[fpc-devel] Registers reloading

[J. Gareth Moreton]

Hi Marco,

That is actually one thing I've been researching myself. Currently the 
peephole optimiser doesn't have the facilities to detect if a register 
already contains a value that is being assigned to it, something I call 
the "Deep Optimizer" but is otherwise just basic data flow analysis.  I 
have been overhauling the peephole optimiser for i386 and x86-64 to be 
more efficient with fewer passes, although it's taking time to get it 
into a form that the administrators are happy to merge into the trunk, 
since it is quite a major change.

Yours sincerely,

J. Gareth "Kit" Moreton


On 18/07/2019 14:33, Marco Borsari via fpc-devel wrote:
> Hi all,
> does the compiler have the optimization to avoid reloading into
> a register of an address (or a value) whether it is already
> present?
> I ask this because of this code fragment compiled with the -O4
> option:
>
> .Lj380:
>     # Register eax allocated
> # [726] T := T-1;
>     subw    $1,2(%ebx)
>     # Register ax released
>     # Register edx allocated
> # [727] H6 := @STK^[T];
>     movl    U_$GLOBALS_$$_STK,%edx
>     # Register eax allocated
>     movzwl    2(%ebx),%eax
>     leal    -8(%edx,%eax,8),%eax
>     # Register edx released
>     movl    %eax,-256(%ebp)
> # [728] H7 := H6;
>     movl    %eax,-260(%ebp) <-(1)
>     # Register eax released
> # [729] Inc(H7, 1);
>     addl    $8,-260(%ebp)
>     # Register eax allocated
> # [730] H6^.I := H6^.I+H7^.I;
>     movl    -256(%ebp),%eax <-(2)
>     # Register ecx allocated
>     movswl    4(%eax),%ecx
>     movl    -260(%ebp),%eax
>     # Register edx allocated
>     movswl    4(%eax),%edx
>     leal    (%ecx,%edx),%eax
>     # Register ecx released
>     movw    %ax,%dx
>     movl    -256(%ebp),%eax <-(3)
>     movw    %dx,4(%eax)
>     # Register eax,dx released
>
> In (1) the compiler assumes that eax has the correct
> address but in (2) does not omit to reload it (eax released
> too soon).
> Both seem to be of the same level of the analysis complexity,
> more deeper should be takes into account of (3) with the use of
> a supplementar registers as si or di (they are used in other parts of 
> the code).
> Thank you for any response, Marco Borsari
> _______________________________________________
> fpc-devel maillist  -  fpc-devel at lists.freepascal.org
> https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
>
>

---
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus