On Mon, Nov 26, 2018 at 04:20:18PM -0600, Segher Boessenkool wrote:
> > --- gcc/config/rs6000/rs6000.c.jj 2018-11-16 21:35:42.497982764 +0100
> > +++ gcc/config/rs6000/rs6000.c 2018-11-26 10:56:28.079516458 +0100
> > @@ -20644,7 +20644,11 @@ print_operand (FILE *file, rtx x, int co
> >
> > case 'D':
> > /* Like 'J' but get to the GT bit only. */
> > - gcc_assert (REG_P (x));
> > + if (!REG_P (x))
> > + {
> > + output_operand_lossage ("invalid %%D value");
> > + return;
> > + }
> >
> > /* Bit 1 is GT bit. */
> > i = 4 * (REGNO (x) - CR0_REGNO) + 1;
>
> 'D' just blows up if you pass it something not a CR field (well, it compiles
> into something that doesn't assemble).
I can add || !CR_REGNO_P (REGNO (x)) .
> 'J' uses ccr_bit. This actually explodes (a gcc_assert, an ICE).
Oops, I thought ccr_bit will just return -1.
So, shall we change all the gcc_assert/gcc_unreachable in ccr_bit to
(conditional) return -1; and perhaps add gcc_assert in the non-print_operand
callers?
> > @@ -20900,7 +20904,11 @@ print_operand (FILE *file, rtx x, int co
> >
> > case 't':
> > /* Like 'J' but get to the OVERFLOW/UNORDERED bit. */
> > - gcc_assert (REG_P (x) && GET_MODE (x) == CCmode);
> > + if (!REG_P (x) || GET_MODE (x) != CCmode)
> > + {
> > + output_operand_lossage ("invalid %%t value");
> > + return;
> > + }
> >
> > /* Bit 3 is OV bit. */
> > i = 4 * (REGNO (x) - CR0_REGNO) + 3;
>
> 't' does not seem easy to blow up, heh.
>
> Both 'D' and 't', as well as ccr_bit, should use cc_reg_operand I think?
> That allows some pseudos as well, but that doesn't matter.
It also allows SUBREGs and the code will crash badly if they appear.
Testing just REG_P (x) && CR_REGNO_P (REGNO (x)) seems better to me.
> > @@ -21059,7 +21067,11 @@ print_operand (FILE *file, rtx x, int co
> > names. If we are configured for System V (or the embedded ABI) on
> > the PowerPC, do not emit the period, since those systems do not use
> > TOCs and the like. */
> > - gcc_assert (GET_CODE (x) == SYMBOL_REF);
> > + if (GET_CODE (x) != SYMBOL_REF)
>
> SYMBOL_REF_P please.
Ok.
Jakub
