Andreas noted we were getting an uninit warning after the recent constant
synthesis changes. Essentially there's no way for the uninit analysis code to
know the first entry in the CODES array is a UNKNOWN which will set X before
its first use.
So trivial initialization with NULL_RTX is the obvious fix.
Robin Dapp [Wed, 24 Jul 2024 07:08:00 +0000 (09:08 +0200)]
RISC-V: Error early with V and no M extension.
For calculating the value of a poly_int at runtime we use a
multiplication instruction that requires the M extension.
Instead of just asserting and ICEing this patch emits an early
error at option-parsing time.
gcc/ChangeLog:
PR target/116036
* config/riscv/riscv.cc (riscv_override_options_internal): Error
with TARGET_VECTOR && !TARGET_MUL.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/arch-31.c: Add m to arch string and expect it.
* gcc.target/riscv/arch-32.c: Ditto.
* gcc.target/riscv/predef-14.c: Ditto.
* gcc.target/riscv/predef-15.c: Ditto.
* gcc.target/riscv/predef-16.c: Ditto.
* gcc.target/riscv/predef-26.c: Ditto.
* gcc.target/riscv/predef-27.c: Ditto.
* gcc.target/riscv/predef-32.c: Ditto.
* gcc.target/riscv/predef-33.c: Ditto.
* gcc.target/riscv/rvv/autovec/pr111486.c: Add m to arch string.
* gcc.target/riscv/compare-debug-1.c: Ditto.
* gcc.target/riscv/compare-debug-2.c: Ditto.
* gcc.target/riscv/rvv/base/pr116036.c: New test.
Robin Dapp [Wed, 31 Jul 2024 14:54:03 +0000 (16:54 +0200)]
RISC-V: Correct mode_idx attribute for viwalu wx variants [PR116149].
In PR116149 we choose a wrong vector length which causes wrong values in
a reduction. The problem happens in avlprop where we choose the
number of units in the instruction's mode as vector length. For the
non-scalar variants the respective operand has the correct non-widened
mode. For the scalar variants, however, the same operand has a scalar
mode which obviously only has one unit. This makes us choose VL = 1
leaving three elements undisturbed (so potentially -1). Those end up
in the reduction causing the wrong result.
This patch adjusts the mode_idx just for the scalar variants of the
affected instruction patterns.
gcc/ChangeLog:
PR target/116149
* config/riscv/vector.md: Fix mode_idx attribute of scalar
widen add/sub variants.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr116149.c: New test.
Jeff Law [Thu, 8 Aug 2024 13:42:26 +0000 (07:42 -0600)]
[RISC-V][PR target/116240] Ensure object is a comparison before extracting arguments
This was supposed to go out the door yesterday, but I kept getting interrupted.
The target bits for rtx costing can't assume the rtl they're given actually
matches a target pattern. It's just kind of inherent in how the costing
routines get called in various places.
In this particular case we're trying to cost a conditional move:
(set (dest) (if_then_else (cond) (true) (false))
On the RISC-V port the backend only allows actual conditionals for COND. So
something like (eq (reg) (const_int 0)). In the costing code for if-then-else
we did something like
(XEXP (XEXP (cond, 0), 0)))
Which fails miserably if COND is a terminal node like (reg) rather than (ne
(reg) (const_int 0)
So this patch tightens up the RTL scanning to ensure that we have a comparison
before we start looking at the comparison's arguments.
Run through my tester without incident, but I'll wait for the pre-commit tester
to run through a cycle before pushing to the trunk.
Jeff
ps. We probably could support a naked REG for the condition and internally convert it to (ne (reg) (const_int 0)), but I don't think it likely happens with any regularity.
PR target/116240
gcc/
* config/riscv/riscv.cc (riscv_rtx_costs): Ensure object is a
comparison before looking at its arguments.
gcc/testsuite
* gcc.target/riscv/pr116240.c: New test.
曾治金 [Wed, 14 Aug 2024 06:06:23 +0000 (14:06 +0800)]
RISC-V: Fix factor in dwarf_poly_indeterminate_value [PR116305]
This patch is to fix the bug (BugId:116305) introduced by the commit
bd93ef for risc-v target.
The commit bd93ef changes the chunk_num from 1 to TARGET_MIN_VLEN/128
if TARGET_MIN_VLEN is larger than 128 in riscv_convert_vector_bits. So
it changes the value of BYTES_PER_RISCV_VECTOR. For example, before
merging the commit bd93ef and if TARGET_MIN_VLEN is 256, the value
of BYTES_PER_RISCV_VECTOR should be [8, 8], but now [16, 16]. The value
of riscv_bytes_per_vector_chunk and BYTES_PER_RISCV_VECTOR are no longer
equal.
Prologue will use BYTES_PER_RISCV_VECTOR.coeffs[1] to estimate the vlenb
register value in riscv_legitimize_poly_move, and dwarf2cfi will also
get the estimated vlenb register value in riscv_dwarf_poly_indeterminate_value
to calculate the number of times to multiply the vlenb register value.
So need to change the factor from riscv_bytes_per_vector_chunk to
BYTES_PER_RISCV_VECTOR, otherwise we will get the incorrect dwarf
information. The incorrect example as follow:
The sequence '0x92,0xa2,0x38,0' means the vlenb register, '0x34' means
the literal 4, '0x1e' means the multiply operation. But in fact, the
vlenb register value just need to multiply the literal 2.
PR target/116305
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_dwarf_poly_indeterminate_value): Take
BYTES_PER_RISCV_VECTOR for *factor instead of riscv_bytes_per_vector_chunk.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/base/scalable_vector_cfi.c: New test.
Robin Dapp [Tue, 27 Aug 2024 08:25:34 +0000 (10:25 +0200)]
RISC-V: Fix subreg of VLS modes larger than a vector [PR116086].
When the source mode is potentially larger than one vector (e.g. an
LMUL2 mode for VLEN=128) we don't know which vector the subreg actually
refers to. For zvl128b and LMUL=2 the subreg in (subreg:V2DI (reg:V4DI))
could actually be the a full (high) vector register of a two-register
group (at VLEN=128) or the higher part of a single register (at VLEN>128).
As the subreg is statically ambiguous we prevent such situations in
can_change_mode_class.
The culprit in PR116086 is
_12 = BIT_FIELD_REF <vect_cst__42, 128, 128>;
which can be expanded with a vector-vector extract (from V4DI to V2DI).
This patch adds a VLS-mode vector-vector extract that handles "halving"
cases like this one by sliding down the source vector, thus making sure
the correct part is used.
