[Bug target/86771] [9 Regression] gfortran.dg/actual_array_constructor_1.f90 fails on arm after combine 2 insns to 2 insns patch

[wilco at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86771

Wilco <wilco at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2018-08-20
                 CC|                            |wilco at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #13 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to nsz from comment #12)
> the wrong string seems to be caused by a missing ldm

Yes what happens is that it decides to combine instruction 150 and 190 and
split an ldm2:

insn_cost 4 for   150: r197:SI=sfp:SI-0x24
...
insn_cost 0 for   190: {r0:SI=[r197:SI];r1:SI=[r197:SI+0x4];}


allowing combination of insns 150 and 190
original costs 4 + 0 = 0
replacement costs 8 + 8 = 16
modifying insn i2   150: r0:SI=[sfp:SI-0x24]
deferring rescan insn with uid = 150.
modifying insn i3   190: r1:SI=[sfp:SI-0x20]
deferring rescan insn with uid = 190.

There are lots of instructions between 150 and 190, including another ldm2
which writes r0:

150: r0:SI=[sfp:SI-0x24]
...
163: {r0:SI=[r206:SI];r1:SI=[r206:SI+0x4];}
...
190: r1:SI=[sfp:SI-0x20]

The combination ends up corrupting r0 because one of the loads is lifted across
many instructions without checking liveness. It seems the safest way to split
an instruction is to place the new instructions next to each other.