Use error_mark_node after error in convert

While working on PR c/111903, I Noticed that
convert will convert integer_zero_node to that
type after an error instead of returning error_mark_node.
From what I can tell this was the old way of not having
error recovery since other places in this file does return
error_mark_node and the places I am replacing date from
when the file was imported into the repro (either via a gcc2 merge
or earlier).

I also had to update the objc front-end to allow for the error_mark_node
change, I suspect you could hit the ICE without this change though.

Bootstrapped and tested on x86_64-linux-gnu with no regressions.

gcc/ChangeLog:

* convert.cc (convert_to_pointer_1): Return error_mark_node
after an error.
(convert_to_real_1): Likewise.
(convert_to_integer_1): Likewise.
(convert_to_complex_1): Likewise.

gcc/objc/ChangeLog:

* objc-gnu-runtime-abi-01.cc (build_objc_method_call): Allow
for error_operand after call to build_c_cast.
* objc-next-runtime-abi-01.cc (build_objc_method_call): Likewise.
* objc-next-runtime-abi-02.cc (build_v2_build_objc_method_call): Likewise.

convert_to_complex vs invalid_conversion [PR111903]

convert_to_complex when creating a COMPLEX_EXPR does
not currently check if either the real or imag parts
was not error_mark_node. This later on confuses the gimpilfier
when there was a SAVE_EXPR wrapped around that COMPLEX_EXPR.
The simple fix is after calling convert inside convert_to_complex_1,
check that the either result was an error_operand and return
an error_mark_node in that case.

Bootstrapped and tested on x86_64-linux-gnu with no regressions.

PR c/111903

gcc/ChangeLog:

* convert.cc (convert_to_complex_1): Return
error_mark_node if either convert was an error
when converting from a scalar.

gcc/testsuite/ChangeLog:

* gcc.target/i386/float16-8.c: New test.

tree-optimization/111917 - bougs IL after guard hoisting

The unswitching code to hoist guards inserts conditions in wrong
places.  The following fixes this, simplifying code.

PR tree-optimization/111917
* tree-ssa-loop-unswitch.cc (hoist_guard): Always insert
new conditional after last stmt.

* gcc.dg/torture/pr111917.c: New testcase.

RISC-V: Fix ICE for the fusion case from vsetvl to scalar move[PR111927]

ICE:

during RTL pass: vsetvl
<source>: In function 'riscv_lms_f32':
<source>:240:1: internal compiler error: in merge, at config/riscv/riscv-vsetvl.cc:1997
  240 | }

In general compatible_p (avl_equal_p) has:

    if (next.has_vl () && next.vl_used_by_non_rvv_insn_p ())
      return false;

Don't fuse AVL of vsetvl if the VL operand is used by non-RVV instructions.

It is reasonable to add it into 'can_use_next_avl_p' since we don't want to
fuse AVL of vsetvl into a scalar move instruction which doesn't demand AVL.
And after the fusion, we will alway use compatible_p to check whether the demand
is correct or not.

PR target/111927

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc: Fix bug.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/pr111927.c: New test.

RISC-V: Remove unnecessary asm check for vec cvt

The vsetvl asm check is unnecessary for the vector convert. We
should be focus for constrait and leave the vsetvl test to the
vsetvl pass.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/cvt-0.c: Remove the vsetvl
asm check from func body.
* gcc.target/riscv/rvv/autovec/unop/cvt-1.c: Ditto.

Signed-off-by: Pan Li <pan2.li@intel.com>

libatomic: drop redundant all-multi command

./multilib.am already specifies this same command, and make warns about
the earlier one being ignored when seeing the later one. All that needs
retaining to still satisfy the preceding comment is the extra
dependency.

libatomic/

* Makefile.am (all-multi): Drop commands.
* Makefile.in: Update accordingly.

RISC-V: Bugfix for merging undef tmp register for trunc

For trunc function autovec, there will be one step like below take MU
for the merge operand.

rtx tmp = gen_reg_rtx (vec_int_mode);
emit_vec_cvt_x_f_rtz (tmp, op_1, mask, vec_fp_mode);

The MU will leave the tmp (aka dest register) register unmasked elements
unchanged and it is undefined here. This patch would like to adjust the
MU to MA.

gcc/ChangeLog:

* config/riscv/riscv-v.cc (emit_vec_cvt_x_f_rtz): Add insn type
arg.
(expand_vec_trunc): Take MA instead of MU for cvt_x_f_rtz.

Signed-off-by: Pan Li <pan2.li@intel.com>

LoongArch: Document -mexplicit-relocs={auto,none,always}

gcc/ChangeLog:

* doc/invoke.texi (-mexplicit-relocs=style): Document.
(-mexplicit-relocs): Document as an alias of
-mexplicit-relocs=always.
(-mno-explicit-relocs): Document as an alias of
-mexplicit-relocs=none.
(-mcmodel=extreme): Mention -mexplicit-relocs=always instead of
-mexplicit-relocs.

LoongArch: Use explicit relocs for addresses only used for one load or store with -mexplicit-relocs=auto and -mcmodel={normal,medium}

In these cases, if we use explicit relocs, we end up with 2
instructions:

    pcalau12i    t0, %pc_hi20(x)
    ld.d         t0, t0, %pc_lo12(x)

If we use la.local pseudo-op, in the best scenario (x is in +/- 2MiB
range) we still have 2 instructions:

    pcaddi       t0, %pcrel_20(x)
    ld.d         t0, t0, 0

If x is out of the range we'll have 3 instructions.  So for these cases
just emit machine instructions with explicit relocs.

gcc/ChangeLog:

* config/loongarch/predicates.md (symbolic_pcrel_operand): New
predicate.
* config/loongarch/loongarch.md (define_peephole2): Optimize
la.local + ld/st to pcalau12i + ld/st if the address is only used
once if -mexplicit-relocs=auto and -mcmodel=normal or medium.

gcc/testsuite/ChangeLog:

* gcc.target/loongarch/explicit-relocs-auto-single-load-store.c:
New test.
* gcc.target/loongarch/explicit-relocs-auto-single-load-store-no-anchor.c:
New test.

LoongArch: Use explicit relocs for TLS access with -mexplicit-relocs=auto

The linker does not know how to relax TLS access for LoongArch, so let's
emit machine instructions with explicit relocs for TLS.

gcc/ChangeLog:

* config/loongarch/loongarch.cc (loongarch_explicit_relocs_p):
Return true for TLS symbol types if -mexplicit-relocs=auto.
(loongarch_call_tls_get_addr): Replace TARGET_EXPLICIT_RELOCS
with la_opt_explicit_relocs != EXPLICIT_RELOCS_NONE.
(loongarch_legitimize_tls_address): Likewise.
* config/loongarch/loongarch.md (@tls_low<mode>): Remove
TARGET_EXPLICIT_RELOCS from insn condition.

gcc/testsuite/ChangeLog:

* gcc.target/loongarch/explicit-relocs-auto-tls-ld-gd.c: New
test.
* gcc.target/loongarch/explicit-relocs-auto-tls-le-ie.c: New
test.

LoongArch: Use explicit relocs for GOT access when -mexplicit-relocs=auto and LTO during a final link with linker plugin

If we are performing LTO for a final link and linker plugin is enabled,
then we are sure any GOT access may resolve to a symbol out of the link
unit (otherwise the linker plugin will tell us the symbol should be
resolved locally and we'll use PC-relative access instead).

Produce machine instructions with explicit relocs instead of la.global
for better scheduling.

gcc/ChangeLog:

* config/loongarch/loongarch-protos.h
(loongarch_explicit_relocs_p): Declare new function.
* config/loongarch/loongarch.cc (loongarch_explicit_relocs_p):
Implement.
(loongarch_symbol_insns): Call loongarch_explicit_relocs_p for
SYMBOL_GOT_DISP, instead of using TARGET_EXPLICIT_RELOCS.
(loongarch_split_symbol): Call loongarch_explicit_relocs_p for
deciding if return early, instead of using
TARGET_EXPLICIT_RELOCS.
(loongarch_output_move): CAll loongarch_explicit_relocs_p
instead of using TARGET_EXPLICIT_RELOCS.
* config/loongarch/loongarch.md (*low<mode>): Remove
TARGET_EXPLICIT_RELOCS from insn condition.
(@ld_from_got<mode>): Likewise.
* config/loongarch/predicates.md (move_operand): Call
loongarch_explicit_relocs_p instead of using
TARGET_EXPLICIT_RELOCS.

gcc/testsuite/ChangeLog:

* gcc.target/loongarch/explicit-relocs-auto-lto.c: New test.

LoongArch: Add enum-style -mexplicit-relocs= option

To take a better balance between scheduling and relaxation when -flto is
enabled, add three-way -mexplicit-relocs={auto,none,always} options.
The old -mexplicit-relocs and -mno-explicit-relocs options are still
supported, they are mapped to -mexplicit-relocs=always and
-mexplicit-relocs=none.

The default choice is determined by probing assembler capabilities at
build time.  If the assembler does not supports explicit relocs at all,
the default will be none; if it supports explicit relocs but not
relaxation, the default will be always; if both explicit relocs and
relaxation are supported, the default will be auto.

Currently auto is same as none.  We will make auto more clever in
following changes.

gcc/ChangeLog:

* config/loongarch/genopts/loongarch-strings: Add strings for
-mexplicit-relocs={auto,none,always}.
* config/loongarch/genopts/loongarch.opt.in: Add options for
-mexplicit-relocs={auto,none,always}.
* config/loongarch/loongarch-str.h: Regenerate.
* config/loongarch/loongarch.opt: Regenerate.
* config/loongarch/loongarch-def.h
(EXPLICIT_RELOCS_AUTO): Define.
(EXPLICIT_RELOCS_NONE): Define.
(EXPLICIT_RELOCS_ALWAYS): Define.
(N_EXPLICIT_RELOCS_TYPES): Define.
* config/loongarch/loongarch.cc
(loongarch_option_override_internal): Error out if the old-style
-m[no-]explicit-relocs option is used with
-mexplicit-relocs={auto,none,always} together.  Map
-mno-explicit-relocs to -mexplicit-relocs=none and
-mexplicit-relocs to -mexplicit-relocs=always for backward
compatibility.  Set a proper default for -mexplicit-relocs=
based on configure-time probed linker capability.  Update a
diagnostic message to mention -mexplicit-relocs=always instead
of the old-style -mexplicit-relocs.
(loongarch_handle_model_attribute): Update a diagnostic message
to mention -mexplicit-relocs=always instead of the old-style
-mexplicit-relocs.
* config/loongarch/loongarch.h (TARGET_EXPLICIT_RELOCS): Define.

RISC-V: Fix typo[VSETVL PASS]

When fixing an issue, I find there is a typo in VSETVL PASS.

Change 'use_by' into 'used_by'.

Committed it as it is very obvious.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (pre_vsetvl::fuse_local_vsetvl_info): Fix typo.
(pre_vsetvl::pre_global_vsetvl_info): Ditto.

gcc.c-torture/execute/builtins/pr93262-chk.c: Remove return statement

The main_test function returns void, so return with an expression
is a constraint violation.  The test case still fails with this
change applied before the fix for PR 93262 in r14-4813.

gcc/testsuite/

* gcc.c-torture/execute/builtins/pr93262-chk.c (main_test):
Remove unnecessary return statement.

RISC-V: Remove unnecessary asm check for binop constraint

The vsetvl asm check is unnecessary for the binop constraint. We
should be focus for constrait and leave the vsetvl test to the
vsetvl pass.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/binop_vv_constraint-1.c: Remove the
vsetvl asm check from func body.
* gcc.target/riscv/rvv/base/binop_vx_constraint-1.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-10.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-11.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-12.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-129.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-13.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-130.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-131.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-133.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-134.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-135.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-14.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-15.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-153.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-154.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-155.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-158.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-16.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-17.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-171.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-172.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-173.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-174.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-18.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-19.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-2.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-20.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-21.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-22.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-23.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-24.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-25.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-26.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-27.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-28.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-29.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-3.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-30.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-31.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-32.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-33.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-34.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-35.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-36.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-37.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-38.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-39.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-4.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-40.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-41.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-42.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-43.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-44.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-5.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-6.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-7.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-8.c: Ditto.
* gcc.target/riscv/rvv/base/binop_vx_constraint-9.c: Ditto.
* gcc.target/riscv/rvv/base/shift_vx_constraint-1.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-1.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-2.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-3.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-4.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-5.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vv_constraint-6.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vx_constraint-1.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vx_constraint-8.c: Ditto.
* gcc.target/riscv/rvv/base/ternop_vx_constraint-9.c: Ditto.
* gcc.target/riscv/rvv/base/unop_v_constraint-1.c: Ditto.
* gcc.target/riscv/rvv/base/unop_v_constraint-2.c: Ditto.

Signed-off-by: Pan Li <pan2.li@intel.com>

RISC-V: Remove unnecessary asm check for rounding autovec

The vsetvl asm check is unnecessary for the rounding function autovec.
These rounding test cases should focus on the rounding insn sequence.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/bswap16-0.c: Remove the
vsetvl check.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-iceil-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-ifloor-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-irint-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-iround-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lceil-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lceil-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lfloor-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lfloor-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-llceil-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-llfloor-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-llrint-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-llround-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lrint-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lrint-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lround-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-lround-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-3.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-3.c: Ditto.

Signed-off-by: Pan Li <pan2.li@intel.com>

RISC-V: Fix AVL_TYPE attribute of tuple mode mov<mode>

The tuple mode mov<mode> pattern doesn't have avl_type so it is invalid attribute.

gcc/ChangeLog:

* config/riscv/vector.md: Fix avl_type attribute of tuple mov<mode>.

vect: Cost adjacent vector loads/stores together [PR111784]

As comments[1][2], this patch is to change the costing way
on some adjacent vector loads/stores from costing one by
one to costing them together with the total number once.

It helps to fix the exposed regression PR111784 on aarch64,
as aarch64 specific costing could make different decisions
according to the different costing ways (counting with total
number vs. counting one by one).  Based on a reduced test
case from PR111784, only considering vec_num can fix the
regression already, but vector loads/stores in regard to
ncopies are also adjacent accesses, so they are considered
as well.

btw, this patch leaves the costing on dr_explicit_realign
and dr_explicit_realign_optimized alone to make it simple.
The costing way change can cause the differences for them
since there is one costing depending on targetm.vectorize.
builtin_mask_for_load and it's costed according to the
calling times.  IIUC, these two dr_alignment_support are
mainly used for old Power? (only having 16 bytes aligned
vector load/store but no unaligned vector load/store).

