compiler thoughts

Friday July 27, 2007

GCC vs GCCfss vs Studio performance on SPARC

Here is a relative performance chart for several versions of plain GCC including the latest GCC 4.2, just released GCCfss 4.0.4, previous GCCfss 4.0.3, upcoming GCCfss 4.2 and Studio 12 with the most recent patches.

All runs were done on UltraSPARC 4+ box on SPEC CPU2006 FP base with C and C++ benchmarks.

GCC flags:

 -O3 -mcpu=ultrasparc3 -funroll-loops -ffast-math -fprefetch-loop-arrays 

GCCfss and Studio flags:

-fast -xipo=2 -xpagesize=4M -xprefetch_level=2 -xalias_level=std/layout/compatible

The runs are not SPEC conforming, since Fortran benchmarks were not run, but very representative of any floating point heavy application on SPARC. So consider them an 'estimate'.

Some observations:

• GCCfss 4.0.4 is using Studio 12 backend, so the difference between GCCfss 4.0.4 and Studio 12 can only be explained by differences in IR generated by front-ends.
• GCCfss 4.0.3 allows most of the gcc tree-ssa optimizers to run, whereas GCCfss 4.0.4 has completely different 'gimplification' process, so no tree-ssa passes can run, but outgoing IR is much better suited for Studio 12 backend to work on. As demonstrated by the jump from GCCfss 4.0.3 to 4.0.4
• GCCfss 4.2 is 'work in progress', which should be released by the end of this year

Tuesday June 05, 2007

GCCfss 4.0.4 is LIVE!

We got our next version of GCCfss 4.0.4 available for download here

Main differences vs GCCfss 4.0.3:

• based on plain GCC version 4.0.4


• supplied SUNW0scgfss backend is essentially the same backend found in Sun Studio 12
  The users can use either one, however I recommend to use SUNW0scgfss archive since its download size is much smaller. It's easier to install and has only needed binaries, whereas Studio 12 is huge.


• fully relocatable!
  Just download front-end tarball (SUNW0gccfss.tar.bz2) and backend tarball (SUNW0scgfss.tar.bz2), bunzip2 and untar them and run 'gcc'. The driver will pick up correct compiler components for Solaris 9 or 10, include files, libraries, etc regardless of compiler location. Compare that to plain GCC!


• much higher benchmark performance.
  gcc middle-end tree optimizations are no longer run. gimplification is done in a different way to preserve high level structures needed for Studio 12 optimizer to perform complex array and loop transformations.


• better debuggability


• besides ATS, BIT and Thread Analyzer tools SUNW0scgfss archive includes new Discover Tool that detects programming errors related to the allocation and use of program memory at runtime on SPARC systems.


• all the issues mentioned on the forum as of today have been addressed

Friday April 13, 2007

dirt bike

Recently went to Rich Oliver's MSF dirt bike school and I should tell you that was a ton of fun! My first ever bike experience was very exciting. Sliding the bike is as much fun as sliding some awd car. Highly recommended.

Monday April 09, 2007

relocatable GCCfss

Upcoming GCCfss (GCC for Sparc Systems) 4.0.4 is going to relocatable.

• It's going to be released in two tarballs instead of packages, so root permissions are no longer necessary to try it out and downloads are quicker
• Besides the obvious, there are several ways to relocate both front-end and backend indepedently. More details will be in the manual
• /usr/sfw/lib path is going to be baked into the executable, just like plain GCC shipped with Solaris is doing and since GCCfss is 100% compatible with plain G++ the users of GCCfss will have a lesser need to ship libstdc++.so with their built c++ binaries

Monday February 26, 2007

interesting bug

While working on building OpenSolaris with gcc4/gccfss I encountered one interesting bug which can be demonstrated by the following simple C program:

void foo (long long *);

void bar (long a, long long b)
{
   typeof (a+b) c;
   foo (&c);
}

compiled with gcc 3.4 it will print the warning:

$ gcc3 -c simple.c -m64
simple.c: In function `bar':
simple.c:6: warning: passing arg 1 of `foo' from incompatible pointer type

whereas gcc 4.x will compile this testcase without any warnings.

The reason for that is GCC 3.x consider 'long + long_long' expression as 'long' type in 64-bit mode which is incorrect.

Why would I care ?

There are multiple places in OpenSolaris where it needs to print 64-bit value and it is done with '%llx' printf specifier. Due to this gcc 3.4 bug the Solaris sources were using '%lx' specifier to print expressions involving long and longlong types and now our changes to compile ON with gcc 4.x would break ON build with /usr/sfw/bin/gcc 3.4 unless the fix can be backported into "csl-sol210-3_4" Solaris specific gcc branch.

All in all, Solaris seems to be overdue to upgrade to some newer version of GCC.

Tuesday January 16, 2007

enumerations

By working on gcc4/gccfss for OpenSolaris project I found one interesting difference between Studio and GCC compilers.
Studio treats all enumration types in C as signed integers, whereas GCC can put them in unsigned type depending on the values of enumerators.
At first it may look like a minor difference, but in Solaris there was this piece of code:

typedef enum {
  A = 0,
  B,
  C
} en_t;

en_t var;

for (;var >= A; var--);

From GCC point of view the expression "var>=A" is always true, so the loop is collapsed into for(;;) even at -O0 optimization level.
From Studio point of view this code is fine.
Both are correct, because the choice of type is implementation defined by C standard.

Sunday December 17, 2006

Ultra3 vs Ultra4

One user reported an interesting problem: the same application was giving one result on Ultra III and different result on Ultra IV and Ultra II.

Turned out that some operations on floating point were trapping into the OS on ultra3 and Solaris fp emulator executed them in IEEE compliant way whereas Ultra4 and Ultra2 were executing the same code natively on cpu and because 'non-standard' bit in fsr register was set, Ultra4/2 chips produced different result than Solaris emulation.

Of course this behaviour of Ultra3 is documented, but you really need to be a chip guru to know/remember such details.

That's just something to be aware of if you use floating point a lot
and compile with -ffast-math or -fsimple or -fns flags with GCCfss.

Friday November 03, 2006

GCCfss 4.0.3 patch update is available

We've just released a patch for GCCfss 4.0.3. It's located at the same place place on SDLC website.
Please download it, remove old 4.0.3 packages and install new 4.0.3 packages.
This patch contains several fixes for the problems reported on the cooltools forum.
You would need to download both packages (SUNW0scgfss and SUNW0gccfss), since the bugfixes went into both.

Please give it a try and send us your feedback. We're constantly working on improving GCCfss and the user feedback makes it even better.

Monday October 23, 2006

Zoomzoomlive

This year Mazda did yet another zoomzoom event in Alameda,CA. My best laptime in MazdaSpeed6 was 41.546 sec which is 7th overall. Got lucky I guess.

Thursday September 14, 2006

GCCfss 4.0.3 is LIVE!

We got our next version of GCCfss 4.0.3 available for download here

Main differences vs GCCfss 4.0.2:

• based on plain GCC version 4.0.3


• based on the same Sun Compiler backend found in Studio Express


• better c++ debuggability


• includes Data Race Detection Tool


• all the issues mentioned on the forum before we created a release branch in mid August have been addressed.
  Including flag mapping between Studio-like flags and gcc flags.


• includes automatic Solaris detection.
  You download just one front-end package (SUNW0gccfss) and one backend package (SUNW0scgfss), install them and run it as normal: /opt/gcc/bin/gcc. The driver will pick up the compiler components for Solaris 9 or Solaris 10 in the run-time. So you don't need to deal with weird error messages that plain GCC for Solaris 10 will print if run on Solaris 9.