[Bug rtl-optimization/69447] [5/6 Regression] wrong code with -O2 -fno-schedule-insns and mixed 8/16/32/64bit arithmetics @ armv7a

[vmakarov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69447

--- Comment #14 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
(In reply to Richard Henderson from comment #13)
> (In reply to Jakub Jelinek from comment #11)
> > Without knowing the lra-remat code at all, I just wonder if subreg_regs
> > needs to be one per the whole function, rather than say per extended basic
> > block or basic block, with the patch any uses in multi-reg subregs anywhere
> > in the function will affect remat of all other spots where it is used.
> 
> I started with subreg_regs being per-block.  But since IRA has
> some global component, I was concerned that there would be some
> edge case that would be missed, and switched to a global bitmap.
> 
> Perhaps someone who knows the allocator better can comment.

Richard, thanks for working on it.  You used the right approach.  I believe
some edge cases are possible, e.g. sched1 could move one subreg insn into
another block.  So just non-global approach would be vulnerable.

The patch looks ok to me.  So you can commit it if the test results are ok.

Thanks again.  I had no time to work on it as I am too busy with other RA bugs.