[1] https://gcc.gnu.org/pipermail/gcc-patches/2023-September/630742.html
[2] https://gcc.gnu.org/pipermail/gcc-patches/2023-September/630744.html

PR tree-optimization/111784

gcc/ChangeLog:

* tree-vect-stmts.cc (vectorizable_store): Adjust costing way for
adjacent vector stores, by costing them with the total number
rather than costing them one by one.
(vectorizable_load): Adjust costing way for adjacent vector
loads, by costing them with the total number rather than costing
them one by one.

i386: Prevent splitting to xmm16+ when !TARGET_AVX512VL

Currently, there will be a chance in split to use x/ymm16+ w/o AVX512VL,
which finally leads to an ICE as pr111753 does.

This patch aims to fix that.

gcc/ChangeLog:

PR target/111753
* config/i386/i386.cc (ix86_standard_x87sse_constant_load_p):
Do not split to xmm16+ when !TARGET_AVX512VL.

gcc/testsuite/ChangeLog:

PR target/111753
* gcc.target/i386/pr111753.c: New test.

RISC-V: Bugfix for merging undefined tmp register in math

For math function autovec, there will be one step like

rtx tmp = gen_reg_rtx (vec_int_mode);
emit_vec_cvt_x_f (tmp, op_1, mask, UNARY_OP_TAMU_FRM_DYN, vec_fp_mode);

The MU will leave the tmp (aka dest register) register unmasked elements
unchanged and it is undefined here. This patch would like to adjust the
MU to MA.

gcc/ChangeLog:

* config/riscv/riscv-protos.h (enum insn_type): Add new type
values.
* config/riscv/riscv-v.cc (emit_vec_cvt_x_f): Add undef merge
operand handling.
(expand_vec_ceil): Take MA instead of MU for tmp register.
(expand_vec_floor): Ditto.
(expand_vec_nearbyint): Ditto.
(expand_vec_rint): Ditto.
(expand_vec_round): Ditto.
(expand_vec_roundeven): Ditto.

Signed-off-by: Pan Li <pan2.li@intel.com>

compiler: remove traverse_assignments pass

The last caller was removed in https://go.dev/cl/18261 in 2016.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/536637

compiler: remove name_ field from Type_switch_statement

It's not used for anything.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/536636

compiler: pass Gogo to determine types pass

Also pass Gogo to the type verification pass.

This is a refactoring that does not change the compiler behavior.
This is in preparation for future CLs that rearrange the pass ordering.

This introduces one new call to go_get_gogo, which will be removed in
a future CL.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/536635

LoongArch: Define macro CLEAR_INSN_CACHE.

LoongArch's microstructure ensures cache consistency by hardware.
Due to out-of-order execution, "ibar" is required to ensure the visibility of the
store (invalidated icache) executed by this CPU before "ibar" (to the instance).
"ibar" will not invalidate the icache, so the start and end parameters are not Affect
"ibar" performance.

gcc/ChangeLog:

* config/loongarch/loongarch.h (CLEAR_INSN_CACHE): New definition.

Expand: Enable vector mode for by pieces compares

Vector mode compare instructions are efficient for equality compare on
rs6000. This patch refactors the codes of by pieces operation to enable
vector mode for compare.

gcc/
PR target/111449
* expr.cc (can_use_qi_vectors): New function to return true if
we know how to implement OP using vectors of bytes.
(qi_vector_mode_supported_p): New function to check if optabs
exists for the mode and certain by pieces operations.
(widest_fixed_size_mode_for_size): Replace the second argument
with the type of by pieces operations.  Call can_use_qi_vectors
and qi_vector_mode_supported_p to do the check.  Call
scalar_mode_supported_p to check if the scalar mode is supported.
(by_pieces_ninsns): Pass the type of by pieces operation to
widest_fixed_size_mode_for_size.
(class op_by_pieces_d): Remove m_qi_vector_mode.  Add m_op to
record the type of by pieces operations.
(op_by_pieces_d::op_by_pieces_d): Change last argument to the
type of by pieces operations, initialize m_op with it.  Pass
m_op to function widest_fixed_size_mode_for_size.
(op_by_pieces_d::get_usable_mode): Pass m_op to function
widest_fixed_size_mode_for_size.
(op_by_pieces_d::smallest_fixed_size_mode_for_size): Call
can_use_qi_vectors and qi_vector_mode_supported_p to do the
check.
(op_by_pieces_d::run): Pass m_op to function
widest_fixed_size_mode_for_size.
(move_by_pieces_d::move_by_pieces_d): Set m_op to MOVE_BY_PIECES.
(store_by_pieces_d::store_by_pieces_d): Set m_op with the op.
(can_store_by_pieces): Pass the type of by pieces operations to
widest_fixed_size_mode_for_size.
(clear_by_pieces): Initialize class store_by_pieces_d with
CLEAR_BY_PIECES.
(compare_by_pieces_d::compare_by_pieces_d): Set m_op to
COMPARE_BY_PIECES.

Avoid compile time hog on vect_peel_nonlinear_iv_init for nonlinear induction vec_step_op_mul when iteration count is too big.

There's loop in vect_peel_nonlinear_iv_init to get init_expr *
pow (step_expr, skip_niters). When skipn_iters is too big, compile time
hogs. To avoid that, optimize init_expr * pow (step_expr, skip_niters) to
init_expr << (exact_log2 (step_expr) * skip_niters) when step_expr is
pow of 2, otherwise give up vectorization when skip_niters >=
TYPE_PRECISION (TREE_TYPE (init_expr)).

Also give up vectorization when niters_skip is negative which will be
used for fully masked loop.

gcc/ChangeLog:

PR tree-optimization/111820
PR tree-optimization/111833
* tree-vect-loop-manip.cc (vect_can_peel_nonlinear_iv_p): Give
up vectorization for nonlinear iv vect_step_op_mul when
step_expr is not exact_log2 and niters is greater than
TYPE_PRECISION (TREE_TYPE (step_expr)). Also don't vectorize
for nagative niters_skip which will be used by fully masked
loop.
(vect_can_advance_ivs_p): Pass whole phi_info to
vect_can_peel_nonlinear_iv_p.
* tree-vect-loop.cc (vect_peel_nonlinear_iv_init): Optimize
init_expr * pow (step_expr, skipn) to init_expr
<< (log2 (step_expr) * skipn) when step_expr is exact_log2.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr111820-1.c: New test.
* gcc.target/i386/pr111820-2.c: New test.
* gcc.target/i386/pr111820-3.c: New test.
* gcc.target/i386/pr103144-mul-1.c: Adjust testcase.
* gcc.target/i386/pr103144-mul-2.c: Adjust testcase.

Remove unused mmx_pinsrw.

gcc/ChangeLog:

* config/i386/mmx.md (mmx_pinsrw): Remove.

Daily bump.

aarch64: Emit csinv again for `a ? ~b : b` [PR110986]

After r14-3110-g7fb65f10285, the canonical form for
`a ? ~b : b` changed to be `-(a) ^ b` that means
for aarch64 we need to add a few new insn patterns
to be able to catch this and change it to be
what is the canonical form for the aarch64 backend.
A secondary pattern was needed to support a zero_extended
form too; this adds a testcase for all 3 cases.

Bootstrapped and tested on aarch64-linux-gnu with no regressions.

PR target/110986

gcc/ChangeLog:

* config/aarch64/aarch64.md (*cmov<mode>_insn_insv): New pattern.
(*cmov_uxtw_insn_insv): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/cond_op-1.c: New test.

d: Merge upstream dmd f4be7f6f7b.

D front-end changes:

    - Fix bootstrap failure with i686-darwin9.
      ```
      Undefined symbols for architecture i386:
          "gendocfile", referenced from:
          __ZL20d_generate_ddoc_fileP6ModuleR9OutBuffer in d-lang.o
      ld: symbol(s) not found for architecture i386
      ```
gcc/d/ChangeLog:

* dmd/MERGE: Merge upstream dmd f4be7f6f7b.
* Make-lang.in (D_FRONTEND_OBJS): Rename d/root-rootobject.o to
d/rootobject.o.

objc++: type/expr tsubst conflation [PR111920]

After r14-4796-g3e3d73ed5e85e7, tsubst_copy_and_build (now named
tsubst_expr) no longer dispatches to tsubst for type trees, and
callers have to do it themselves if appropriate.  This patch makes
some overlooked adjustments to Objective-C++-specific code paths.

PR objc++/111920

gcc/cp/ChangeLog:

* pt.cc (tsubst_expr) <case AT_ENCODE_EXPR>: Use tsubst instead
of tsubst_expr.

gcc/objcp/ChangeLog:

* objcp-lang.cc (objcp_tsubst_expr) <case CLASS_REFERENCE_EXPR>:
Use tsubst instead of tsubst_expr for type operands.

Doc: document the new Darwin options

gcc/ChangeLog:

* doc/invoke.texi: Document the new -nodefaultrpaths option.
* doc/install.texi: Document the new --with-darwin-extra-rpath
option.

Testsuite: allow non-installed testing on darwin

DYLD_LIBRARY_PATH is now removed from the environment for all system
tools, including the shell. Adapt the testsuite and pass the right
options to allow testing, even when the compiler and libraries have not
been installed.

gcc/ChangeLog:

* Makefile.in: set ENABLE_DARWIN_AT_RPATH in site.tmp.

gcc/testsuite/ChangeLog:

* gfortran.dg/coarray/caf.exp: Correctly set
libatomic flags.
* gfortran.dg/dg.exp: Likewise.
* lib/asan-dg.exp: Set correct -B flags.
* lib/atomic-dg.exp: Likewise.
* lib/target-libpath.exp: Handle ENABLE_DARWIN_AT_RPATH.

libatomic/ChangeLog:

* testsuite/lib/libatomic.exp: Pass correct flags on darwin.

libffi/ChangeLog:

* testsuite/lib/libffi.exp: Likewise.

libitm/ChangeLog:

* testsuite/lib/libitm.exp: Likewise.
* testsuite/libitm.c++/c++.exp: Likewise.

Darwin, rpaths: Add --with-darwin-extra-rpath.

This is provided to allow distributions to add a single additional
runpath to allow for cases where the installed GCC library directories
are then symlinked to a common dirctory outside of any of the GCC
installations.

For example:

/opt/distro/lib:
  libgfortran.dylib -> /opt/distro/lib/gcc-11.3/lib/libgfortran.dylib

So that libraries which are designed to be found in the runpath we would then
add --with-darwin-add-rpath=/opt/distro/lib to the configure line.

This patch makes the configuration a little more forgiving of using
--disable-darwin-at-rpath (although for platform versions >= 10.11 this will
result in misconfigured target libraries).

gcc/ChangeLog:

* configure.ac: Add --with-darwin-extra-rpath option.
* config/darwin.h: Handle DARWIN_EXTRA_RPATH.
* config.in: Regenerate.
* configure: Regenerate.

Config,Darwin: Allow for configuring Darwin to use embedded runpath.

Recent Darwin versions place contraints on the use of run paths
specified in environment variables.  This breaks some assumptions
in the GCC build.

This change allows the user to configure a Darwin build to use
'@rpath/libraryname.dylib' in library names and then to add an
embedded runpath to executables (and libraries with dependents).

The embedded runpath is added by default unless the user adds
'-nodefaultrpaths' to the link line.

For an installed compiler, it means that any executable built with
that compiler will reference the runtimes installed with the
compiler (equivalent to hard-coding the library path into the name
of the library).

During build-time configurations  any "-B" entries will be added to
the runpath thus the newly-built libraries will be found by exes.

Since the install name is set in libtool, that decision needs to be
available here (but might also cause dependent ones in Makefiles,
so we need to export a conditional).

This facility is not available for Darwin 8 or earlier, however the
existing environment variable runpath does work there.

We default this on for systems where the external DYLD_LIBRARY_PATH
does not work and off for Darwin 8 or earlier.  For systems that can
use either method, if the value is unset, we use the default (which
is currently DYLD_LIBRARY_PATH).

ChangeLog:

* configure: Regenerate.
* configure.ac: Do not add default runpaths to GCC exes
when we are building -static-libstdc++/-static-libgcc (the
default).
* libtool.m4: Add 'enable-darwin-at-runpath'.  Act  on the
enable flag to alter Darwin libraries to use @rpath names.

gcc/ChangeLog:

* aclocal.m4: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.
* config/darwin.h: Handle Darwin rpaths.
* config/darwin.opt: Handle Darwin rpaths.
* Makefile.in:  Handle Darwin rpaths.

gcc/ada/ChangeLog:

* gcc-interface/Makefile.in: Handle Darwin rpaths.

gcc/jit/ChangeLog:
* Make-lang.in: Handle Darwin rpaths.

libatomic/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libbacktrace/ChangeLog:

* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libcc1/ChangeLog:

* configure: Regenerate.

libffi/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.

libgcc/ChangeLog:

* config/t-slibgcc-darwin: Generate libgcc_s
with an @rpath name.
* config.host: Handle Darwin rpaths.

libgfortran/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths

libgm2/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* aclocal.m4: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.
* libm2cor/Makefile.am: Handle Darwin rpaths.
* libm2cor/Makefile.in: Regenerate.
* libm2iso/Makefile.am: Handle Darwin rpaths.
* libm2iso/Makefile.in: Regenerate.
* libm2log/Makefile.am: Handle Darwin rpaths.
* libm2log/Makefile.in: Regenerate.
* libm2min/Makefile.am: Handle Darwin rpaths.
* libm2min/Makefile.in: Regenerate.
* libm2pim/Makefile.am: Handle Darwin rpaths.
* libm2pim/Makefile.in: Regenerate.

libgomp/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths

libitm/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libobjc/ChangeLog:

* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libphobos/ChangeLog:

* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.
* libdruntime/Makefile.am: Handle Darwin rpaths.
* libdruntime/Makefile.in: Regenerate.
* src/Makefile.am: Handle Darwin rpaths.
* src/Makefile.in: Regenerate.

libquadmath/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libsanitizer/ChangeLog:

* asan/Makefile.am: Handle Darwin rpaths.
* asan/Makefile.in: Regenerate.
* configure: Regenerate.
* hwasan/Makefile.am: Handle Darwin rpaths.
* hwasan/Makefile.in: Regenerate.
* lsan/Makefile.am: Handle Darwin rpaths.
* lsan/Makefile.in: Regenerate.
* tsan/Makefile.am: Handle Darwin rpaths.
* tsan/Makefile.in: Regenerate.
* ubsan/Makefile.am: Handle Darwin rpaths.
* ubsan/Makefile.in: Regenerate.

libssp/ChangeLog:

* Makefile.am: Handle Darwin rpaths.
* Makefile.in: Regenerate.
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

libstdc++-v3/ChangeLog:

* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.
* src/Makefile.am: Handle Darwin rpaths.
* src/Makefile.in: Regenerate.

libvtv/ChangeLog:

* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

lto-plugin/ChangeLog:
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

zlib/ChangeLog:
* configure: Regenerate.
* configure.ac: Handle Darwin rpaths.

