On Tue, Apr 10, 2018 at 03:50:33PM +0200, Jakub Jelinek wrote:
> The following testcase ICEs, because subst first creates invalid RTL
> (in this case (float:SF (const_int 256))) and only later on simplifies it
> using on the-side op0_mode. In the end try_combine succeeds, so we don't
> undo anything from the undo buffer, but call propagate_for_debug with the
> patterns that were partially modified by subst and ICE on those.
> subst already has code to deal with ZERO_EXTEND, this patch just handles
> FLOAT and UNSIGNED_FLOAT the same; that way we don't create the invalid
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Yes please. And thanks for the explanation!
> 2018-04-10 Jakub Jelinek <ja...@redhat.com>
> PR rtl-optimization/85300
> * combine.c (subst): Handle subst of CONST_SCALAR_INT_P new_rtx also
> into FLOAT and UNSIGNED_FLOAT like ZERO_EXTEND, return a CLOBBER if
> simplify_unary_operation fails.
> * gcc.dg/pr85300.c: New test.