Re: [Valgrind-developers] Sub-optimal code gen for 'leave'?

[Nicholas N. <nj...@ca...>]

On Tue, 28 Jan 2003, Julian Seward wrote:

> > I just noticed that the UCode for the x86 instruction 'leave' is made
> > like this:
> >
> >    case 0xC9: /* LEAVE */
> >       t1 = newTemp(cb); t2 = newTemp(cb);
> >       uInstr2(cb, GET,  4, ArchReg, R_EBP, TempReg, t1);
> >       uInstr2(cb, PUT,  4, TempReg, t1, ArchReg, R_ESP);     // (1)
> >       uInstr2(cb, LOAD, 4, TempReg, t1, TempReg, t2);
> >       uInstr2(cb, PUT,  4, TempReg, t2, ArchReg, R_EBP);
> >       uInstr2(cb, ADD,  4, Literal, 0, TempReg, t1);
> >       uLiteral(cb, 4);
> >       uInstr2(cb, PUT,  4, TempReg, t1, ArchReg, R_ESP);     // (2)
> >       if (dis) VG_(printf)("leave");
> >       break;
> >
> > AFAICT, UInstruction (1) is redundant due to (2)... right?

> Interesting.  I think you're right.  Did you find that with your
> spot-multiple-writes-without-intervening-reads skin, that you
> mentioned on Friday?

No, it only looks at LOADs and STOREs.  Just spotted it when looking
through some UCode in detail.

> I guess this should be fixed; I can do so if you like.

I'll do it... it's not hard.

N