[Bug rtl-optimization/77416] New: LRA rematerializing use of CA reg across function call

[pthaugen at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77416

            Bug ID: 77416
           Summary: LRA rematerializing use of CA reg across function call
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pthaugen at gcc dot gnu.org
                CC: bergner at gcc dot gnu.org, dje.gcc at gmail dot com,
                    vmakarov at gcc dot gnu.org, wschmidt at gcc dot gnu.org
  Target Milestone: ---
              Host: powerpc64-unknown-linux-gnu
            Target: powerpc64-unknown-linux-gnu
             Build: powerpc64-unknown-linux-gnu

Created attachment 39521
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=39521&action=edit
reduced testcase

CPU2006 benchmark 445.gobmk started failing verification of output starting
with r239105 (switch rs6000 port to LRA). I tracked it down to code where LRA
is rematerializing the use of the CA reg across a function call, which is not
valid since a call can clobber it.

Before the switch to LRA we saw the following code:

        addic 9,9,-1
        addze 9,16
        stw 9,492(1)
... code with function call(s)
        lwz 9,492(1)


After LRA switch the code looks like this:

        addic 9,9,-1
        addze 9,16
        stw 9,492(1)
... code with function call(s)
        addze 9,16


Compile options for reduced testcase: gcc -S -O2 -m32 -mcpu=power7