Dear G++ developers,
*PLATFORM*
GCC 4.2.3, Intel Core 1 (32 bit), Kubuntu 8.04-alpha (Debian-based).
*DESCRIPTION OF THE PROBLEM*
When using nested templates such as
A<A<A<A<x,y>, z>, t>, u>
the optimizer seems to give up some useful optimizations when the depth is
greater than 17.
I know that the C++ standard only requires the compiler to support depth<=17,
but it still is very useful to be able to use deeper recursive templates. Also,
that's what g++'s option -ftemplate-depth-N seems to control. Unfortunately, it
doesn't seem to have any effect here.
*TESTCASE*
See attachment to this bug report (test.cpp). This is a very simple example of
Expression Templates.
$ g++ -O3 -DDEPTH17 test.cpp -o test
$ time ./test
real 0m1.810s
$ g++ -O3 -DDEPTH18 test.cpp -o test
$ time ./test
real 1m9.941s
$ g++ -O3 -DDEPTH18 -ftemplate-depth-20 test.cpp -o test
$ time ./test
real 1m10.751s
*A MORE GENERAL QUESTION*
I hit this bug while developing an expression-templates library for linear
algebra.
This bug might have little cousins. I have observed smaller performance issues
in cases where the template depth was lower than 17, but where more complicated
things were being done.
In test.cpp, what really matters for performance is that g++ realizes that it
needs not actually construct the objects of classes Sum and Int -- it only
needs to emit the code of the value() method.
My general question would be, is there a way to hint the g++ that a given class
is a "small" "helper" class whose objects need not be actually constructed?
Something like "inline" except it'd be for a class, not a function.
g++ already does this "inlining" perfectly unless too complex things are being
done. Is there some --param that I could try passing to g++ in order to make it
handle more complex situations?
--
Summary: Deep templates get 40x slower when depth goes from 17 to
18.
Product: gcc
Version: 4.2.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jacob at math dot jussieu dot fr
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35393