PR target/116086
gcc/ChangeLog:
* config/riscv/autovec.md (vec_extract<mode><v_half>): Add
vector-vector extract for VLS modes.
* config/riscv/riscv.cc (riscv_can_change_mode_class): Forbid
VLS modes larger than one vector.
* config/riscv/vector-iterators.md: Add vector-vector extract
iterators.
gcc/testsuite/ChangeLog:
* lib/target-supports.exp: Add effective target checks for
zvl256b and zvl512b.
* gcc.target/riscv/rvv/autovec/pr116086-2-run.c: New test.
* gcc.target/riscv/rvv/autovec/pr116086-2.c: New test.
* gcc.target/riscv/rvv/autovec/pr116086.c: New test.
RISC-V: Fix vl_used_by_non_rvv_insn logic of vsetvl pass
This patch fixes a bug in the current vsetvl pass. The current pass uses
`m_vl` to determine whether the dest operand has been used by non-RVV
instructions. However, `m_vl` may have been modified as a result of an
`update_avl` call, and thus would be no longer the dest operand of the
original instruction. This can lead to incorrect vsetvl eliminations, as is
shown in the testcase. In this patch, we create a `dest_vl` variable for
this scenerio.
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc: Use `dest_vl` for dest VL operand
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/vsetvl/vsetvl_bug-3.c: New test.
Andreas Schwab [Thu, 12 Sep 2024 11:55:09 +0000 (13:55 +0200)]
riscv: Fix duplicate assmbler label in @tlsdesc<mode> insn
Use %= instead of maintaining a sequence number manually, so that it
doesn't result in a duplicate assembler label when the insn is duplicated.
PR target/116693
* config/riscv/riscv.cc (riscv_legitimize_tls_address): Don't pass
seqno to gen_tlsdesc and remove it.
* config/riscv/riscv.md (@tlsdesc<mode>): Remove operand 1. Use
%= instead of %1 in template.
Bohan Lei [Wed, 18 Sep 2024 13:20:23 +0000 (07:20 -0600)]
[PATCH] RISC-V: Allow zero operand for DI variants of vssubu.vx
The RISC-V vector machine description relies on the helper function
`sew64_scalar_helper` to emit actual insns for the DI variants of
vssub.vx and vssubu.vx. This works with vssub.vx, but can cause
problems with vssubu.vx with the scalar operand being constant zero,
because `has_vi_variant_p` returns false, and the operand will be taken
without being loaded into a reg. The attached testcases can cause an
internal compiler error as a result.
Allowing a constant zero operand in those insns seems to be a simple
solution that only affects minimum existing code.
gcc/ChangeLog:
* config/riscv/vector.md: Allow zero operand for DI variants of
vssubu.vx
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/base/vssubu-1.c: New test.
* gcc.target/riscv/rvv/base/vssubu-2.c: New test.
[PATCH] RISC-V: Fix th.extu operands exceeding range on rv32.
The Combine Pass may generate zero_extract instructions that are out of range.
Drawing from other architectures like AArch64, we should impose restrictions
on the "*th_extu<mode>4" pattern.
gcc/
* config/riscv/thead.md (*th_extu<mode>4): Fix th.extu
operands exceeding range on rv32.
Robin Dapp [Thu, 21 Nov 2024 13:49:53 +0000 (14:49 +0100)]
RISC-V: Ensure vtype for full-register moves [PR117544].
As discussed in PR117544 the VTYPE register is not preserved across
function calls. Even though vmv1r-like instructions operate
independently of the actual vtype they still require a valid vtype. As
we cannot guarantee that the vtype is valid we must make sure to emit a
vsetvl between a function call and a vmv1r.v.
This patch makes the necessary changes by splitting the full-reg-move
insns into patterns that use the vtype register and adding vmov to the
types of instructions requiring a vset.
Pan Li [Wed, 4 Dec 2024 05:53:52 +0000 (13:53 +0800)]
RISC-V: Add assert for insn operand out of range access [PR117878][NFC]
According to the the initial analysis of PR117878, the ice comes from
the out-of-range operand access for recog_data.operand[]. Thus, add
one assert here to expose this explicitly.
PR target/117878
gcc/ChangeLog:
* config/riscv/riscv-v.cc (vlmax_avl_type_p): Add assert for
out of range access.
(nonvlmax_avl_type_p): Ditto.
Jeff Law [Mon, 30 Dec 2024 20:51:55 +0000 (13:51 -0700)]
[RISC-V][PR target/106544] Avoid ICEs due to bogus asms
This is a fix for a bug Andrew P filed a while back where essentially a poorly
crafted asm statement could trigger a ICE during assembly output. Various
cases will use INTVAL (op) without verifying the operand is a CONST_INT node
first.
The usual way to handle this is via output_operand_lossage, which this patch
implements.
I focused primarily on the CONST_INT cases, there could well be other problems
in this space, if so they should get distinct bugs with testcases.
Tested in my tester on rv32 and rv64. Waiting for pre-commit testing before
moving forward.
PR target/106544
gcc/
* config/riscv/riscv.cc (riscv_print_operand): Issue an error for
invalid operands rather than invalidly accessing INTVAL of an
object that is not a CONST_INT. Fix one error string for 'N'.
gcc/testsuite
* gcc.target/riscv/pr106544.c: New test.
with negative step expecting wraparound semantics due to -fwrapv.
For building interleaved patterns we have an optimization that
does e.g.
{1, 209, ...} = { 1, 0, 209, 0, ...}
and
{201, 25, ...} >> 8 = { 0, 201, 0, 25, ...}
and IORs those.
The optimization only works if the lowpart bits are zero. When
overflowing e.g. with a negative step we cannot guarantee this.
This patch makes us fall back to the generic merge handling for negative
steps.
I'm not 100% certain we're good even for positive steps. If the
step or the vector length is large enough we'd still overflow and
have non-zero lower bits. I haven't seen this happen during my
testing, though and the patch doesn't make things worse, so...
Regtested on rv64gcv_zvl512b. Let's see what the CI says.
Regards
Robin
PR target/117682
gcc/ChangeLog:
* config/riscv/riscv-v.cc (expand_const_vector): Fall back to
merging if either step is negative.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr117682.c: New test.
In PR118154 we emit strided stores but the first of those does not
always have the proper VTYPE. That's because we erroneously delete
a necessary vsetvl.
Kito Cheng [Mon, 23 Dec 2024 15:23:44 +0000 (23:23 +0800)]
RISC-V: Fix code gen for reduction with length 0 [PR118182]
`.MASK_LEN_FOLD_LEFT_PLUS`(or `mask_len_fold_left_plus_m`) is expecting the
return value will be the start value even if the length is 0.
