https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120276
ktkachov at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at gcc dot gnu.org
--- Comment #2 from ktkachov at gcc dot gnu.org ---
Had a quick look. The ICE through regcprop comes in partial_subreg_p:
/* Modes involved in a subreg must be ordered. In particular, we must
always know at compile time whether the subreg is paradoxical. */
poly_int64 outer_prec = GET_MODE_PRECISION (outermode);
poly_int64 inner_prec = GET_MODE_PRECISION (innermode);
gcc_checking_assert (ordered_p (outer_prec, inner_prec));
Where the two modes are VNx4QI and V8QI.
The regcprop code that calls this is:
/* If a narrower value is copied using wider mode, the upper bits
are undefined (could be e.g. a former paradoxical subreg). Signal
in that case we've only copied value using the narrower mode.
Consider:
(set (reg:DI r14) (mem:DI ...))
(set (reg:QI si) (reg:QI r14))
(set (reg:DI bp) (reg:DI r14))
(set (reg:DI r14) (const_int ...))
(set (reg:DI dx) (reg:DI si))
(set (reg:DI si) (const_int ...))
(set (reg:DI dx) (reg:DI bp))
The last set is not redundant, while the low 8 bits of dx are already
equal to low 8 bits of bp, the other bits are undefined. */
else if (partial_subreg_p (vd->e[sr].mode, GET_MODE (src)))
I'm not super familiar with that pass, but could it be that we need another
check on the modes before calling partial_subreg_p here or maybe do some other
comparison between the modes?
