Skip to contentSkip to conversation

Detect and use GCC atomic builtins for locking

PATCH In review Loïc Minier 3 participants 10 messages Feb 19, 2010 – Feb 28, 2010

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 Detect and use GCC atomic builtins for locking

---
 configure   |   17 +++++++++++++++++
 qemu-lock.h |   13 +++++++++++++
 2 files changed, 30 insertions(+), 0 deletions(-)
configure +17 -0
index 0a84b0e..b33f045 100755
@@ -1823,6 +1823,20 @@ if compile_prog "" ""; then
 fi
 
 ##########################################
+# check if we have gcc atomic built-ins
+gcc_atomic_builtins=no
+cat > $TMPC << EOF
+int main(void) {
+    int i;
+    __sync_lock_test_and_set(&i, 1);
+    __sync_lock_release(&i);
+}
+EOF
+if compile_prog "" ""; then
+    gcc_atomic_builtins=yes
+fi
+
+##########################################
 # check if we have fdatasync
 
 fdatasync=no
@@ -2168,6 +2182,9 @@ fi
 if test "$gcc_attribute_warn_unused_result" = "yes" ; then
   echo "CONFIG_GCC_ATTRIBUTE_WARN_UNUSED_RESULT=y" >> $config_host_mak
 fi
+if test "$gcc_atomic_builtins" = "yes" ; then
+  echo "CONFIG_GCC_ATOMIC_BUILTINS=y" >> $config_host_mak
+fi
 if test "$fdatasync" = "yes" ; then
   echo "CONFIG_FDATASYNC=y" >> $config_host_mak
 fi
qemu-lock.h +13 -0
index 9a3e6ac..5c8eb34 100644
@@ -33,6 +33,14 @@
 
 #else
 
+#ifdef CONFIG_GCC_ATOMIC_BUILTINS
+typedef int spinlock_t;
+
+#define SPIN_LOCK_UNLOCKED 0
+
+#define resetlock(p) __sync_lock_release((p))
+#else /* CONFIG_GCC_ATOMIC_BUILTINS */
+
 #if defined(__hppa__)
 
 typedef int spinlock_t[4];
@@ -56,7 +64,11 @@ static inline void resetlock (spinlock_t *p)
 }
 
 #endif
+#endif /* !CONFIG_GCC_ATOMIC_BUILTINS */
 
+#ifdef CONFIG_GCC_ATOMIC_BUILTINS
+#define testandset(p) __sync_lock_test_and_set((p), 1)
+#else /* CONFIG_GCC_ATOMIC_BUILTINS */
 #if defined(_ARCH_PPC)
 static inline int testandset (int *p)
 {
@@ -213,6 +225,7 @@ static inline int testandset (int *p)
 #else
 #error unimplemented CPU support
 #endif
+#endif /* !CONFIG_GCC_ATOMIC_BUILTINS */
 
 #if defined(CONFIG_USER_ONLY)
 static inline void spin_lock(spinlock_t *lock)
Signature · 1 line
1.7.0
#1 lore raw

PATCH 2/2 Make spinlock_t types volatile

---
 qemu-lock.h |    6 +++---
 1 files changed, 3 insertions(+), 3 deletions(-)
qemu-lock.h +3 -3
index 5c8eb34..23e3442 100644
@@ -34,7 +34,7 @@
 #else
 
 #ifdef CONFIG_GCC_ATOMIC_BUILTINS
-typedef int spinlock_t;
+typedef volatile int spinlock_t;
 
 #define SPIN_LOCK_UNLOCKED 0
 
@@ -43,7 +43,7 @@ typedef int spinlock_t;
 
 #if defined(__hppa__)
 
-typedef int spinlock_t[4];
+typedef volatile int spinlock_t[4];
 
 #define SPIN_LOCK_UNLOCKED { 1, 1, 1, 1 }
 
@@ -54,7 +54,7 @@ static inline void resetlock (spinlock_t *p)
 
 #else
 
-typedef int spinlock_t;
+typedef volatile int spinlock_t;
 
 #define SPIN_LOCK_UNLOCKED 0
 
Signature · 1 line
1.7.0
#2 lore raw
In reply to Loïc Minier

On Fri, 19 Feb 2010, Lo?c Minier wrote:

Please look up "gcc 4.1 implements compiler builtins for atomic ops" thread for reasons why this might not be the best idea.

[..snip..]

Signature · 1 line
mailto:[email protected]
#3 lore raw

NB: Addition of these builtins was prompted by qemu failing to build on armel in Ubuntu; this is because we default to Thumb 2 mode which doesn't have the assembly instructions in question.
http://launchpadlibrarian.net/38837077/buildlog_ubuntu-lucid-armel.qemu-kvm_0.12.2-0ubuntu6_FAILEDTOBUILD.txt.gz
[...]
CC arm-softmmu/syborg_virtio.o
CC arm-softmmu/exec.o
/tmp/cc24C9yx.s: Assembler messages:
/tmp/cc24C9yx.s:5392: Error: selected processor does not support swp r4,r4,[r3]' /tmp/cc24C9yx.s:6599: Error: selected processor does not support swp r6,r6,[r3]' make[2]: *** [exec.o] Error 1
make[1]: *** [subdir-arm-softmmu] Error 2
[...]

Quoted 2 lines · malc

Please look up "gcc 4.1 implements compiler builtins for atomic ops" thread for reasons why this might not be the best idea.