Driver: Provide a spec to insert rpaths for compiler lib dirs.

This provides a spec to insert "-rpath DDD" for each DDD corresponding
to a compiler startfile directory.  This allows a target to use @rpath
as the install path for libraries, and have the compiler provide the
necessary rpath to handle this.

Embed real paths, not relative ones.

We embed a runpath for every path in which libraries might be found.  This
change ensures that we embed the actual real path and not a relative one from
the compiler's version-specific directory.

e.g.
/opt/distro/gcc-11-3Dr0/lib

instead of:
/opt/distro/gcc-11-3Dr0/lib/gcc/x86_64-apple-darwin19/11.3.0/../../..

This ensures that if we install, for example, 11.4.0 (and delete the 11.3.0
installation) exes built by 11.3 would continue to function (providing, of course
that 11.4 does not bump any SO names).

gcc/ChangeLog:
* gcc.cc (RUNPATH_OPTION): New.
(do_spec_1): Provide '%P' as a spec to insert rpaths for
each compiler startfile path.

libgcc: support heap-based trampolines

Add support for heap-based trampolines on x86_64-linux, aarch64-linux,
and x86_64-darwin. Implement the __builtin_nested_func_ptr_created and
__builtin_nested_func_ptr_deleted functions for these targets.

Co-Authored-By: Maxim Blinov <maxim.blinov@embecosm.com>
Co-Authored-By: Iain Sandoe <iain@sandoe.co.uk>
Co-Authored-By: Francois-Xavier Coudert <fxcoudert@gcc.gnu.org>
libgcc/ChangeLog:

* libgcc2.h (__builtin_nested_func_ptr_created): Declare.
(__builtin_nested_func_ptr_deleted): Declare.
* libgcc-std.ver.in: Add the new symbols.
* config/aarch64/heap-trampoline.c: Implement heap-based
trampolines for aarch64.
* config/aarch64/t-heap-trampoline: Add rule to build
config/aarch64/heap-trampoline.c
* config/i386/heap-trampoline.c: Implement heap-based
trampolines for x86_64.
* config/i386/t-heap-trampoline: Add rule to build
config/i386/heap-trampoline.cc
* config.host: Handle --enable-heap-trampolines on
x86_64-*-linux*, aarch64-*-linux*, x86_64-*-darwin*.

target: Support heap-based trampolines

Enable -ftrampoline-impl=heap by default if we are on macOS 11
or later.

Co-Authored-By: Maxim Blinov <maxim.blinov@embecosm.com>
Co-Authored-By: Francois-Xavier Coudert <fxcoudert@gcc.gnu.org>
Co-Authored-By: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* config.gcc: Default to heap trampolines on macOS 11 and above.
* config/i386/darwin.h: Define X86_CUSTOM_FUNCTION_TEST.
* config/i386/i386.h: Define X86_CUSTOM_FUNCTION_TEST.
* config/i386/i386.cc: Use X86_CUSTOM_FUNCTION_TEST.

core: Support heap-based trampolines

Generate heap-based nested function trampolines

Add support for allocating nested function trampolines on an
executable heap rather than on the stack. This is motivated by targets
such as AArch64 Darwin, which globally prohibit executing code on the
stack.

The target-specific routines for allocating and writing trampolines are
to be provided in libgcc.

The gcc flag -ftrampoline-impl controls whether to generate code
that instantiates trampolines on the stack, or to emit calls to
__builtin_nested_func_ptr_created and
__builtin_nested_func_ptr_deleted. Note that this flag is completely
independent of libgcc: If libgcc is for any reason missing those
symbols, you will get a link failure.

This implementation imposes some implicit restrictions as compared to
stack trampolines. longjmp'ing back to a state before a trampoline was
created will cause us to skip over the corresponding
__builtin_nested_func_ptr_deleted, which will leak trampolines
starting from the beginning of the linked list of allocated
trampolines. There may be scope for instrumenting longjmp/setjmp to
trigger cleanups of trampolines.

Co-Authored-By: Maxim Blinov <maxim.blinov@embecosm.com>
Co-Authored-By: Iain Sandoe <iain@sandoe.co.uk>
Co-Authored-By: Francois-Xavier Coudert <fxcoudert@gcc.gnu.org>
gcc/ChangeLog:

* builtins.def (BUILT_IN_NESTED_PTR_CREATED): Define.
(BUILT_IN_NESTED_PTR_DELETED): Ditto.
* common.opt (ftrampoline-impl): Add option to control
generation of trampoline instantiation (heap or stack).
* coretypes.h: Define enum trampoline_impl.
* tree-nested.cc (convert_tramp_reference_op): Don't bother calling
__builtin_adjust_trampoline for heap trampolines.
(finalize_nesting_tree_1): Emit calls to
__builtin_nested_...{created,deleted} if we're generating with
-ftrampoline-impl=heap.
* tree.cc (build_common_builtin_nodes): Build
__builtin_nested_...{created,deleted}.
* doc/invoke.texi (-ftrampoline-impl): Document.

RISC-V: Prohibit combination of 'E' and 'H'

According to the ratified privileged specification (version 20211203),
it says:

> The hypervisor extension depends on an "I" base integer ISA with 32 x
> registers (RV32I or RV64I), not RV32E, which has only 16 x registers.

Also in the latest draft, it also prohibits RV64E with the 'H' extension.
This commit prohibits the combination of 'E' and 'H' extensions.

gcc/ChangeLog:

* common/config/riscv/riscv-common.cc (riscv_subset_list::parse):
Prohibit 'E' and 'H' combinations.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/arch-26.c: New test.

RISC-V: 'Zfa' extension is now ratified

Since this extension is ratified, it now has the version number 1.0.

Reference:
<https://github.com/riscv/riscv-isa-manual/pull/1096>

gcc/ChangeLog:

* common/config/riscv/riscv-common.cc (riscv_ext_version_table):
Change version number of the 'Zfa' extension to 1.0.

Daily bump.

libstdc++: Split std::basic_string::_M_use_local_data into two functions

This splits out the activate-the-union-member-for-constexpr logic from
_M_use_local_data, so that it can be used separately in cases that don't
need to use std::pointer_traits<pointer>::pointer_to to obtain the
return value.

This leaves only three uses of _M_use_local_data() which are all of the
same form:

  __s._M_data(_M_use_local_data());
  __s._M_set_length(0);

We could remove _M_use_local_data() and change those three places to use
a new _M_reset() function that does:

  _M_init_local_buf();
  _M_data(_M_local_data());
  _M_set_length(0);

This is left for a future change.

libstdc++-v3/ChangeLog:

* include/bits/basic_string.h (_M_init_local_buf()): New
function.
(_M_use_local_data()): Use _M_init_local_buf.
(basic_string(), basic_string(const Alloc&))
(basic_string(basic_string&&))
(basic_string(basic_string&&, const Alloc&)): Use
_M_init_local_buf instead of _M_use_local_data().
* include/bits/basic_string.tcc (swap(basic_string&))
(_M_construct(InIter, InIter, input_iterator_tag))
(_M_construct(InIter, InIter, forward_iterator_tag))
(_M_construct(size_type, CharT), reserve()): Likewise.

libstdc++: Workaround for LLVM-61763 in <ranges>

This patch adds a small workaround that avoids declaring constrained
friends when compiling with Clang, instead making some members public.
MSVC's standard library has implemented a similar workaround.

Signed-off-by: Benjamin Brock <brock@cs.berkeley.edu>
libstdc++-v3/ChangeLog:

* include/std/ranges (zip_view, adjacent_view): Implement
workaround for LLVM-61763.

libstdc++: testsuite: Enhance codecvt_unicode with tests for length()

We can test codecvt::length() with the same data that we test
codecvt::in(). For each call of in() we add another call to length().
Some additional small cosmentic changes are applied.

libstdc++-v3/ChangeLog:

* testsuite/22_locale/codecvt/codecvt_unicode.h: Test length()

libstdc++: Fix formatting of filesystem directory iterators

Fix indentation.

libstdc++-v3/ChangeLog:

* include/bits/fs_dir.h (operator==(default_sentinel_t)): Fix
indentation.

C99 testsuite readiness: Compile more tests with -std=gnu89

gcc/testsuite/

* gcc.c-torture/compile/20000403-1.c: Compile with -std=gnu89.
* gcc.c-torture/compile/20000511-1.c: Likewise.
* gcc.c-torture/compile/20000804-1.c: Likewise.
* gcc.c-torture/compile/20020418-1.c: Likewise.
* gcc.c-torture/compile/20020927-1.c: Likewise.
* gcc.c-torture/compile/20030109-1.c: Likewise.
* gcc.c-torture/compile/20030224-1.c: Likewise.
* gcc.c-torture/compile/20030415-1.c: Likewise.
* gcc.c-torture/compile/20030612-1.c: Likewise.
* gcc.c-torture/compile/20030917-1.c: Likewise.
* gcc.c-torture/compile/20031113-1.c: Likewise.
* gcc.c-torture/compile/20031220-2.c: Likewise.
* gcc.c-torture/compile/20040309-1.c: Likewise.
* gcc.c-torture/compile/20040310-1.c: Likewise.
* gcc.c-torture/compile/20040317-3.c: Likewise.
* gcc.c-torture/compile/20040817-1.c: Likewise.
* gcc.c-torture/compile/20091215-1.c: Likewise.
* gcc.c-torture/compile/86.c: Likewise.
* gcc.c-torture/compile/900216-1.c: Likewise.
* gcc.c-torture/compile/900313-1.c: Likewise.
* gcc.c-torture/compile/900407-1.c: Likewise.
* gcc.c-torture/compile/900516-1.c: Likewise.
* gcc.c-torture/compile/920409-2.c: Likewise.
* gcc.c-torture/compile/920415-1.c: Likewise.
* gcc.c-torture/compile/920428-1.c: Likewise.
* gcc.c-torture/compile/920428-5.c: Likewise.
* gcc.c-torture/compile/920428-7.c: Likewise.
* gcc.c-torture/compile/920501-1.c: Likewise.
* gcc.c-torture/compile/920501-13.c: Likewise.
* gcc.c-torture/compile/920501-15.c: Likewise.
* gcc.c-torture/compile/920501-16.c: Likewise.
* gcc.c-torture/compile/920501-18.c: Likewise.
* gcc.c-torture/compile/920501-20.c: Likewise.
* gcc.c-torture/compile/920501-6.c: Likewise.
* gcc.c-torture/compile/920501-7.c: Likewise.
* gcc.c-torture/compile/920502-1.c: Likewise.
* gcc.c-torture/compile/920502-2.c: Likewise.
* gcc.c-torture/compile/920520-1.c: Likewise.
* gcc.c-torture/compile/920521-1.c: Likewise.
* gcc.c-torture/compile/920608-1.c: Likewise.
* gcc.c-torture/compile/920617-1.c: Likewise.
* gcc.c-torture/compile/920617-2.c: Likewise.
* gcc.c-torture/compile/920625-1.c: Likewise.
* gcc.c-torture/compile/920625-2.c: Likewise.
* gcc.c-torture/compile/920626-1.c: Likewise.
* gcc.c-torture/compile/920706-1.c: Likewise.
* gcc.c-torture/compile/920710-2.c: Likewise.
* gcc.c-torture/compile/920723-1.c: Likewise.
* gcc.c-torture/compile/920808-1.c: Likewise.
* gcc.c-torture/compile/920809-1.c: Likewise.
* gcc.c-torture/compile/920817-1.c: Likewise.
* gcc.c-torture/compile/920831-1.c: Likewise.
* gcc.c-torture/compile/920917-1.c: Likewise.
* gcc.c-torture/compile/920928-2.c: Likewise.
* gcc.c-torture/compile/920928-5.c: Likewise.
* gcc.c-torture/compile/921012-1.c: Likewise.
* gcc.c-torture/compile/921021-1.c: Likewise.
* gcc.c-torture/compile/921024-1.c: Likewise.
* gcc.c-torture/compile/921103-1.c: Likewise.
* gcc.c-torture/compile/921109-1.c: Likewise.
* gcc.c-torture/compile/921111-1.c: Likewise.
* gcc.c-torture/compile/921116-2.c: Likewise.
* gcc.c-torture/compile/921118-1.c: Likewise.
* gcc.c-torture/compile/921202-1.c: Likewise.
* gcc.c-torture/compile/921202-2.c: Likewise.
* gcc.c-torture/compile/921203-1.c: Likewise.
* gcc.c-torture/compile/921203-2.c: Likewise.
* gcc.c-torture/compile/921206-1.c: Likewise.
* gcc.c-torture/compile/930109-1.c: Likewise.
* gcc.c-torture/compile/930111-1.c: Likewise.
* gcc.c-torture/compile/930117-1.c: Likewise.
* gcc.c-torture/compile/930118-1.c: Likewise.
* gcc.c-torture/compile/930120-1.c: Likewise.
* gcc.c-torture/compile/930217-1.c: Likewise.
* gcc.c-torture/compile/930325-1.c: Likewise.
* gcc.c-torture/compile/930411-1.c: Likewise.
* gcc.c-torture/compile/930427-2.c: Likewise.
* gcc.c-torture/compile/930503-2.c: Likewise.
* gcc.c-torture/compile/930506-2.c: Likewise.
* gcc.c-torture/compile/930513-2.c: Likewise.
* gcc.c-torture/compile/930530-1.c: Likewise.
* gcc.c-torture/compile/930602-1.c: Likewise.
* gcc.c-torture/compile/930618-1.c: Likewise.
* gcc.c-torture/compile/930623-1.c: Likewise.
* gcc.c-torture/compile/931003-1.c: Likewise.
* gcc.c-torture/compile/931013-1.c: Likewise.
* gcc.c-torture/compile/931013-2.c: Likewise.
* gcc.c-torture/compile/931102-2.c: Likewise.
* gcc.c-torture/compile/931203-1.c: Likewise.
* gcc.c-torture/compile/940718-1.c: Likewise.
* gcc.c-torture/compile/941014-1.c: Likewise.
* gcc.c-torture/compile/941014-2.c: Likewise.
* gcc.c-torture/compile/941014-3.c: Likewise.
* gcc.c-torture/compile/941014-4.c: Likewise.
* gcc.c-torture/compile/941111-1.c: Likewise.
* gcc.c-torture/compile/941113-1.c: Likewise.
* gcc.c-torture/compile/950124-1.c: Likewise.
* gcc.c-torture/compile/950329-1.c: Likewise.
* gcc.c-torture/compile/950612-1.c: Likewise.
* gcc.c-torture/compile/950618-1.c: Likewise.
* gcc.c-torture/compile/950719-1.c: Likewise.
* gcc.c-torture/compile/950910-1.c: Likewise.
* gcc.c-torture/compile/950922-1.c: Likewise.
* gcc.c-torture/compile/951106-1.c: Likewise.
* gcc.c-torture/compile/951222-1.c: Likewise.
* gcc.c-torture/compile/960106-1.c: Likewise.
* gcc.c-torture/compile/960319-1.c: Likewise.
* gcc.c-torture/compile/960829-1.c: Likewise.
* gcc.c-torture/compile/970206-1.c: Likewise.
* gcc.c-torture/compile/980825-1.c: Likewise.
* gcc.c-torture/compile/990829-1.c: Likewise.
* gcc.c-torture/compile/991213-2.c: Likewise.

