https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117202
Richard Sandiford <rsandifo at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |juzhe.zhong at rivai dot ai,
| |law at gcc dot gnu.org,
| |pan2.li at intel dot com
Target| |riscv64*-*-*
Summary|SLP permutation for VLA |RISC-V: SLP permutation for
|vectors broken |VLA vectors broken
Component|tree-optimization |target
--- Comment #6 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
Ah! I think I see the problem, and RTL checking confirms:
foo.c:10:23: internal compiler error: RTL check: expected code 'const_int',
have 'const_poly_int' in expand_const_vector, at config/riscv/riscv-v.cc:1365
10 | dest[i * 4] = value;
| ~~~~~~~~~~~~^~~~~~~
riscv-v.cc:expand_const_vector assumes that the elements of a stepped
variable-length CONST_VECTOR are simple integer constants, but they can be
poly_ints (wheels in wheels!). E.g.:
/* Step 4: Multiply to step size n. */
HOST_WIDE_INT step_size =
INTVAL (builder.elt (builder.npatterns ()))
- INTVAL (builder.elt (0));
this step size is actually a poly_int64 in general. Similarly here:
/* Step 5: Add starting value to all elements. */
HOST_WIDE_INT init_val = INTVAL (builder.elt (0));
So I think this is a target issue. I haven't checked whether there are other
PRs for this already.
ALthough this particular instance depends on your patch, I imagine there are
other instances that fail in a similar way without it (and perhaps wouldn't
have failed before GCC 15, due to the new SLP permute support).
