[llvm] [AMDGPU] Per-chain MFMA->AGPR conversion (PR #217328)

[Romanov Vlad via llvm-commits]

================
@@ -1399,20 +1457,113 @@ bool RewriteMFMAFormStage::initGCNSchedStage() {
   TII = ST.getInstrInfo();
   SRI = ST.getRegisterInfo();
 
-  std::vector<std::pair<MachineInstr *, unsigned>> RewriteCands;
-  DenseMap<MachineBasicBlock *, std::set<Register>> CopyForUse;
-  SmallPtrSet<MachineInstr *, 8> CopyForDef;
+  // Collect all convertible MFMAs.
+  SmallVector<MachineInstr *, 32> AllCands;
+  for (MachineBasicBlock &MBB : MF)
+    for (MachineInstr &MI : MBB)
+      if (isRewriteCandidate(&MI))
+        AllCands.push_back(&MI);
+
+  if (AllCands.empty())
+    return false;
+
+  // Identify accumulator chains and sort by size descending. We operate at
+  // chain granularity rather than individual MFMAs because:
+  // 1. It avoids converting MFMAs from different chains that would each need
+  //    separate src2 and dst bridge copies (up to 4 copies for 2 MFMAs from
+  //    2 chains) while only reducing VGPR pressure for 2 instructions.
+  // 2. It reduces the search space from O(N_mfmas) to O(N_chains).
+  //
+  // Longer chains benefit more: each chain requires at most one src2 bridge
+  // copy and one dst bridge copy regardless of length, but reduces VGPR
+  // pressure proportionally to the number of chain members.
+  SmallVector<SmallVector<unsigned, 8>, 16> Chains =
+      identifyAccChains(AllCands, TII);
+  llvm::sort(Chains,
+             [](const auto &A, const auto &B) { return A.size() > B.size(); });
+
+  // Evaluate the cost of converting the first N sorted chains to AGPR form.
+  auto EvaluateProbe = [&](int N) -> int64_t {
----------------
romanovvlad wrote:

Moved.

https://github.com/llvm/llvm-project/pull/217328