| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Fix a typo in __gthr_win32_abs_to_rel_time that caused it to return a
relative time in seconds instead of milliseconds. As a consequence,
__gthr_win32_cond_timedwait called SleepConditionVariableCS with a
1000x shorter timeout; this caused ~1000x more spurious wakeups in
CV timed waits such as std::condition_variable::wait_for or wait_until,
resulting generally in much higher CPU usage.
This can be demonstrated by this sample program:
```
int main() {
std::condition_variable cv;
std::mutex mx;
bool pass = false;
auto thread_fn = [&](bool timed) {
int wakeups = 0;
using sc = std::chrono::system_clock;
auto before = sc::now();
std::unique_lock<std::mutex> ml(mx);
if (timed) {
cv.wait_for(ml, std::chrono::seconds(2), [&]{
++wakeups;
return pass;
});
} else {
cv.wait(ml, [&]{
++wakeups;
return pass;
});
}
printf("pass: %d; wakeups: %d; elapsed: %d ms\n", pass, wakeups,
int((sc::now() - before) / std::chrono::milliseconds(1)));
pass = false;
};
{
// timed wait, let expire
std::thread t(thread_fn, true);
t.join();
}
{
// timed wait, wake up explicitly after 1 second
std::thread t(thread_fn, true);
std::this_thread::sleep_for(std::chrono::seconds(1));
{
std::unique_lock<std::mutex> ml(mx);
pass = true;
}
cv.notify_all();
t.join();
}
{
// non-timed wait, wake up explicitly after 1 second
std::thread t(thread_fn, false);
std::this_thread::sleep_for(std::chrono::seconds(1));
{
std::unique_lock<std::mutex> ml(mx);
pass = true;
}
cv.notify_all();
t.join();
}
return 0;
}
```
On builds based on non-affected threading models (e.g. POSIX on Linux,
or winpthreads or MCF on Win32) the output is something like
```
pass: 0; wakeups: 2; elapsed: 2000 ms
pass: 1; wakeups: 2; elapsed: 991 ms
pass: 1; wakeups: 2; elapsed: 996 ms
```
while with the Win32 threading model we get
```
pass: 0; wakeups: 1418; elapsed: 2000 ms
pass: 1; wakeups: 479; elapsed: 988 ms
pass: 1; wakeups: 2; elapsed: 992 ms
```
(notice the huge number of wakeups in the timed wait cases only).
This commit fixes the conversion, adjusting the final division by
NSEC100_PER_SEC to use NSEC100_PER_MSEC instead (already defined in the
file and not used in any other place, so probably just a typo).
libgcc/ChangeLog:
PR libgcc/113850
* config/i386/gthr-win32-cond.c (__gthr_win32_abs_to_rel_time):
fix absolute timespec to relative milliseconds count
conversion (it incorrectly returned seconds instead of
milliseconds); this avoids spurious wakeups in
__gthr_win32_cond_timedwait
|
|
|
|
Add x32 and IBT support to x86 heap trampoline implementation with a
testcase.
2024-02-13 Jakub Jelinek <jakub@redhat.com>
H.J. Lu <hjl.tools@gmail.com>
libgcc/
PR target/113855
* config/i386/heap-trampoline.c (trampoline_insns): Add IBT
support and pad to the multiple of 4 bytes. Use movabsq
instead of movabs in comments. Add -mx32 variant.
gcc/testsuite/
PR target/113855
* gcc.dg/heap-trampoline-1.c: New test.
* lib/target-supports.exp (check_effective_target_heap_trampoline):
New.
|
|
|
|
As I wrote earlier, I was seeing
FAIL: gcc.dg/torture/bitint-24.c -O0 execution test
FAIL: gcc.dg/torture/bitint-24.c -O2 execution test
with the ia32 _BitInt enablement patch on i686-linux. I thought
floatbitintxf.c was miscompiled with -O2 -march=i686 -mtune=generic, but it
turned out to be UB in it.
If a signed _BitInt to be converted to binary floating point has
(after sign extension from possible partial limb to full limb) one or
more most significant limbs equal to all ones and then in the limb below
(the most significant non-~(UBILtype)0 limb) has the most significant limb
cleared, like for 32-bit limbs
0x81582c05U, 0x0a8b01e4U, 0xc1b8b18fU, 0x2aac2a08U, -1U, -1U
then bitint_reduce_prec can't reduce it to that 0x2aac2a08U limb, so
msb is all ones and precision is negative (so it reduced precision from
161 to 192 bits down to 160 bits, in theory could go as low as 129 bits
but that wouldn't change anything on the following behavior).
But still iprec is negative, -160 here.
For that case (i.e. where we are dealing with an negative input), the
code was using 65 - __builtin_clzll (~msb) to compute how many relevant
bits we have from the msb. Unfortunately that invokes UB for msb all ones.
The right number of relevant bits in that case is 1 though (like for
-2 it is 2 and -4 or -3 3 as already computed) - all we care about from that
is that the most significant bit is set (i.e. the number is negative) and
the bits below that should be supplied from the limbs below.
