[llvm-commits] [PATCH] Fix sext(setcc(N0 cc N0)) for vectors

[Jose Fonseca]

----- Original Message -----
> Hi James,
> 
> > We were emitting 0/1 booleans, instead of 0/-1 booleans.
> >
> > Ran full test suite to ensure no regressions.
> 
> +++ b/test/CodeGen/Generic/sext-setcc-self.ll
> @@ -0,0 +1,76 @@
> +; RUN: %lli %s | grep {^0xFFFFFFFF 0xFFFFFFFF 0xFFFFFFFF
> 0xFFFFFFFF$}
> 
> I'm pretty sure you aren't supposed to be executing code like this in
> the
> regression testsuite, instead you should do pattern matching.  Code
> to be
> executed belongs in the nightly testsuite.  I could be wrong about
> this,
> but I think it's the policy.
> 
> Ciao, Duncan.
> 
> PS: Otherwise it looks OK.

There are a few lli tests in test/ExecutionEngine/*

  $ git grep %lli test
  test/ExecutionEngine/2002-12-16-ArgTest.ll:; RUN: %lli %s > /dev/null
  test/ExecutionEngine/2003-01-04-ArgumentBug.ll:; RUN: %lli %s > /dev/null
  test/ExecutionEngine/2003-01-04-LoopTest.ll:; RUN: %lli %s > /dev/null
  [...]

This where we got the idea of using lli in the test.


If there's no better solution, we'll rewrite the test as x86 assembly pattern matching then.


The test will be specific to x86 backend, although the problem was in the generic code, and could happen with arm too, but's better than nothing.


Jose