Re: regmove fix

Re: regmove fix

[Jeffrey A Law]

  In message < 199901200015.BAA01622@quatramaran.ens.fr >you write:
  > It may be a few days before I can tell you the result, from my point
  > of view.
  > Basically, it seems that egcs-current is more aggressive, and broke a
  > few asm statements, that were probably badly formed, e.g.,
  > 
  > __asm __volatile("cld\n\trepne\n\toutsl"  :
  > : "d" (port), "S" (addr), "c" (cnt) : "%esi", "%ecx");
  > 
  > Since I don't know enough about i386 assembler, I'm getting some other
  > developpers to fix them... I hope to have some relevant statistics right
  > after that...
That asm is bogus.  It explicitly clobbers registers which overlap with inputs
and/or outputs.

This asm was wrong regardless of what regmove may have been doing.

jeff

regmove fix

[Jeffrey A Law]

I just checked in a bugfix to regmove that may (or may not) have a positive
effect on the global register allocator.

Basically regmove could incorrectly update the live length of a register,
which is one of the 3 key components which drive what registers have allocation
priority in the global register allocation phase.

regmove should now keep that data more accurate, which should in turn improve
allocation.  The difference may (or may not) be noticable.

jeff

Re: regmove fix

[Marc Espie]

It may be a few days before I can tell you the result, from my point
of view.
Basically, it seems that egcs-current is more aggressive, and broke a
few asm statements, that were probably badly formed, e.g.,

__asm __volatile("cld\n\trepne\n\toutsl"  :
: "d" (port), "S" (addr), "c" (cnt) : "%esi", "%ecx");

Since I don't know enough about i386 assembler, I'm getting some other
developpers to fix them... I hope to have some relevant statistics right
after that...

In article < 11604.916626751@hurl.cygnus.com > you write:

>I just checked in a bugfix to regmove that may (or may not) have a positive
>effect on the global register allocator.