From: Michael Widenius
Date: December 6 2000 12:35am
Subject: Compile options
List-Archive: http://lists.mysql.com/internals/198
Message-Id: <14893.35257.567384.144200@narttu.mysql.fi>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit


Hi!

>>>>> "Antony" == Antony T Curtis <antony@stripped> writes:

Antony> Since most people would be using GCC 2.95.x nowadays, I think it should
Antony> be a good idea to change the default optimisation level in the
Antony> configure.in file from "-O6" to "-O2"... 

Antony> In the 2.95.x compiler, setting -O3 can cause dangerous optimisations
Antony> which would break applications.
Antony> The "-O6" option only made any sense on the old 2.7 compilers.

We have tested MySQL quite extensively under -O6 and -O3, so -O3
should be safe.  Note however that if the user is using pgcc, then -O6
should give you some better performance..

Antony> However, enabling "-fexpensive-optimizations" would appear to be
Antony> relatively safe.

I shall try to do a full test in the REAL near future which
optimization will give MySQL the best speed!

Antony> Also - here are some additional tips for compiling for 64bittyness under
Antony> AIX using GCC 2.95 (I already change the configure.in and run autoconf)

Antony> $ OBJECT_MODE=64 \
>> CC="gcc -pipe -maix64 -Wa,-many" \
>> CXX="g++ -pipe -maix64 -Wa,-many" \
>> CXXFLAGS="-felide-constructors -fno-exceptions -fno-rtti" \
>> MAKE="gmake" \
>> ./configure  --without-docs --without-debug --with-low-memory

Antony> to compile, the libtool shell script has to me modified, adding
Antony> "-maix64" after the two instances where "-shared" occour and then to
Antony> compile, simply run

Antony> $ OBJECT_MODE=64 gmake

Antony> (GNU make must be used, IBM make fails)

Antony> Surprisingly enough, the compiled output is a smaller than the 32-bit
Antony> version. Response times appear to be good however I haven't had much
Antony> time to do objective comparisons with the 32bit binaries.

Tim, do you have any ideas of how to make the libtool change without
causing trouble for others ?

Antony> (IBM's POWER architecture appears to use 32bit fixed length instructions
Antony> which may account for the smaller binaries)

Antony> The latest version of AIX shouldn't need the "-Wa,-many" option but
Antony> since all it does is disable the opcode verification within IBM's
Antony> assembler, it should be harmless and it avoids the bugs in the older AIX
Antony> assemblers.

Thanks for the information!

Regards,
Monty