So, the following patch fixes it by special casing it not to invoke UB.
For msb 0 we already have a special case from before (but that is also
different because msb 0 implies the whole number is 0 given the way
bitint_reduce_prec works - even if we have limbs like ..., 0x80000000U, 0U
the reduction can skip the most significant limb and msb then would be
the one below it), so if iprec > 0, we already don't call __builtin_clzll
on 0.
2024-02-13 Jakub Jelinek <jakub@redhat.com>
* soft-fp/bitint.h (FP_FROM_BITINT): If iprec < 0 and msb is all ones,
just set n to 1 instead of using __builtin_clzll (~msb).
|
|
|
|
The initial heap trampoline implementation was targeting 64b
platforms. As the PR demonstrates this creates an issue where it
is expected that the same symbols are exported for 32 and 64b.
Rather than conditionalize the exports and code-gen on x86_64,
this patch provides a basic implementation of the IA32 trampoline.
This also avoids potential user confusion, when a 32b target has
64b multilibs, and vice versa; which is the case for Darwin.
PR target/113855
gcc/ChangeLog:
* config/i386/darwin.h (DARWIN_HEAP_T_LIB): Moved to be
available to all sub-targets.
* config/i386/darwin32-biarch.h (DARWIN_HEAP_T_LIB): Delete.
* config/i386/darwin64-biarch.h (DARWIN_HEAP_T_LIB): Delete.
libgcc/ChangeLog:
* config.host: Add trampoline support to x?86-linux.
* config/i386/heap-trampoline.c (trampoline_insns): Provide
a variant for IA32.
(union ix86_trampoline): Likewise.
(__gcc_nested_func_ptr_created): Implement a basic trampoline
for IA32.
|
|
|
|
The ia32 _BitInt support revealed a bug in floatbitint?d.c.
As can be even guessed from how the code is written in the loop,
the intention was to set inexact to non-zero whenever the remainder
after division wasn't zero, but I've ended up just checking whether
the 2 least significant limbs of the remainder were non-zero.
Now, in the dfp/bitint-4.c test in one case the remainder happens
to have least significant 64 bits zero and then the higher limbs are
non-zero; with 32-bit limbs that means 2 least significant limbs are zero
and so the code acted as if it was exactly divisible.
Fixed thusly.
2024-02-10 Jakub Jelinek <jakub@redhat.com>
* soft-fp/floatbitintdd.c (__bid_floatbitintdd): Or in all remainder
limbs into inexact rather than just first two.
* soft-fp/floatbitintsd.c (__bid_floatbitintsd): Likewise.
* soft-fp/floatbitinttd.c (__bid_floatbitinttd): Likewise.
|
|
I've tried last night to enable _BitInt support for i?86-linux, and
a few spots in libgcc emitted -Wshift-count-overflow warnings and clearly
didn't do what it was supposed to do.
Fixed thusly.
2024-02-10 Jakub Jelinek <jakub@redhat.com>
* soft-fp/fixddbitint.c (__bid_fixddbitint): Fix up
BIL_TYPE_SIZE == 32 shifts.
* soft-fp/fixsdbitint.c (__bid_fixsdbitint): Likewise.
* soft-fp/fixtdbitint.c (__bid_fixtdbitint): Likewise.
* soft-fp/floatbitintdd.c (__bid_floatbitintdd): Likewise.
* soft-fp/floatbitinttd.c (__bid_floatbitinttd): Likewise.
|
|
|
|
Some exports were missed from the GCC-13 cycle, these are added here
along with the bitint-related ones added in GCC-14.
libgcc/ChangeLog:
* config/i386/libgcc-darwin.ver: Export bf and bitint-related
synbols.
|
|
|
|
As reported in the PR, all libgcc x86 symbol versions added after
GCC_7.0.0 were only added to i386/libgcc-glibc.ver, missing all of
libgcc-sol2.ver, libgcc-bsd.ver, and libgcc-darwin.ver.
This patch fixes this for Solaris/x86, adding all of them
(GCC_1[234].0.0) as GCC_14.0.0 to not retroactively change history.
Since this isn't the first time this happens, I've added a note to the
end of libgcc-glibc.ver to request notifying other maintainers in case
of additions.
Tested on i386-pc-solaris2.11.
2024-02-01 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
libgcc:
PR target/113700
* config/i386/libgcc-sol2.ver (GCC_14.0.0): Added all symbols from
i386/libgcc-glibc.ver (GCC_12.0.0, GCC_13.0.0, GCC_14.0.0).
* config/i386/libgcc-glibc.ver: Request notifications on updates.
|
|
SEH _Unwind_Resume_or_Rethrow invokes abort directly if
_Unwind_RaiseException doesn't manage to find a handler for the rethrown
exception; this is incorrect, as in this case std::terminate should be
invoked, allowing an application-provided terminate handler to handle
the situation instead of straight crashing the application through
abort.
