i386 ELF PIC PLT optimization ?

i386 ELF PIC PLT optimization ?

[Fabrice Bellard]

Hi,

Did anybody try to improve the i386 ELF ABI for PIC code ? By looking at 
how the thinks are handled on PowerPC, I think that several improvements 
are possible, for both smaller and faster code :

1) On i386, the PLT is built by the static linker whereas on PPC it is 
built dynamically by the dynamic linker. The net result is that the 
executables are smaller and that indirect jumps can be avoided in the 
PLT, so the inter DLL/EXEs calls are faster. The PLT can be also smaller 
because only 5 (or 8) bytes per entry could be used (a call opcode can 
be used for lazy binding and a direct jmp after).

2) On i386, it is almost always needed to initialize %ebx properly to 
the GOT address when entering a DLL function (except for functions 
calling only static functions and referencing no global variables). On 
PPC, it is only needed to initialize %r30 when accessing global variables.

By using a dynamically generated PLT on i386, we could avoid the %ebx 
register usage when calling a function by the PLT and so we would need 
to initialize %ebx only in rare cases (only if the function accesses 
global variables).

The net result would be smaller code (because no function prologue in 
most cases) and faster DLL function calls (no indirect jmp and in most 
cases %ebx will be usable in the function).

I am thinking about implementing it in TinyCC (a tiny C compiler) to 
test it. Does anyone have implemented similar ideas in GCC ? Of course, 
a small patch in ld.so would be needed to support that while staying 
compatible with standard DLLs.

Fabrice.