Split gnu/javax/swing/text/html/parser/HTML_401F.java (Was [vta] Add chains from referenced VALUEs to DVs that reference them)

Andrew Haley aph@redhat.com
Wed Jul 1 12:37:00 GMT 2009


Jakub Jelinek wrote:
> On Tue, Jun 30, 2009 at 03:50:11PM +0100, Andrew Haley wrote:
>> Gerald Pfeifer wrote:
>>> On Tue, 30 Jun 2009, Jakub Jelinek wrote:
>>>> Alternatively we could split the huge HTML_401F.java function into
>>>> say 4 smaller ones.
>>> Pleeeeeease! :-)  This one has been causing troubles beyond the context
>>> of just vta (where, for example, on some system one needs to boot with
>>> a special kernel option to provide sufficient amounts of memory to GCJ).

>> I can split the function, but papering over the problem by splitting
>> one function won't make the problem go away for others who use gcc.
>> It'll just mean that gcc developers don't notice it.
> 
> Sure, that's why I've spent last 2 weeks on var-tracking.c improvements
> for this exact testcase.  The question just is, if we really need to include
> so huge testcases as part of everybody's daily bootstrap/regtest cycle, or
> if it is sufficient to keep such testcases on the side, for automated
> testers that track their compile time/memory usage and nag us if we regress
> too much on it.
> 
>> The big function is about 250k lines of GIMPLE.  jc1 uses about 474m of
>> RAM at -O2 on a 64-bit system, 414m at -O1, 536m at -O0.  On what class
>> of machines are you trying to build this?
> 
> From var-tracking POV, especially on VTA branch, the main problem is
> that this function has 10000 basic blocks and on VTA needs to track over
> 15000 of variables/VALUEs across all those bbs.  Vanilla VTA branch needs
> 2.9GB of memory and 25 minutes to compile this at -g -O2, with all the
> patches I've sent it needs just 1.6GB of memory and 8 minutes.
> 
>> Don't we have some sort of heuristic that says "this function is
>> freaking huge, don't do any expensive optimizations." ?
> 
> For var-tracking we just bail out on highly connected large cfgs:
>   if (n_basic_blocks > 500 && n_edges / n_basic_blocks >= 20)
>     return 0;
> 
> I haven't studied how exactly is --enable-java-maintainer-mode
> compiling the classes; if I just gcj -C HTML_401F.java on
> Fedora 11 (GCC 4.4.0, ecj 3.4.2), the compile time with patched
> VTA is only 4:53 with 1.5GB top memory usage, if I patch HTML_401F.java
> with the following patch, it compiles within 0:55 and maxes at 250MB.
> I have no idea whether it will work correctly (what to test it with)
> and whether it is or is not an ABI change.

It's not an ABI change.  This patch is OK iff accompanied by a comment
in the code that explains the problem.

Thanks,
Andrew.



More information about the Java-patches mailing list