,mf2,ta,ma
add a2,a5,t4
vse32.v 13,0(t1)
add a3,a3,t3
vsetvli zero,11,e64,m1,tu,mu
add a5,a5,t3
vle64.v v24,0(t0),v0.t
vse64.v 11,0(a7)
vle64.v v24,0(t6),v0.t
vse64.v 11,0(a6)
vle64.v v24,0(t5),v0.t
vse64.v 11,0(a2),v0.t
vsetvli zero,11,e8,mf4,ta,ma
vle8.v v24,0(a3)
vse8.v 11,0(a5)
bne a1,a4,.L8
.L6:
ret

gcc/ChangeLog:

* config/riscv/riscv-vector-builtins-bases.cc (class loadstore): use
pred_store for vse.
* config/riscv/riscv-vector-builtins.cc
(function_expander::add_mem_operand): Refine function.
(function_expander::use_contiguous_load_insn): Adjust new
implementation.
(function_expander::use_contiguous_store_insn): Ditto.
* config/riscv/riscv-vector-builtins.h: Refine function.
* config/riscv/vector.md (@pred_store<mode>): New pattern.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/vse-constraint-1.c: New test.

21 months agoRISC-V: Fix pointer tree type for store pointer.
Ju-Zhe Zhong [Wed, 28 Dec 2022 05:11:08 +0000 (13:11 +0800)]
RISC-V: Fix pointer tree type for store pointer.

For store intrinsic,
the function type should be void store (T *...) instead of void store (const T *...)

gcc/ChangeLog:

* config/riscv/riscv-vector-builtins.cc: Change to scalar pointer.

21 months agoFortran: fix ICE in check_host_association [PR108544]
Harald Anlauf [Wed, 25 Jan 2023 21:47:26 +0000 (22:47 +0100)]
Fortran: fix ICE in check_host_association [PR108544]

gcc/fortran/ChangeLog:

PR fortran/108544
* resolve.cc (check_host_association): Extend host association check
so that it is not restricted to functions.  Also prevent NULL pointer
dereference.

gcc/testsuite/ChangeLog:

PR fortran/108544
* gfortran.dg/pr108544.f90: New test.
* gfortran.dg/pr96102b.f90: New test.

21 months agoopts: SANITIZE_ADDRESS wrongly cleared [PR108543]
Marek Polacek [Wed, 25 Jan 2023 22:19:54 +0000 (17:19 -0500)]
opts: SANITIZE_ADDRESS wrongly cleared [PR108543]

Here we crash on a null fndecl ultimately because we haven't defined
the built-ins described in sanitizer.def.  So
builtin_decl_explicit (BUILT_IN_ASAN_POINTER_SUBTRACT);
returns NULL_TREE, causing an ICE later.

DEF_SANITIZER_BUILTIN only actually defines the built-ins when flag_sanitize
has SANITIZE_ADDRESS, or some of the other SANITIZE_*, but it doesn't check
SANITIZE_KERNEL_ADDRESS or SANITIZE_USER_ADDRESS.  Unfortunately, with
-fsanitize=address -fno-sanitize=kernel-address
or
-fsanitize=kernel-address -fno-sanitize=address
SANITIZE_ADDRESS ends up being unset from flag_sanitize even though
_USER/_KERNEL are set.  That's because -fsanitize=address means
SANITIZE_ADDRESS | SANITIZE_USER_ADDRESS and -fsanitize=kernel-address
is SANITIZE_ADDRESS | SANITIZE_KERNEL_ADDRESS but parse_sanitizer_options
does
  flags &= ~sanitizer_opts[i].flag;
so the subsequent -fno- unsets SANITIZE_ADDRESS.  Then no sanitizer
built-ins are actually defined.

I'm not sure why SANITIZE_ADDRESS isn't just SANITIZE_USER_ADDRESS |
SANITIZE_KERNEL_ADDRESS, I don't think we need 3 bits.

PR middle-end/108543

gcc/ChangeLog:

* opts.cc (parse_sanitizer_options): Don't always clear SANITIZE_ADDRESS
if it was previously set.

gcc/testsuite/ChangeLog:

* c-c++-common/asan/pointer-subtract-5.c: New test.
* c-c++-common/asan/pointer-subtract-6.c: New test.
* c-c++-common/asan/pointer-subtract-7.c: New test.
* c-c++-common/asan/pointer-subtract-8.c: New test.