RISC-V: Support partial VLS mode when preference fixed-vlmax [PR111857]

Given we have code like below:

typedef char vnx16i __attribute__ ((vector_size (16)));

vnx16i __attribute__ ((noinline, noclone))
test (vnx16i x, vnx16i y)
{
  return __builtin_shufflevector (x, y, 11, 12, 13, 14, 11, 12, 13, 14,
                                        11, 12, 13, 14, 11, 12, 13, 14);
}

It can perform the auto vectorization when

-march=rv64gcv_zvl1024b --param=riscv-autovec-preference=fixed-vlmax

but cannot when

-march=rv64gcv_zvl2048b --param=riscv-autovec-preference=fixed-vlmax

The reason comes from the miniaml machine mode of QI is RVVMF8QI, which
is 1024 / 8 = 128 bits, aka the size of VNx16QI. When we set zvl2048b,
the bit size of RVVMFQI is 2048 / 8 = 256, which is not matching the
bit size of VNx16QI (128 bits).

Thus, this patch would like to enable the VLS mode for such case, aka
VNx16QI vls mode for zvl2048b.

Before this patch:
test:
  srli    a4,a1,40
  andi    a4,a4,0xff
  srli    a3,a1,32
  srli    a5,a1,48
  slli    a0,a4,8
  andi    a3,a3,0xff
  andi    a5,a5,0xff
  slli    a2,a5,16
  or      a0,a3,a0
  srli    a1,a1,56
  or      a0,a0,a2
  slli    a2,a1,24
  slli    a3,a3,32
  or      a0,a0,a2
  slli    a4,a4,40
  or      a0,a0,a3
  slli    a5,a5,48
  or      a0,a0,a4
  or      a0,a0,a5
  slli    a1,a1,56
  or      a0,a0,a1
  mv      a1,a0
  ret

After this patch:
test:
  vsetivli        zero,16,e8,mf8,ta,ma
  vle8.v  v2,0(a1)
  vsetivli        zero,4,e32,mf2,ta,ma
  vrgather.vi     v1,v2,3
  vsetivli        zero,16,e8,mf8,ta,ma
  vse8.v  v1,0(a0)
  ret

PR target/111857

gcc/ChangeLog:

* config/riscv/riscv-opts.h (TARGET_VECTOR_VLS): Remove.
* config/riscv/riscv-protos.h (vls_mode_valid_p): New func decl.
* config/riscv/riscv-v.cc (autovectorize_vector_modes): Replace
macro reference to func.
(vls_mode_valid_p): New func impl for vls mode valid or not.
* config/riscv/riscv-vector-switch.def (VLS_ENTRY): Replace
macro reference to func.
* config/riscv/vector-iterators.md: Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/vls-vlmax/perm-4.c: Adjust checker.
* gcc.target/riscv/rvv/autovec/vls/def.h: Add help define.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-1.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-2.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-3.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-4.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-5.c: New test.
* gcc.target/riscv/rvv/autovec/vls-vlmax/pr111857-6.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>

Daily bump.

PR 106245: Split (x<<31)>>31 as -(x&1) in i386.md

This patch is the backend piece of a solution to PRs 101955 and 106245,
that adds a define_insn_and_split to the i386 backend, to perform sign
extension of a single (least significant) bit using and $1 then neg.

Previously, (x<<31)>>31 would be generated as

        sall    $31, %eax // 3 bytes
        sarl    $31, %eax // 3 bytes

with this patch the backend now generates:

        andl    $1, %eax // 3 bytes
        negl    %eax // 2 bytes

Not only is this smaller in size, but microbenchmarking confirms
that it's a performance win on both Intel and AMD; Intel sees only a
2% improvement (perhaps just a size effect), but AMD sees a 7% win.

2023-10-21  Roger Sayle  <roger@nextmovesoftware.com>
    Uros Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog
PR middle-end/101955
PR tree-optimization/106245
* config/i386/i386.md (*extv<mode>_1_0): New define_insn_and_split.

gcc/testsuite/ChangeLog
PR middle-end/101955
PR tree-optimization/106245
* gcc.target/i386/pr106245-2.c: New test case.
* gcc.target/i386/pr106245-3.c: New 32-bit test case.
* gcc.target/i386/pr106245-4.c: New 64-bit test case.
* gcc.target/i386/pr106245-5.c: Likewise.

c++: abstract class and overload resolution

In my implementation of P0929 I treated a conversion to an rvalue of
abstract class type as a bad conversion, but that's still too soon to check
it; we need to wait until we're done with overload resolution.

gcc/cp/ChangeLog:

* call.cc (implicit_conversion_1): Rename...
(implicit_conversion): ...to this.  Remove the old wrapper.

gcc/testsuite/ChangeLog:

* g++.dg/template/sfinae-dr657.C: Adjust.

testsuite: constexpr-diag1.C and implicit constexpr

This test doesn't break as expected with implicit constexpr.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/constexpr-diag1.C: Add -fno-implicit-constexpr.

c++: fix tourney logic

In r13-3766 I changed the logic at the end of tourney to avoid redundant
comparisons, but the change also meant skipping any less-good matches
between the champ_compared_to_predecessor candidate and champ itself.

This should not be a correctness issue, since we believe that joust is a
partial order.  But it can lead to missed warnings, as in this testcase.

gcc/cp/ChangeLog:

* call.cc (tourney): Only skip champ_compared_to_predecessor.

gcc/testsuite/ChangeLog:

* g++.dg/warn/Wsign-promo1.C: New test.

c++: Constructor streaming [PR105322]

An expresion node's type is streamed after the expression's operands,
because the type can come from some aspect of an operand (for instance
decltype and noexcept). There's a comment in the code explaining that.

But that doesn't work for constructors, which can directly reference
components of their type (eg FIELD_DECLS). If this is a
type-introducing CONSTRUCTOR, we need to ensure the type has been
streamed first. So move CONSTRUCTOR stream to after the type streaming.

The reason things like COMPONENT_REF work is that they stream their
first operand first, and that introduces the type that their second
operand looks up a field in.

gcc/cp/
PR c++/105322
* module.cc (trees_out::core_vals): Stream CONSTRUCTOR operands
after the type.
(trees_in::core_vals): Likewise.
gcc/testsuite/
* g++.dg/modules/decltype-1_a.C: New.
* g++.dg/modules/decltype-1_b.C: New.
* g++.dg/modules/lambda-5_a.C: New.
* g++.dg/modules/lambda-5_b.C: New.

c: -Wint-conversion should cover pointer/integer mismatches in ?:

gcc/c/

PR c/109827
PR other/44209
* c-typeck.cc (build_conditional_expr): Use OPT_Wint_conversion
for pointer/integer mismatch warnings.

gcc/testsuite/

* gcc.dg/Wint-conversion-3.c: New.

c: -Wincompatible-pointer-types should cover mismatches in ?:

gcc/c/

PR c/109826
PR other/44209
* c-typeck.cc (build_conditional_expr): Use
OPT_Wincompatible_pointer_types for pointer mismatches.
Emit location information for the operand.

gcc/testsuite/

* gcc.dg/Wincompatible-pointer-types-2.c: New.
* gcc.dg/Wincompatible-pointer-types-3.c: New.
* gcc.dg/Wincompatible-pointer-types-4.c: New.

bootstrap: tm_p.h requires memmodel.h on SPARC.

SPARC target header references memmodel, which requires memmodel.h.

gcc/ChangeLog:

* gimple-harden-control-flow.cc: Include memmodel.h.

Signed-off-by: David Edelsohn <dje.gcc@gmail.com>

c-family: char8_t and aliasing in C vs C++ [PR111884]

In the PR, Joseph says that in C char8_t is not a distinct type.  So
we should behave as if it can alias anything, like ordinary char.
In C, unsigned_char_type_node == char8_type_node, so with this patch
we return 0 instead of -1.  And the following comment says:

  /* The C standard guarantees that any object may be accessed via an
     lvalue that has narrow character type (except char8_t).  */
  if (t == char_type_node
      || t == signed_char_type_node
      || t == unsigned_char_type_node)
    return 0;

Which appears to be wrong, so I'm adjusting that as well.

PR c/111884

gcc/c-family/ChangeLog:

* c-common.cc (c_common_get_alias_set): Return -1 for char8_t only
in C++.

gcc/testsuite/ChangeLog:

* c-c++-common/alias-1.c: New test.

bootstrap: Include tm_p.h

ASM_GENERATE_INTERNAL_LABEL may use target-specific functions, so
tm_p.h must be included.  This fixes bootstrap breakage on AIX, at least.

gcc/ChangeLog

* gimple-harden-control-flow.cc: Include tm_p.h.

Signed-off-by: David Edelsohn <dje.gcc@gmail.com>

rust: build failure after NON_DEPENDENT_EXPR removal [PR111899]

This patch removes stray NON_DEPENDENT_EXPR checks following the removal
of this tree code from the C++ FE.  (Since this restores the build I
supppose it means the Rust FE never creates NON_DEPENDENT_EXPR trees in
the first place, so no further analysis is needed.)

PR rust/111899

gcc/rust/ChangeLog:

* backend/rust-constexpr.cc (potential_constant_expression_1):
Remove NON_DEPENDENT_EXPR handling.
* backend/rust-tree.cc (mark_exp_read): Likewise.
(mark_use): Likewise.
(lvalue_kind): Likewise.

libstdc++: add casts to from_chars in <charconv> [PR111883]

This fixes

.../charconv: In function 'std::from_chars_result std::from_chars(const char*, const char*, _Float16&, chars_format)':
.../charconv:687:17: warning: converting to '_Float16' from 'float' with greater conversion rank
  687 |       __value = __val;
      |                 ^~~~~
.../charconv: In function 'std::from_chars_result std::from_chars(const char*, const char*, __gnu_cxx::__bfloat16_t&, chars_format)':
.../charconv:763:17: warning: converting to '__gnu_cxx::__bfloat16_t' {aka '__bf16'} from 'float' with greater conversion rank
  763 |       __value = __val;
      |                 ^~~~~

which was breaking a test:

FAIL: g++.dg/warn/Wstringop-overflow-6.C  -std=gnu++26 (test for excess errors)

PR testsuite/111883

libstdc++-v3/ChangeLog:

* include/std/charconv (from_chars): Add explicit casts.

ifcvt: Don't lower bitfields with non-constant offsets [PR 111882]

This patch stops lowering of bitfields by ifcvt when they have non-constant
offsets as we are not likely to be able to do anything useful with those during
vectorization.  That also fixes the issue reported in PR 111882, which was
being caused by an offset with a side-effect being lowered, but constants have
no side-effects so we will no longer run into that problem.

gcc/ChangeLog:

PR tree-optimization/111882
* tree-if-conv.cc (get_bitfield_rep): Return NULL_TREE for bitfields
with non-constant offsets.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/pr111882.c: New test.

c++: rename tsubst_copy_and_build and tsubst_expr

After the previous patch, we now only have two tsubst entry points for
expression trees: tsubst_copy_and_build and tsubst_expr.  The former
despite its unwieldy name is the main entry point, and the latter is
just a superset of the former that also handles statement trees.  We
could merge them so that we just have tsubst_expr, but it seems natural
to distinguish statement trees from expression trees and to maintain a
separate entry point for them.

To that end, this this patch renames tsubst_copy_and_build to
tsubst_expr, and renames the current tsubst_expr to tsubst_stmt, which
continues to be a superset of the former (which is convenient since
sometimes expression trees appear in statement contexts, e.g. a branch
of an IF_STMT could be NOP_EXPR).  (Making tsubst_stmt disjoint from
tsubst_expr is left as future work if deemed desirable.)

This patch in turn renames suitable existing uses of tsubst_expr (that
expect to take statement trees) to use tsubst_stmt.  Thus untouched
tsubst_expr calls are implicitly strengthened to expect only expression
trees after this patch.  For the tsubst_omp_* routines I opted to rename
all existing uses to ensure no unintended functional change.  This patch
also moves the handling of CO_YIELD_EXPR and CO_AWAIT_EXPR from tsubst_stmt
to tsubst_expr since they're indeed expression trees.

gcc/cp/ChangeLog:

* cp-lang.cc (objcp_tsubst_copy_and_build): Rename to ...
(objcp_tsubst_expr): ... this.
* cp-objcp-common.h (objcp_tsubst_copy_and_build): Rename to ...
(objcp_tsubst_expr): ... this.
* cp-tree.h (tsubst_copy_and_build): Remove declaration.
* init.cc (maybe_instantiate_nsdmi_init): Use tsubst_expr
instead of tsubst_copy_and_build.
* pt.cc (expand_integer_pack): Likewise.
(instantiate_non_dependent_expr_internal): Likewise.
(instantiate_class_template): Use tsubst_stmt instead of
tsubst_expr for STATIC_ASSERT.
(tsubst_function_decl): Adjust tsubst_copy_and_build uses.
(tsubst_arg_types): Likewise.
(tsubst_exception_specification): Likewise.
(tsubst_tree_list): Likewise.
(tsubst): Likewise.
(tsubst_name): Likewise.
(tsubst_omp_clause_decl): Use tsubst_stmt instead of tsubst_expr.
(tsubst_omp_clauses): Likewise.
(tsubst_copy_asm_operands): Adjust tsubst_copy_and_build use.
(tsubst_omp_for_iterator): Use tsubst_stmt instead of tsubst_expr.
(tsubst_expr): Rename to ...
(tsubst_stmt): ... this.
<case CO_YIELD_EXPR, CO_AWAIT_EXPR>: Move to tsubst_expr.
(tsubst_omp_udr): Use tsubst_stmt instead of tsubst_expr.
(tsubst_non_call_postfix_expression): Adjust tsubst_copy_and_build
use.
(tsubst_lambda_expr): Likewise.  Use tsubst_stmt instead of
tsubst_expr for the body of a lambda.
(tsubst_copy_and_build_call_args): Rename to ...
(tsubst_call_args): ... this.  Adjust tsubst_copy_and_build use.
(tsubst_copy_and_build): Rename to tsubst_expr.  Adjust
tsubst_copy_and_build and tsubst_copy_and_build_call_args use.
<case TRANSACTION_EXPR>: Use tsubst_stmt instead of tsubst_expr.
(maybe_instantiate_noexcept): Adjust tsubst_copy_and_build use.
(instantiate_body): Use tsubst_stmt instead of tsubst_expr for
substituting the function body.
(tsubst_initializer_list): Adjust tsubst_copy_and_build use.

gcc/objcp/ChangeLog:

* objcp-lang.cc (objcp_tsubst_copy_and_build): Rename to ...
(objcp_tsubst_expr): ... this.  Adjust tsubst_copy_and_build
uses.

