[Bug target/105624] [13 Regression] ICE in final_scan_insn_1, at final.cc:2861 (error: could not split insn)

["ubizjak at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105624

--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to rsandifo@gcc.gnu.org from comment #5)
> FWIW, I think the problem is specific to operands that are
> commutative with a non-constant operand.  For example,
> suppose the pre-RA instruction had a pseudo register R matching
> a register_operand and a constant C matching a const_int_operand.
> If R does not get allocated, and so gets replaced by a stack slot M,
> the % would allow the RA to try mapping C to the register_operand
> and M to the const_int_operand.  Without a constraint on the latter,
> the M mapping would seem to be valid, and reloading C into a register
> might seem less costly than reloading M into a register.
> 
> The intent of the patch seemed good otherwise (and a nice clean-up).
> I don't think the whole thing needed to be reverted.

I was afraid I don't understood the reason of the failure well, although it
happened very rarely (actually, no failures were detected during the build or
testsuite run). The patch obviously triggered some inconsistency in the
infrastructure, so without some assurances, I took the safe way and reverted
everything.

I would gladly revert the revert. The reload is just doing unnecessary work
when multiple constraints are the same; all necessary information could be
retreived from the predicate.