incorrect speculation on ia64-redhat-linux (GCSE pass)

[Paolo Bonzini <bonzini at wbkst15 dot mach dot uni-karlsruhe dot de>]

In the test case I attach (compiled with 2.96-ia64-000717 snap 000717)
there is an incorrect speculated load of selfCache->object in function
interpret.

The snapshot I'm using puts selfCache in r32 then does an incorrect

    ld8 r29 = [r32]

before the computed goto (`br b4') rather than in the blocks that really
access selfCache->object.  This incorrect speculation is done
after every `selfCache = self' statement, as far as I can tell from a
quick look at the generated assembly language.

I was told that GCSE and computed gotos interact badly by Brad Lucier (on
gcc@gcc.gnu.org) -- and indeed interp.i uses them a lot.

The bug appears with

   gcc -S -O2 -fno-inline-functions interp.i

and not with

   gcc -S -O2 -fno-inline-functions -fno-gcse interp.i

(note that while compiling interp.i cc1 can grow to 70+ megabytes of
resident set)

interp.i has three warnings, but they are not important.

Paolo

interp.i.gz