[PATCH] Native spinlock support on RISC-V

Marek Szuba <marecki@gentoo.org> writes:

Tested against PostgreSQL 13.3 on a physical rv64gc system (BeagleV
Starlight beta board) - builds and installs fine, all tests pass.

Cool ... I had hoped to acquire one of those myself, but I didn't
make the cut.

regards, tom lane

Hi,

On 2021-08-13 11:09:04 -0400, Tom Lane wrote:

Marek Szuba <marecki@gentoo.org> writes:

Tested against PostgreSQL 13.3 on a physical rv64gc system (BeagleV
Starlight beta board) - builds and installs fine, all tests pass.

Seems like a good idea to me.

Cool ... I had hoped to acquire one of those myself, but I didn't
make the cut.

Should we backpatch this? It's not like we're going to break existing
risc-v systems by enabling spinlock support...

Greetings,

Andres Freund

Andres Freund <andres@anarazel.de> writes:

On 2021-08-13 11:09:04 -0400, Tom Lane wrote:

Marek Szuba <marecki@gentoo.org> writes:

Tested against PostgreSQL 13.3 on a physical rv64gc system (BeagleV
Starlight beta board) - builds and installs fine, all tests pass.

Should we backpatch this? It's not like we're going to break existing
risc-v systems by enabling spinlock support...

Yeah, why not? If you were building with --disable-spinlocks before,
this shouldn't change anything for you.

(I haven't actually looked at the patch, mind you, but in principle
it shouldn't break anything that worked before.)

regards, tom lane

I wrote:

Andres Freund <andres@anarazel.de> writes:

Should we backpatch this? It's not like we're going to break existing
risc-v systems by enabling spinlock support...

Yeah, why not? If you were building with --disable-spinlocks before,
this shouldn't change anything for you.
(I haven't actually looked at the patch, mind you, but in principle
it shouldn't break anything that worked before.)

I now have looked at the patch, and it seems good as far as it goes,
but I wonder whether some effort ought to be expended in
src/include/port/atomics/.

regards, tom lane

Hi,

On Fri, Aug 13, 2021, at 19:25, Tom Lane wrote:

I wrote:

Andres Freund <andres@anarazel.de> writes:

Should we backpatch this? It's not like we're going to break existing
risc-v systems by enabling spinlock support...

Yeah, why not? If you were building with --disable-spinlocks before,
this shouldn't change anything for you.
(I haven't actually looked at the patch, mind you, but in principle
it shouldn't break anything that worked before.)

I now have looked at the patch, and it seems good as far as it goes,
but I wonder whether some effort ought to be expended in
src/include/port/atomics/.

That should automatically pick up the intrinsic. I think we should do the same on modern compilers for spinlocks, but that's a separate discussion I guess.

Address

"Andres Freund" <andres@anarazel.de> writes:

On Fri, Aug 13, 2021, at 19:25, Tom Lane wrote:

I now have looked at the patch, and it seems good as far as it goes,
but I wonder whether some effort ought to be expended in
src/include/port/atomics/.

That should automatically pick up the intrinsic. I think we should do the same on modern compilers for spinlocks, but that's a separate discussion I guess.

I was looking at the comment in atomics.h about

* Provide a full fallback of the pg_*_barrier(), pg_atomic**_flag and
* pg_atomic_* APIs for platforms without sufficient spinlock and/or atomics
* support. In the case of spinlock backed atomics the emulation is expected
* to be efficient, although less so than native atomics support.

so it seems like someday we might want to expend some effort on native
atomics. I agree that that day need not be today, though. This patch
seems sufficient until we get to the point of (at least) having some
RISC-V in the buildfarm.

regards, tom lane

Hi,

On 2021-08-13 13:37:02 -0400, Tom Lane wrote:

"Andres Freund" <andres@anarazel.de> writes:

On Fri, Aug 13, 2021, at 19:25, Tom Lane wrote:

I now have looked at the patch, and it seems good as far as it goes,
but I wonder whether some effort ought to be expended in
src/include/port/atomics/.

That should automatically pick up the intrinsic. I think we should do the same on modern compilers for spinlocks, but that's a separate discussion I guess.