However current code gen in RISC-V backend is not meet that semantic, it will
result a random garbage value if length is 0.
Let example by current code gen for MASK_LEN_FOLD_LEFT_PLUS with f64:
# _148 = .MASK_LEN_FOLD_LEFT_PLUS (stmp__148.33_134, vect__70.32_138, { -1, ... }, loop_len_161, 0);
vsetvli zero,a5,e64,m1,ta,ma
vfmv.s.f v2,fa5 # insn 1
vfredosum.vs v1,v1,v2 # insn 2
vfmv.f.s fa5,v1 # insn 3
insn 1:
- vfmv.s.f won't do anything if VL=0, which means v2 will contain garbage value.
insn 2:
- vfredosum.vs won't do anything if VL=0, and keep vd unchanged even TA.
(v-spec say: `If vl=0, no operation is performed and the destination register
is not updated.`)
insn 3:
- vfmv.f.s will move the value from v1 even VL=0, so this is safe.
So how we fix that? we need two fix for that:
1. insn 1: need always execute with VL=1, so that we can guarantee it will
always work as expect.
2. insn 2: Add new pattern to force `vd` use same reg as `vs1` (start value) for
all reduction patterns, then we can guarantee vd[0] will contain the
start value when vl=0
For 1, it's just a simple change to riscv_vector::expand_reduction, but for 2,
we have to add _VL0_SAFE variant reduction to force `vd` use same reg as `vs1`
(start value).
Change since V3:
- Rename _AV to _VL0_SAFE for readability.
- Use non-VL0_SAFE version if VL is const or VLMAX.
- Only force VL=1 for vfmv.s.f when VL is non-const and non-VLMAX.
- Two more testcase.
Jin Ma [Sat, 18 Jan 2025 14:43:17 +0000 (07:43 -0700)]
[PR target/118357] RISC-V: Disable fusing vsetvl instructions by VSETVL_VTYPE_CHANGE_ONLY for XTheadVector.
In RVV 1.0, the instruction "vsetvli zero,zero,*" indicates that the
available vector length (avl) does not change. However, in XTheadVector,
this same instruction signifies that the avl should take the maximum value.
Consequently, when fusing vsetvl instructions, the optimization labeled
"VSETVL_VTYPE_CHANGE_ONLY" is disabled for XTheadVector.
PR target/118357
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc: Function change_vtype_only_p always
returns false for XTheadVector.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/xtheadvector/pr118357.c: New test.
Jeff Law [Sat, 18 Jan 2025 20:44:33 +0000 (13:44 -0700)]
[RISC-V][PR target/116308] Fix generation of initial RTL for atomics
While this wasn't originally marked as a regression, it almost certainly is
given that older versions of GCC would have used libatomic and would not have
ICE'd on this code.
Basically this is another case where we directly used simplify_gen_subreg when
we should have used gen_lowpart.
When I fixed a similar bug a while back I noted the code in question as needing
another looksie. I think at that time my brain saw the mixed modes (SI & QI)
and locked up. But the QI stuff is just the shift count, not some deeper
issue. So fixing is trivial.
We just replace the simplify_gen_subreg with a gen_lowpart and get on with our
lives.
Tested on rv64 and rv32 in my tester. Waiting on pre-commit testing for final
verdict.
PR target/116308
gcc/
* config/riscv/riscv.cc (riscv_lshift_subword): Use gen_lowpart
rather than simplify_gen_subreg.
The right shift is always going to produce 0. 0 + 1 = 1 which is a power of 2.
So exact_log2 returns 0 and we get a true result rather than a false result.
The fix is trivial. "<=". While inside we might as well fix the formatting.
Tested on rv32 and rv64 in my tester. Waiting on upstream pre-commit testing
to render a verdict.
Jonathan Wakely [Mon, 7 Apr 2025 18:52:55 +0000 (19:52 +0100)]
libstdc++: Fix use-after-free in std::format [PR119671]
When formatting floating-point values to wide strings there's a case
where we invalidate a std::wstring buffer while a std::wstring_view is
still referring to it.
libstdc++-v3/ChangeLog:
PR libstdc++/119671
* include/std/format (__formatter_fp::format): Do not invalidate
__wstr unless _M_localized returns a valid string.
* testsuite/std/format/functions/format.cc: Check wide string
formatting of floating-point types with classic locale.
Jonathan Wakely [Wed, 26 Mar 2025 11:47:05 +0000 (11:47 +0000)]
libstdc++: Replace use of __mindist in ranges::uninitialized_xxx algos [PR101587]
In r15-8980-gf4b6acfc36fb1f I introduced a new function object for
finding the smaller of two distances. In bugzilla Hewill Kang pointed
out that we still need to explicitly convert the result back to the
right difference type, because the result might be an integer-like class
type that doesn't convert to an integral type explicitly.
Rather than doing that conversion in the __mindist function object, I
think it's simpler to remove it again and just do a comparison and
assignment. We always want the result to have a specific type, so we can
just check if the value of the other type is smaller, and then convert
that to the other type if so.
libstdc++-v3/ChangeLog:
PR libstdc++/101587
* include/bits/ranges_uninitialized.h (__detail::__mindist):
Remove.
(ranges::uninitialized_copy, ranges::uninitialized_copy_n)
(ranges::uninitialized_move, ranges::uninitialized_move_n): Use
comparison and assignment instead of __mindist.
* testsuite/20_util/specialized_algorithms/uninitialized_copy/constrained.cc:
Check with ranges that use integer-like class type for
difference type.
* testsuite/20_util/specialized_algorithms/uninitialized_move/constrained.cc:
Likewise.
Reviewed-by: Tomasz Kaminski <tkaminsk@redhat.com> Reviewed-by: Hewill Kang <hewillk@gmail.com>
(cherry picked from commit 03ac8886e5c1fa16da90276fd721a57fa9435f4f)
Jonathan Wakely [Wed, 26 Mar 2025 11:47:05 +0000 (11:47 +0000)]
libstdc++: Replace use of std::min in ranges::uninitialized_xxx algos [PR101587]
Because ranges can have any signed integer-like type as difference_type,
it's not valid to use std::min(diff1, diff2). Instead of calling
std::min with an explicit template argument, this adds a new __mindist
helper that determines the common type and uses that with std::min.
libstdc++-v3/ChangeLog:
PR libstdc++/101587
* include/bits/ranges_uninitialized.h (__detail::__mindist):
New function object.
