[llvm] unpack packed instructions overlapped by MFMAs post-RA scheduling (PR #157968)

[Jeffrey Byrnes via llvm-commits]

================
@@ -461,7 +727,21 @@ bool SIPreEmitPeephole::run(MachineFunction &MF) {
     // and limit the distance to 20 instructions for compile time purposes.
     // Note: this needs to work on bundles as S_SET_GPR_IDX* instructions
     // may be bundled with the instructions they modify.
+    //
+    // Unpack packed instructions overlapped by MFMAs. This allows the compiler
+    // to co-issue unpacked instructions with MFMA
+    uint16_t NumMFMACycles = 0;
+    auto SchedModel = TII->getSchedModel();
+    SetVector<MachineInstr *> InstrsToUnpack;
+
     for (auto &MI : make_early_inc_range(MBB.instrs())) {
+      if (SIInstrInfo::isMFMA(MI)) {
+        const MCSchedClassDesc *SchedClassDesc =
+            SchedModel.resolveSchedClass(&MI);
+        NumMFMACycles =
----------------
jrbyrnes wrote:

Do we need to subtract for the issue latency of MFMA?

https://github.com/llvm/llvm-project/pull/157968