[llvm] [X86][CodeGen] Support using NF instructions for flag copy lowering (PR #93508)

[Phoebe Wang via llvm-commits]

================
@@ -254,14 +268,102 @@ bool X86FlagsCopyLoweringPass::runOnMachineFunction(MachineFunction &MF) {
   // turn copied again we visit the first one first. This ensures we can find
   // viable locations for testing the original EFLAGS that dominate all the
   // uses across complex CFGs.
-  SmallVector<MachineInstr *, 4> Copies;
+  SmallSetVector<MachineInstr *, 4> Copies;
   ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
   for (MachineBasicBlock *MBB : RPOT)
     for (MachineInstr &MI : *MBB)
       if (MI.getOpcode() == TargetOpcode::COPY &&
           MI.getOperand(0).getReg() == X86::EFLAGS)
-        Copies.push_back(&MI);
+        Copies.insert(&MI);
+
+  // Try to elminate the copys by transform the instructions between copy and
+  // copydef to the NF (no flags update) variants, e.g.
+  //
+  // %1:gr64 = COPY $eflags
+  // OP1 implicit-def dead $eflags
+  // $eflags = COPY %1
+  // OP2 cc, implicit $eflags
+  //
+  // ->
+  //
+  // OP1_NF
+  // OP2 implicit $eflags
+  if (Subtarget->hasNF()) {
+    SmallSetVector<MachineInstr *, 4> RemovedCopies;
+    // CopyIIt may be invalidated by removing copies.
+    auto CopyIIt = Copies.begin(), CopyIEnd = Copies.end();
+    while (CopyIIt != CopyIEnd) {
+      auto NCopyIIt = std::next(CopyIIt);
+      SmallSetVector<MachineInstr *, 4> EvitableClobbers;
+      MachineInstr *CopyI = *CopyIIt;
+      MachineOperand &VOp = CopyI->getOperand(1);
+      MachineInstr *CopyDefI = MRI->getVRegDef(VOp.getReg());
+      MachineBasicBlock *CopyIMBB = CopyI->getParent();
+      MachineBasicBlock *CopyDefIMBB = CopyDefI->getParent();
+      // Walk all basic blocks reachable in depth-first iteration on the inverse
+      // CFG from CopyIMBB to CopyDefIMBB. These blocks are all the blocks that
+      // may be executed between the execution of CopyDefIMBB and CopyIMBB. On
+      // all execution paths, instructions from CopyDefI to CopyI (exclusive)
+      // has to be NF-convertible if it clobbers flags.
+      for (auto BI = idf_begin(CopyIMBB), BE = idf_end(CopyDefIMBB); BI != BE;
+           ++BI) {
+        MachineBasicBlock *MBB = *BI;
+        for (auto I = (MBB != CopyDefIMBB)
+                          ? MBB->begin()
+                          : std::next(MachineBasicBlock::iterator(CopyDefI)),
+                  E = (MBB != CopyIMBB) ? MBB->end()
+                                        : MachineBasicBlock::iterator(CopyI);
+             I != E; ++I) {
+          MachineInstr &MI = *I;
+          EFLAGSClobber ClobberType = getClobberType(MI);
+          if (ClobberType == NoClobber)
+            continue;
+
+          if (ClobberType == InevitableClobber)
+            goto ProcessNextCopyI;
+
+          assert(ClobberType == EvitableClobber && "unexpected workflow");
----------------
phoebewang wrote:

Can't this be failed? I think we cannot transform another `$eflags = COPY %x` in `MBB` at least, right?

https://github.com/llvm/llvm-project/pull/93508