Skip to contentSkip to conversation

Pre-release LTP build

LTP Cyril Hrubis 2 participants 9 messages Apr 11, 2016 – Apr 13, 2016 Library code

lore
Share & tools
lore
#0 lore raw

Hi!
I've tried to build LTP source for several distributions and as I expected the support for compiler atomic operations is missing in old compilers. Particulary gcc-4.1 on x86 (32bit RHEL5) and gcc-3.4 (SLES9) and older which were released in 2007 and 2006.

The questions is: do we care enough to provide fallback inline assembler?

I would be inclined to say that we don't. So does anybody out there still uses LTP with a compiler old enough that it does not support atomic builtins?

Signature · 2 lines
Cyril Hrubis
[email protected]
#1 lore raw

----- Original Message -----

Quoted 11 lines

From: "Cyril Hrubis" <[email protected]>
To: [email protected]
Sent: Monday, 11 April, 2016 7:23:05 PM
Subject: [LTP] Pre-release LTP build

Hi!
I've tried to build LTP source for several distributions and as I expected the support for compiler atomic operations is missing in old compilers. Particulary gcc-4.1 on x86 (32bit RHEL5) and gcc-3.4 (SLES9) and older which were released in 2007 and 2006.

Quoted 5 lines

The questions is: do we care enough to provide fallback inline assembler?

I would be inclined to say that we don't. So does anybody out there still uses LTP with a compiler old enough that it does not support atomic builtins?

I do. At least RHEL5.6 is going to be around for another year.

I have repeatedly built your patch-series on RHEL5.6 x86_64, which works fine. But as you pointed out 32bit version fails, other arches possibly too.

My first thought was some kind of lock, so we don't have to care about each architecture separately.

#2 lore raw

Hi!

Quoted 7 lines

I do. At least RHEL5.6 is going to be around for another year.

I have repeatedly built your patch-series on RHEL5.6 x86_64, which works fine. But as you pointed out 32bit version fails, other arches possibly too.

My first thought was some kind of lock, so we don't have to care about each architecture separately.

I think that easiest solution would be copying the assembler the builtin produces and using it as a fallback. That would have to be done per arch, but the rest of the code could be left untouched.

Signature · 2 lines
Cyril Hrubis
[email protected]
#3 lore raw

----- Original Message -----

Quoted 13 lines

From: "Cyril Hrubis" <[email protected]>
To: "Jan Stancek" <[email protected]>
Cc: [email protected]
Sent: Tuesday, 12 April, 2016 10:40:49 AM
Subject: Re: [LTP] Pre-release LTP build

Hi!

Quoted 6 lines

I do. At least RHEL5.6 is going to be around for another year.

I have repeatedly built your patch-series on RHEL5.6 x86_64, which works fine. But as you pointed out 32bit version fails, other arches possibly too.

Quoted 6 lines
Quoted 2 lines

My first thought was some kind of lock, so we don't have to care about each architecture separately.

I think that easiest solution would be copying the assembler the builtin produces and using it as a fallback. That would have to be done per arch, but the rest of the code could be left untouched.

As for RHEL5.6 it's only i386. I was able to compile x86_64, ia64, ppc/ppc64 and s390/s390x with no changes.

Any of these two allow it to compile on x86:
1. -march=i486 (or higher)
2.

lib/tst_atomic.c new file +8 -0
new file mode 100644
index 0000000..601fd6c
@@ -0,0 +1,8 @@
+#if defined(__i386__)
+unsigned int __sync_add_and_fetch_4(unsigned int *v, unsigned value)
+{
+       register int val = value;
+       __asm__ __volatile__ ( "lock xadd %1,%0" : "=m" (*v), "=r" (val) : "1" (val) : "memory");
+       return val + value;
+}
+#endif

According to gcc docs it should only make external call to above function, when it's not provided natively, so it works as fallback.

#4 lore raw

Hi!

Quoted 21 lines

As for RHEL5.6 it's only i386. I was able to compile x86_64, ia64, ppc/ppc64 and s390/s390x with no changes.

Any of these two allow it to compile on x86:
1. -march=i486 (or higher)
2.

lib/tst_atomic.c new file +8 -0
new file mode 100644
index 0000000..601fd6c
@@ -0,0 +1,8 @@
+#if defined(__i386__)
+unsigned int __sync_add_and_fetch_4(unsigned int *v, unsigned value)
+{
+       register int val = value;
+       __asm__ __volatile__ ( "lock xadd %1,%0" : "=m" (*v), "=r" (val) : "1" (val) : "memory");
+       return val + value;
+}
+#endif
Quoted 2 lines

According to gcc docs it should only make external call to above function, when it's not provided natively, so it works as fallback.

Accordingly to:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=34115

It does not work with gcc 4.1 when the builtin is returned directly at the end of the function.

So storing the return value to a variable may fix that as well.

But I would prefer having fallback definition over working around compiler bugs.

Signature · 2 lines
Cyril Hrubis
[email protected]
#5 lore raw

----- Original Message -----

Quoted 36 lines

From: "Cyril Hrubis" <[email protected]>
To: "Jan Stancek" <[email protected]>
Cc: [email protected]
Sent: Tuesday, 12 April, 2016 2:23:07 PM
Subject: Re: [LTP] Pre-release LTP build

Hi!