I was looking at the comment in atomics.h about

* Provide a full fallback of the pg_*_barrier(), pg_atomic**_flag and
* pg_atomic_* APIs for platforms without sufficient spinlock and/or atomics
* support. In the case of spinlock backed atomics the emulation is expected
* to be efficient, although less so than native atomics support.

so it seems like someday we might want to expend some effort on native
atomics. I agree that that day need not be today, though. This patch
seems sufficient until we get to the point of (at least) having some
RISC-V in the buildfarm.

For gcc compatible compilers the relevant comments would be

* There exist generic, hardware independent, implementations for several
* compilers which might be sufficient, although possibly not optimal, for a
* new platform. If no such generic implementation is available spinlocks (or
* even OS provided semaphores) will be used to implement the API.

and

/*
* Compiler specific, but architecture independent implementations.
*
* Provide architecture independent implementations of the atomic
* facilities. At the very least compiler barriers should be provided, but a
* full implementation of
* * pg_compiler_barrier(), pg_write_barrier(), pg_read_barrier()
* * pg_atomic_compare_exchange_u32(), pg_atomic_fetch_add_u32()
* using compiler intrinsics are a good idea.
*/

Gcc's intriniscs are pretty good these days (and if not, a *lot* of
projects just outright break).

What's more, It turns out that using intrinsics with compilers of the
last ~5 years often generates *better* code than inline assembly
(e.g. because the compiler can utilize condition codes more effectively
and has more detailed information about clobbers). So for new platforms
that'll only have support by new compilers it doesn't really make sense
to add inline assembler imo.

Greetings,

Andres Freund

Andres Freund <andres@anarazel.de> writes:

On 2021-08-13 13:37:02 -0400, Tom Lane wrote:

so it seems like someday we might want to expend some effort on native
atomics. I agree that that day need not be today, though. This patch
seems sufficient until we get to the point of (at least) having some
RISC-V in the buildfarm.

Gcc's intriniscs are pretty good these days (and if not, a *lot* of
projects just outright break).
What's more, It turns out that using intrinsics with compilers of the
last ~5 years often generates *better* code than inline assembly
(e.g. because the compiler can utilize condition codes more effectively
and has more detailed information about clobbers). So for new platforms
that'll only have support by new compilers it doesn't really make sense
to add inline assembler imo.

I didn't say it had to be __asm blocks ;-). I was thinking more of the
sort of stuff we have in e.g. arch-arm.h and arch-ia64.h, where we know
some things about what is efficient or less efficient on a particular
architecture. gcc will do its best to provide implementations of
its builtins, but that doesn't mean that using a particular one of
them is always the most sane thing to do.

But anyway, that seems like minor optimization that maybe someday
somebody will be motivated to do. We're on the same page about this
being enough for today.

I did not like confusing the RISC-V case with the ARM case: duplicating
the code block seems better, in case there's ever reason to add
arch-specific stuff like SPIN_DELAY. So I split it off to its own
code block and pushed it.

regards, tom lane

On 13 Aug 2021, at 20:09, Tom Lane <tgl@sss.pgh.pa.us> wrote:

..I split it off to its own code block and pushed it.

There didn’t seem to be anything left here (at least until there is hardware in
the buildfarm) so I took the liberty to close it as committed in the CF app.

--
Daniel Gustafsson https://vmware.com/

Daniel Gustafsson <daniel@yesql.se> writes:

There didn’t seem to be anything left here (at least until there is hardware in
the buildfarm) so I took the liberty to close it as committed in the CF app.

Ah, sorry, I did not realize there was a CF entry.

regards, tom lane

Re: Tom Lane

I did not like confusing the RISC-V case with the ARM case: duplicating
the code block seems better, in case there's ever reason to add
arch-specific stuff like SPIN_DELAY. So I split it off to its own
code block and pushed it.

