Hi!
On 2025-03-14T11:39:20+0100, I wrote:
> As the first of a few patches to enable libstdc++ for GCN, nvptx targets,
> [...]
> some more fine-tuning is to follow later on.)
Any comments before I push the attached
"GCN, nvptx libstdc++: Force use of '__atomic' builtins [PR119645]"?
Jonathan, please put a sharp eye on the
'libstdc++-v3/acinclude.m4:GLIBCXX_ENABLE_LOCK_POLICY' change; to make
sure this only affects GCN, nvptx, but nothing else.
Grüße
Thomas
>From 1d3278050f9560666f6debcd2ead711660bebd4e Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <tschwi...@baylibre.com>
Date: Sat, 5 Apr 2025 23:11:23 +0200
Subject: [PATCH] GCN, nvptx libstdc++: Force use of '__atomic' builtins
[PR119645]
For both GCN, nvptx, this gets rid of 'configure'-time:
configure: WARNING: No native atomic operations are provided for this platform.
configure: WARNING: They will be faked using a mutex.
configure: WARNING: Performance of certain classes will degrade as a result.
..., and changes:
-checking for lock policy for shared_ptr reference counts... mutex
+checking for lock policy for shared_ptr reference counts... atomic
That means, '[...]/[target]/libstdc++-v3/', 'Makefile's change:
-ATOMICITY_SRCDIR = config/cpu/generic/atomicity_mutex
+ATOMICITY_SRCDIR = config/cpu/generic/atomicity_builtins
..., and '[...]/[target]/libstdc++-v3/config.h' changes:
/* Defined if shared_ptr reference counting should use atomic operations. */
-/* #undef HAVE_ATOMIC_LOCK_POLICY */
+#define HAVE_ATOMIC_LOCK_POLICY 1
/* Define if the compiler supports C++11 atomics. */
-/* #undef _GLIBCXX_ATOMIC_BUILTINS */
+#define _GLIBCXX_ATOMIC_BUILTINS 1
..., and '[...]/[target]/libstdc++-v3/include/[target]/bits/c++config.h'
changes:
/* Defined if shared_ptr reference counting should use atomic operations. */
-/* #undef _GLIBCXX_HAVE_ATOMIC_LOCK_POLICY */
+#define _GLIBCXX_HAVE_ATOMIC_LOCK_POLICY 1
/* Define if the compiler supports C++11 atomics. */
-/* #undef _GLIBCXX_ATOMIC_BUILTINS */
+#define _GLIBCXX_ATOMIC_BUILTINS 1
This means that '[...]/[target]/libstdc++-v3/libsupc++/atomicity.cc',
'[...]/[target]/libstdc++-v3/libsupc++/atomicity.o' then uses atomic
instructions for synchronization instead of C++ static local variables, which
in turn for their guard variables, via 'libstdc++-v3/libsupc++/guard.cc', used
'libgcc/gthr.h' recursive mutexes, which currently are unsupported for GCN.
For GCN, this turns ~500 libstdc++ execution test FAILs into PASSes, and also
progresses:
PASS: g++.dg/tree-ssa/pr20458.C -std=gnu++17 (test for excess errors)
[-FAIL:-]{+PASS:+} g++.dg/tree-ssa/pr20458.C -std=gnu++17 execution test
PASS: g++.dg/tree-ssa/pr20458.C -std=gnu++26 (test for excess errors)
[-FAIL:-]{+PASS:+} g++.dg/tree-ssa/pr20458.C -std=gnu++26 execution test
UNSUPPORTED: g++.dg/tree-ssa/pr20458.C -std=gnu++98: exception handling not supported
(For nvptx, there is no effective change, due to other misconfiguration.)
PR target/119645
libstdc++-v3/
* acinclude.m4 (GLIBCXX_ENABLE_LOCK_POLICY) [GCN, nvptx]:
Hard-code results.
* configure: Regenerate.
* configure.host [GCN, nvptx] (atomicity_dir): Set to
'cpu/generic/atomicity_builtins'.
