On 5/18/21 3:22 AM, Richard Biener wrote:
On Tue, May 18, 2021 at 1:23 AM Andrew MacLeod via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
The code in PR 100512 triggers an interaction between ranger and the
propagation engine related to undefined values.
I put the detailed analysis in the PR, but it boils down to the early
VRP pass has concluded that something is a constant and can be replaced,
and removes the definition expecting the constant to be propagated
everywhere.
If the code is in an undefined region that the CFG is going to remove,
we can find impossible situations,a nd ranger then changes that value ot
UNDEFINED.. because, well, it is. But then the propagation engine
panics because it doesnt have a constant any more, so odesnt replace it,
and now we have a used but not defined value.
Once we get to a globally constant range where further refinements can
only end up in an UNDEFINED state, stop further evaluating the range.
This is typically in places which are about to be removed by CFG cleanup
anyway, and it will make the propagation engine happy with no surprises.
Yeah, the propagation engine and EVRP as I know it relies on not visiting
"unexecutable" (as figured by anaysis) paths in the CFG and thus considering
edges coming from such regions not contributing conditions/values/etc. that
would cause such "undefinedness" to appear. Not sure how it works with
ranger, maybe that can as well get a mode where it does only traverse
EDGE_EXECUTABLE edges. Might be a bit difficult since IIRC it works
with SSA edges and not CFG edges.
Bootstraps on x86_64-pc-linux-gnu with no regressions, and fixes the PR.
So that means the lattice isn't an optimistic lattice, right? EVRPs wasn't
optimistic either, but VRPs is/was. Whatever this means in this context ;)
Richard.
Comparing in a gcc build, and out of the 4586 extra cases ranger finds
across 400 files, this patch does cost us 1 of them. So I suspect it
probably costs a few more compared to VRP proper. May have to deal with
the propagation issue better when we get to that point.
Andrew