Fwiw I can report success on Debian's riscv port with that commit
cherry-picked onto 13.4:

https://buildd.debian.org/status/fetch.php?pkg=postgresql-13&amp;arch=riscv64&amp;ver=13.4-3&amp;stamp=1630411643&amp;raw=0

Christoph

On Wed, Sep 1, 2021 at 9:22 PM Christoph Berg <myon@debian.org> wrote:

Re: Tom Lane

I did not like confusing the RISC-V case with the ARM case: duplicating
the code block seems better, in case there's ever reason to add
arch-specific stuff like SPIN_DELAY. So I split it off to its own
code block and pushed it.

Fwiw I can report success on Debian's riscv port with that commit
cherry-picked onto 13.4:

https://buildd.debian.org/status/fetch.php?pkg=postgresql-13&amp;arch=riscv64&amp;ver=13.4-3&amp;stamp=1630411643&amp;raw=0

A couple of things I noticed on this architecture:

1. Even though we're using generic built-ins for atomics, I guess we
could still use a src/include/port/atomics/arch-riscv.h file so we
have a place to define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY when
building for 64 bit. We'd need to find the chapter-and-verse to
justify that, of course, which I can try to do; if someone more
familiar with the ISA/spec/manual can point to it I'm all ears.

2. When configured with all options on FreeBSD 13, everything looks
good so far except LLVM JIT, which fails with various "Cannot select"
errors. Clang works fine for compiling PostgreSQL itself. Tested
with LLVM 12 (LLVM has supported RISCV since 9). Example:

+FATAL:  fatal llvm error: Cannot select: 0x4f772068: ch = brcond
0x4f770f70, 0x4f772208, BasicBlock:ch< 0x4f76d600>
+  0x4f772208: i64 = setcc 0x4f7723a8, Constant:i64<0>, setlt:ch
+    0x4f7723a8: i64,ch = load<(load 4 from `i32* inttoptr (i64
1260491408 to i32*)`, align 16), sext from i32> 0x4fdee058,
Constant:i64<1260491408>, undef:i64
+      0x4f770a90: i64 = Constant<1260491408>
+      0x4f7703a8: i64 = undef
+    0x4f7701a0: i64 = Constant<0>
+In function: evalexpr_0_0

Hi,

On November 2, 2021 3:55:58 PM PDT, Thomas Munro <thomas.munro@gmail.com> wrote:

2. When configured with all options on FreeBSD 13, everything looks
good so far except LLVM JIT, which fails with various "Cannot select"
errors. Clang works fine for compiling PostgreSQL itself. Tested
with LLVM 12 (LLVM has supported RISCV since 9). Example:

+FATAL:  fatal llvm error: Cannot select: 0x4f772068: ch = brcond
0x4f770f70, 0x4f772208, BasicBlock:ch< 0x4f76d600>
+  0x4f772208: i64 = setcc 0x4f7723a8, Constant:i64<0>, setlt:ch
+    0x4f7723a8: i64,ch = load<(load 4 from `i32* inttoptr (i64
1260491408 to i32*)`, align 16), sext from i32> 0x4fdee058,
Constant:i64<1260491408>, undef:i64
+      0x4f770a90: i64 = Constant<1260491408>
+      0x4f7703a8: i64 = undef
+    0x4f7701a0: i64 = Constant<0>
+In function: evalexpr_0_0

Any chance you could enable jit_dump_bitcode and manually try a failing query? That should dump. bc files in the data directory. That'd might allow debugging this outside the emulated environment.

I don't see where the undef is originating from, but I think it might suggest that something lead to that code being considered unreachable.

Andres
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

On Wed, Nov 3, 2021 at 5:13 PM Andres Freund <andres@anarazel.de> wrote:

Any chance you could enable jit_dump_bitcode and manually try a failing query? That should dump. bc files in the data directory. That'd might allow debugging this outside the emulated environment.