(ranges::uninitialized_copy, ranges::uninitialized_copy_n)
(ranges::uninitialized_move, ranges::uninitialized_move_n): Use
__mindist instead of std::min.
* testsuite/20_util/specialized_algorithms/uninitialized_copy/constrained.cc:
Check ranges with difference difference types.
* testsuite/20_util/specialized_algorithms/uninitialized_move/constrained.cc:
Likewise.
LoongArch: Add LoongArch architecture detection to __float128 support in libgfortran and libquadmath [PR119408].
In GCC14, LoongArch added __float128 as an alias for _Float128.
In commit r15-8962, support for q/Q suffixes for 128-bit floating point
numbers. This will cause the compiler to automatically link libquadmath
when compiling Fortran programs. But on LoongArch `long double` is
IEEE quad, so there is no need to implement libquadmath.
This causes link failure.
PR target/119408
libgfortran/ChangeLog:
* acinclude.m4: When checking for __float128 support, determine
whether the current architecture is LoongArch. If so, return false.
* configure: Regenerate.
libquadmath/ChangeLog:
* configure.ac: When checking for __float128 support, determine
whether the current architecture is LoongArch. If so, return false.
* configure: Regenerate.
Sigend-off-by: Xi Ruoyao <xry111@xry111.site> Sigend-off-by: Jakub Jelinek <jakub@redhat.com>
(cherry picked from commit 1534f0099c98ea14c08a401302b05edf2231f411)
Patrick Palka [Thu, 13 Mar 2025 23:55:00 +0000 (19:55 -0400)]
libstdc++: Work around C++20 tuple<tuple<any>> constraint recursion [PR116440]
The type tuple<tuple<any>> is clearly copy/move constructible, but for
reasons that are not yet completely understood checking this triggers
constraint recursion with our C++20 tuple implementation (but not the
C++17 implementation).
It turns out this recursion stems from considering the non-template
tuple(const _Elements&) constructor during the copy/move constructibility
check. Considering this constructor is ultimately redundant, since the
defaulted copy/move constructors are better matches.
GCC has a non-standard "perfect candidate" optimization[1] that causes
overload resolution to shortcut considering template candidates if we
find a (non-template) perfect candidate. So to work around this issue
(and as a general compile-time optimization) this patch turns the
problematic constructor into a template so that GCC doesn't consider it
when checking for copy/move constructibility of this tuple type.
Changing the template-ness of a constructor can affect overload
resolution (since template-ness is a tiebreaker) so there's a risk this
change could e.g. introduce overload resolution ambiguities. But the
original C++17 implementation has long defined this constructor as a
template (in order to constrain it etc), so doing the same thing in the
C++20 mode should naturally be quite safe.
The testcase still fails with Clang (in C++20 mode) since it doesn't
implement said optimization.
Jason Merrill [Mon, 7 Apr 2025 15:49:19 +0000 (11:49 -0400)]
c++: constinit and value-initialization [PR119652]
Value-initialization built an AGGR_INIT_EXPR to set AGGR_INIT_ZERO_FIRST on.
Passing that AGGR_INIT_EXPR to maybe_constant_value returned a TARGET_EXPR,
which potential_constant_expression_1 mistook for a temporary.
We shouldn't add a TARGET_EXPR to the AGGR_INIT_EXPR in this case, just like
we already avoid adding it to CONSTRUCTOR or CALL_EXPR.
PR c++/119652
gcc/cp/ChangeLog:
* constexpr.cc (cxx_eval_outermost_constant_expr): Also don't add a
TARGET_EXPR around AGGR_INIT_EXPR.
Jason Merrill [Fri, 4 Apr 2025 21:34:08 +0000 (17:34 -0400)]
c++: __FUNCTION__ in lambda return type [PR118629]
In this testcase, the use of __FUNCTION__ is within a function parameter
scope, the lambda's. And P1787 changed __func__ to live in the parameter
scope. But [basic.scope.pdecl] says that the point of declaration of
__func__ is immediately before {, so in the trailing return type it isn't in
scope yet, so this __FUNCTION__ should refer to foo().
Looking first for a block scope, then a function parameter scope, gives us
the right result.
PR c++/118629
gcc/cp/ChangeLog:
* name-lookup.cc (pushdecl_outermost_localscope): Look for an
sk_block.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/lambda/lambda-__func__3.C: New test.
Jason Merrill [Wed, 29 Jan 2025 10:15:00 +0000 (05:15 -0500)]
c++: lambda in requires outside template [PR99546]
Since r10-7441 we set processing_template_decl in a requires-expression so
that we can use tsubst_expr to evaluate the requirements, but that confuses
lambdas terribly; begin_lambda_type silently returns error_mark_node and we
continue into other failures. This patch clears processing_template_decl
again while we're defining the closure and op() function, so it only remains
set while parsing the introducer (i.e. any init-captures) and building the
resulting object. This properly avoids trying to create another lambda in
tsubst_lambda_expr.
* g++.dg/cpp2a/lambda-requires2.C: New test.
* g++.dg/cpp2a/lambda-requires3.C: New test.
* g++.dg/cpp2a/lambda-requires4.C: New test.
* g++.dg/cpp2a/lambda-requires5.C: New test.
Patrick Palka [Fri, 4 Apr 2025 18:03:58 +0000 (14:03 -0400)]
c++: constraint variable used in evaluated context [PR117849]
Here we wrongly reject the type-requirement at parse time due to its use
of the constraint variable 't' within a template argument (an evaluated
context). Fix this simply by refining the "use of parameter outside
function body" error path to exclude constraint variables.
PR c++/104255 tracks the same issue for function parameters, but fixing
that would be more involved, requiring changes to the PARM_DECL case of
tsubst_expr.
PR c++/117849
gcc/cp/ChangeLog:
* semantics.cc (finish_id_expression_1): Allow use of constraint
variable outside an unevaluated context.
Patrick Palka [Thu, 3 Apr 2025 20:33:46 +0000 (16:33 -0400)]
c++: P2280R4 and speculative constexpr folding [PR119387]
Compiling the testcase in this PR uses 2.5x more memory and 6x more
time ever since r14-5979 which implements P2280R4. This is because
our speculative constexpr folding now does a lot more work trying to
fold ultimately non-constant calls to constexpr functions, and in turn
produces a lot of garbage. We do sometimes successfully fold more
thanks to P2280R4, but it seems to be trivial stuff like calls to
std::array::size or std::addressof. The benefit of P2280 therefore
doesn't seem worth the cost during speculative constexpr folding, so
this patch restricts the paper to only manifestly-constant evaluation.
