> One of the main differences on some architectures is that there are
> hooks into the pmap for mapping pool pages.  On the MIPS and Alpha,
> KSEG is used.  On the PowerPC, the direct-mapped area mapped by the
> BATs is used.  Etc.  This means that on these platforms, TLB usage
> is greatly reduced for things allocated with the pool allocator.

Hmm.  Is it still possible this technic is used for malloc()?
I remember I've changed all malloc() whose size is less than or equal to
page size returns KSEG0 address instead of KSEG2 address, and the
performance for compiling kernel was improved over 10% at the SVR4
for MIPS cpu.

Atsushi Onoe