I don't see where the undef is originating from, but I think it might suggest that something lead to that code being considered unreachable.

postgres=# set jit_above_cost = 0;
SET
postgres=# select 1 + 1;
FATAL: fatal llvm error: Cannot select: 0x4b3ec1a0: i64,ch =
load<(load 8 from %ir.14)> 0x41ef6fe8, 0x4b3ec138, undef:i64
0x4b3ec138: i64 = add nuw 0x4b3eab60, Constant:i64<24>
0x4b3eab60: i64,ch = load<(load 8 from %ir.7)> 0x41ef6fe8,
0x4b3eaaf8, undef:i64
0x4b3eaaf8: i64 = add nuw 0x4b3ea068, Constant:i64<16>
0x4b3ea068: i64,ch = CopyFromReg 0x41ef6fe8, Register:i64 %4
0x4b3ea000: i64 = Register %4
0x4b3ea888: i64 = Constant<16>
0x4b3ea6e8: i64 = undef
0x4b3ea9c0: i64 = Constant<24>
0x4b3ea6e8: i64 = undef
In function: evalexpr_0_0

Ah, I hadn't noticed this in the log before:

'generic' is not a recognized processor for this target (ignoring processor)

Sounds kinda serious :-)

Resulting .bc files and .ll files (produced by llvm-dis) attached.

On Wed, Nov 3, 2021 at 5:41 PM Thomas Munro <thomas.munro@gmail.com> wrote:

'generic' is not a recognized processor for this target (ignoring processor)

I still don't know what's wrong but I spent 20 minutes searching for
more clues this morning...

First, 'generic' is coming from LLVMGetHostCPUName(), and yet it's not
accepted by LLVMCreateTargetMachine(). Ok then, let's try something
else:

$ clang12 --print-supported-cpus

Target: riscv64-portbld-freebsd13.0
Thread model: posix
InstalledDir: /usr/local/llvm12/bin
Available CPUs for this target:

generic-rv32
generic-rv64
rocket-rv32
rocket-rv64
sifive-7-rv32
sifive-7-rv64
sifive-e31
sifive-e76
sifive-u54
sifive-u74

So I hacked my copy of PostgreSQL thusly:

+ cpu = LLVMCreateMessage("generic-rv64");

... and now I get:

2021-11-03 20:27:28.487 UTC [26880] FATAL: fatal llvm error:
Relocation type not implemented yet!

Taking that at face value, instead of LLVMRelocDefault, I tried a
couple of other values like LLVMRelocDynamicNoPic, but same result.
So I attached a debugger to see the stack producing the error, and
huh, I see a clue that it's confusing this architecture with SystemZ
(= IBM mainframe).

(gdb) break fatal_llvm_error_handler
Function "fatal_llvm_error_handler" not defined.
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (fatal_llvm_error_handler) pending.
(gdb) cont
Continuing.

