[Bug target/106303] [13 Regression] ICE on valid code at -O3 with -fno-inline-small-functions on x86_64-linux-gnu: in extract_insn, at recog.cc:2791 since r13-1607-gc3ed9e0d6e96d869

["roger at nextmovesoftware dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106303

--- Comment #2 from Roger Sayle <roger at nextmovesoftware dot com> ---
Grr... This problem is at (currently just beyond) the limits of my abilities. 
The TImode STV pass uses data flow to split conversion opportunities up into
chains, effectively performing live range splitting, as some of uses/defs of a
pseudo register may be in one chain, and other (independent) uses/defs of that
pseudo in another chain.  In the case reported in the PR, this results in some
occurences of (reg 90) being converted to V1TImode, and other occurrences
remaining in TImode.  Worse, thanks to RTL sharing, calling PUT_MODE on a
register in one chain, implicitly affects the mode of that register in other
chains/insns.  I must admit that I'd assumed that either all instances of a
pseudo (like (reg 90)) either get converted or all don't get converted [i.e.
chains ~ pseudos].  Fixing this will require either some form of regno mapping
or unsharing... (or??).

I'll continue investigating, but if anyone can see/knows the correct way to fix
this, please do let me know.