I found a very old thred (2005) on libc-alpha with this subject; is this the one you mean?

People participating in the libc-alpha were not really constructive and were presenting some bogus arguments; let me try to go over the various arguments (long):
- some people wanted atomic builtins to be inline for performance and others wanted them to be library calls to allow changing their behavior later (e.g. to support a new CPU); the implementation actually uses both: inline assembly when supported on the architecture, or calls into libgcc which will call into the kernel otherwise (or direct calls into the kernel when building statically obviously), such as when the ISA doesn't offer sufficient primitives. Because *any* instruction might be gotten wrong in hardware, I don't see a need to special case locking inline assembly.
- userspace apps abusing atomic builtins for locking; this is actually the case of qemu, but I think using gcc primitives will actually make it easier to get it right and will allow coverage of more architectures; in my opinion, there's no need to maintain hand-written assembly for locks in 2010.
- someone claimed that modern architectures can do these operations in assembly without calling into a library; that's what the atomic builtins do, and actually some modern architectures can't do all operations in assembly.
- there were arguments over where such functions belong and the semantics of each call; these are in my eyes purely political and don't relate to qemu.

Which are your concerns with atomic builtins and are these in the above list?

Signature · 1 line
Loïc Minier
#4 lore raw
Quoted 14 lines · Lo?c Minier

NB: Addition of these builtins was prompted by qemu failing to build on armel in Ubuntu; this is because we default to Thumb 2 mode which doesn't have the assembly instructions in question.
http://launchpadlibrarian.net/38837077/buildlog_ubuntu-lucid-armel.qemu-kvm_0.12.2-0ubuntu6_FAILEDTOBUILD.txt.gz
[...]
CC arm-softmmu/syborg_virtio.o
CC arm-softmmu/exec.o
/tmp/cc24C9yx.s: Assembler messages:
/tmp/cc24C9yx.s:5392: Error: selected processor does not support swp r4,r4,[r3]' /tmp/cc24C9yx.s:6599: Error: selected processor does not support swp r6,r6,[r3]' make[2]: *** [exec.o] Error 1
make[1]: *** [subdir-arm-softmmu] Error 2
[...]

Quoted 6 lines
Quoted 2 lines · malc

Please look up "gcc 4.1 implements compiler builtins for atomic ops" thread for reasons why this might not be the best idea.

I found a very old thred (2005) on libc-alpha with this subject; is this the one you mean?

Yes.

Quoted 26 lines

People participating in the libc-alpha were not really constructive and were presenting some bogus arguments; let me try to go over the various arguments (long):
- some people wanted atomic builtins to be inline for performance and others wanted them to be library calls to allow changing their behavior later (e.g. to support a new CPU); the implementation actually uses both: inline assembly when supported on the architecture, or calls into libgcc which will call into the kernel otherwise (or direct calls into the kernel when building statically obviously), such as when the ISA doesn't offer sufficient primitives. Because *any* instruction might be gotten wrong in hardware, I don't see a need to special case locking inline assembly.
- userspace apps abusing atomic builtins for locking; this is actually the case of qemu, but I think using gcc primitives will actually make it easier to get it right and will allow coverage of more architectures; in my opinion, there's no need to maintain hand-written assembly for locks in 2010.
- someone claimed that modern architectures can do these operations in assembly without calling into a library; that's what the atomic builtins do, and actually some modern architectures can't do all operations in assembly.
- there were arguments over where such functions belong and the semantics of each call; these are in my eyes purely political and don't relate to qemu.

Quoted 2 lines

Which are your concerns with atomic builtins and are these in the above list?

For instance this:
http://sources.redhat.com/ml/libc-alpha/2005-06/msg00112.html

The builtins are too coarse grained and will do more stuff than strictly necessary.

Signature · 1 line
mailto:[email protected]
#5 lore raw
Quoted 5 lines · malc

For instance this:
http://sources.redhat.com/ml/libc-alpha/2005-06/msg00112.html

The builtins are too coarse grained and will do more stuff than strictly necessary.

Is this the case of the builtins I'm proposing to use? We could ask for new ones without the drawbacks if any.

Do you have another option to implement locking on thumb-2?

Signature · 1 line
Loïc Minier
#6 lore raw
Quoted 11 lines · Lo?c Minier
Quoted 5 lines · malc

For instance this:
http://sources.redhat.com/ml/libc-alpha/2005-06/msg00112.html

The builtins are too coarse grained and will do more stuff than strictly necessary.

Is this the case of the builtins I'm proposing to use? We could ask for new ones without the drawbacks if any.

Do you have another option to implement locking on thumb-2?

No, i'm against using locking GCC builtins for all the other targets (well PPC), if they provide satisfactory results for thumb-2 please do submit an updated patch which uses them only for this host.

Signature · 1 line
mailto:[email protected]
#7 lore raw
Quoted 2 lines · malc

No, i'm against using locking GCC builtins for all the other targets (well PPC)

Do you have benchmarks with/without GCC atomic builtins?

Signature · 1 line
Loïc Minier
#8 lore raw
Quoted 5 lines · Lo?c Minier
Quoted 2 lines · malc

No, i'm against using locking GCC builtins for all the other targets (well PPC)

Do you have benchmarks with/without GCC atomic builtins?

No.

Signature · 1 line
mailto:[email protected]