Reviewed-by: Jason Merrill <jason@redhat.com>

c++: merge tsubst_copy into tsubst_copy_and_build

The relationship between tsubst_copy_and_build and tsubst_copy (two of
the main template argument substitution routines for expression trees)
is rather hazy.  The former is mostly a superset of the latter, with
some differences.

The main apparent difference is their handling of various tree codes,
but much of the tree code handling in tsubst_copy appears to be dead
code.  This is because tsubst_copy mostly gets (directly) called on
id-expressions rather than on arbitrary expressions.  The interesting
tree codes are PARM_DECL, VAR_DECL, BIT_NOT_EXPR, SCOPE_REF,
TEMPLATE_ID_EXPR and IDENTIFIER_NODE:

 * for PARM_DECL and VAR_DECL, tsubst_copy_and_build calls tsubst_copy
   followed by doing some extra handling of its own
 * for BIT_NOT_EXPR tsubst_copy implicitly handles unresolved destructor
   calls (i.e. the first operand is an identifier or a type)
 * for SCOPE_REF, TEMPLATE_ID_EXPR and IDENTIFIER_NODE tsubst_copy
   refrains from doing name lookup of the terminal name

Other more minor differences are that tsubst_copy exits early when
'args' is null, and it calls maybe_dependent_member_ref, and finally
it dispatches to tsubst for type trees.[1]

Thus tsubst_copy is similar enough to tsubst_copy_and_build that it
makes sense to merge the two functions, with the main difference we
want to preserve is tsubst_copy's lack of name lookup for id-expressions.
This patch achieves this via a new tsubst flag tf_no_name_lookup which
controls name lookup and resolution of a (top-level) id-expression.

[1]: Exiting early for null 'args' doesn't seem right since it means we
return templated trees even when !processing_template_decl.  And
dispatching to tsubst for type trees muddles the distinction between
type and expressions which makes things less clear at the call site.
So these properties of tsubst_copy don't seem worth preserving.

N.B. the diff for this patch looks much cleaner when generated using
the "patience diff" algorithm via Git's --patience flag.

gcc/cp/ChangeLog:

* cp-tree.h (enum tsubst_flags): Add tf_no_name_lookup.
* pt.cc (tsubst_pack_expansion): Use tsubst for substituting
BASES_TYPE.
(tsubst_decl) <case USING_DECL>: Use tsubst_name instead of
tsubst_copy.
(tsubst) <case TEMPLATE_TYPE_PARM>: Use tsubst_copy_and_build
instead of tsubst_copy for substituting
CLASS_PLACEHOLDER_TEMPLATE.
<case TYPENAME_TYPE>: Use tsubst_name instead of tsubst_copy for
substituting TYPENAME_TYPE_FULLNAME.
(tsubst_name): Define.
(tsubst_qualified_id): Use tsubst_name instead of tsubst_copy
for substituting the component name of a SCOPE_REF.
(tsubst_copy): Remove.
(tsubst_copy_and_build): Clear tf_no_name_lookup at the start,
and remember if it was set.  Call maybe_dependent_member_ref if
tf_no_name_lookup was not set.
<case IDENTIFIER_NODE>: Don't do name lookup if tf_no_name_lookup
was set.
<case TEMPLATE_ID_EXPR>: If tf_no_name_lookup was set, use
tsubst_name instead of tsubst_copy_and_build to substitute the
template and don't finish the template-id.
<case BIT_NOT_EXPR>: Handle identifier and type operand (if
tf_no_name_lookup was set).
<case SCOPE_REF>: Avoid trying to resolve a SCOPE_REF if
tf_no_name_lookup was set by calling build_qualified_name directly
instead of tsubst_qualified_id.
<case SIZEOF_EXPR>: Handling of sizeof...  copied from tsubst_copy.
<case CALL_EXPR>: Use tsubst_name instead of tsubst_copy to
substitute a TEMPLATE_ID_EXPR callee naming an unresolved template.
<case COMPONENT_REF>: Likewise to substitute the member.
<case FUNCTION_DECL>: Copied from tsubst_copy and merged with ...
<case VAR_DECL, PARM_DECL>: ... these.  Initial handling copied
from tsubst_copy.  Optimize local variable substitution by
trying retrieve_local_specialization before checking
uses_template_parms.
<case CONST_DECL>: Copied from tsubst_copy.
<case FIELD_DECL>: Likewise.
<case NAMESPACE_DECL>: Likewise.
<case OVERLOAD>: Likewise.
<case TEMPLATE_DECL>: Likewise.
<case TEMPLATE_PARM_INDEX>: Likewise.
<case TYPE_DECL>: Likewise.
<case CLEANUP_POINT_EXPR>: Likewise.
<case OFFSET_REF>: Likewise.
<case EXPR_PACK_EXPANSION>: Likewise.
<case NONTYPE_ARGUMENT_PACK>: Likewise.
<case *_CST>: Likewise.
<case *_*_FOLD_EXPR>: Likewise.
<case DEBUG_BEGIN_STMT>: Likewise.
<case CO_AWAIT_EXPR>: Likewise.
<case TRAIT_EXPR>: Use tsubst and tsubst_copy_and_build instead
of tsubst_copy.
<default>: Copied from tsubst_copy.
(tsubst_initializer_list): Use tsubst and tsubst_copy_and_build
instead of tsubst_copy.

Reviewed-by: Jason Merrill <jason@redhat.com>

c++: non-static memfn call dependence cleanup [PR106086]

In cp_parser_postfix_expression, and in the CALL_EXPR case of
tsubst_copy_and_build, we essentially repeat the type-dependent and
COMPONENT_REF callee cases of finish_call_expr.  This patch deduplicates
this logic by making both spots consistently go through finish_call_expr.

This allows us to easily fix PR106086 -- which is about us neglecting to
capture 'this' when we resolve a use of a non-static member function of
the current instantiation only at lambda regeneration time -- by moving
the call to maybe_generic_this_capture from the parser to finish_call_expr
so that we consider capturing 'this' at regeneration time as well.

PR c++/106086

gcc/cp/ChangeLog:

* parser.cc (cp_parser_postfix_expression): Consolidate three
calls to finish_call_expr, one to build_new_method_call and
one to build_min_nt_call_vec into one call to finish_call_expr.
Don't call maybe_generic_this_capture here.
* pt.cc (tsubst_copy_and_build) <case CALL_EXPR>: Remove
COMPONENT_REF callee handling.
(type_dependent_expression_p): Use t_d_object_e_p instead of
t_d_e_p for COMPONENT_REF and OFFSET_REF.
* semantics.cc (finish_call_expr): In the type-dependent case,
call maybe_generic_this_capture here instead.

gcc/testsuite/ChangeLog:

* g++.dg/template/crash127.C: Expect additional error due to
being able to check the member access expression ahead of time.
Strengthen the test by not instantiating the class template.
* g++.dg/cpp1y/lambda-generic-this5.C: New test.

c++: remove NON_DEPENDENT_EXPR, part 2

This follow-up patch removes build_non_dependent_expr (and
make_args_non_dependent) and calls thereof, no functional change.

gcc/cp/ChangeLog:

* call.cc (build_new_method_call): Remove calls to
build_non_dependent_expr and/or make_args_non_dependent.
* coroutines.cc (finish_co_return_stmt): Likewise.
* cp-tree.h (build_non_dependent_expr): Remove.
(make_args_non_dependent): Remove.
* decl2.cc (grok_array_decl): Remove calls to
build_non_dependent_expr and/or make_args_non_dependent.
(build_offset_ref_call_from_tree): Likewise.
* init.cc (build_new): Likewise.
* pt.cc (make_args_non_dependent): Remove.
(test_build_non_dependent_expr): Remove.
(cp_pt_cc_tests): Adjust.
* semantics.cc (finish_expr_stmt): Remove calls to
build_non_dependent_expr and/or make_args_non_dependent.
(finish_for_expr): Likewise.
(finish_call_expr): Likewise.
(finish_omp_atomic): Likewise.
* typeck.cc (finish_class_member_access_expr): Likewise.
(build_x_indirect_ref): Likewise.
(build_x_binary_op): Likewise.
(build_x_array_ref): Likewise.
(build_x_vec_perm_expr): Likewise.
(build_x_shufflevector): Likewise.
(build_x_unary_op): Likewise.
(cp_build_addressof): Likewise.
(build_x_conditional_expr): Likewise.
(build_x_compound_expr): Likewise.
(build_static_cast): Likewise.
(build_x_modify_expr): Likewise.
(check_return_expr): Likewise.
* typeck2.cc (build_x_arrow): Likewise.

Reviewed-by: Jason Merrill <jason@redhat.com>

c++: remove NON_DEPENDENT_EXPR, part 1

This tree code dates all the way back to r69130[1] which implemented
typing of non-dependent expressions.  Its motivation was never clear (to
me at least) since its documentation in e.g. cp-tree.def doesn't seem
accurate anymore.  build_non_dependent_expr has since gained a bunch of
edge cases about whether or how to wrap certain templated trees, making
it hard to reason about in general.

So this patch removes this tree code, and temporarily turns
build_non_dependent_expr into the identity function.  The subsequent
patch will remove build_non_dependent_expr and adjust its callers
appropriately.

We now need to more thoroughly handle templated (sub)trees in a couple
of places which previously didn't need to since they didn't look through
NON_DEPENDENT_EXPR.

[1]: https://gcc.gnu.org/pipermail/gcc-patches/2003-July/109355.html

gcc/c-family/ChangeLog:

* c-warn.cc (check_address_or_pointer_of_packed_member): Handle
type-dependent callee of CALL_EXPR.

gcc/cp/ChangeLog:

* class.cc (instantiate_type): Remove NON_DEPENDENT_EXPR
handling.
* constexpr.cc (cxx_eval_constant_expression): Likewise.
(potential_constant_expression_1): Likewise.
* coroutines.cc (coro_validate_builtin_call): Don't
expect ALIGNOF_EXPR to be wrapped in NON_DEPENDENT_EXPR.
* cp-objcp-common.cc (cp_common_init_ts): Remove
NON_DEPENDENT_EXPR handling.
* cp-tree.def (NON_DEPENDENT_EXPR): Remove.
* cp-tree.h (build_non_dependent_expr): Temporarily redefine as
the identity function.
* cvt.cc (maybe_warn_nodiscard): Handle type-dependent and
variable callee of CALL_EXPR.
* cxx-pretty-print.cc (cxx_pretty_printer::expression): Remove
NON_DEPENDENT_EXPR handling.
* error.cc (dump_decl): Likewise.
(dump_expr): Likewise.
* expr.cc (mark_use): Likewise.
(mark_exp_read): Likewise.
* pt.cc (build_non_dependent_expr): Remove.
* tree.cc (lvalue_kind): Remove NON_DEPENDENT_EXPR handling.
(cp_stabilize_reference): Likewise.
* typeck.cc (warn_for_null_address): Likewise.
(cp_build_binary_op): Handle type-dependent SIZEOF_EXPR operands.
(cp_build_unary_op) <case TRUTH_NOT_EXPR>: Don't fold inside a
template.

gcc/testsuite/ChangeLog:

* g++.dg/concepts/var-concept3.C: Adjust expected diagnostic
for attempting to call a variable concept.

Reviewed-by: Jason Merrill <jason@redhat.com>

middle-end: don't pass loop_vinfo to vect_set_loop_condition during prolog peeling

During the refactoring I had passed loop_vinfo on to vect_set_loop_condition
during prolog peeling.  This parameter is unused in most cases except for in
vect_set_loop_condition_partial_vectors where it's behaviour depends on whether
loop_vinfo is NULL or not.  Apparently this code expect it to be NULL and it
reads the structures from a different location.

This fixes the failing testcase which was not using the lens values determined
earlier in vectorizable_store because it was looking it up in the given
loop_vinfo instead.

gcc/ChangeLog:

PR tree-optimization/111866
* tree-vect-loop-manip.cc (vect_do_peeling): Pass null as vinfo to
vect_set_loop_condition during prolog peeling.

tree-optimization/111383 - testcase for fixed PR

PR tree-optimization/111383
PR tree-optimization/110243
gcc/testsuite/
* gcc.dg/torture/pr111383.c: New testcase.

tree-optimization/111445 - simple_iv simplification fault

The following fixes a missed check in the simple_iv attempt
to simplify (signed T)((unsigned T) base + step) where it
allows a truncating inner conversion leading to wrong code.

PR tree-optimization/111445
* tree-scalar-evolution.cc (simple_iv_with_niters):
Add missing check for a sign-conversion.

* gcc.dg/torture/pr111445.c: New testcase.

tree-optimization/110243 - IVOPTs introducing undefined overflow

The following addresses IVOPTs rewriting expressions in its
strip_offset without caring for definedness of overflow.  Rather
than the earlier attempt of just using the proper
split_constant_offset from data-ref analysis the following adjusts
IVOPTs helper trying to minimize changes from this fix, possibly
easing backports.

PR tree-optimization/110243
PR tree-optimization/111336
* tree-ssa-loop-ivopts.cc (strip_offset_1): Rewrite
operations with undefined behavior on overflow to
unsigned arithmetic.

* gcc.dg/torture/pr110243.c: New testcase.
* gcc.dg/torture/pr111336.c: Likewise.

tree-optimization/111891 - fix assert in vectorizable_simd_clone_call

The following fixes the assert in vectorizable_simd_clone_call to
assert we have a vector type during transform.  Whether we have
one during analysis depends on whether another SLP user decided
on the type of a constant/external already.  When we end up with
a mismatch in desire the updating will fail and make vectorization
fail.

PR tree-optimization/111891
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Fix
assert.

