This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Register Pressure in Instruction Level Parallelism (fwd)
- From: dewar at gnat dot com (Robert Dewar)
- To: drow at mvista dot com, duraid at fl dot net dot au
- Cc: gcc at gcc dot gnu dot org, matzmich at cs dot tu-berlin dot de
- Date: Sat, 13 Jul 2002 21:42:51 -0400 (EDT)
- Subject: Re: Register Pressure in Instruction Level Parallelism (fwd)
Daniel said
<<What isn't promising about it is that integer programming is a
technique for exact solutions to NP-complete problems. Since these are
hard problems, exact solutions tend to be very computation-intensive.
GCC is slow enough as it is; integer programming based allocation is
exceptionally computation-intensive, especially for large or
pathological programs.
>>
Perhaps we will have to introduce the switch that was in the original
graph coloring IBM compiler. The switch was called something like No_NP
and it told the compiler *not* to go into an exponential time algorithm
to find an optimal coloring of the graph in cases where a simple attempt
at coloring failed (without this switch they had cases of short programs
that took days to compile :-)
Of course machines are faster now, but exponential algorithms get out of
hand pretty quickly, even with Moore's law hard at work :-)