PR c++/119387
gcc/cp/ChangeLog:
* constexpr.cc (p2280_active_p): New.
(cxx_eval_constant_expression) <case VAR_DECL>: Use it to
restrict P2280 relaxations.
<case PARM_DECL>: Likewise.
Eric Botcazou [Fri, 4 Apr 2025 09:45:23 +0000 (11:45 +0200)]
Ada: Fix thinko in Eigensystem for complex Hermitian matrices
The implementation solves the eigensystem for a NxN complex Hermitian matrix
by first solving it for a 2Nx2N real symmetric matrix and then interpreting
the 2Nx1 real vectors as Nx1 complex ones, but the last step does not work.
The patch fixes the last step and also performs a small cleanup throughout
the implementation, mostly in the commentary and without functional changes.
gcc/ada/
* libgnat/a-ngcoar.adb (Eigensystem): Adjust notation and fix the
layout of the real symmetric matrix in the main comment. Adjust
the layout of the associated code accordingly and correctly turn
the 2Nx1 real vectors into Nx1 complex ones.
(Eigenvalues): Minor similar tweaks.
* libgnat/a-ngrear.adb (Jacobi): Minor tweaks in the main comment.
Adjust notation and corresponding parameter names of functions.
Fix call to Unit_Matrix routine. Adjust the comment describing
the various kinds of iterations to match the implementation.
Using specific SSA names in pattern matching in `dg-final' makes tests
"unstable", in that changes in passes prior to the pass whose dump is
analyzed in the particular test may change the numbering of the SSA
variables, causing the test to start failing spuriously.
We thus switch from specific SSA names to the use of a multi-line
regular expression making use of capture groups for matching particular
variables across different statements, ensuring the test will pass
more consistently across different versions of GCC.
PR testsuite/118597
gcc/testsuite/ChangeLog:
* gcc.dg/vect/vect-fncall-mask.c: Update test directives.
Tomasz Kamiński [Thu, 3 Apr 2025 08:23:45 +0000 (10:23 +0200)]
libstdc++: Fix handling of field width for wide strings and characters [PR119593]
This patch corrects handling of UTF-32LE and UTF32-BE in
__unicode::__literal_encoding_is_unicode<_CharT>, so they are
recognized as unicode and functions produces correct result for wchar_t.
Use `__unicode::__field_width` to compute the estimated witdh
of the charcter for unicode wide encoding.
PR libstdc++/119593
libstdc++-v3/ChangeLog:
* include/bits/unicode.h
(__unicode::__literal_encoding_is_unicode<_CharT>):
Corrected handing for UTF-16 and UTF-32 with "LE" or "BE" suffix.
* include/std/format (__formatter_str::_S_character_width):
Define.
(__formatter_str::_S_character_width): Updated passed char
length.
* testsuite/std/format/functions/format.cc: Test for wchar_t.
Reviewed-by: Jonathan Wakely <jwakely@redhat.com> Signed-off-by: Tomasz Kamiński <tkaminsk@redhat.com>
Richard Biener [Fri, 7 Mar 2025 09:15:20 +0000 (10:15 +0100)]
tree-optimization/119145 - avoid stray .MASK_CALL after vectorization
When we BB vectorize an if-converted loop body we make sure to not
leave around .MASK_LOAD or .MASK_STORE created by if-conversion but
we failed to check for .MASK_CALL.
PR tree-optimization/119145
* tree-vectorizer.cc (try_vectorize_loop_1): Avoid BB
vectorizing an if-converted loop body when there's a .MASK_CALL
in the loop body.
Richard Biener [Thu, 6 Mar 2025 08:08:07 +0000 (09:08 +0100)]
middle-end/119119 - re-gimplification of empty CTOR assignments
The following testcase runs into a re-gimplification issue during
inlining when processing
MEM[(struct e *)this_2(D)].a = {};
where re-gimplification does not handle assignments in the same
way than the gimplifier but instead relies on rhs_predicate_for
and gimplifying the RHS standalone. This fails to handle
special-casing of CTORs. The is_gimple_mem_rhs_or_call predicate
already handles clobbers but not empty CTORs so we end up in
the fallback code trying to force the CTOR into a separate stmt
using a temporary - but as we have a non-copyable type here that ICEs.
The following generalizes empty CTORs in is_gimple_mem_rhs_or_call
since those need no additional re-gimplification.
PR middle-end/119119
* gimplify.cc (is_gimple_mem_rhs_or_call): All empty CTORs
are OK when not a register type.
When we vectorize a .COND_ADD reduction and apply the single-use-def
cycle optimization we can end up chosing the wrong else value for
subsequent .COND_ADD. The following rectifies this.
PR tree-optimization/119096
* tree-vect-loop.cc (vect_transform_reduction): Use the
correct else value for .COND_fn.
Richard Biener [Mon, 3 Mar 2025 08:54:15 +0000 (09:54 +0100)]
ipa/119067 - bogus TYPE_PRECISION check on VECTOR_TYPE
odr_types_equivalent_p can end up using TYPE_PRECISION on vector
types which is a no-go. The following instead uses TYPE_VECTOR_SUBPARTS
for vector types so we also end up comparing the number of vector elements.
PR ipa/119067
* ipa-devirt.cc (odr_types_equivalent_p): Check
TYPE_VECTOR_SUBPARTS for vectors.
* g++.dg/lto/pr119067_0.C: New testcase.
* g++.dg/lto/pr119067_1.C: Likewise.
We are detecting a cycle as double reduction where the inner loop
cycle has extra out-of-loop uses. This clashes at least with
assumptions from the SLP discovery code which says the cycle
isn't reachable from another SLP instance. It also was not intended
to support this case, in fact with GCC 14 we seem to generate wrong
code here.
PR tree-optimization/119057
* tree-vect-loop.cc (check_reduction_path): Add argument
specifying whether we're analyzing the inner loop of a
double reduction. Do not allow extra uses outside of the
double reduction cycle in this case.
(vect_is_simple_reduction): Adjust.
Richard Biener [Thu, 6 Mar 2025 12:48:16 +0000 (13:48 +0100)]
lto/114501 - missed free-lang-data for CONSTRUCTOR index
The following makes sure to also walk CONSTRUCTOR element indexes
which can be FIELD_DECLs, referencing otherwise unused types we
need to clean. walk_tree only walks CONSTRUCTOR element data.
PR lto/114501
* ipa-free-lang-data.cc (find_decls_types_r): Explicitly
handle CONSTRUCTORs as walk_tree handling of those is
incomplete.
