We use "preserved" CPU registers.
According to ABI called functions have to keep them unchanged. Of course
they may save and then restore them back.
I didn't see any problems yet, however I know that some old GCC may
generate invalid code when using global registers.
Most probably we will have to limit GCC versions.
Thanks. Dmitry.
On Tue, Mar 17, 2015 at 10:25 PM, Rasmus Lerdorf <ras...@lerdorf.com> wrote:
> > Diff:
> > diff --git a/Zend/Zend.m4 b/Zend/Zend.m4
> > index 16f2d5f..e12b00d 100644
> > --- a/Zend/Zend.m4
> > +++ b/Zend/Zend.m4
> > @@ -409,3 +409,48 @@ else
> > AC_MSG_RESULT(no)
> > fi
> > fi
> > +
> > +AC_ARG_ENABLE(gcc-global-regs,
> > +[ --disable-gcc-global-regs
> > + whether to enable GCC global register
> > variables],[
> > + ZEND_GCC_GLOBAL_REGS=$enableval
> > +],[
> > + ZEND_GCC_GLOBAL_REGS=yes
> > +])
> > +AC_MSG_CHECKING(for global register variables support)
> > +if test "$ZEND_GCC_GLOBAL_REGS" != "no"; then
> > + AC_TRY_COMPILE([
> > + ],[
> > +#if defined(__GNUC__) && defined(i386)
> > +# define ZEND_VM_FP_GLOBAL_REG "%esi"
> > +# define ZEND_VM_IP_GLOBAL_REG "%edi"
> > +#elif defined(__GNUC__) && defined(__x86_64__)
> > +# define ZEND_VM_FP_GLOBAL_REG "%r14"
> > +# define ZEND_VM_IP_GLOBAL_REG "%r15"
> > +#else
> > +# error "global register variables are not supported"
> > +#endif
> > +typedef int (*opcode_handler_t)(void);
> > +register void *FP __asm__(ZEND_VM_FP_GLOBAL_REG);
> > +register const opcode_handler_t *IP __asm__(ZEND_VM_IP_GLOBAL_REG);
> > +int emu(const opcode_handler_t *ip, void *fp) {
> > + const opcode_handler_t *orig_ip = IP;
> > + void *orig_fp = FP;
> > + IP = ip;
> > + FP = fp;
> > + while ((*ip)());
> > + FP = orig_fp;
> > + IP = orig_ip;
> > +}
> > + ], [
> > + ZEND_GCC_GLOBAL_REGS=yes
> > + ], [
> > + ZEND_GCC_GLOBAL_REGS=no
> > + ])
> > +fi
> > +if test "$ZEND_GCC_GLOBAL_REGS" = "yes"; then
> > + AC_DEFINE([HAVE_GCC_GLOBAL_REGS], 1, [Define if the target system has
> > support for global register variables])
> > +else
> > + HAVE_GCC_GLOBAL_REGS=no
> > +fi
> > +AC_MSG_RESULT(ZEND_GCC_GLOBAL_REGS)
>
> Dmitry, the perf boost of this is awesome, but is it completely safe?
> Won't a signal potentially overwrite a register variable here? Like on a
> timeout, for example?
>
> -Rasmus
>
>