The bug can be demonstrated with this simple test case:
===
static void custom_terminate_handler() {
fprintf(stderr, "custom_terminate_handler invoked\n");
std::exit(1);
}
int main(int argc, char *argv[]) {
std::set_terminate(&custom_terminate_handler);
if (argc < 2) return 1;
const char *mode = argv[1];
fprintf(stderr, "%s\n", mode);
if (strcmp(mode, "throw") == 0) {
throw std::exception();
} else if (strcmp(mode, "rethrow") == 0) {
try {
throw std::exception();
} catch (...) {
throw;
}
} else {
return 1;
}
return 0;
}
===
On all gcc builds with non-SEH exceptions, this will print
"custom_terminate_handler invoked" both if launched as ./a.out throw or
as ./a.out rethrow, on SEH builds instead if will work as expected only
with ./a.exe throw, but will crash with the "built-in" abort message
with ./a.exe rethrow.
This patch fixes the problem, forwarding back the error code to the
caller (__cxa_rethrow), that calls std::terminate if
_Unwind_Resume_or_Rethrow returns.
The change makes the code path coherent with SEH _Unwind_RaiseException,
and with the generic _Unwind_Resume_or_Rethrow from libgcc/unwind.inc
(used for SjLj and Dw2 exception backend).
libgcc/ChangeLog:
PR libgcc/113337
* unwind-seh.c (_Unwind_Resume_or_Rethrow): forward
_Unwind_RaiseException return code back to caller instead of
calling abort, allowing __cxa_rethrow to invoke std::terminate
in case of uncaught rethrown exception
|
|
|
|
The following testcase ends up with SIGFPE in __divmodbitint4.
The problem is a thinko in my attempt to implement Knuth's algorithm.
The algorithm does (where b is 65536, i.e. one larger than what
fits in their unsigned short word):
// Compute estimate qhat of q[j].
qhat = (un[j+n]*b + un[j+n-1])/vn[n-1];
rhat = (un[j+n]*b + un[j+n-1]) - qhat*vn[n-1];
again:
if (qhat >= b || qhat*vn[n-2] > b*rhat + un[j+n-2])
{ qhat = qhat - 1;
rhat = rhat + vn[n-1];
if (rhat < b) goto again;
}
The problem is that it uses a double-word / word -> double-word
division (and modulo), while all we have is udiv_qrnnd unless
we'd want to do further library calls, and udiv_qrnnd is a
double-word / word -> word division and modulo.
Now, as the algorithm description says, it can produce at most
word bits + 1 bit quotient. And I believe that actually the
highest qhat the original algorithm can produce is
(1 << word_bits) + 1. The algorithm performs earlier canonicalization
where both the divisor and dividend are shifted left such that divisor
has msb set. If it has msb set already before, no shifting occurs but
we start with added 0 limb, so in the first uv1:uv0 double-word uv1
is 0 and so we can't get too high qhat, if shifting occurs, the first
limb of dividend is shifted right by UWtype bits - shift count into
a new limb, so again in the first iteration in the uv1:uv0 double-word
uv1 doesn't have msb set while vv1 does and qhat has to fit into word.
In the following iterations, previous iteration should guarantee that
the previous quotient digit is correct. Even if the divisor was the
maximal possible vv1:all_ones_in_all_lower_limbs, if the old uv0:lower_limbs
would be larger or equal to the divisor, the previous quotient digit
would increase and another divisor would be subtracted, which I think
implies that in the next iteration in uv1:uv0 double-word uv1 <= vv1,
but uv0 could be up to all ones, e.g. in case of all lower limbs
of divisor being all ones and at least one dividend limb below uv0
being not all ones. So, we can e.g. for 64-bit UWtype see
uv1:uv0 / vv1 0x8000000000000000UL:0xffffffffffffffffUL / 0x8000000000000000UL
or 0xffffffffffffffffUL:0xffffffffffffffffUL / 0xffffffffffffffffUL
In all these cases (when uv1 == vv1 && uv0 >= uv1), qhat is
0x10000000000000001UL, i.e. 2 more than fits into UWtype result,
if uv1 == vv1 && uv0 < uv1 it would be 0x10000000000000000UL, i.e.
1 more than fits into UWtype result.
Because we only have udiv_qrnnd which can't deal with those too large
cases (SIGFPEs or otherwise invokes undefined behavior on those), I've
tried to handle the uv1 >= vv1 case separately, but for one thing
I thought it would be at most 1 larger than what fits, and for two
have actually subtracted vv1:vv1 from uv1:uv0 instead of subtracting
0:vv1 from uv1:uv0.
For the uv1 < vv1 case, the implementation already performs roughly
what the algorithm does.
Now, let's see what happens with the two possible extra cases in
the original algorithm.
