[llvm] [GVN] Share equality propagation for assume and condition (PR #161639)

[Nikita Popov via llvm-commits]

================
@@ -2652,26 +2558,45 @@ bool GVNPass::propagateEquality(Value *LHS, Value *RHS,
       // If the number we were assigned was brand new then there is no point in
       // looking for an instruction realizing it: there cannot be one!
       if (Num < NextNum) {
-        Value *NotCmp = findLeader(Root.getEnd(), Num);
-        if (NotCmp && isa<Instruction>(NotCmp)) {
-          unsigned NumReplacements =
-              DominatesByEdge
-                  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
-                  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
-                                             Root.getStart());
-          Changed |= NumReplacements > 0;
-          NumGVNEqProp += NumReplacements;
-          // Cached information for anything that uses NotCmp will be invalid.
-          if (MD)
-            MD->invalidateCachedPointerInfo(NotCmp);
+        for (const auto &Entry : LeaderTable.getLeaders(Num)) {
----------------
nikic wrote:

To explain what is going on here: I initially changed this to look for leaders in DominatedBlocks instead, but that caused regressions for the case where the edge is non-dominating, but there is a phi use of the inverted condition.

So instead I explicitly loop over the leaders and inspect the ones that could potentially have replaceable uses. Worth noting that what findLeader() does internally is also loop over the leaders and perform a dominance check, we just need more flexibility here for the dominance condition to use.

This change is what accounts for the improvements in llvm-opt-benchmark. 

https://github.com/llvm/llvm-project/pull/161639