[Bug rtl-optimization/62151] [5 Regression] wrong code at -O2 and -O3 on x86_64-linux-gnu

[amker.cheng at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62151

--- Comment #15 from bin.cheng <amker.cheng at gmail dot com> ---
Hmm, words on tem_insn issue at the end of comment #12 isn't mature.  It's more
complicated than that.

Turns out live range of register which is noted as DEAD in i1/i2 can be
extended because we propagate its use into i3, like below example:

ix: rx <- frags:CCZ!=0
i1: r1 <- r0 & 0xfffe
          REG_DEAD (r0)
i2: r2 <- r1 | rx
          REG_DEAD (r1)
          REG_DEAD (rx)
i3: r3 <- r2 & 0xfffd
          REG_DEAD (r2)

Which can be combined into:

ix: rx <- frags:CCZ!=0
i1: INSN_DELETED
i2: r2 <- r0 & 0xfffc
          REG_DEAD (r0)
i3: r3 <- r2 | rx
          REG_DEAD (rx)

Apparently, live ranges of both r0 and rx are extended.  The REG_DEAD(r0) in
new i2 and REG_DEAD(rx) in new i3 are distributed from the original i1 and i2. 
In this case, if we set tem_insn to from_insn, gcc will delete the previous
definition of r0/rx, resulting in wrong code.

I think the case I given in #12 may still hold.  Interesting thing is, the
comment in distribute_notes is describing that case, rather than what it
handles as in this comment:
Quoted from source code:

          /* ...
         If the register is not used as an input in either I3 or I2
         and it is not one of the registers we were supposed to eliminate,
         there are two possibilities.  We might have a non-adjacent I2
         or we might have somehow eliminated an additional register
         from a computation.  For example, we might have had A & B where
         we discover that B will always be zero.  In this case we will
         eliminate the reference to A.

         In both cases, we must search to see if we can find a previous
         use of A and put the death note there.  */

I will rephrase the comments along with code changes.