If uv1 == vv1 && uv0 < uv1, qhat above would be b, so we take
if (qhat >= b, decrement qhat by 1 (it becomes b - 1), add
vn[n-1] aka vv1 to rhat and goto again if rhat < b (but because
qhat already fits we can goto to the again label in the uv1 < vv1
code). rhat in this case is uv0 and rhat + vv1 can but doesn't
have to overflow, say for uv0 42UL and vv1 0x8000000000000000UL
it will not (and so we should goto again), while for uv0
0x8000000000000000UL and vv1 0x8000000000000001UL it will (and
we shouldn't goto again).
If uv1 == vv1 && uv0 >= uv1, qhat above would be b + 1, so we
take if (qhat >= b, decrement qhat by 1 (it becomes b), add
vn[n-1] aka vv1 to rhat. But because vv1 has msb set and
rhat in this case is uv0 - vv1, the rhat + vv1 addition
certainly doesn't overflow, because (uv0 - vv1) + vv1 is uv0,
so in the algorithm we goto again, again take if (qhat >= b and
decrement qhat so it finally becomes b - 1, and add vn[n-1]
aka vv1 to rhat again. But this time I believe it must always
overflow, simply because we added (uv0 - vv1) + vv1 + vv1 and
vv1 has msb set, so already vv1 + vv1 must overflow. And
because it overflowed, it will not goto again.
So, I believe the following patch implements this correctly, by
subtracting vv1 from uv1:uv0 double-word once, then comparing
again if uv1 >= vv1. If that is true, subtract vv1 from uv1:uv0
again and add 2 * vv1 to rhat, no __builtin_add_overflow is needed
as we know it always overflowed and so won't goto again.
If after the first subtraction uv1 < vv1, use __builtin_add_overflow
when adding vv1 to rhat, because it can but doesn't have to overflow.
I've added an extra testcase which tests the behavior of all the changed
cases, so it has a case where uv1:uv0 / vv1 is 1:1, where it is
1:0 and rhat + vv1 overflows and where it is 1:0 and rhat + vv1 does not
overflow, and includes tests also from Zdenek's other failing tests.
2024-02-02 Jakub Jelinek <jakub@redhat.com>
PR libgcc/113604
* libgcc2.c (__divmodbitint4): If uv1 >= vv1, subtract
vv1 from uv1:uv0 once or twice as needed, rather than
subtracting vv1:vv1.
* gcc.dg/torture/bitint-53.c: New test.
* gcc.dg/torture/bitint-55.c: New test.
|
|
libgcc/
* config/i386/enable-execute-stack-mingw32.c: Include
stdlib.h for abort() definition.
|
|
libgcc_s.so.1 [PR113700]
Rainer pointed out that __PFX__ and __FIXPTPFX__ prefix replacement is done
solely for libgcc-std.ver.in and not for the *.ver files in config.
I've used the __PFX__ prefix even in config/i386/libgcc-glibc.ver because it
was used for similar symbols in libgcc-std.ver.in, and that results in those
symbols being STB_LOCAL in libgcc_s.so.1. Tests still work because gcc by
default uses -static-libgcc when linking (unlike g++ etc.), but would
have failed when using -shared-libgcc (but I see nothing in the testsuite
actually testing with -shared-libgcc, so am not adding tests).
With the patch, libgcc_s.so.1 now exports
__fixtfbitint@@GCC_14.0.0 FUNC GLOBAL DEFAULT
__fixxfbitint@@GCC_14.0.0 FUNC GLOBAL DEFAULT
__floatbitintbf@@GCC_14.0.0 FUNC GLOBAL DEFAULT
__floatbitinthf@@GCC_14.0.0 FUNC GLOBAL DEFAULT
__floatbitinttf@@GCC_14.0.0 FUNC GLOBAL DEFAULT
__floatbitintxf@@GCC_14.0.0 FUNC GLOBAL DEFAULT
on x86_64-linux which it wasn't before.
2024-02-02 Jakub Jelinek <jakub@redhat.com>
PR target/113700
* config/i386/libgcc-glibc.ver (GCC_14.0.0): Remove __PFX prefixes
from symbol names.
|
|
|
|
I'm seeing hundreds of
In file included from ../../../libgcc/libgcc2.c:56:
../../../libgcc/libgcc2.h:32:13: warning: conflicting types for built-in function ‘__gcc_nested_func_ptr_created’; expected ‘void(void *, void *, void *)’
+[-Wbuiltin-declaration-mismatch]
32 | extern void __gcc_nested_func_ptr_created (void *, void *, void **);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
warnings.
Either we need to add like in r14-6218
#pragma GCC diagnostic ignored "-Wbuiltin-declaration-mismatch"
(but in that case because of the libgcc2.h prototype (why is it there?)
it would need to be also with #pragma GCC diagnostic push/pop around),
or we could go with just following how the builtins are prototyped on the
compiler side and only cast to void ** when dereferencing (which is in
a single spot in each TU).
2024-02-01 Jakub Jelinek <jakub@redhat.com>
PR libgcc/113402
* libgcc2.h (__gcc_nested_func_ptr_created): Change type of last
argument from void ** to void *.
