[fpc-devel] -Cg and x86_64-linux

[Daniël Mantione]

Op Sat, 6 Jan 2007, schreef Florian Klaempfl:

> Dani=EBl Mantione schrieb:
> > =

> > Op Sat, 6 Jan 2007, schreef Florian Klaempfl:
> > =

> > > As of revision 5831, I've enabled -Cg for x86_64-linux building. This
> > > means
> > > all code is built with PIC so you can build libraries without
> > > recompiling the
> > > fpc rtl etc. However, this might have some downsides: first, pic'ed
> > > code is
> > > slightly slower, second, pic support still might contain bugs so
> > > please share
> > > your findings and opinions.
> > =

> > How difficult do you think it is to use RIP relative addresses to acces=
s =

> =

> How should this look like?

Assume we have:

var global:byte;

procedure z;

begin
  global:=3D1;
end;

Referencing to z without PIC would be an absolute memory access, =

preventing us to place z in memory where we want without breaking the =

code. Normally with -Cg we do something like:

      call  FPC_GETEIPINEBX
      add   rbx,dword global_offset_table
      mov   [rbx + offset_of_global_in_datasegment].byte,1

... to get access to the variable. With x64_64 you can do:

      mov   [global wrt rip].byte,1

The assembler must translate this to:

      mov   [rip-data_segment_distance+offset_of_global_in_datasegment].byt=
e,1

... where offset_of_global_in_datasegment is a constant and =

data_segment_distance symbol that is resolved by the linker.

RIP relative addresses are more efficient than absolute addresses, because
an absolute address is always stored as 64 bit, while the RIP offset is =

encoded as a 32 bit signed value (8 bit in x86 tradition possible as =

well, but usually the datasegment is further away than 128 bytes), the =

processor can decode them faster.

Dani=EBl