Breakpoint 1, fatal_llvm_error_handler (user_data=0x0,
reason="Relocation type not implemented yet!", gen_crash_diag=true) at
llvmjit_error.cpp:147
147 ereport(FATAL,
(gdb) bt
#0 fatal_llvm_error_handler (user_data=0x0, reason="Relocation type
not implemented yet!", gen_crash_diag=true) at llvmjit_error.cpp:147
#1 0x000000004dc3729a in llvm::report_fatal_error(llvm::Twine const&,
bool) () from /usr/local/llvm12/lib/libLLVM-12.so
#2 0x000000004dc371d0 in llvm::report_fatal_error(char const*, bool)
() from /usr/local/llvm12/lib/libLLVM-12.so
#3 0x000000004ef61010 in
llvm::RuntimeDyldELF::resolveSystemZRelocation(llvm::SectionEntry
const&, unsigned long, unsigned long, unsigned int, long) () from
/usr/local/llvm12/lib/libLLVM-12.so
#4 0x000000004ef547c6 in
llvm::RuntimeDyldImpl::applyExternalSymbolRelocations(llvm::StringMap<llvm::JITEvaluatedSymbol,
llvm::MallocAllocator>) () from /usr/local/llvm12/lib/libLLVM-12.so
#5 0x000000004ef54fb4 in ?? () from /usr/local/llvm12/lib/libLLVM-12.so
#6 0x000000004ef54be6 in
llvm::RuntimeDyldImpl::finalizeAsync(std::__1::unique_ptr<llvm::RuntimeDyldImpl,
std::__1::default_delete<llvm::RuntimeDyldImpl> >,
llvm::unique_function<void
(llvm::object::OwningBinary<llvm::object::ObjectFile>,
std::__1::unique_ptr<llvm::RuntimeDyld::LoadedObjectInfo,
std::__1::default_delete<llvm::RuntimeDyld::LoadedObjectInfo> >,
llvm::Error)>, llvm::object::OwningBinary<llvm::object::ObjectFile>,
std::__1::unique_ptr<llvm::RuntimeDyld::LoadedObjectInfo,
std::__1::default_delete<llvm::RuntimeDyld::LoadedObjectInfo> >) ()
from /usr/local/llvm12/lib/libLLVM-12.so
#7 0x000000004ef55e2a in
llvm::jitLinkForORC(llvm::object::OwningBinary<llvm::object::ObjectFile>,
llvm::RuntimeDyld::MemoryManager&, llvm::JITSymbolResolver&, bool,
llvm::unique_function<llvm::Error (llvm::object::ObjectFile const&,
llvm::RuntimeDyld::LoadedObjectInfo&, std::__1::map<llvm::StringRef,
llvm::JITEvaluatedSymbol, std::__1::less<llvm::StringRef>,
std::__1::allocator<std::__1::pair<llvm::StringRef const,
llvm::JITEvaluatedSymbol> > >)>, llvm::unique_function<void
(llvm::object::OwningBinary<llvm::object::ObjectFile>,
std::__1::unique_ptr<llvm::RuntimeDyld::LoadedObjectInfo,
std::__1::default_delete<llvm::RuntimeDyld::LoadedObjectInfo> >,
llvm::Error)>) ()
from /usr/local/llvm12/lib/libLLVM-12.so
#8 0x000000004ef421ea in
llvm::orc::RTDyldObjectLinkingLayer::emit(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >,
std::__1::unique_ptr<llvm::MemoryBuffer,
std::__1::default_delete<llvm::MemoryBuffer> >) () from
/usr/local/llvm12/lib/libLLVM-12.so
#9 0x000000004ef3e946 in
llvm::orc::ObjectTransformLayer::emit(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >,
std::__1::unique_ptr<llvm::MemoryBuffer,
std::__1::default_delete<llvm::MemoryBuffer> >) () from
/usr/local/llvm12/lib/libLLVM-12.so
#10 0x000000004ef25c0c in
llvm::orc::IRCompileLayer::emit(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >,
llvm::orc::ThreadSafeModule) () from
/usr/local/llvm12/lib/libLLVM-12.so
#11 0x000000004ef25fb2 in
llvm::orc::IRTransformLayer::emit(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >,
llvm::orc::ThreadSafeModule) () from
/usr/local/llvm12/lib/libLLVM-12.so
#12 0x000000004ef25fb2 in
llvm::orc::IRTransformLayer::emit(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >,
llvm::orc::ThreadSafeModule) () from
/usr/local/llvm12/lib/libLLVM-12.so
#13 0x000000004ef2a614 in
llvm::orc::BasicIRLayerMaterializationUnit::materialize(std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >)
() from /usr/local/llvm12/lib/libLLVM-12.so
#14 0x000000004ef08dc2 in ?? () from /usr/local/llvm12/lib/libLLVM-12.so
#15 0x000000004ef0f61c in std::__1::__function::__func<void
(*)(std::__1::unique_ptr<llvm::orc::MaterializationUnit,
std::__1::default_delete<llvm::orc::MaterializationUnit> >,
std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility> >),
std::__1::allocator<void
(*)(std::__1::unique_ptr<llvm::orc::MaterializationUnit,
std::__1::default_delete<llvm::orc::MaterializationUnit> >,
std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility>

)>, void (std::__1::unique_ptr<llvm::orc::MaterializationUnit,

std::__1::default_delete<llvm::orc::MaterializationUnit> >,
std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility>

)>::operator()(std::__1::unique_ptr<llvm::orc::MaterializationUnit,

std::__1::default_delete<llvm::orc::MaterializationUnit> >&&,
std::__1::unique_ptr<llvm::orc::MaterializationResponsibility,
std::__1::default_delete<llvm::orc::MaterializationResponsibility>

