[Openmcl-devel] Porting the OpenMCL Compiler

[Gary Byers]

On Thu, 7 Jul 2005, Hamilton Link wrote:

> As far as register usage goes, and in particular the number of immediates vs. 
> caller-save vs. callee-save etc. registers, how much of that could be 
> parameterized (vs. having to be written in an architecture-specific portion 
> of the back end) if the compiler back end was revised to be more cleanly 
> separated from the compiler front end?  Just curious, since we've idly talked 
> about separating syntax tree manipulation from register allocation and code 
> generation.
>
> h
>

If the compiler generated better (or worse, or different ...) code,
the static register-usage frequences would probably be at least a
little different.  If it was better (in most ways in which I can
think of it being better), I'd expect to see fewer references to
arg_z and imm0 and more references to other node/imm regs, fewer
references to stack pointers (sp, vsp, tsp), and a little less
consing (especially consing of intermediate floating-point values)
and therefore fewer references to allocptr/allocbase.

The most popular registers (arg_z and imm0) would probably still
be most popular by significant margins, even if those margins
dipped a bit.