21 months agoModula-2: Remove debug code [PR108553].
Iain Sandoe [Thu, 26 Jan 2023 09:46:32 +0000 (09:46 +0000)]
Modula-2: Remove debug code [PR108553].

Remove debugging code accidentally left in place in r13-5373-g80cf2c5e8f496b.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
PR modula2/108553

gcc/m2/ChangeLog:

* gm2-lang.cc (gm2_langhook_init_options): Remove debug code.

21 months agofrange: Fix up foperator_{,not_}equal::fold_range for signed zeros [PR108540]
Jakub Jelinek [Thu, 26 Jan 2023 16:21:22 +0000 (17:21 +0100)]
frange: Fix up foperator_{,not_}equal::fold_range for signed zeros [PR108540]

The following testcases are miscompiled, because threader sees some
SSA_NAME would have -0.0 value and when computing range of SSA_NAME == 0.0
foperator_equal::fold_range sees one operand has [-0.0, -0.0] singleton
range, the other [0.0, 0.0], they aren't equal (frange operator== uses
real_identical etc. rather than real comparisons) and so it thinks they
compare unequal.  With signed zeros -0.0 == 0.0 is true though, so we
need to special case the both ranges singleton code.
Similarly, if we see op1 range being say [-42.0, -0.0] and op2 range
[0.0, 42.0], we'd check that the intersection of the two ranges is empty
(that is correct) and fold the result of == between such operands to
[0, 0] which is wrong, because -0.0 == 0.0, it needs to be [0, 1].
Similarly for foperator_not_equal::fold_range.

2023-01-26  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/108540
* range-op-float.cc (foperator_equal::fold_range): If both op1 and op2
are singletons, use range_true even if op1 != op2
when one range is [-0.0, -0.0] and another [0.0, 0.0].  Similarly,
even if intersection of the ranges is empty and one has
zero low bound and another zero high bound, use range_true_and_false
rather than range_false.
(foperator_not_equal::fold_range): If both op1 and op2
are singletons, use range_false even if op1 != op2
when one range is [-0.0, -0.0] and another [0.0, 0.0].  Similarly,
even if intersection of the ranges is empty and one has
zero low bound and another zero high bound, use range_true_and_false
rather than range_true.

* gcc.c-torture/execute/ieee/pr108540-1.c: New test.
* gcc.c-torture/execute/ieee/pr108540-2.c: New test.

21 months agovalue-relation: Small tweaks to tables
Jakub Jelinek [Thu, 26 Jan 2023 16:20:23 +0000 (17:20 +0100)]
value-relation: Small tweaks to tables

As I said earlier, all these tables are used solely in value-relation.cc
and never modified, plus because VREL_LAST is small especially the
two-dimensional arrays are vast a lot of .data (or .rodata) space
- 576 bytes each.  The following patch makes those arrays static const
and uses unsigned char instead of relation_kind so that the
two-dimensional arrays shrink to 144 bytes.

2023-01-26  Jakub Jelinek  <jakub@redhat.com>

* value-relation.cc (kind_string): Add const.
(rr_negate_table, rr_swap_table, rr_intersect_table,
rr_union_table, rr_transitive_table): Add static const, change
element type from relation_kind to unsigned char.
(relation_negate, relation_swap, relation_intersect, relation_union,
relation_transitive): Cast rr_*_table element to relation_kind.
(relation_to_code): Add static const.
(relation_tests): Assert VREL_LAST is smaller than UCHAR_MAX.

21 months agotestsuite: Fix hwasan/arguments-3.c failures
Richard Sandiford [Thu, 26 Jan 2023 16:01:09 +0000 (16:01 +0000)]
testsuite: Fix hwasan/arguments-3.c failures

This testcase had three dg-error tests for ".*<message>.*".
But since . matches \n in Tcl regexps, the first dg-error
ate all the output, leaving the other two to fail.

The regexp is eventually embedded in a larger one, so we
can't prefix it with (?n).  But the .*s aren't necessary,
since dg-error tests for a partial rather than a full match.

gcc/testsuite/
* c-c++-common/hwasan/arguments-3.c: Remove extraneous .*s.

21 months agoaarch64: Remove expected error for compound literals
Richard Sandiford [Thu, 26 Jan 2023 15:51:00 +0000 (15:51 +0000)]
aarch64: Remove expected error for compound literals

GCC no longer treats empty compound literals as an error
(see 14cfa01755a66afbae2539f8b5796c960ddcecc6).

gcc/testsuite/
* gcc.target/aarch64/bfloat16_scalar_typecheck.c: Accept empty
compound literals.