[llvm] [InstCombinePHI] Enhance PHI CSE to remove redundant phis (PR #163453)

[via llvm-commits]

================
@@ -1621,11 +1621,90 @@ Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {
     // Note that even though we've just canonicalized this PHI, due to the
     // worklist visitation order, there are no guarantess that *every* PHI
     // has been canonicalized, so we can't just compare operands ranges.
-    if (!PN.isIdenticalToWhenDefined(&IdenticalPN))
-      continue;
-    // Just use that PHI instead then.
-    ++NumPHICSEs;
-    return replaceInstUsesWith(PN, &IdenticalPN);
+    if (PN.isIdenticalToWhenDefined(&IdenticalPN)) {
+      // Just use that PHI instead then.
+      ++NumPHICSEs;
+      return replaceInstUsesWith(PN, &IdenticalPN);
+    }
+
+    // Look for the following pattern and do PHI CSE to clean up the
+    // redundant %phi. Here %phi, %1 and %phi.next perform the same
+    // functionality as %identicalPhi and hence %phi can be eliminated.
+    //
+    // BB1:
+    //   %identicalPhi = phi [ X, %BB0 ], [ %identicalPhi.next, %BB1 ]
+    //   %phi = phi [ X, %BB0 ], [ %phi.next, %BB1 ]
+    //   ...
+    //   %identicalPhi.next = select %cmp, %val, %identicalPhi
+    //   %1 = select %cmp2, %identicalPhi, float %phi
+    //   %phi.next = select %cmp, %val, %1
+    //
+    // Prove that %phi and %identicalPhi are the same by induction:
+    //
+    // Base case: Both %phi and %identicalPhi are equal on entry to the loop.
+    // Inductive case:
+    // Suppose %phi and %identicalPhi are equal at iteration i.
+    // We look at their values at iteration i+1 which are %phi.next and
+    // %identicalPhi.next. They would have become different only when %cmp is
+    // false and the corresponding values %1 and %identicalPhi differ.
+    //
+    // The only condition when %1 and %identicalPh could differ is when %cmp2
+    // is false and %1 is %phi, which contradicts our inductive hypothesis
+    // that %phi and %identicalPhi are equal. Thus %phi and %identicalPhi are
+    // always equal at iteration i+1.
+
+    if (PN.getNumIncomingValues() == 2 && PN.getNumUses() == 1) {
----------------
CongzheUalberta wrote:

Here it is supposed to check for the target pattern that is the redundant cycle of phi (%phi.to.remove) and two select instructions (%phi.to.remove.next, %same.as.v1).

I've now improved the code and I'm checking `SI->getNumUses() == 1` on line 1865 instead, to make sure that `%phi.to.remove.next` (which is checked to be a select) has only one use that is the backedge incoming value of `%phi.to.remove`. Hope it is more readable now.

https://github.com/llvm/llvm-project/pull/163453