[Bug target/102211] [12 regression] ICE introduced by r12-3277

["wilson at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102211

Jim Wilson <wilson at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |wilson at gcc dot gnu.org

--- Comment #4 from Jim Wilson <wilson at gcc dot gnu.org> ---
Yes, moving SI/DI values to FP regs is OK.  However, RISC-V requires that FP
values in FP registers be stored NaN-boxed.  So an SFmode value in a 64-bit FP
reg has the upper 32-bits of all ones, and the lower 32-bits is the value. 
Thus if accessed as a 64-bit value, you get a NaN.  The hardware may trap if
you access a 32-bit value which is not properly NaN-boxed.  Using qemu to check
this may not be good enough, as last time I looked at qemu it wasn't handling
NaN-boxing correctly, but this was over a year ago, so maybe it has been fixed
since.  I don't know.

Anyways, this code sequence is OK
foo:
        fmv.w.x fa0,a0
        ret
because we are moving a 32-bit SImode value to an FP reg and then treating it
as SFmode, and the 32-bit move will properly NaN-box the SFmode value.

This code sequence is not OK
foo:
        fmv.d.x fa5,a0
        fmul.s  fa0,fa0,fa5
because we are moving a 64-bit DImode value to an FP reg and then treating it
as SFmode, which is not OK because the value won't be NaN-boxed and may trap at
run time.

validate_subreg used to prevent the bad subreg from being created.

I would think that TARGET_CAN_CHANGE_MODE_CLASS could help here, but it isn't
being called inside general_operand when called from fwprop1 where the bad
substitution happens.  Because we have a pseudo-register, and it is only called
for hard registers.

I don't see a way to fix this as a backend change with current validate_subreg,
other than by replacing register_operand with riscv_register_operand, and
putting the subreg check I need inside riscv_register_operand.  And likewise
for any other affected predicate, like move_operand.  This will be a big
change, though a lot of it will be mechanical.  As an optimization, we can
continue to use register_operand in any pattern that can't use FP registers.

As a middle end change, I need a new hook in general_operand to reject subregs
that we can't support on RISC-V.

Or maybe re-add the check I need to validate_subreg as a hook, so it can be
conditionally enabled for RISC-V.

We can allow (subreg:SF (reg:DI)) if it gets allocated to an integer register. 
It is only when it is allocated to an FP register that it can't work.  I don't
know offhand if that can be described.  But disallowing the subreg always for
RISC-V is simpler and also works.