&&) ()

from /usr/local/llvm12/lib/libLLVM-12.so
#16 0x000000004ef09c12 in
llvm::orc::ExecutionSession::dispatchOutstandingMUs() () from
/usr/local/llvm12/lib/libLLVM-12.so
#17 0x000000004ef0b684 in
llvm::orc::ExecutionSession::OL_completeLookup(std::__1::unique_ptr<llvm::orc::InProgressLookupState,
std::__1::default_delete<llvm::orc::InProgressLookupState> >,
std::__1::shared_ptr<llvm::orc::AsynchronousSymbolQuery>,
std::__1::function<void (llvm::DenseMap<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> >,
llvm::DenseMapInfo<llvm::orc::JITDylib*>,
llvm::detail::DenseMapPair<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> > > > const&)>) () from
/usr/local/llvm12/lib/libLLVM-12.so
#18 0x000000004ef141ee in ?? () from /usr/local/llvm12/lib/libLLVM-12.so
#19 0x000000004ef01c1e in
llvm::orc::ExecutionSession::OL_applyQueryPhase1(std::__1::unique_ptr<llvm::orc::InProgressLookupState,
std::__1::default_delete<llvm::orc::InProgressLookupState> >,
llvm::Error) () from /usr/local/llvm12/lib/libLLVM-12.so
#20 0x000000004ef0096c in
llvm::orc::ExecutionSession::lookup(llvm::orc::LookupKind,
std::__1::vector<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags>,
std::__1::allocator<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags> > > const&,
llvm::orc::SymbolLookupSet, llvm::orc::SymbolState,
llvm::unique_function<void
(llvm::Expected<llvm::DenseMap<llvm::orc::SymbolStringPtr,
llvm::JITEvaluatedSymbol,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr>,
llvm::detail::DenseMapPair<llvm::orc::SymbolStringPtr,
llvm::JITEvaluatedSy--Type <RET> for more, q to quit, c to continue
without paging--
mbol> > >)>, std::__1::function<void
(llvm::DenseMap<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> >,
llvm::DenseMapInfo<llvm::orc::JITDylib*>,
llvm::detail::DenseMapPair<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> > > > const&)>) () from
/usr/local/llvm12/lib/libLLVM-12.so
#21 0x000000004ef09f8a in
llvm::orc::ExecutionSession::lookup(std::__1::vector<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags>,
std::__1::allocator<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags> > > const&, llvm::orc::SymbolLookupSet
const&, llvm::orc::LookupKind, llvm::orc::SymbolState,
std::__1::function<void (llvm::DenseMap<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> >,
llvm::DenseMapInfo<llvm::orc::JITDylib*>,
llvm::detail::DenseMapPair<llvm::orc::JITDylib*,
llvm::DenseSet<llvm::orc::SymbolStringPtr,
llvm::DenseMapInfo<llvm::orc::SymbolStringPtr> > > > const&)>) ()
from /usr/local/llvm12/lib/libLLVM-12.so
#22 0x000000004ef0a2ba in
llvm::orc::ExecutionSession::lookup(std::__1::vector<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags>,
std::__1::allocator<std::__1::pair<llvm::orc::JITDylib*,
llvm::orc::JITDylibLookupFlags> > > const&,
llvm::orc::SymbolStringPtr, llvm::orc::SymbolState) () from
/usr/local/llvm12/lib/libLLVM-12.so
#23 0x000000004ef2b724 in
llvm::orc::LLJIT::lookupLinkerMangled(llvm::orc::JITDylib&,
llvm::orc::SymbolStringPtr) ()
from /usr/local/llvm12/lib/libLLVM-12.so
#24 0x000000004ef41364 in LLVMOrcLLJITLookup () from
/usr/local/llvm12/lib/libLLVM-12.so
#25 0x0000000041a67c06 in llvm_get_function (context=0x41f75530,
funcname=0x41ff8070 "evalexpr_0_0") at llvmjit.c:311
#26 0x0000000041a73b9c in ExecRunCompiledExpr (state=0x41ff7770,
econtext=0x41ff7470, isNull=0x3fffffd5b7) at llvmjit_expr.c:2402
#27 0x00000000004f70ea in ExecEvalExprSwitchContext (state=0x41ff7770,
econtext=0x41ff7470, isNull=0x3fffffd5b7)
at ../../../src/include/executor/executor.h:339
#28 ExecProject (projInfo=0x41ff7768) at
../../../src/include/executor/executor.h:373
#29 ExecResult (pstate=<optimized out>) at nodeResult.c:136
#30 0x00000000004c90b0 in ExecProcNode (node=0x41ff7358) at
../../../src/include/executor/executor.h:257
#31 ExecutePlan (estate=0x41ff7120, planstate=0x41ff7358,
use_parallel_mode=<optimized out>, operation=<optimized out>,
numberTuples=<optimized out>, direction=<optimized out>,
dest=0x41ff5d80, execute_once=<optimized out>, sendTuples=<optimized
out>)
at execMain.c:1551
#32 standard_ExecutorRun (queryDesc=0x41e9c520, direction=<optimized
out>, count=<optimized out>, execute_once=<optimized out>)
at execMain.c:361
#33 0x000000000064e356 in PortalRunSelect (portal=0x41f24120,
forward=<optimized out>, count=0, dest=<optimized out>) at
pquery.c:921
#34 0x000000000064df50 in PortalRun (portal=0x41f24120,
count=9223372036854775807, isTopLevel=true, run_once=<optimized out>,
dest=0x41ff5d80, altdest=0x41ff5d80, qc=0x3fffffd9b8) at pquery.c:765
#35 0x000000000064d040 in exec_simple_query (query_string=0x41e1f120
"select 1 + 1;") at postgres.c:1214
#36 0x000000000064a674 in PostgresMain (dbname=<optimized out>,
username=<optimized out>) at postgres.c:4313
#37 0x00000000005b62d6 in BackendRun (port=<optimized out>) at postmaster.c:4560
#38 0x00000000005b5ab2 in BackendStartup (port=0x41ec2000) at postmaster.c:4288
#39 ServerLoop () at postmaster.c:1801
#40 0x00000000005b2fec in PostmasterMain (argc=3, argv=0x3fffffeac0)
at postmaster.c:1473
#41 0x000000000051d8ca in main (argc=3, argv=0x3fffffeac0) at main.c:198