Richard Biener [Fri, 28 Feb 2025 10:44:26 +0000 (11:44 +0100)]
ipa/111245 - bogus modref analysis for store in call that might throw
We currently record a kill for
*x_4(D) = always_throws ();
because we consider the store always executing since the appropriate
check for whether the stmt could throw is guarded by
!cfun->can_throw_non_call_exceptions.
PR ipa/111245
* ipa-modref.cc (modref_access_analysis::analyze_store): Do
not guard the check of whether the stmt could throw by
cfun->can_throw_non_call_exceptions.
Jonathan Wakely [Wed, 26 Mar 2025 11:21:32 +0000 (11:21 +0000)]
libstdc++: Fix std::ranges::iter_move for function references [PR119469]
The result of std::move (or a cast to an rvalue reference) on a function
reference is always an lvalue. Because std::ranges::iter_move was using
the type std::remove_reference_t<X>&& as the result of std::move, it was
giving the wrong type for function references. Use a decltype-specifier
with declval<remove_reference_t<X>>() instead of just using the
remove_reference_t<X>&& type directly. This gives the right result,
while still avoiding the cost of doing overload resolution for
std::move.
libstdc++-v3/ChangeLog:
PR libstdc++/119469
* include/bits/iterator_concepts.h (_IterMove::__result): Use
decltype-specifier instead of an explicit type.
* testsuite/24_iterators/customization_points/iter_move.cc:
Check results for function references.
Jonathan Wakely [Fri, 28 Feb 2025 21:44:41 +0000 (21:44 +0000)]
libstdc++: Fix ranges::iter_move handling of rvalues [PR106612]
The specification for std::ranges::iter_move apparently requires us to
handle types which do not satisfy std::indirectly_readable, for example
with overloaded operator* which behaves differently for different value
categories.
libstdc++-v3/ChangeLog:
PR libstdc++/106612
* include/bits/iterator_concepts.h (_IterMove::__iter_ref_t):
New alias template.
(_IterMove::__result): Use __iter_ref_t instead of
std::iter_reference_t.
(_IterMove::__type): Remove incorrect __dereferenceable
constraint.
(_IterMove::operator()): Likewise. Add correct constraints. Use
__iter_ref_t instead of std::iter_reference_t. Forward parameter
as correct value category.
(iter_swap): Add comments.
* testsuite/24_iterators/customization_points/iter_move.cc: Test
that iter_move is found by ADL and that rvalue arguments are
handled correctly.
Jonathan Wakely [Fri, 28 Mar 2025 15:41:41 +0000 (15:41 +0000)]
libstdc++: Fix -Warray-bounds warning in std::vector<bool> [PR110498]
In this case, we need to tell the compiler that the current size is not
larger than the new size so that all the existing elements can be copied
to the new storage. This avoids bogus warnings about overflowing the new
storage when the compiler can't tell that that cannot happen.
We might as well also hoist the loads of begin() and end() before the
allocation too. All callers will have loaded at least begin() before
calling _M_reallocate.
libstdc++-v3/ChangeLog:
PR libstdc++/110498
* include/bits/vector.tcc (vector<bool, A>::_M_reallocate):
Hoist loads of begin() and end() before allocation and use them
to state an unreachable condition.
* testsuite/23_containers/vector/bool/capacity/110498.cc: New
test.
Jonathan Wakely [Fri, 28 Mar 2025 15:41:41 +0000 (15:41 +0000)]
libstdc++: Fix -Wstringop-overread warning in std::vector<bool> [PR114758]
As in r13-4393-gcca06f0d6d76b0 and a few other commits, we can avoid
bogus warnings in std::vector<bool> by hoisting some loads to before the
allocation that calls operator new. This means that the compiler has
enough info to remove the dead branches that trigger bogus warnings.
On trunk this is only needed with -fno-assume-sane-operators-new-delete
but it will help on the branches where that option doesn't exist.
libstdc++-v3/ChangeLog:
PR libstdc++/114758
* include/bits/vector.tcc (vector<bool, A>::_M_fill_insert):
Hoist loads of begin() and end() before allocation.
* testsuite/23_containers/vector/bool/capacity/114758.cc: New
test.
Jonathan Wakely [Fri, 28 Mar 2025 22:00:38 +0000 (22:00 +0000)]
libstdc++: Fix bogus -Wstringop-overflow in std::vector::insert [PR117983]
This was fixed on trunk by r15-4473-g3abe751ea86e34, but that isn't
suitable for backporting. Instead, just add another unreachable
condition in std::vector::_M_range_insert so the compiler knows this
memcpy doesn't use a length originating from a negative ptrdiff_t
converted to a very positive size_t.
libstdc++-v3/ChangeLog:
PR libstdc++/117983
* include/bits/vector.tcc (vector::_M_range_insert): Add
unreachable condition to tell the compiler begin() <= end().
* testsuite/23_containers/vector/modifiers/insert/117983.cc: New
test.
Jonathan Wakely [Mon, 31 Mar 2025 11:30:44 +0000 (12:30 +0100)]
libstdc++: Fix -Warray-bounds warning in std::vector::resize [PR114945]
This is yet another false positive warning fix. This time the compiler
can't prove that when the vector has sufficient excess capacity to
append new elements, the pointer to the existing storage is not null.
libstdc++-v3/ChangeLog:
PR libstdc++/114945
* include/bits/vector.tcc (vector::_M_default_append): Add
unreachable condition so the compiler knows that _M_finish is
not null.
* testsuite/23_containers/vector/capacity/114945.cc: New test.
Richard Biener [Wed, 5 Mar 2025 13:24:50 +0000 (14:24 +0100)]
debug/101533 - ICE with variant typedef DIE generation
There's a sanity check in gen_type_die_with_usage that trips
unnecessarily for a case where the relevant DIE has already been
generated successfully in other ways. The following keys the
existing TREE_ASM_WRITTEN check on the correct object, honoring
this and does nothing instead of ICEing for the testcase at hand.
PR debug/101533
* dwarf2out.cc (gen_type_die_with_usage): When we have
output the typedef already do nothing for a typedef variant.
Do not set TREE_ASM_WRITTEN on the type.
Richard Biener [Wed, 31 Jul 2024 08:07:45 +0000 (10:07 +0200)]
middle-end/101478 - ICE with degenerate address during gimplification
When we gimplify &MEM[0B + 4] we are re-folding the address in case
types are not canonical which ends up with a constant address that
recompute_tree_invariant_for_addr_expr ICEs on. Properly guard
that call.
