[llvm] [IR2Vec] Added fixme for cyclic dependency in Flow-Aware embedding computation (PR #162522)

[S. VenkataKeerthy via llvm-commits]

https://github.com/svkeerthy created https://github.com/llvm/llvm-project/pull/162522

None

>From e4c6990185c714c2538d90ea790eb1f6cedec7d5 Mon Sep 17 00:00:00 2001
From: svkeerthy <venkatakeerthy at google.com>
Date: Wed, 8 Oct 2025 18:20:17 +0000
Subject: [PATCH] IR2Vec Flow-aware fix

---
 llvm/lib/Analysis/IR2Vec.cpp | 19 +++++++++++++++----
 1 file changed, 15 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Analysis/IR2Vec.cpp b/llvm/lib/Analysis/IR2Vec.cpp
index 688535161d4b9..1794a604b991d 100644
--- a/llvm/lib/Analysis/IR2Vec.cpp
+++ b/llvm/lib/Analysis/IR2Vec.cpp
@@ -239,10 +239,21 @@ void FlowAwareEmbedder::computeEmbeddings(const BasicBlock &BB) const {
       // If the operand is defined elsewhere, we use its embedding
       if (const auto *DefInst = dyn_cast<Instruction>(Op)) {
         auto DefIt = InstVecMap.find(DefInst);
-        assert(DefIt != InstVecMap.end() &&
-               "Instruction should have been processed before its operands");
-        ArgEmb += DefIt->second;
-        continue;
+        // Fixme (#159171): Ideally we should never miss an instruction
+        // embedding here.
+        // But when we have cyclic dependencies (e.g., phi
+        // nodes), we might miss the embedding. In such cases, we fall back to
+        // using the vocabulary embedding. This can be fixed by iterating to a
+        // fixed-point, or by using a simple solver for the set of simultaneous
+        // equations.
+        // Another case when we might miss an instruction embedding is when
+        // the operand instruction is in a different basic block that has not
+        // been processed yet. This can be fixed by processing the basic blocks
+        // in a topological order.
+        if (DefIt != InstVecMap.end())
+          ArgEmb += DefIt->second;
+        else
+          ArgEmb += Vocab[*Op];
       }
       // If the operand is not defined by an instruction, we use the vocabulary
       else {