On Wed, Oct 31, 2018 at 11:04:32AM +0100, Martin Liška wrote:
> Hi.
>
> As Jakub pointed out we should not ICE when last argument
> of __builtin_expect_with_probability is not a real cst.
> Plus I documented the behavior.
That is not what you've implemented. The documentation says that
it must be a compile time constant, but nothing enforces it, just silently
ignores it. So, either diagnose it at the time when
__builtin_expect_with_probability is removed from the IL, i.e.
the code you've touched? and expand_builtin_expect_with_probability
and the argument would be really a late constant (constant after
inlining and optimizations), or diagnose it in fold_builtin_expect
and replace with the argument at that point. And, please diagnose also the
cases where the probability is not >= 0.0 and <= 1.0.
> * doc/extend.texi: Update constrain about the last arguemnt
s/arguemnt/argument/
Jakub