---
libstdc++-v3/acinclude.m4 | 7 ++++---
libstdc++-v3/configure | 11 ++++++-----
libstdc++-v3/configure.host | 11 +++++++++++
3 files changed, 21 insertions(+), 8 deletions(-)
diff --git a/libstdc++-v3/acinclude.m4 b/libstdc++-v3/acinclude.m4
index 912e85a46679..5b9f0c5ee89d 100644
--- a/libstdc++-v3/acinclude.m4
+++ b/libstdc++-v3/acinclude.m4
@@ -4007,10 +4007,11 @@ AC_DEFUN([GLIBCXX_ENABLE_LOCK_POLICY], [
dnl Why don't we check 8-byte CAS for sparc64, where _Atomic_word is long?!
dnl New targets should only check for CAS for the _Atomic_word type.
AC_TRY_COMPILE([
- #if defined __riscv
+ #if defined __AMDGCN__ || defined __nvptx__
+ /* Yes, please. */
+ #elif defined __riscv
# error "Defaulting to mutex-based locks for ABI compatibility"
- #endif
- #if ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_2
+ #elif ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_2
# error "No 2-byte compare-and-swap"
#elif ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
# error "No 4-byte compare-and-swap"
diff --git a/libstdc++-v3/configure b/libstdc++-v3/configure
index 0b4481f0739c..2b2eedeb2e71 100755
--- a/libstdc++-v3/configure
+++ b/libstdc++-v3/configure
@@ -16393,10 +16393,11 @@ ac_compiler_gnu=$ac_cv_cxx_compiler_gnu
cat confdefs.h - <<_ACEOF >conftest.$ac_ext
/* end confdefs.h. */
- #if defined __riscv
+ #if defined __AMDGCN__ || defined __nvptx__
+ /* Yes, please. */
+ #elif defined __riscv
# error "Defaulting to mutex-based locks for ABI compatibility"
- #endif
- #if ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_2
+ #elif ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_2
# error "No 2-byte compare-and-swap"
#elif ! defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
# error "No 4-byte compare-and-swap"
@@ -16443,7 +16444,7 @@ $as_echo "mutex" >&6; }
# unnecessary for this test.
cat > conftest.$ac_ext << EOF
-#line 16446 "configure"
+#line 16447 "configure"
int main()
{
_Decimal32 d1;
@@ -16485,7 +16486,7 @@ ac_compiler_gnu=$ac_cv_cxx_compiler_gnu
# unnecessary for this test.
cat > conftest.$ac_ext << EOF
-#line 16488 "configure"
+#line 16489 "configure"
template<typename T1, typename T2>
struct same
{ typedef T2 type; };
diff --git a/libstdc++-v3/configure.host b/libstdc++-v3/configure.host
index 8375764bf4dc..05ba8fc797c7 100644
--- a/libstdc++-v3/configure.host
+++ b/libstdc++-v3/configure.host
@@ -370,10 +370,21 @@ case "${host}" in
;;
esac
;;
+ amdgcn-*-amdhsa)
+ # To avoid greater pain elsewhere, force use of '__atomic' builtins,
+ # irregardless of outcome of 'configure' checks; see PR119645
+ # "GCN, nvptx: libstdc++ 'checking for atomic builtins [...]... no'".
+ atomicity_dir=cpu/generic/atomicity_builtins
+ ;;
arm*-*-freebsd*)
port_specific_symbol_files="\$(srcdir)/../config/os/gnu-linux/arm-eabi-extra.ver"
;;
nvptx-*-none)
+ # To avoid greater pain elsewhere, force use of '__atomic' builtins,
+ # irregardless of outcome of 'configure' checks; see PR119645
+ # "GCN, nvptx: libstdc++ 'checking for atomic builtins [...]... no'".
+ atomicity_dir=cpu/generic/atomicity_builtins
+
# For 'make all-target-libstdc++-v3', re 'alloca'/VLA usage:
EXTRA_CFLAGS="${EXTRA_CFLAGS} -mfake-ptx-alloca"
OPTIMIZE_CXXFLAGS="${OPTIMIZE_CXXFLAGS} -mfake-ptx-alloca"
--
2.34.1
