[llvm-commits] [llvm] r144290 - /llvm/trunk/test/CodeGen/X86/lsr-loop-exit-cond.ll

[Andrew Trick]

On Nov 10, 2011, at 6:48 AM, NAKAMURA Takumi wrote:

> 2011/11/10 Duncan Sands <baldrick at free.fr>:
>> Hi Takumi, is this correct?  I think this test is checking that the compare
>> occurs next to the branch.  If that's not working on some platforms shouldn't
>> that be fixed rather than hidden?
> 
> As far as I checked freebsd bots, They say;
> "movl (" expected, but "movl XXX(" seen after "decq".
> It is just as same as on linux.
> 
> FYI, it could be reproduced with -mtriple=x86_64-{linux|freebsd}.
> 
> ...Ah, it might be possible simply to relax 2nd CHECK.
> Feel free to rewrite. gotta sleep.

This test was not intentionally testing the schedule. It just wants to
know that decq+jne was emitted.  My scheduler optimization made that
check more reliable so I was able to reenable the test. But Evan's
checkin (r144267) relaxed the schedule so the test is no longer
reliable. The check for movl is a side effect of the current
schedule. It's fine to either add a triple or regex to match any
platform's movl.

-Andy