* config/i386/heap-trampoline.c (__gcc_nested_func_ptr_created):
Change type of dst from void ** to void * and cast dst to void **
before dereferencing it.
* config/aarch64/heap-trampoline.c (__gcc_nested_func_ptr_created):
Likewise.
|
|
I'm seeing
../../../libgcc/shared-object.mk:14: warning: overriding recipe for target 'heap-trampoline.o'
../../../libgcc/shared-object.mk:14: warning: ignoring old recipe for target 'heap-trampoline.o'
../../../libgcc/shared-object.mk:17: warning: overriding recipe for target 'heap-trampoline_s.o'
../../../libgcc/shared-object.mk:17: warning: ignoring old recipe for target 'heap-trampoline_s.o'
This patch fixes that.
2024-02-01 Jakub Jelinek <jakub@redhat.com>
PR libgcc/113403
* config/i386/t-heap-trampoline: Add to LIB2ADDEHSHARED
i386/heap-trampoline.c rather than aarch64/heap-trampoline.c.
|
|
|
|
Use aarch64-asm.h in asm code consistently, this was started in
commit c608ada288ced0268bbbbc1fd4136f56c34b24d4
Author: Zac Walker <zacwalker@microsoft.com>
CommitDate: 2024-01-23 15:32:30 +0000
Ifdef `.hidden`, `.type`, and `.size` pseudo-ops for `aarch64-w64-mingw32` target
But that commit failed to remove some existing markings from asm files,
which means some objects got double marked with gnu property notes.
libgcc/ChangeLog:
* config/aarch64/crti.S: Remove stack marking.
* config/aarch64/crtn.S: Remove stack marking, include aarch64-asm.h
* config/aarch64/lse.S: Remove stack and GNU property markings.
|
|
|
|
In order to handle system security constraints during GCC build
and test and that most platform versions cannot link to libgcc_eh
since the unwinder there is incompatible with the system one.
1. We make the support functions weak definitions.
2. We include them as a CRT for platform conditions that do not
allow libgcc_eh.
3. We ensure that the weak symbols are exported from DSOs (which
includes exes on Darwin) so that the dynamic linker will
pick one instance (which avoids duplication of trampoline
caches).
PR libgcc/113403
gcc/ChangeLog:
* config/darwin.h (DARWIN_SHARED_WEAK_ADDS, DARWIN_WEAK_CRTS): New.
(REAL_LIBGCC_SPEC): Move weak CRT handling to separate spec.
* config/i386/darwin.h (DARWIN_HEAP_T_LIB): New.
* config/i386/darwin32-biarch.h (DARWIN_HEAP_T_LIB): New.
* config/i386/darwin64-biarch.h (DARWIN_HEAP_T_LIB): New.
* config/rs6000/darwin.h (DARWIN_HEAP_T_LIB): New.
libgcc/ChangeLog:
* config.host: Build libheap_t.a for i686/x86_64 Darwin.
* config/aarch64/heap-trampoline.c (HEAP_T_ATTR): New.
(allocate_tramp_ctrl): Allow a target to build this as a weak def.
(__gcc_nested_func_ptr_created): Likewise.
* config/i386/heap-trampoline.c (HEAP_T_ATTR): New.
(allocate_tramp_ctrl): Allow a target to build this as a weak def.
(__gcc_nested_func_ptr_created): Likewise.
* config/t-darwin: Build libheap_t.a (a CRT with heap trampoline
support).
|
|
This removes the heap trampoline support functions from libgcc.a and
adds them to libgcc_eh.a. They are also present in libgcc_s.
PR libgcc/113403
libgcc/ChangeLog:
* config/aarch64/t-heap-trampoline: Move the heap trampoline
support functions from libgcc.a to libgcc_eh.a.
* config/i386/t-heap-trampoline: Likewise.
|
|
|
|
The symbols for the functions supporting heap-based trampolines were
exported at an incorrect symbol version, the following patch fixes that.
As requested in the PR, this also renames __builtin_nested_func_ptr* to
__gcc_nested_func_ptr*. In carrying our the rename, we move the builtins
to use DEF_EXT_LIB_BUILTIN.
PR libgcc/113402
gcc/ChangeLog:
* builtins.cc (expand_builtin): Handle BUILT_IN_GCC_NESTED_PTR_CREATED
and BUILT_IN_GCC_NESTED_PTR_DELETED.
* builtins.def (BUILT_IN_GCC_NESTED_PTR_CREATED,
BUILT_IN_GCC_NESTED_PTR_DELETED): Make these builtins LIB-EXT and
rename the library fallbacks to __gcc_nested_func_ptr_created and
__gcc_nested_func_ptr_deleted.
* doc/invoke.texi: Rename these to __gcc_nested_func_ptr_created
and __gcc_nested_func_ptr_deleted.
* tree-nested.cc (finalize_nesting_tree_1): Use builtin_explicit for
BUILT_IN_GCC_NESTED_PTR_CREATED and BUILT_IN_GCC_NESTED_PTR_DELETED.