Quoted 23 lines

As for RHEL5.6 it's only i386. I was able to compile x86_64, ia64, ppc/ppc64
and s390/s390x with no changes.

Any of these two allow it to compile on x86:
1. -march=i486 (or higher)
2.

lib/tst_atomic.c new file +5 -0
new file mode 100644
index 0000000..601fd6c
@@ -0,0 +1,8 @@
+#if defined(__i386__)
+unsigned int __sync_add_and_fetch_4(unsigned int *v, unsigned value)
+{
+       register int val = value;
+       __asm__ __volatile__ ( "lock xadd %1,%0" : "=m" (*v), "=r" (val) :

"1" (val) : "memory");

lib/tst_atomic.c new file +3 -0
+       return val + value;
+}
+#endif
Quoted 2 lines

According to gcc docs it should only make external call to above function, when it's not provided natively, so it works as fallback.

Accordingly to:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=34115

Quoted 4 lines

It does not work with gcc 4.1 when the builtin is returned directly at the end of the function.

So storing the return value to a variable may fix that as well.

It made no difference for me:

../../lib/libltp.a(tst_test.o): In function `tst_atomic_inc':
/root/ltp/lib/../include/tst_atomic.h:23: undefined reference to __sync_add_and_fetch_4' /root/ltp/lib/../include/tst_atomic.h:23: undefined reference to __sync_add_and_fetch_4' /root/ltp/lib/../include/tst_atomic.h:23: undefined reference to `__sync_add_and_fetch_4'
collect2: ld returned 1 exit status
make[1]: *** [test01] Error 1
make[1]: Leaving directory `/root/ltp/lib/newlib_tests'
make: *** [all] Error 2

Quoted 2 lines

But I would prefer having fallback definition over working around compiler bugs.

Agreed.

#6 lore raw

Hi!

Quoted 12 lines
Quoted 12 lines
Quoted 12 lines

+++ b/lib/tst_atomic.c

lib/tst_atomic.c new file +5 -0
@@ -0,0 +1,8 @@
+#if defined(__i386__)
+unsigned int __sync_add_and_fetch_4(unsigned int *v, unsigned value)
+{
+       register int val = value;
+       __asm__ __volatile__ ( "lock xadd %1,%0" : "=m" (*v), "=r" (val) :

"1" (val) : "memory");

lib/tst_atomic.c new file +3 -0
+       return val + value;
+}
+#endif
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

According to gcc docs it should only make external call to above function, when it's not provided natively, so it works as fallback.

This does not work with gcc-3.3 where no redirection from __sync_add_and_fetch -> __sync_add_and_fetch_4 happens. I would say that if we are adding fallback definition it should work with 3.3 as well. So I guess we should go with configure check and defining sync_add_and_fetch() directly. We can also print nice #error message in case that there is no fallback implementation if we decide to go this way.

Signature · 2 lines
Cyril Hrubis
[email protected]
#7 lore raw

----- Original Message -----

Quoted 20 lines

From: "Cyril Hrubis" <[email protected]>
To: "Jan Stancek" <[email protected]>
Cc: [email protected]
Sent: Tuesday, 12 April, 2016 4:24:50 PM
Subject: Re: [LTP] Pre-release LTP build

Hi!

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

+++ b/lib/tst_atomic.c

lib/tst_atomic.c new file +5 -0
@@ -0,0 +1,8 @@
+#if defined(__i386__)
+unsigned int __sync_add_and_fetch_4(unsigned int *v, unsigned value)
+{
+       register int val = value;
+       __asm__ __volatile__ ( "lock xadd %1,%0" : "=m" (*v), "=r"

(val) :
"1" (val) : "memory");

lib/tst_atomic.c new file +3 -0
+       return val + value;
+}
+#endif
Quoted 11 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

According to gcc docs it should only make external call to above function,
when it's not provided natively, so it works as fallback.

This does not work with gcc-3.3 where no redirection from __sync_add_and_fetch -> __sync_add_and_fetch_4 happens. I would say that if we are adding fallback definition it should work with 3.3 as well. So I guess we should go with configure check and defining sync_add_and_fetch() directly. We can also print nice #error message in case that there is no fallback implementation if we decide to go this way.

I'll post a series that goes this way shortly. I've based it on kernel atomic_add_return for architectures that are accessible for me. I tested it only on RHEL5.6 (gcc 4.1.2), which I think is the oldest active RHEL release at the moment.

What arches do you care about, that have gcc-3.3?

#8 lore raw

Hi!

Quoted 6 lines

I'll post a series that goes this way shortly. I've based it on kernel atomic_add_return for architectures that are accessible for me. I tested it only on RHEL5.6 (gcc 4.1.2), which I think is the oldest active RHEL release at the moment.

What arches do you care about, that have gcc-3.3?

I would say that we should work fine on x86 and x86_64. The rest of the archs are either new ones that does not have this problem or not easy to get. Well there may be old mac machines that runs Linux just fine but if you are runing LTP on Linux on ppc mac you should be able to write a few lines of inline assembler...

Signature · 2 lines
Cyril Hrubis
[email protected]