* gfortran.dg/pr111891.f90: New testcase.

amdgcn: add -march=gfx1030 EXPERIMENTAL

Accept the architecture configure option and resolve build failures.  This is
enough to build binaries, but I've not got a device to test it on, so there
are probably runtime issues to fix.  The cache control instructions might be
unsafe (or too conservative), and the kernel metadata might be off.  Vector
reductions will need to be reworked for RDNA2.  In principle, it would be
better to use wavefrontsize32 for this architecture, but that would mean
switching everything to allow SImode masks, so wavefrontsize64 it is.

The multilib is not included in the default configuration so either configure
--with-arch=gfx1030 or include it in --with-multilib-list=gfx1030,....

The majority of this patch has no effect on other devices, but changing from
using scalar writes for the exit value to vector writes means we don't need
the scalar cache write-back instruction anywhere (which doesn't exist in RDNA2).

gcc/ChangeLog:

* config.gcc: Allow --with-arch=gfx1030.
* config/gcn/gcn-hsa.h (NO_XNACK): gfx1030 does not support xnack.
(ASM_SPEC): gfx1030 needs -mattr=+wavefrontsize64 set.
* config/gcn/gcn-opts.h (enum processor_type): Add PROCESSOR_GFX1030.
(TARGET_GFX1030): New.
(TARGET_RDNA2): New.
* config/gcn/gcn-valu.md (@dpp_move<mode>): Disable for RDNA2.
(addc<mode>3<exec_vcc>): Add RDNA2 syntax variant.
(subc<mode>3<exec_vcc>): Likewise.
(<convop><mode><vndi>2_exec): Add RDNA2 alternatives.
(vec_cmp<mode>di): Likewise.
(vec_cmp<u><mode>di): Likewise.
(vec_cmp<mode>di_exec): Likewise.
(vec_cmp<u><mode>di_exec): Likewise.
(vec_cmp<mode>di_dup): Likewise.
(vec_cmp<mode>di_dup_exec): Likewise.
(reduc_<reduc_op>_scal_<mode>): Disable for RDNA2.
(*<reduc_op>_dpp_shr_<mode>): Likewise.
(*plus_carry_dpp_shr_<mode>): Likewise.
(*plus_carry_in_dpp_shr_<mode>): Likewise.
* config/gcn/gcn.cc (gcn_option_override): Recognise gfx1030.
(gcn_global_address_p): RDNA2 only allows smaller offsets.
(gcn_addr_space_legitimate_address_p): Likewise.
(gcn_omp_device_kind_arch_isa): Recognise gfx1030.
(gcn_expand_epilogue): Use VGPRs instead of SGPRs.
(output_file_start): Configure gfx1030.
* config/gcn/gcn.h (TARGET_CPU_CPP_BUILTINS): Add __RDNA2__;
(ASSEMBLER_DIALECT): New.
* config/gcn/gcn.md (rdna): New define_attr.
(enabled): Use "rdna" attribute.
(gcn_return): Remove s_dcache_wb.
(addcsi3_scalar): Add RDNA2 syntax variant.
(addcsi3_scalar_zero): Likewise.
(addptrdi3): Likewise.
(mulsi3): v_mul_lo_i32 should be v_mul_lo_u32 on all ISA.
(*memory_barrier): Add RDNA2 syntax variant.
(atomic_load<mode>): Add RDNA2 cache control variants, and disable
scalar atomics for RDNA2.
(atomic_store<mode>): Likewise.
(atomic_exchange<mode>): Likewise.
* config/gcn/gcn.opt (gpu_type): Add gfx1030.
* config/gcn/mkoffload.cc (EF_AMDGPU_MACH_AMDGCN_GFX1030): New.
(main): Recognise -march=gfx1030.
* config/gcn/t-omp-device: Add gfx1030 isa.

libgcc/ChangeLog:

* config/gcn/amdgcn_veclib.h (CDNA3_PLUS): Set false for __RDNA2__.

libgomp/ChangeLog:

* plugin/plugin-gcn.c (EF_AMDGPU_MACH_AMDGCN_GFX1030): New.
(isa_hsa_name): Recognise gfx1030.
(isa_code): Likewise.
* team.c (defined): Remove s_endpgm.

tree-optimization/111000 - restrict invariant motion of shifts

The following restricts moving variable shifts to when they are
always executed in the loop as we currently do not have an efficient
way to rewrite them to something that is unconditionally
well-defined and value range analysis will otherwise compute
invalid ranges for the shift operand.

PR tree-optimization/111000
* stor-layout.h (element_precision): Move ..
* tree.h (element_precision): .. here.
* tree-ssa-loop-im.cc (movement_possibility_1): Restrict
motion of shifts and rotates.

* gcc.dg/torture/pr111000.c: New testcase.

Control flow redundancy hardening

This patch introduces an optional hardening pass to catch unexpected
execution flows.  Functions are transformed so that basic blocks set a
bit in an automatic array, and (non-exceptional) function exit edges
check that the bits in the array represent an expected execution path
in the CFG.

Functions with multiple exit edges, or with too many blocks, call an
out-of-line checker builtin implemented in libgcc.  For simpler
functions, the verification is performed in-line.

-fharden-control-flow-redundancy enables the pass for eligible
functions, --param hardcfr-max-blocks sets a block count limit for
functions to be eligible, and --param hardcfr-max-inline-blocks
tunes the "too many blocks" limit for in-line verification.
-fhardcfr-skip-leaf makes leaf functions non-eligible.

Additional -fhardcfr-check-* options are added to enable checking at
exception escape points, before potential sibcalls, hereby dubbed
returning calls, and before noreturn calls and exception raises.  A
notable case is the distinction between noreturn calls expected to
throw and those expected to terminate or loop forever: the default
setting for -fhardcfr-check-noreturn-calls, no-xthrow, performs
checking before the latter, but the former only gets checking in the
exception handler.  GCC can only tell between them by explicit marking
noreturn functions expected to raise with the newly-introduced
expected_throw attribute, and corresponding ECF_XTHROW flag.

for  gcc/ChangeLog

* tree-core.h (ECF_XTHROW): New macro.
* tree.cc (set_call_expr): Add expected_throw attribute when
ECF_XTHROW is set.
(build_common_builtin_node): Add ECF_XTHROW to
__cxa_end_cleanup and _Unwind_Resume or _Unwind_SjLj_Resume.
* calls.cc (flags_from_decl_or_type): Check for expected_throw
attribute to set ECF_XTHROW.
* gimple.cc (gimple_build_call_from_tree): Propagate
ECF_XTHROW from decl flags to gimple call...
(gimple_call_flags): ... and back.
* gimple.h (GF_CALL_XTHROW): New gf_mask flag.
(gimple_call_set_expected_throw): New.
(gimple_call_expected_throw_p): New.
* Makefile.in (OBJS): Add gimple-harden-control-flow.o.
* builtins.def (BUILT_IN___HARDCFR_CHECK): New.
* common.opt (fharden-control-flow-redundancy): New.
(-fhardcfr-check-returning-calls): New.
(-fhardcfr-check-exceptions): New.
(-fhardcfr-check-noreturn-calls=*): New.
(Enum hardcfr_check_noreturn_calls): New.
(fhardcfr-skip-leaf): New.
* doc/invoke.texi: Document them.
(hardcfr-max-blocks, hardcfr-max-inline-blocks): New params.
* flag-types.h (enum hardcfr_noret): New.
* gimple-harden-control-flow.cc: New.
* params.opt (-param=hardcfr-max-blocks=): New.
(-param=hradcfr-max-inline-blocks=): New.
* passes.def (pass_harden_control_flow_redundancy): Add.
* tree-pass.h (make_pass_harden_control_flow_redundancy):
Declare.
* doc/extend.texi: Document expected_throw attribute.

for  gcc/ada/ChangeLog

* gcc-interface/trans.cc (gigi): Mark __gnat_reraise_zcx with
ECF_XTHROW.
(build_raise_check): Likewise for all rcheck subprograms.

for  gcc/c-family/ChangeLog

* c-attribs.cc (handle_expected_throw_attribute): New.
(c_common_attribute_table): Add expected_throw.

for  gcc/cp/ChangeLog

* decl.cc (push_throw_library_fn): Mark with ECF_XTHROW.
* except.cc (build_throw): Likewise __cxa_throw,
_ITM_cxa_throw, __cxa_rethrow.

for  gcc/testsuite/ChangeLog

* c-c++-common/torture/harden-cfr.c: New.
* c-c++-common/harden-cfr-noret-never-O0.c: New.
* c-c++-common/torture/harden-cfr-noret-never.c: New.
* c-c++-common/torture/harden-cfr-noret-noexcept.c: New.
* c-c++-common/torture/harden-cfr-noret-nothrow.c: New.
* c-c++-common/torture/harden-cfr-noret.c: New.
* c-c++-common/torture/harden-cfr-notail.c: New.
* c-c++-common/torture/harden-cfr-returning.c: New.
* c-c++-common/torture/harden-cfr-tail.c: New.
* c-c++-common/torture/harden-cfr-abrt-always.c: New.
* c-c++-common/torture/harden-cfr-abrt-never.c: New.
* c-c++-common/torture/harden-cfr-abrt-no-xthrow.c: New.
* c-c++-common/torture/harden-cfr-abrt-nothrow.c: New.
* c-c++-common/torture/harden-cfr-abrt.c: New.
* c-c++-common/torture/harden-cfr-always.c: New.
* c-c++-common/torture/harden-cfr-never.c: New.
* c-c++-common/torture/harden-cfr-no-xthrow.c: New.
* c-c++-common/torture/harden-cfr-nothrow.c: New.
* c-c++-common/torture/harden-cfr-bret-always.c: New.
* c-c++-common/torture/harden-cfr-bret-never.c: New.
* c-c++-common/torture/harden-cfr-bret-noopt.c: New.
* c-c++-common/torture/harden-cfr-bret-noret.c: New.
* c-c++-common/torture/harden-cfr-bret-no-xthrow.c: New.
* c-c++-common/torture/harden-cfr-bret-nothrow.c: New.
* c-c++-common/torture/harden-cfr-bret-retcl.c: New.
* c-c++-common/torture/harden-cfr-bret.c: New.
* g++.dg/harden-cfr-throw-always-O0.C: New.
* g++.dg/harden-cfr-throw-returning-O0.C: New.
* g++.dg/torture/harden-cfr-noret-always-no-nothrow.C: New.
* g++.dg/torture/harden-cfr-noret-never-no-nothrow.C: New.
* g++.dg/torture/harden-cfr-noret-no-nothrow.C: New.
* g++.dg/torture/harden-cfr-throw-always.C: New.
* g++.dg/torture/harden-cfr-throw-never.C: New.
* g++.dg/torture/harden-cfr-throw-no-xthrow.C: New.
* g++.dg/torture/harden-cfr-throw-no-xthrow-expected.C: New.
* g++.dg/torture/harden-cfr-throw-nothrow.C: New.
* g++.dg/torture/harden-cfr-throw-nocleanup.C: New.
* g++.dg/torture/harden-cfr-throw-returning.C: New.
* g++.dg/torture/harden-cfr-throw.C: New.
* gcc.dg/torture/harden-cfr-noret-no-nothrow.c: New.
* gcc.dg/torture/harden-cfr-tail-ub.c: New.
* gnat.dg/hardcfr.adb: New.

for  libgcc/ChangeLog

* Makefile.in (LIB2ADD): Add hardcfr.c.
* hardcfr.c: New.

rtl-ssa: Don't leave NOTE_INSN_DELETED around

This patch tweaks change_insns to also call ::remove_insn to ensure the
underlying RTL insn gets removed from the insn chain in the case of a
deletion.

This avoids leaving NOTE_INSN_DELETED around after deleting insns.

For movement, the RTL insn chain is updated earlier in change_insns with
the call to move_insn.  For deletion, it seems reasonable to do it here.

gcc/ChangeLog:

* rtl-ssa/changes.cc (function_info::change_insns): Ensure we call
::remove_insn on deleted insns.

Document {L,R}ROTATE_EXPR

The following amends the {L,R}SHIFT_EXPR documentation with
documentation about the {L,R}ROTATE_EXPR case.

* doc/generic.texi ({L,R}ROTATE_EXPR): Document.

SH: Fix PR 101177

Fix accidentally inverted comparison.

gcc/ChangeLog:

PR target/101177
* config/sh/sh.md (unnamed split pattern): Fix comparison of
find_regno_note result.

Rewrite more refs for epilogue vectorization

The following makes sure to rewrite all gather/scatter detected by
dataref analysis plus stmts classified as VMAT_GATHER_SCATTER.  Maybe
we need to rewrite all refs, the following covers the cases I've
run into now.

* tree-vect-loop.cc (update_epilogue_loop_vinfo): Rewrite
both STMT_VINFO_GATHER_SCATTER_P and VMAT_GATHER_SCATTER
stmt refs.

Fixup vect_get_and_check_slp_defs for gathers and .MASK_LOAD

I went a little bit too simple with implementing SLP gather support
for emulated and builtin based gathers.  The following fixes the
conflict that appears when running into .MASK_LOAD where we rely
on vect_get_operand_map and the bolted-on STMT_VINFO_GATHER_SCATTER_P
checking wrecks that.  The following properly integrates this with
vect_get_operand_map, adding another special index refering to
the vect_check_gather_scatter analyzed offset.

This unbreaks aarch64 (and hopefully riscv), I'll followup with
more fixes and testsuite coverage for x86 where I think I got
masked gather SLP support wrong.

* tree-vect-slp.cc (off_map, off_op0_map, off_arg2_map,
off_arg3_arg2_map): New.
(vect_get_operand_map): Get flag whether the stmt was
recognized as gather or scatter and use the above
accordingly.
(vect_get_and_check_slp_defs): Adjust.
(vect_build_slp_tree_2): Likewise.

omp_lib.f90.in: Deprecate omp_lock_hint_* for OpenMP 5.0

The omp_lock_hint_* parameters were deprecated in favor of
omp_sync_hint_*.  While omp.h contained deprecation markers for those,
the omp_lib module only contained them for omp_{g,s}_nested.

Note: The -Wdeprecated-declarations warning will only become active once
openmp_version / _OPENMP is bumped from 201511 (4.5) to 201811 (5.0).

libgomp/ChangeLog:

* omp_lib.f90.in: Tag omp_lock_hint_* as being deprecated when
_OPENMP >= 201811.

RISC-V: Rename some variables of vector_block_info[NFC]

1. Remove "m_" prefix as they are not private members.
2. Rename infos -> local_infos, info -> global_info to clarify their meaning.

Pushed as it is obvious.