* tree.cc (build_common_builtin_nodes): Build the
BUILT_IN_GCC_NESTED_PTR_CREATED and BUILT_IN_GCC_NESTED_PTR_DELETED local
builtins only for non-explicit.
libgcc/ChangeLog:
* config/aarch64/heap-trampoline.c: Rename
__builtin_nested_func_ptr_created to __gcc_nested_func_ptr_created and
__builtin_nested_func_ptr_deleted to __gcc_nested_func_ptr_deleted.
* config/i386/heap-trampoline.c: Likewise.
* libgcc2.h: Likewise.
* libgcc-std.ver.in (GCC_7.0.0): Likewise and then move
__gcc_nested_func_ptr_created and
__gcc_nested_func_ptr_deleted from this symbol version to ...
(GCC_14.0.0): ... this one.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
Co-authored-by: Jakub Jelinek <jakub@redhat.com>
|
|
|
|
This is enough to get gfx1030 and gfx1100 working; there are still some test
failures to investigate, and probably some tuning to do.
gcc/ChangeLog:
* config/gcn/gcn-opts.h (TARGET_PACKED_WORK_ITEMS): Add TARGET_RDNA3.
* config/gcn/gcn-valu.md (all_convert): New iterator.
(<convop><V_INT_1REG_ALT:mode><V_INT_1REG:mode>2<exec>): New
define_expand, and rename the old one to ...
(*<convop><V_INT_1REG_ALT:mode><V_INT_1REG:mode>_sdwa<exec>): ... this.
(extend<V_INT_1REG_ALT:mode><V_INT_1REG:mode>2<exec>): Likewise, to ...
(extend<V_INT_1REG_ALT:mode><V_INT_1REG:mode>_sdwa<exec>): .. this.
(*<convop><V_INT_1REG_ALT:mode><V_INT_1REG:mode>_shift<exec>): New.
* config/gcn/gcn.cc (gcn_global_address_p): Use "offsetbits" correctly.
(gcn_hsa_declare_function_name): Update the vgpr counting for gfx1100.
* config/gcn/gcn.md (<u>mulhisi3): Disable on RDNA3.
(<u>mulqihi3_scalar): Likewise.
libgcc/ChangeLog:
* config/gcn/amdgcn_veclib.h (CDNA3_PLUS): Handle RDNA3.
libgomp/ChangeLog:
* config/gcn/time.c (RTC_TICKS): Configure RDNA3.
(omp_get_wtime): Add RDNA3-compatible variant.
* plugin/plugin-gcn.c (max_isa_vgprs): Tune for gfx1030 and gfx1100.
Signed-off-by: Andrew Stubbs <ams@baylibre.com>
|
|
|
|
target
Recent
change (https://gcc.gnu.org/pipermail/gcc-cvs/2023-December/394915.html)
added a generic SME support using `.hidden`, `.type`, and ``.size`
pseudo-ops in the assembly sources, `aarch64-w64-mingw32` does not
support the pseudo-ops though. This patch wraps usage of those
pseudo-ops using macros and ifdefs them for `__ELF__` define.
libgcc/
* config/aarch64/aarch64-asm.h (HIDDEN, SYMBOL_SIZE, SYMBOL_TYPE)
(ENTRY_ALIGN, GNU_PROPERTY): New macros.
* config/aarch64/__arm_sme_state.S: Use them.
* config/aarch64/__arm_tpidr2_save.S: Likewise.
* config/aarch64/__arm_za_disable.S: Likewise.
* config/aarch64/crti.S: Likewise.
* config/aarch64/lse.S: Likewise.
|
|
|
|
As discussed on IRC, the following patch uses may_alias attribute, so that
on targets like aarch64 where abi_limb_mode != limb_mode the library
accesses the limbs (half limbs of the ABI) in the arrays with conservative
alias set.
2024-01-12 Jakub Jelinek <jakub@redhat.com>
* libgcc2.h (UBILtype): New typedef with may_alias attribute.
(__mulbitint3, __divmodbitint4): Use UBILtype * instead of
UWtype * and const UBILtype * instead of const UWtype *.
* libgcc2.c (bitint_reduce_prec, bitint_mul_1, bitint_addmul_1,
__mulbitint3, bitint_negate, bitint_submul_1, __divmodbitint4):
Likewise.
* soft-fp/bitint.h (UBILtype): Change define into a typedef with
may_alias attribute.
|
|
Exception handling on nios2-linux-gnu with -fpic has been broken since
revision 790854ea7670f11c14d431c102a49181d2915965, "Use _dl_find_object
in _Unwind_Find_FDE". For whatever reason, this doesn't work on nios2.
Nios2 uses the GOT address as the base for DW_EH_PE_datarel
relocations in PIC; see my previous fix to make this work, revision
2d33dcfe9f0494c9b56a8d704c3d27c5a4329ebc, "Support for GOT-relative
DW_EH_PE_datarel encoding". So this may be a horrible bug in the ABI
or in my interpretation of it or just glibc's implementation of
_dl_find_object for this target, but there's existing code out there
that does things this way; and realistically, nobody is going to
re-engineer this now that the vendor has EOL'ed the nios2
architecture. So, just skip over the code trying to use
_dl_find_object on this target and fall back to the way that works.