On Wed, Nov 3, 2021 at 11:55 AM Thomas Munro <thomas.munro@gmail.com> wrote:

1. Even though we're using generic built-ins for atomics, I guess we
could still use a src/include/port/atomics/arch-riscv.h file so we
have a place to define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY when
building for 64 bit. We'd need to find the chapter-and-verse to
justify that, of course, which I can try to do; if someone more
familiar with the ISA/spec/manual can point to it I'm all ears.

I randomly remembered this topic after seeing an s390x announcement
from Christoph[1]https://www.postgresql.org/about/news/postgresql-on-s390x-on-debian-and-ubuntu-2752/, and figured he or someone else might be interested
in the same observation about that platform. That is, we finally got
around to defining this for ARM, but I bet one internet point that
RISCV64 and s390x have that property too (but would need to find a
written reference, or perhaps a similar declaration in the Linux
sources, etc):

$ git grep '#define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY'
src/include/port/atomics/arch-arm.h:#define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
src/include/port/atomics/arch-ppc.h:#define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
src/include/port/atomics/arch-x86.h:#define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY

[1]: https://www.postgresql.org/about/news/postgresql-on-s390x-on-debian-and-ubuntu-2752/

Re: Thomas Munro

I randomly remembered this topic after seeing an s390x announcement
from Christoph[1], and figured he or someone else might be interested
in the same observation about that platform. That is, we finally got
around to defining this for ARM, but I bet one internet point that
RISCV64 and s390x have that property too (but would need to find a
written reference, or perhaps a similar declaration in the Linux
sources, etc):

Fwiw, while s390x is admittedly more of historical interest, riscv64
is well on the way of becoming an official release architecture with
the next Debian stable release, so having it supported properly by
PostgreSQL would be appreciated.

Christoph