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (pre_vsetvl::fuse_local_vsetvl_info): Rename variables.
(pre_vsetvl::pre_global_vsetvl_info): Ditto.
(pre_vsetvl::emit_vsetvl): Ditto.

ifcvt: Support bitfield lowering of multiple-exit loops

With the patch enabling the vectorization of early-breaks, we'd like to allow
bitfield lowering in such loops, which requires the relaxation of allowing
multiple exits when doing so.  In order to avoid a similar issue to PR107275,
the code that rejects loops with certain types of gimple_stmts was hoisted from
'if_convertible_loop_p_1' to 'get_loop_body_in_if_conv_order', to avoid trying
to lower bitfields in loops we are not going to vectorize anyway.

This also ensures 'ifcvt_local_dec' doesn't accidentally remove statements it
shouldn't as it will never come across them.  I made sure to add a comment to
make clear that there is a direct connection between the two and if we were to
enable vectorization of any other gimple statement we should make sure both
handle it.

gcc/ChangeLog:

* tree-if-conv.cc (if_convertible_loop_p_1): Move check from here ...
(get_loop_body_if_conv_order): ... to here.
(if_convertible_loop_p): Remove single_exit check.
(tree_if_conversion): Move single_exit check to if-conversion part and
support multiple exits.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/vect-bitfield-read-1-not.c: New test.
* gcc.dg/vect/vect-bitfield-read-2-not.c: New test.
* gcc.dg/vect/vect-bitfield-read-8.c: New test.
* gcc.dg/vect/vect-bitfield-read-9.c: New test.

Co-Authored-By: Andre Vieira <andre.simoesdiasvieira@arm.com>

middle-end: Enable bit-field vectorization to work correctly when we're vectoring inside conds

The bitfield vectorization support does not currently recognize bitfields inside
gconds. This means they can't be used as conditions for early break
vectorization which is a functionality we require.

This adds support for them by explicitly matching and handling gcond as a
source.

Testcases are added in the testsuite update patch as the only way to get there
is with the early break vectorization.   See tests:

  - vect-early-break_20.c
  - vect-early-break_21.c

gcc/ChangeLog:

* tree-vect-patterns.cc (vect_init_pattern_stmt): Copy STMT_VINFO_TYPE
from original statement.
(vect_recog_bitfield_ref_pattern): Support bitfields in gcond.

Co-Authored-By: Andre Vieira <andre.simoesdiasvieira@arm.com>

Fix testcases that are raised by support -mevex512

Hi, all

This patch aims to fix some scan-asm fail of pr89229-{5,6,7}b.c since we emit
scalar vmov{s,d} here, when trying to use x/ymm 16+ w/o avx512vl but with
avx512f+evex512.

If everyone has no objection to the modification of this behavior, then we tend
to solve these failures by modifying these testcases.

BRs,
Lin

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr89229-5b.c: Modify test.
* gcc.target/i386/pr89229-6b.c: Ditto.
* gcc.target/i386/pr89229-7b.c: Ditto.

RISC-V: Fix failed hoist in LICM of vmv.v.x instruction

Confirm dynamic LMUL algorithm works well for choosing LMUL = 4 for the PR:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111848

But it generate horrible register spillings.

The root cause is that we didn't hoist the vmv.v.x outside the loop which
increase the SLP loop register pressure.

So, change the COSNT_VECTOR move into vec_duplicate splitter that we can gain better optimizations:

1. better LICM.
2. More opportunities of transforming 'vv' into 'vx' in the future.

Before this patch:

f3:
        ble     a4,zero,.L8
        csrr    t0,vlenb
        slli    t1,t0,4
        csrr    a6,vlenb
        sub     sp,sp,t1
        csrr    a5,vlenb
        slli    a6,a6,3
        slli    a5,a5,2
        add     a6,a6,sp
        vsetvli a7,zero,e16,m8,ta,ma
        slli    a4,a4,3
        vid.v   v8
        addi    t6,a5,-1
        vand.vi v8,v8,-2
        neg     t5,a5
        vs8r.v  v8,0(sp)
        vadd.vi v8,v8,1
        vs8r.v  v8,0(a6)
        j       .L4
.L12:
        vsetvli a7,zero,e16,m8,ta,ma
.L4:
        csrr    t0,vlenb
        slli    t0,t0,3
        vl8re16.v       v16,0(sp)
        add     t0,t0,sp
        vmv.v.x v8,t6
        mv      t1,a4
        vand.vv v24,v16,v8
        mv      a6,a4
        vl8re16.v       v16,0(t0)
        vand.vv v8,v16,v8
        bleu    a4,a5,.L3
        mv      a6,a5
.L3:
        vsetvli zero,a6,e8,m4,ta,ma
        vle8.v  v20,0(a2)
        vle8.v  v16,0(a3)
        vsetvli a7,zero,e8,m4,ta,ma
        vrgatherei16.vv v4,v20,v24
        vadd.vv v4,v16,v4
        vsetvli zero,a6,e8,m4,ta,ma
        vse8.v  v4,0(a0)
        vle8.v  v20,0(a2)
        vsetvli a7,zero,e8,m4,ta,ma
        vrgatherei16.vv v4,v20,v8
        vadd.vv v4,v4,v16
        vsetvli zero,a6,e8,m4,ta,ma
        vse8.v  v4,0(a1)
        add     a4,a4,t5
        add     a0,a0,a5
        add     a3,a3,a5
        add     a1,a1,a5
        add     a2,a2,a5
        bgtu    t1,a5,.L12
        csrr    t0,vlenb
        slli    t1,t0,4
        add     sp,sp,t1
        jr      ra
.L8:
        ret

After this patch:

f3:
ble a4,zero,.L6
csrr a6,vlenb
csrr a5,vlenb
slli a6,a6,2
slli a5,a5,2
addi a6,a6,-1
slli a4,a4,3
neg t5,a5
vsetvli t1,zero,e16,m8,ta,ma
vmv.v.x v24,a6
vid.v v8
vand.vi v8,v8,-2
vadd.vi v16,v8,1
vand.vv v8,v8,v24
vand.vv v16,v16,v24
.L4:
mv t1,a4
mv a6,a4
bleu a4,a5,.L3
mv a6,a5
.L3:
vsetvli zero,a6,e8,m4,ta,ma
vle8.v v28,0(a2)
vle8.v v24,0(a3)
vsetvli a7,zero,e8,m4,ta,ma
vrgatherei16.vv v4,v28,v8
vadd.vv v4,v24,v4
vsetvli zero,a6,e8,m4,ta,ma
vse8.v v4,0(a0)
vle8.v v28,0(a2)
vsetvli a7,zero,e8,m4,ta,ma
vrgatherei16.vv v4,v28,v16
vadd.vv v4,v4,v24
vsetvli zero,a6,e8,m4,ta,ma
vse8.v v4,0(a1)
add a4,a4,t5
add a0,a0,a5
add a3,a3,a5
add a1,a1,a5
add a2,a2,a5
bgtu t1,a5,.L4
.L6:
ret

Note that this patch triggers multiple FAILs:
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-3.c execution test
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-3.c execution test
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-4.c execution test
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-4.c execution test
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-8.c execution test
FAIL: gcc.target/riscv/rvv/autovec/cond/cond_arith_run-8.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_load_run-1.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_load_run-1.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_load_run-2.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_load_run-2.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_store_run-1.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_store_run-1.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_store_run-2.c execution test
FAIL: gcc.target/riscv/rvv/autovec/gather-scatter/strided_store_run-2.c execution test

They failed are all because of bugs on VSETVL PASS:

10dd4:       0c707057                vsetvli zero,zero,e8,mf2,ta,ma
   10dd8:       5e06b8d7                vmv.v.i v17,13
   10ddc:       9ed030d7                vmv1r.v v1,v13
   10de0:       b21040d7                vncvt.x.x.w     v1,v1           ----> raise illegal instruction since we don't have SEW = 8 -> SEW = 4 narrowing.
   10de4:       5e0785d7                vmv.v.v v11,v15

Confirm the recent VSETVL refactor patch: https://gcc.gnu.org/pipermail/gcc-patches/2023-October/633231.html fixed all of them.

So this patch should be committed after the VSETVL refactor patch.

PR target/111848

gcc/ChangeLog:

* config/riscv/riscv-selftests.cc (run_const_vector_selftests): Adapt selftest.
* config/riscv/riscv-v.cc (expand_const_vector): Change it into vec_duplicate splitter.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/costmodel/riscv/rvv/dynamic-lmul2-7.c: Adapt test.
* gcc.dg/vect/costmodel/riscv/rvv/pr111848.c: New test.

RISC-V: Refactor and cleanup vsetvl pass

This patch refactors and cleanups the vsetvl pass in order to make the code
easier to modify and understand. This patch does several things:

1. Introducing a virtual CFG for vsetvl infos and Phase 1, 2 and 3 only maintain
   and modify this virtual CFG. Phase 4 performs insertion, modification and
   deletion of vsetvl insns based on the virtual CFG. The basic block in the
   virtual CFG is called vsetvl_block_info and the vsetvl information inside
   is called vsetvl_info.
2. Combine Phase 1 and 2 into a single Phase 1 and unified the demand system,
   this phase only fuse local vsetvl info in forward direction.
3. Refactor Phase 3, change the logic for determining whether to uplift vsetvl
   info to a pred basic block to a more unified method that there is a vsetvl
   info in the vsetvl defintion reaching in compatible with it.
4. Place all modification operations to the RTL in Phase 4 and Phase 5.
   Phase 4 is responsible for inserting, modifying and deleting vsetvl
   instructions based on fully optimized vsetvl infos. Phase 5 removes the avl
   operand from the RVV instruction and removes the unused dest operand
   register from the vsetvl insns.

These modifications resulted in some testcases needing to be updated. The reasons
for updating are summarized below:

1. more optimized
   vlmax_back_prop-{25,26}.c
   vlmax_conflict-{3,12}.c/vsetvl-{13,23}.c/vsetvl-23.c/
   avl_single-{23,84,95}.c/pr109773-1.c
2. less unnecessary fusion
   avl_single-46.c/imm_bb_prop-1.c/pr109743-2.c/vsetvl-18.c
3. local fuse direction (backward -> forward)
   scalar_move-1.c
4. add some bugfix testcases.
   pr111037-{3,4}.c/pr111037-4.c
   avl_single-{89,104,105,106,107,108,109}.c

PR target/111037
PR target/111234
PR target/111725

gcc/ChangeLog:

* config/riscv/riscv-vsetvl.cc (bitmap_union_of_preds_with_entry): New.
(debug): Removed.
(compute_reaching_defintion): New.
(enum vsetvl_type): Moved.
(vlmax_avl_p): Moved.
(enum emit_type): Moved.
(vlmul_to_str): Moved.
(vlmax_avl_insn_p): Removed.
(policy_to_str): Moved.
(loop_basic_block_p): Removed.
(valid_sew_p): Removed.
(vsetvl_insn_p): Moved.
(vsetvl_vtype_change_only_p): Removed.
(after_or_same_p): Removed.
(before_p): Removed.
(anticipatable_occurrence_p): Removed.
(available_occurrence_p): Removed.
(insn_should_be_added_p): Removed.
(get_all_sets): Moved.
(get_same_bb_set): Moved.
(gen_vsetvl_pat): Removed.
(calculate_vlmul): Moved.
(get_max_int_sew): New.
(emit_vsetvl_insn): Removed.
(get_max_float_sew): New.
(eliminate_insn): Removed.
(insert_vsetvl): Removed.
(count_regno_occurrences): Moved.
(get_vl_vtype_info): Removed.
(enum def_type): Moved.
(validate_change_or_fail): Moved.
(change_insn): Removed.
(get_all_real_uses): Moved.
(get_forward_read_vl_insn): Removed.
(get_backward_fault_first_load_insn): Removed.
(change_vsetvl_insn): Removed.
(avl_source_has_vsetvl_p): Removed.
(source_equal_p): Moved.
(calculate_sew): Removed.
(same_equiv_note_p): Moved.
(get_expr_id): New.
(incompatible_avl_p): Removed.
(get_regno): New.
(different_sew_p): Removed.
(get_bb_index): New.
(different_lmul_p): Removed.
(has_no_uses): Moved.
(different_ratio_p): Removed.
(different_tail_policy_p): Removed.
(different_mask_policy_p): Removed.
(possible_zero_avl_p): Removed.
(enum demand_flags): New.
(second_ratio_invalid_for_first_sew_p): Removed.
(second_ratio_invalid_for_first_lmul_p): Removed.
(enum class): New.
(float_insn_valid_sew_p): Removed.
(second_sew_less_than_first_sew_p): Removed.
(first_sew_less_than_second_sew_p): Removed.
(class vsetvl_info): New.
(compare_lmul): Removed.
(second_lmul_less_than_first_lmul_p): Removed.
(second_ratio_less_than_first_ratio_p): Removed.
(DEF_INCOMPATIBLE_COND): Removed.
(greatest_sew): Removed.
(first_sew): Removed.
(second_sew): Removed.
(first_vlmul): Removed.
(second_vlmul): Removed.
(first_ratio): Removed.
(second_ratio): Removed.
(vlmul_for_first_sew_second_ratio): Removed.
(vlmul_for_greatest_sew_second_ratio): Removed.
(ratio_for_second_sew_first_vlmul): Removed.
(class vsetvl_block_info): New.
(DEF_SEW_LMUL_FUSE_RULE): New.
(always_unavailable): Removed.
(avl_unavailable_p): Removed.
(class demand_system): New.
(sew_unavailable_p): Removed.
(lmul_unavailable_p): Removed.
(ge_sew_unavailable_p): Removed.
(ge_sew_lmul_unavailable_p): Removed.
(ge_sew_ratio_unavailable_p): Removed.
(DEF_UNAVAILABLE_COND): Removed.
(same_sew_lmul_demand_p): Removed.
(propagate_avl_across_demands_p): Removed.
(reg_available_p): Removed.
(support_relaxed_compatible_p): Removed.
(demands_can_be_fused_p): Removed.
(earliest_pred_can_be_fused_p): Removed.
(vsetvl_dominated_by_p): Removed.
(avl_info::avl_info): Removed.
(avl_info::single_source_equal_p): Removed.
(avl_info::multiple_source_equal_p): Removed.
(DEF_SEW_LMUL_RULE): New.
(avl_info::operator=): Removed.
(avl_info::operator==): Removed.
(DEF_POLICY_RULE): New.
(avl_info::operator!=): Removed.
(avl_info::has_non_zero_avl): Removed.
(vl_vtype_info::vl_vtype_info): Removed.
(vl_vtype_info::operator==): Removed.
(DEF_AVL_RULE): New.
(vl_vtype_info::operator!=): Removed.
(vl_vtype_info::same_avl_p): Removed.
(vl_vtype_info::same_vtype_p): Removed.
(vl_vtype_info::same_vlmax_p): Removed.
(vector_insn_info::operator>=): Removed.
(vector_insn_info::operator==): Removed.
(class pre_vsetvl): New.
(vector_insn_info::parse_insn): Removed.
(vector_insn_info::compatible_p): Removed.
(vector_insn_info::skip_avl_compatible_p): Removed.
(vector_insn_info::compatible_avl_p): Removed.
(vector_insn_info::compatible_vtype_p): Removed.
(vector_insn_info::available_p): Removed.
(vector_insn_info::fuse_avl): Removed.
(vector_insn_info::fuse_sew_lmul): Removed.
(vector_insn_info::fuse_tail_policy): Removed.
(vector_insn_info::fuse_mask_policy): Removed.
(vector_insn_info::local_merge): Removed.
(vector_insn_info::global_merge): Removed.
(vector_insn_info::get_avl_or_vl_reg): Removed.
(vector_insn_info::update_fault_first_load_avl): Removed.
(vector_insn_info::dump): Removed.
(vector_infos_manager::vector_infos_manager): Removed.
(vector_infos_manager::create_expr): Removed.
(vector_infos_manager::get_expr_id): Removed.
(vector_infos_manager::all_same_ratio_p): Removed.
(vector_infos_manager::all_avail_in_compatible_p): Removed.
(vector_infos_manager::all_same_avl_p): Removed.
(vector_infos_manager::expr_set_num): Removed.
(vector_infos_manager::release): Removed.
(vector_infos_manager::create_bitmap_vectors): Removed.
(vector_infos_manager::free_bitmap_vectors): Removed.
(vector_infos_manager::dump): Removed.
(class pass_vsetvl): Adjust.
(pass_vsetvl::get_vector_info): Removed.
(pass_vsetvl::get_block_info): Removed.
(pass_vsetvl::update_vector_info): Removed.
(pass_vsetvl::update_block_info): Removed.
(pre_vsetvl::compute_avl_def_data): New.
(pass_vsetvl::simple_vsetvl): Removed.
(pass_vsetvl::compute_local_backward_infos): Removed.
(pass_vsetvl::need_vsetvl): Removed.
(pass_vsetvl::transfer_before): Removed.
(pass_vsetvl::transfer_after): Removed.
(pre_vsetvl::compute_vsetvl_def_data): New.
(pass_vsetvl::emit_local_forward_vsetvls): Removed.
(pass_vsetvl::prune_expressions): Removed.
(pass_vsetvl::compute_local_properties): Removed.
(pre_vsetvl::compute_lcm_local_properties): New.
(pass_vsetvl::earliest_fusion): Removed.
(pre_vsetvl::fuse_local_vsetvl_info): New.
(pass_vsetvl::vsetvl_fusion): Removed.
(pass_vsetvl::can_refine_vsetvl_p): Removed.
(pre_vsetvl::earliest_fuse_vsetvl_info): New.
(pass_vsetvl::refine_vsetvls): Removed.
(pass_vsetvl::cleanup_vsetvls): Removed.
(pass_vsetvl::commit_vsetvls): Removed.
(pass_vsetvl::pre_vsetvl): Removed.
(pass_vsetvl::get_vsetvl_at_end): Removed.
(local_avl_compatible_p): Removed.
(pass_vsetvl::local_eliminate_vsetvl_insn): Removed.
(pre_vsetvl::pre_global_vsetvl_info): New.
(get_first_vsetvl_before_rvv_insns): Removed.
(pass_vsetvl::global_eliminate_vsetvl_insn): Removed.
(pre_vsetvl::emit_vsetvl): New.
(pass_vsetvl::ssa_post_optimization): Removed.
(pre_vsetvl::cleaup): New.
(pre_vsetvl::remove_avl_operand): New.
(pass_vsetvl::df_post_optimization): Removed.
(pre_vsetvl::remove_unused_dest_operand): New.
(pass_vsetvl::init): Removed.
(pass_vsetvl::done): Removed.
(pass_vsetvl::compute_probabilities): Removed.
(pass_vsetvl::lazy_vsetvl): Adjust.
(pass_vsetvl::execute): Adjust.
* config/riscv/riscv-vsetvl.def (DEF_INCOMPATIBLE_COND): Removed.
(DEF_SEW_LMUL_RULE): New.
(DEF_SEW_LMUL_FUSE_RULE): Removed.
(DEF_POLICY_RULE): New.
(DEF_UNAVAILABLE_COND): Removed
(DEF_AVL_RULE): New demand type.
(sew_lmul): New demand type.
(ratio_only): New demand type.
(sew_only): New demand type.
(ge_sew): New demand type.
(ratio_and_ge_sew): New demand type.
(tail_mask_policy): New demand type.
(tail_policy_only): New demand type.
(mask_policy_only): New demand type.
(ignore_policy): New demand type.
(avl): New demand type.
(non_zero_avl): New demand type.
(ignore_avl): New demand type.
* config/riscv/t-riscv: Removed riscv-vsetvl.h
* config/riscv/riscv-vsetvl.h: Removed.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/scalar_move-1.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-23.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-46.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-84.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-89.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-95.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/imm_bb_prop-1.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/pr109743-2.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/pr109773-1.c: Adjust.
* gcc.target/riscv/rvv/base/pr111037-1.c: Moved to...
* gcc.target/riscv/rvv/vsetvl/pr111037-1.c: ...here.
* gcc.target/riscv/rvv/base/pr111037-2.c: Moved to...
* gcc.target/riscv/rvv/vsetvl/pr111037-2.c: ...here.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-25.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-26.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vlmax_conflict-12.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vlmax_conflict-3.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vsetvl-13.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vsetvl-18.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/vsetvl-23.c: Adjust.
* gcc.target/riscv/rvv/vsetvl/avl_single-104.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-105.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-106.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-107.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-108.c: New test.
* gcc.target/riscv/rvv/vsetvl/avl_single-109.c: New test.
* gcc.target/riscv/rvv/vsetvl/pr111037-3.c: New test.
* gcc.target/riscv/rvv/vsetvl/pr111037-4.c: New test.

return edge in make_eh_edges

The need to initialize edge probabilities has made make_eh_edges
undesirably hard to use.  I suppose we don't want make_eh_edges to
initialize the probability of the newly-added edge itself, so that the
caller takes care of it, but identifying the added edge in need of
adjustments is inefficient and cumbersome.  Change make_eh_edges so
that it returns the added edge.

for  gcc/ChangeLog

* tree-eh.cc (make_eh_edges): Return the new edge.
* tree-eh.h (make_eh_edges): Likewise.

c++: indirect change of active union member in constexpr [PR101631,PR102286]

This patch adds checks for attempting to change the active member of a
union by methods other than a member access expression.

To be able to properly distinguish `*(&u.a) = ` from `u.a = `, this
patch redoes the solution for c++/59950 to avoid extranneous *&; it
seems that the only case that needed the workaround was when copying
empty classes.

This patch also ensures that constructors for a union field mark that
field as the active member before entering the call itself; this ensures
that modifications of the field within the constructor's body don't
cause false positives (as these will not appear to be member access
expressions). This means that we no longer need to start the lifetime of
empty union members after the constructor body completes.

As a drive-by fix, this patch also ensures that value-initialised unions
are considered to have activated their initial member for the purpose of
checking stores and accesses, which catches some additional mistakes
pre-C++20.

PR c++/101631
PR c++/102286

gcc/cp/ChangeLog:

* call.cc (build_over_call): Fold more indirect refs for trivial
assignment op.
* class.cc (type_has_non_deleted_trivial_default_ctor): Create.
* constexpr.cc (cxx_eval_call_expression): Start lifetime of
union member before entering constructor.
(cxx_eval_component_reference): Check against first member of
value-initialised union.
(cxx_eval_store_expression): Activate member for
value-initialised union. Check for accessing inactive union
member indirectly.
* cp-tree.h (type_has_non_deleted_trivial_default_ctor):
Forward declare.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/constexpr-89336-3.C: Fix union initialisation.
* g++.dg/cpp1y/constexpr-union6.C: New test.
* g++.dg/cpp1y/constexpr-union7.C: New test.
* g++.dg/cpp2a/constexpr-union2.C: New test.
* g++.dg/cpp2a/constexpr-union3.C: New test.
* g++.dg/cpp2a/constexpr-union4.C: New test.
* g++.dg/cpp2a/constexpr-union5.C: New test.
* g++.dg/cpp2a/constexpr-union6.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Jason Merrill <jason@redhat.com>

c++: Improve diagnostics for constexpr cast from void*

This patch improves the errors given when casting from void* in C++26 to
include the expected type if the types of the pointed-to objects were
not similar. It also ensures (for all standard modes) that void* casts
are checked even for DECL_ARTIFICIAL declarations, such as
lifetime-extended temporaries, and is only ignored for cases where we
know it's OK (e.g. source_location::current) or have no other choice
(heap-allocated data).

gcc/cp/ChangeLog:

* constexpr.cc (is_std_source_location_current): New.
(cxx_eval_constant_expression): Only ignore cast from void* for
specific cases and improve other diagnostics.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/constexpr-cast4.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Marek Polacek <polacek@redhat.com>
Reviewed-by: Jason Merrill <jason@redhat.com>

Daily bump.

c++: small tweak for cp_fold_r

This patch is an optimization tweak for cp_fold_r.  If we cp_fold_r the
COND_EXPR's op0 first, we may be able to evaluate it to a constant if -O.
cp_fold has:

3143         if (callee && DECL_DECLARED_CONSTEXPR_P (callee)
3144             && !flag_no_inline)
...
3151             r = maybe_constant_value (x, /*decl=*/NULL_TREE,

flag_no_inline is 1 for -O0:

1124   if (opts->x_optimize == 0)
1125     {
1126       /* Inlining does not work if not optimizing,
1127          so force it not to be done.  */
1128       opts->x_warn_inline = 0;
1129       opts->x_flag_no_inline = 1;
1130     }

but otherwise it's 0 and cp_fold will maybe_constant_value calls to
constexpr functions.  And if it doesn't, then folding the COND_EXPR
will keep both arms, and we can avoid calling maybe_constant_value.

gcc/cp/ChangeLog:

* cp-gimplify.cc (cp_fold_r): Don't call maybe_constant_value.

doc: Update contrib.texi

I noticed that Patrick is missing here.

gcc/ChangeLog:

* doc/contrib.texi: Add entry for Patrick Palka.

vect: Use inbranch simdclones in masked loops

This patch enables the compiler to use inbranch simdclones when generating
masked loops in autovectorization.

gcc/ChangeLog:

* omp-simd-clone.cc (simd_clone_adjust_argument_types): Make function
compatible with mask parameters in clone.
* tree-vect-stmts.cc (vect_build_all_ones_mask): Allow vector boolean
typed masks.
(vectorizable_simd_clone_call): Enable the use of masked clones in
fully masked loops.

vect: don't allow fully masked loops with non-masked simd clones [PR 110485]

When analyzing a loop and choosing a simdclone to use it is possible to choose
a simdclone that cannot be used 'inbranch' for a loop that can use partial
vectors.  This may lead to the vectorizer deciding to use partial vectors which
are not supported for notinbranch simd clones.  This patch fixes that by
disabling the use of partial vectors once a notinbranch simd clone has been
selected.

gcc/ChangeLog:

PR tree-optimization/110485
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Disable partial
vectors usage if a notinbranch simdclone has been selected.

gcc/testsuite/ChangeLog:

* gcc.dg/gomp/pr110485.c: New test.

vect: Fix vect_get_smallest_scalar_type for simd clones

The vect_get_smallest_scalar_type helper function was using any argument to a
simd clone call when trying to determine the smallest scalar type that would be
vectorized.  This included the function pointer type in a MASK_CALL for
instance, and would result in the wrong type being selected.  Instead this
patch special cases simd_clone_call's and uses only scalar types of the
original function that get transformed into vector types.

gcc/ChangeLog:

* tree-vect-data-refs.cc (vect_get_smallest_scalar_type): Special case
simd clone calls and only use types that are mapped to vectors.
(simd_clone_call_p): New helper function.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/vect-simd-clone-16f.c: Remove unnecessary differentation
between targets with different pointer sizes.
* gcc.dg/vect/vect-simd-clone-17f.c: Likewise.
* gcc.dg/vect/vect-simd-clone-18f.c: Likewise.

parloops: Allow poly nit and bound

Teach parloops how to handle a poly nit and bound e ahead of the changes to
enable non-constant simdlen.

gcc/ChangeLog:

* tree-parloops.cc (try_transform_to_exit_first_loop_alt): Accept
poly NIT and ALT_BOUND.

parloops: Copy target and optimizations when creating a function clone

SVE simd clones require to be compiled with a SVE target enabled or the argument
types will not be created properly. To achieve this we need to copy
DECL_FUNCTION_SPECIFIC_TARGET from the original function declaration to the
clones.  I decided it was probably also a good idea to copy
DECL_FUNCTION_SPECIFIC_OPTIMIZATION in case the original function is meant to
be compiled with specific optimization options.

gcc/ChangeLog:

* tree-parloops.cc (create_loop_fn): Copy specific target and
optimization options to clone.

omp: Replace simd_clone_subparts with TYPE_VECTOR_SUBPARTS

Refactor simd clone handling code ahead of support for poly simdlen.

gcc/ChangeLog:

* omp-simd-clone.cc (simd_clone_subparts): Remove.
(simd_clone_init_simd_arrays): Replace simd_clone_supbarts with
TYPE_VECTOR_SUBPARTS.
(ipa_simd_modify_function_body): Likewise.
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Likewise.
(simd_clone_subparts): Remove.

libstdc++: [_Hashtable] Do not reuse untrusted cached hash code

On merge, reuse a merged node's possibly cached hash code only if we are on the
same type of hash and this hash is stateless.

Usage of function pointers or std::function as hash functor will prevent reusing
cached hash code.

libstdc++-v3/ChangeLog

* include/bits/hashtable_policy.h
(_Hash_code_base::_M_hash_code(const _Hash&, const _Hash_node_value<>&)): Remove.
(_Hash_code_base::_M_hash_code<_H2>(const _H2&, const _Hash_node_value<>&)): Remove.
* include/bits/hashtable.h
(_M_src_hash_code<_H2>(const _H2&, const key_type&, const __node_value_type&)): New.
(_M_merge_unique<>, _M_merge_multi<>): Use latter.
* testsuite/23_containers/unordered_map/modifiers/merge.cc
(test04, test05, test06): New test cases.