I plan to backport this patch to the GCC 12 and GCC 13 branches as well.
libgcc/ChangeLog
* unwind-dw2-fde-dip.c (_Unwind_Find_FDE): Do not try to use
_dl_find_object on nios2; it doesn't work.
|
|
|
|
2023 -> 2024
|
|
|
|
variables support
For now, for single-threaded GCN, nvptx target use only; extension for
multi-threaded offloading use is to follow later. Eventually switch to
libstdc++-v3/libsupc++ proper.
libgcc/
* c++-minimal/README: New.
* c++-minimal/guard.c: New.
* config/gcn/t-amdgcn (LIB2ADD): Add it.
* config/nvptx/t-nvptx (LIB2ADD): Likewise.
|
|
|
|
If we allow __strub_leave to allocate a frame on sparc, it will
overlap with a lot of the stack range we're supposed to scrub, because
of the large fixed-size outgoing args and register save area.
Unfortunately, setting up the PIC register seems to prevent the frame
pointer from being omitted.
Since the strub runtime doesn't issue calls or use global variables,
at least on sparc, disabling PIC to compile strub.c seems to do the
right thing.
for libgcc/ChangeLog
PR middle-end/112917
* config.host (sparc, sparc64): Enable...
* config/sparc/t-sparc: ... this new fragment.
|
|
|
|
The strub builtins are not suited for cross-unit inlining, they should
only be inlined by the builtin expanders, if at all. While testing on
sparc64, it occurred to me that, if libgcc was built with LTO enabled,
lto1 might inline them, and that would likely break things. So, make
sure they're clearly marked as not inlinable.
for libgcc/ChangeLog
* strub.c (ATTRIBUTE_NOINLINE): New.
(ATTRIBUTE_STRUB_CALLABLE): Add it.
(__strub_dummy_force_no_leaf): Drop it.
|
|
|
|
This adds initial support for function multiversioning on aarch64 using
the target_version and target_clones attributes. This loosely follows
the Beta specification in the ACLE [1], although with some differences
that still need to be resolved (possibly as follow-up patches).
Existing function multiversioning implementations are broken in various
ways when used across translation units. This includes placing
resolvers in the wrong translation units, and using symbol mangling that
callers to unintentionally bypass the resolver in some circumstances.
Fixing these issues for aarch64 will require modifications to our ACLE
specification. It will also require further adjustments to existing
middle end code, to facilitate different mangling and resolver
placement while preserving existing target behaviours.
The list of function multiversioning features specified in the ACLE is
also inconsistent with the list of features supported in target option
extensions. I intend to resolve some or all of these inconsistencies at
a later stage.
The target_version attribute is currently only supported in C++, since
this is the only frontend with existing support for multiversioning
using the target attribute. On the other hand, this patch happens to
enable multiversioning with the target_clones attribute in Ada and D, as
well as the entire C family, using their existing frontend support.
This patch also does not support the following aspects of the Beta
specification:
- The target_clones attribute should allow an implicit unlisted
"default" version.
- There should be an option to disable function multiversioning at
compile time.
- Unrecognised target names in a target_clones attribute should be
ignored (with an optional warning). This current patch raises an
error instead.
[1] https://github.com/ARM-software/acle/blob/main/main/acle.md#function-multi-versioning
gcc/ChangeLog:
* config/aarch64/aarch64-feature-deps.h (fmv_deps_<FEAT_NAME>):
Define aarch64_feature_flags mask foreach FMV feature.
* config/aarch64/aarch64-option-extensions.def: Use new macros
to define FMV feature extensions.
* config/aarch64/aarch64.cc (aarch64_option_valid_attribute_p):
Check for target_version attribute after processing target
attribute.
(aarch64_fmv_feature_data): New.
(aarch64_parse_fmv_features): New.
(aarch64_process_target_version_attr): New.
(aarch64_option_valid_version_attribute_p): New.
(get_feature_mask_for_version): New.
(compare_feature_masks): New.
(aarch64_compare_version_priority): New.
(build_ifunc_arg_type): New.
(make_resolver_func): New.
(add_condition_to_bb): New.
(dispatch_function_versions): New.
(aarch64_generate_version_dispatcher_body): New.
(aarch64_get_function_versions_dispatcher): New.
(aarch64_common_function_versions): New.
(aarch64_mangle_decl_assembler_name): New.
(TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P): New implementation.
(TARGET_OPTION_EXPANDED_CLONES_ATTRIBUTE): New implementation.
(TARGET_OPTION_FUNCTION_VERSIONS): New implementation.
(TARGET_COMPARE_VERSION_PRIORITY): New implementation.
(TARGET_GENERATE_VERSION_DISPATCHER_BODY): New implementation.
