On 08/12/2015 12:32 PM, Richard Biener wrote:
On August 12, 2015 8:07:13 PM GMT+02:00, Jeff Law <l...@redhat.com> wrote:
On 08/12/2015 11:12 AM, Richard Biener wrote:
Prec is almost never a constant and is heavily used from wide-int.
We are not exploiting this undefined ness in C so I object to making
this so much slower.
Can we instead do what we do for abs_hwi and add a checking assert so
we can move the tests to the callers that misbehave instead?
Given that ISO C++ is moving away from making shifting 1 into the sign
bit undefined behaviour, maybe we should make UBSan less strict in its
warning. That may eliminate the need for Mikael's patch.
We can also use an logical left shift followed by an arithmetic right shift. Or
is the latter invoking undefined behaviour as well in some cases we hit?
Hmm, why aren't we using logicals for the left shift to begin with?
That's the problem area. I don't think the right shifts are an issue at
all.
It's strange that when I was researching this, consistently I saw folks
suggesting the LEFT-SHIFT followed by RIGHT-SHIFT, then it'd get shot
down as UB, then folks went to either Mikael's approach or another that
is very similar. Nobody suggested twidding the types to get a
left-logical-shift, then doing a right-arithmetic-shift.
Jeff
