You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Previously, after the algorithm fixpointed, the state was manually
patched by emitting BDVs for EE instructions earlier, while marking some
(but not all) vector and vector<->scalar instructions as conflict. This
causes issues as not all instructions that required BDVs had them
emitted and due to after-fixpoint patching, the extra BDVs did not
propagate to their users.
This change fixes both by rewriting the logic for BDV insertion &
patching. Instead of inserting the BDV for EE earlier, it merely marks
every EE instruction as a conflict. The two phase insertion algorithm
(first insert empty instructions and patch the BDVState, then actually
connect the BDV instructions to their input bases) then ensures correct
propagation to all its users. Furthermore the shufflevector instruction
as well as all instances of IE instruction are conservatively marked as
conflicts as well, fixing the second problem.
This change does not fix the handling of constant values and vectors in
the BDV.
---------
Co-authored-by: Petr Maj <[email protected]>
; FIXME: this is a extract element of constant vector and as such should not be relocated, but our current code relocates it since the extractelement will conservatively be marked as conflict during the BDV insertion algorithm. In many cases the extra instruction will be identical to the one already present and will be removed by later phases.
269
+
entry:
270
+
%val = extractelement <2 x ptraddrspace(1)> <ptraddrspace(1) inttoptr (i645toptraddrspace(1)),
0 commit comments