(TARGET_GET_FUNCTION_VERSIONS_DISPATCHER): New implementation.
(TARGET_MANGLE_DECL_ASSEMBLER_NAME): New implementation.
* config/aarch64/aarch64.h (TARGET_HAS_FMV_TARGET_ATTRIBUTE):
Set target macro.
* config/arm/aarch-common.h (enum aarch_parse_opt_result): Add
new value to report duplicate FMV feature.
* common/config/aarch64/cpuinfo.h: New file.
libgcc/ChangeLog:
* config/aarch64/cpuinfo.c (enum CPUFeatures): Move to shared
copy in gcc/common
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/options_set_17.c: Reorder expected flags.
* gcc.target/aarch64/cpunative/native_cpu_0.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_13.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_16.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_17.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_18.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_19.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_20.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_21.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_22.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_6.c: Ditto.
* gcc.target/aarch64/cpunative/native_cpu_7.c: Ditto.
|
|
This is added to enable function multiversioning, but can also be used
directly. The interface is chosen to match that used in LLVM's
compiler-rt, to facilitate cross-compiler compatibility.
The content of the patch is derived almost entirely from Pavel's prior
contributions to compiler-rt/lib/builtins/cpu_model.c. I have made minor
changes to align more closely with GCC coding style, and to exclude any code
from other LLVM contributors, and am adding this to GCC with Pavel's approval.
libgcc/ChangeLog:
* config/aarch64/t-aarch64: Include cpuinfo.c
* config/aarch64/cpuinfo.c: New file
(__init_cpu_features_constructor) New.
(__init_cpu_features_resolver) New.
(__init_cpu_features) New.
Co-authored-by: Pavel Iliin <Pavel.Iliin@arm.com>
|
|
|
|
This patch try to introduce the rwlock and split the read/write to
unit_root tree and unit_cache with rwlock instead of the mutex to
increase CPU efficiency. In the get_gfc_unit function, the percentage
to step into the insert_unit function is around 30%, in most instances,
we can get the unit in the phase of reading the unit_cache or unit_root
tree. So split the read/write phase by rwlock would be an approach to
make it more parallel.
BTW, the IPC metrics can gain around 9x in our test
server with 220 cores. The benchmark we used is
https://github.com/rwesson/NEAT
libgcc/ChangeLog:
* gthr-posix.h (__GTHREAD_RWLOCK_INIT): New macro.
(__gthrw): New function.
(__gthread_rwlock_rdlock): New function.
(__gthread_rwlock_tryrdlock): New function.
(__gthread_rwlock_wrlock): New function.
(__gthread_rwlock_trywrlock): New function.
(__gthread_rwlock_unlock): New function.
libgfortran/ChangeLog:
* io/async.c (DEBUG_LINE): New macro.
* io/async.h (RWLOCK_DEBUG_ADD): New macro.
(CHECK_RDLOCK): New macro.
(CHECK_WRLOCK): New macro.
(TAIL_RWLOCK_DEBUG_QUEUE): New macro.
(IN_RWLOCK_DEBUG_QUEUE): New macro.
(RDLOCK): New macro.
(WRLOCK): New macro.
(RWUNLOCK): New macro.
(RD_TO_WRLOCK): New macro.
(INTERN_RDLOCK): New macro.
(INTERN_WRLOCK): New macro.
(INTERN_RWUNLOCK): New macro.
* io/io.h (struct gfc_unit): Change UNIT_LOCK to UNIT_RWLOCK in
a comment.
(unit_lock): Remove including associated internal_proto.
(unit_rwlock): New declarations including associated internal_proto.
(dec_waiting_unlocked): Use WRLOCK and RWUNLOCK on unit_rwlock
instead of __gthread_mutex_lock and __gthread_mutex_unlock on
unit_lock.
* io/transfer.c (st_read_done_worker): Use WRLOCK and RWUNLOCK on
unit_rwlock instead of LOCK and UNLOCK on unit_lock.
(st_write_done_worker): Likewise.
* io/unit.c: Change UNIT_LOCK to UNIT_RWLOCK in 'IO locking rules'
comment. Use unit_rwlock variable instead of unit_lock variable.
(get_gfc_unit_from_unit_root): New function.
(get_gfc_unit): Use RDLOCK, WRLOCK and RWUNLOCK on unit_rwlock
instead of LOCK and UNLOCK on unit_lock.
(close_unit_1): Use WRLOCK and RWUNLOCK on unit_rwlock instead of
LOCK and UNLOCK on unit_lock.
(close_units): Likewise.
(newunit_alloc): Use RWUNLOCK on unit_rwlock instead of UNLOCK on
unit_lock.
* io/unix.c (find_file): Use RDLOCK and RWUNLOCK on unit_rwlock
instead of LOCK and UNLOCK on unit_lock.
(flush_all_units): Use WRLOCK and RWUNLOCK on unit_rwlock instead
of LOCK and UNLOCK on unit_lock.
|
|
|
