[llvm] AMDGPU: Allow operand folding between loop body and its preheader (PR #137022)

[Matt Arsenault via llvm-commits]

================
@@ -9743,6 +9743,90 @@ MachineInstr *llvm::getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P,
   return nullptr;
 }
 
+// helper function to checkIfExecMayBeModifiedBeforeUseAcrossBB and
+// execMayBeModifiedBeforeUse. This checks possible EXEC register modifications
+// for a straight-line sequence of instructions between BeginIterator and
+// EndIterator (both inclusive) upto a pre-defined limit MaxInstScan
+bool execMayBeModifiedBeforeUseUtil(
+    const TargetRegisterInfo *TRI,
+    const MachineInstrBundleIterator<const MachineInstr> BeginIterator,
+    const MachineInstrBundleIterator<const MachineInstr> EndIterator,
+    const int MaxInstScan) {
+
+  int NumInst = 0;
+  for (auto I = BeginIterator; I != EndIterator; ++I) {
+    if (I->isDebugInstr())
+      continue;
+
+    if (++NumInst > MaxInstScan) {
+      dbgs() << "## maxinst\n";
+      return true;
+    }
+
+    if (I->modifiesRegister(AMDGPU::EXEC, TRI))
+      return true;
+  }
+  return false;
+}
+
+// Variant of execMayBeModifiedBeforeUse(), where DefMI and UseMI belong to
+// different basic blocks. Current code is limited to a very simple case: DefMI
+// in the predecessor BB of the single BB loop where UseMI resides.
+bool llvm::checkIfExecMayBeModifiedBeforeUseAcrossBB(
+    const MachineRegisterInfo &MRI, Register VReg, const MachineInstr &DefMI,
+    const MachineInstr &UseMI, const bool SIFoldOperandsPreheader,
+    const int SIFoldOperandsPreheaderThreshold) {
+
+  assert(MRI.isSSA() && "Must be run on SSA");
+  auto *TRI = MRI.getTargetRegisterInfo();
+  auto *DefBB = DefMI.getParent();
+  const int MaxInstScan = (SIFoldOperandsPreheaderThreshold > 10000)
+                              ? 10000
+                              : SIFoldOperandsPreheaderThreshold;
+
+  // Check whether EXEC is modified along all possible control flow between
+  // DefMI and UseMI, which may include loop backedge
+  // 1. UseBB is the only successor of DefBB
+  // 2. UseBB is a single basic block loop (only two predecessor blocks: DefBB
+  // and UseBB)
+  // 3. check if EXEC is modified
+  auto *UseBB = UseMI.getParent();
+  if (UseBB != DefBB) {
+    if (SIFoldOperandsPreheader) {
----------------
arsenm wrote:

Option is making this more confusing, shouldn't have entered the function if it were false (but also it shouldn't be an option) 

https://github.com/llvm/llvm-project/pull/137022