[llvm] [CodeGen] Ignore requiresStructuredCFG check in canSplitCriticalEdge if successor is loop header (PR #154063)

[Wenju He via llvm-commits]

https://github.com/wenju-he created https://github.com/llvm/llvm-project/pull/154063

This addresses a performance issue for our downstream GPU target that sets requiresStructuredCFG to true. The issue is that EarlyMachineLICM pass does not hoist loop invariants because a critical edge is not split.

>From 1c2a9e2b97910e1d59852999ae5b2196d1976ebe Mon Sep 17 00:00:00 2001
From: Wenju He <wenju.he at intel.com>
Date: Mon, 18 Aug 2025 06:37:03 +0200
Subject: [PATCH] [CodeGen] Ignore requiresStructuredCFG check in
 canSplitCriticalEdge if successor is loop header

This addresses a performance issue for our downstream GPU target that
sets requiresStructuredCFG to true. The issue is that EarlyMachineLICM
pass does not hoist loop invariants because a critical edge is not split.
---
 llvm/include/llvm/CodeGen/MachineBasicBlock.h |  4 +++-
 llvm/lib/CodeGen/MachineBasicBlock.cpp        | 15 ++++++++++++---
 2 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/MachineBasicBlock.h b/llvm/include/llvm/CodeGen/MachineBasicBlock.h
index 9e3d9196cc184..78e5dd99eab06 100644
--- a/llvm/include/llvm/CodeGen/MachineBasicBlock.h
+++ b/llvm/include/llvm/CodeGen/MachineBasicBlock.h
@@ -1035,7 +1035,9 @@ class MachineBasicBlock
   /// Succ, can be split. If this returns true a subsequent call to
   /// SplitCriticalEdge is guaranteed to return a valid basic block if
   /// no changes occurred in the meantime.
-  LLVM_ABI bool canSplitCriticalEdge(const MachineBasicBlock *Succ) const;
+  LLVM_ABI bool
+  canSplitCriticalEdge(const MachineBasicBlock *Succ,
+                       const SplitCriticalEdgeAnalyses &Analyses = {}) const;
 
   void pop_front() { Insts.pop_front(); }
   void pop_back() { Insts.pop_back(); }
diff --git a/llvm/lib/CodeGen/MachineBasicBlock.cpp b/llvm/lib/CodeGen/MachineBasicBlock.cpp
index c3c5a0f5102d7..82c852bb0f93b 100644
--- a/llvm/lib/CodeGen/MachineBasicBlock.cpp
+++ b/llvm/lib/CodeGen/MachineBasicBlock.cpp
@@ -1160,7 +1160,7 @@ MachineBasicBlock *MachineBasicBlock::SplitCriticalEdge(
 MachineBasicBlock *MachineBasicBlock::SplitCriticalEdge(
     MachineBasicBlock *Succ, const SplitCriticalEdgeAnalyses &Analyses,
     std::vector<SparseBitVector<>> *LiveInSets, MachineDomTreeUpdater *MDTU) {
-  if (!canSplitCriticalEdge(Succ))
+  if (!canSplitCriticalEdge(Succ, Analyses))
     return nullptr;
 
   MachineFunction *MF = getParent();
@@ -1389,7 +1389,8 @@ MachineBasicBlock *MachineBasicBlock::SplitCriticalEdge(
 }
 
 bool MachineBasicBlock::canSplitCriticalEdge(
-    const MachineBasicBlock *Succ) const {
+    const MachineBasicBlock *Succ,
+    const SplitCriticalEdgeAnalyses &Analyses) const {
   // Splitting the critical edge to a landing pad block is non-trivial. Don't do
   // it in this generic function.
   if (Succ->isEHPad())
@@ -1403,7 +1404,15 @@ bool MachineBasicBlock::canSplitCriticalEdge(
   const MachineFunction *MF = getParent();
   // Performance might be harmed on HW that implements branching using exec mask
   // where both sides of the branches are always executed.
-  if (MF->getTarget().requiresStructuredCFG())
+  // However, if `Succ` is a loop header, splitting the critical edge will not
+  // break structured CFG.
+  auto SuccIsLoopHeader = [&]() {
+    if (MachineLoopInfo *MLI = Analyses.MLI)
+      if (MachineLoop *L = MLI->getLoopFor(Succ); L && L->getHeader() == Succ)
+        return true;
+    return false;
+  };
+  if (MF->getTarget().requiresStructuredCFG() && !SuccIsLoopHeader())
     return false;
 
   // Do we have an Indirect jump with a jumptable that we can rewrite?