Re: Re: [PATCH] Use CC atomic builtins if available [was: Re: TAS patch for building on armel/armhf thumb]

On 03.02.2012 02:48, Bruce Momjian wrote:
> Sorry for the late reply, but Heikki, can you get this Itanium
> information into s_lock.h as a comment, particularly the information
> about the +Ovolatile=__unordered flag?

Good idea. I came up with the attached, hope that explains it.

Looking back at the discussions, we concluded that the current code is
safe on gcc, because it implicitly adds the .rel/.acq opcodes to
volatile accesses, and HP's compiler does the same as long as you don't
explicitly disable it with +Ovolatile=__unordered. But what about
Intel's icc compiler? Presumably it's also safe, but looking at Intel's
manuals that I found, I'm not completely sure about it. There's an
option, -m[no-]serialize-volatile, that controls it, but I couldn't
figure out which is the default. Looking at the docs on that from Intel
that I found [1], it seems to me that on Linux, the default is *not*
safe, but on Windows it is.

Sergey, you have dugong in the buildfarm that uses Intel's compiler on
Itanium. Could you verify whether the -mno-serialize-volatile is the
default? If you could for example extract the assembler code generated
by icc for xlog.c, and send it over. Whether it's generating the
.rel/.acq opcodes should be easy to see in the generated code of the
XLogGetLastRemoved() function, for example, which doesn't do much else
than grab a spinlock. On gcc, the -s flag generates the assembly files,
I presume it's the same on icc.

Perhaps we should set those compiler flags explicitly in configure,
regardless of the defaults.

[1]
http://software.intel.com/sites/products/documentation/hpc/compilerpro/en-us/cpp/lin/compiler_c/copts/ccpp_options/option_qserialize-volatile.htm

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com