PR middle-end/101478
* gimplify.cc (gimplify_addr_expr): Check we still have an
ADDR_EXPR before calling recompute_tree_invariant_for_addr_expr.
Richard Biener [Mon, 17 Feb 2025 14:53:11 +0000 (15:53 +0100)]
tree-optimization/98845 - ICE with tail-merging and DCE/DSE disabled
The following shows that tail-merging will make dead SSA defs live
in paths where it wasn't before, possibly introducing UB or as
in this case, uses of abnormals that eventually fail coalescing
later. The fix is to register such defs for stmt comparison.
PR tree-optimization/98845
* tree-ssa-tail-merge.cc (stmt_local_def): Consider a
def with no uses not local.
* gcc.dg/pr98845.c: New testcase.
* gcc.dg/pr81192.c: Adjust.
Richard Biener [Fri, 28 Feb 2025 13:09:29 +0000 (14:09 +0100)]
lto/91299 - weak definition inlined with LTO
The following fixes a thinko in the handling of interposed weak
definitions which confused the interposition check in
get_availability by setting DECL_EXTERNAL too early.
PR lto/91299
gcc/lto/
* lto-symtab.cc (lto_symtab_merge_symbols): Set DECL_EXTERNAL
only after calling get_availability.
gcc/testsuite/
* gcc.dg/lto/pr91299_0.c: New testcase.
* gcc.dg/lto/pr91299_1.c: Likewise.
Richard Biener [Fri, 28 Feb 2025 09:36:11 +0000 (10:36 +0100)]
tree-optimization/87984 - hard register assignments not preserved
The following disables redundant store elimination to hard register
variables which isn't valid.
PR tree-optimization/87984
* tree-ssa-dom.cc (dom_opt_dom_walker::optimize_stmt): Do
not perform redundant store elimination to hard register
variables.
* tree-ssa-sccvn.cc (eliminate_dom_walker::eliminate_stmt):
Likewise.
When the C++ frontend clones a CTOR we do not copy ASM_EXPR constraints
fully as walk_tree does not recurse to TREE_PURPOSE of TREE_LIST nodes.
At this point doing that seems too dangerous so the following instead
avoids gimplification of ASM_EXPRs to clobber the shared constraints
and unshares it there, like it also unshares TREE_VALUE when it
re-writes a "+" output constraint to separate "=" output and matching
input constraint.
PR middle-end/66279
* gimplify.cc (gimplify_asm_expr): Copy TREE_PURPOSE before
rewriting it for "+" processing.
Marek Polacek [Tue, 25 Mar 2025 17:36:24 +0000 (13:36 -0400)]
c++: fix missing lifetime extension [PR119383]
Since r15-8011 cp_build_indirect_ref_1 won't do the *&TARGET_EXPR ->
TARGET_EXPR folding not to change its value category. That fix seems
correct but it made us stop extending the lifetime in this testcase,
causing a wrong-code issue -- extend_ref_init_temps_1 did not see
through the extra *& because it doesn't use a tree walk.
This patch reverts r15-8011 and instead handles the problem in
build_over_call by calling force_lvalue in the is_really_empty_class
case as well as in the general case.
PR c++/119383
gcc/cp/ChangeLog:
* call.cc (build_over_call): Use force_lvalue to ensure op= returns
an lvalue.
* cp-tree.h (force_lvalue): Declare.
* cvt.cc (force_lvalue): New.
* typeck.cc (cp_build_indirect_ref_1): Revert r15-8011.
Jonathan Wakely [Tue, 1 Apr 2025 10:02:43 +0000 (11:02 +0100)]
libstdc++: Avoid bogus -Walloc-size-larger-than warning in test [PR116212]
The compiler can't tell that the vector size fits in int, so it thinks
it might overflow to a negative value, which would then be a huge
positive size_t. In reality, the vector size never exceeds five.
There's no warning on trunk, so just change the local variable to use
type unsigned so that we get rid of the warning on the branches.
libstdc++-v3/ChangeLog:
PR libstdc++/116212
* testsuite/20_util/specialized_algorithms/uninitialized_move/constrained.cc:
Use unsigned for vector size.
Martin Jambor [Fri, 14 Mar 2025 15:07:01 +0000 (16:07 +0100)]
ipa: Do not modify cgraph edges from thunk clones during inlining (PR116572)
In PR 116572 we hit an assert that a thunk which does not have a body
looks like it has one. It does not, but the call_stmt of its outgoing
edge points to a statement, which should not. In fact it has several
outgoing call graph edges, which cannot be. The problem is that the
code updating the edges to reflect inlining into the master clone (an
ex-thunk, unlike the clone, which is still an unexpanded thunk) is
being updated during inling into the master clone. This patch simply
makes the code to skip unexpanded thunk clones.
gcc/ChangeLog:
2025-03-13 Martin Jambor <mjambor@suse.cz>
PR ipa/116572
* cgraph.cc (cgraph_update_edges_for_call_stmt): Do not update
edges of clones that are unexpanded thunk. Assert that the node
passed as the parameter is not an unexpanded thunk.
Jonathan Wakely [Mon, 10 Mar 2025 14:29:36 +0000 (14:29 +0000)]
libstdc++: Add static_assert to std::packaged_task::packaged_task(F&&)
LWG 4154 (approved in Wrocław, November 2024) fixed the Mandates:
precondition for std::packaged_task::packaged_task(F&&) to match what
the implementation actually requires. We already gave a diagnostic in
the right cases as required by the issue resolution, so strictly
speaking we don't need to do anything. But the current diagnostic comes
from inside the implementation of std::__invoke_r and could be more
user-friendly.
For C++17 (when std::is_invocable_r_v is available) add a static_assert
to the constructor, so the error is clear:
Iain Buclaw [Sat, 29 Mar 2025 22:16:25 +0000 (23:16 +0100)]
d: Fix error with -Warray-bounds and -O2 [PR117002]
The record layout of class types in D don't get any tail padding, so it
is possible for the `classInstanceSize' to not be a multiple of the
`classInstanceAlignment'.
Rather than setting the instance alignment on the underlying
RECORD_TYPE, instead give the type an alignment of 1, which will mark it
as TYPE_PACKED. The value of `classInstanceAlignment' is instead
applied to the DECL_ALIGN of both the static `init' symbol, and the
stack allocated variable used when generating `new' for a `scope' class.
PR d/117002
gcc/d/ChangeLog:
* decl.cc (aggregate_initializer_decl): Set explicit decl alignment of
class instance.
* expr.cc (ExprVisitor::visit (NewExp *)): Likewise.
* types.cc (TypeVisitor::visit (TypeClass *)): Mark the record type of
classes as packed.
Jonathan Wakely [Thu, 6 Mar 2025 21:18:21 +0000 (21:18 +0000)]
libstdc++: Make std::erase for linked lists convert to bool
LWG 4135 (approved in Wrocław, November 2024) fixes the lambda
expressions used by std::erase for std::list and std::forward_list.
Previously they attempted to copy something that isn't required to be
copyable. Instead they should convert it to bool right away.
The issue resolution also changes the lambda's parameter to be const, so
that it can't modify the elements while comparing them.
libstdc++-v3/ChangeLog:
* include/std/forward_list (erase): Change lambda to have
explicit return type and const parameter type.
* include/std/list (erase): Likewise.
* testsuite/23_containers/forward_list/erasure.cc: Check lambda
is correct.
* testsuite/23_containers/list/erasure.cc: Likewise.
Jonathan Wakely [Tue, 25 Mar 2025 13:24:08 +0000 (13:24 +0000)]
libstdc++: Optimize std::vector construction from input iterators [PR108487]
LWG 3291 make std::ranges::iota_view's iterator have input_iterator_tag
as its iterator_category, even though it satisfies the C++20
std::forward_iterator concept. This means that the traditional
std::vector::vector(InputIterator, InputIterator) constructor treats
iota_view iterators as input iterators, because it only understands the
C++17 iterator requirements, not the C++20 iterator concepts. This
results in a loop that calls emplace_back for each individual element of
the iota_view, requiring the vector to reallocate repeatedly as the
values are inserted. This makes it unnecessarily slow to construct a
vector from an iota_view.
This change adds a new _M_range_initialize_n function for initializing a
vector from a range (which doesn't have to be common) and a size. This
new function can be used by vector(InputIterator, InputIterator) when
std::ranges::distance can be used to get the size. It can also be used
by the _M_range_initialize overload that gets the size for a
Cpp17ForwardIterator pair using std::distance, and by the
vector(initializer_list) constructor.
With this new function constructing a std::vector from iota_view does
a single allocation of the correct size and so doesn't need to
reallocate in a loop.
libstdc++-v3/ChangeLog:
PR libstdc++/108487
* include/bits/stl_vector.h (vector(initializer_list)): Call
_M_range_initialize_n instead of _M_range_initialize.
(vector(InputIterator, InputIterator)): Use _M_range_initialize_n
for C++20 sized sentinels and forward iterators.
(vector::_M_range_initialize(FwIt, FwIt, forward_iterator_tag)):
Use _M_range_initialize_n.
(vector::_M_range_initialize_n): New function.
* testsuite/23_containers/vector/cons/108487.cc: New test.
gcc/testsuite/ChangeLog:
* g++.dg/tree-ssa/initlist-opt1.C: Match _M_range_initialize_n
instead of _M_range_initialize.
* g++.dg/tree-ssa/initlist-opt2.C: Likewise.
Jonathan Wakely [Fri, 12 Jan 2024 16:57:41 +0000 (16:57 +0000)]
libstdc++: Update tzdata to 2025a
Import the new 2025a tzdata.zi file. The leapseconds file was also
updated to have a new expiry (no new leap seconds were added).
libstdc++-v3/ChangeLog:
* include/std/chrono (__detail::__get_leap_second_info): Update
expiry date for leap seconds list.
* src/c++20/tzdata.zi: Import new file from 2025a release.
* src/c++20/tzdb.cc (tzdb_list::_Node::_S_read_leap_seconds)
Update expiry date for leap seconds list.
Martin Uecker [Sat, 23 Nov 2024 07:04:05 +0000 (08:04 +0100)]
Fix type compatibility for types with flexible array member 2/2 [PR113688,PR114713,PR117724]
For checking or computing TYPE_CANONICAL, ignore the array size when it is
the last element of a structure or union. To not get errors because of
an inconsistent number of members, zero-sized arrays which are the last
element are not ignored anymore when checking the fields of a struct.
LRA substitute all scratches with new pseudos, so we have:
(insn 484 483 485 72 (parallel [
(set (reg/v:SI 143 [ __q1 ])
(plus:SI (reg/v:SI 143 [ __q1 ])
(const_int -2 [0xfffffffffffffffe])))
(clobber (reg:QI 619))
]) "/mnt/d/avr-lra/udivmoddi.c":163:405 discrim 5 186 {addsi3}
(expr_list:REG_UNUSED (reg:QI 619)
(nil)))
Pseudo 619 is a special scratch register generated by LRA which is marked in `scratch_bitmap' and can be tested by call `ira_former_scratch_p(regno)'.
In dump file (udivmoddi.c.317r.reload) we have:
Creating newreg=619
Removing SCRATCH to p619 in insn #484 (nop 3)
rescanning insn with uid = 484.
After that LRA tries to spill (reg:QI 619)
It's a bug because (reg:QI 619) is an output scratch register which is already something like spill register.
Fragment from udivmoddi.c.317r.reload:
Choosing alt 2 in insn 484: (0) r (1) 0 (2) nYnn (3) &d {addsi3}
Creating newreg=728 from oldreg=619, assigning class LD_REGS to r728
IMHO: the bug is in lra-constraints.cc in function `get_reload_reg'
fragment of `get_reload_reg':
if (type == OP_OUT)
{
/* Output reload registers tend to start out with a conservative
choice of register class. Usually this is ALL_REGS, although
a target might narrow it (for performance reasons) through
targetm.preferred_reload_class. It's therefore quite common
for a reload instruction to require a more restrictive class
than the class that was originally assigned to the reload register.
In these situations, it's more efficient to refine the choice
of register class rather than create a second reload register.
This also helps to avoid cycling for registers that are only
used by reload instructions. */
if (REG_P (original)
&& (int) REGNO (original) >= new_regno_start
&& INSN_UID (curr_insn) >= new_insn_uid_start
__________________________________^^
&& in_class_p (original, rclass, &new_class, true))
{
unsigned int regno = REGNO (original);
if (lra_dump_file != NULL)
{
fprintf (lra_dump_file, " Reuse r%d for output ", regno);
dump_value_slim (lra_dump_file, original, 1);
}
This condition incorrectly limits register reuse to ONLY newly generated instructions.
i.e. LRA can reuse registers only from insns generated by himself.
IMHO:It's wrong.
Scratch registers generated by LRA also have to be reused.
The patch is very simple.
On x86_64, it bootstraps+regtests fine.
projects / thirdparty / gcc.git / log
