[llvm] [AMDGPU] Revert "Preliminary patch for divergence driven instruction selection. Operands Folding 1." (PR #71710)

[Jay Foad via llvm-commits]

================
@@ -735,33 +735,6 @@ void SIFoldOperands::foldOperand(
       return;
 
     const TargetRegisterClass *DestRC = TRI->getRegClassForReg(*MRI, DestReg);
-    if (!DestReg.isPhysical()) {
-      if (TRI->isSGPRClass(SrcRC) && TRI->hasVectorRegisters(DestRC)) {
-        SmallVector<FoldCandidate, 4> CopyUses;
-        for (auto &Use : MRI->use_nodbg_operands(DestReg)) {
-          // There's no point trying to fold into an implicit operand.
-          if (Use.isImplicit())
-            continue;
-
-          CopyUses.emplace_back(Use.getParent(),
-                                Use.getParent()->getOperandNo(&Use),
-                                &UseMI->getOperand(1));
-        }
-
-        for (auto &F : CopyUses) {
-          foldOperand(*F.OpToFold, F.UseMI, F.UseOpNo, FoldList,
----------------
jayfoad wrote:

The motivation is to remove these recursive calls to `foldOperand` which are one cause of #71685.

https://github.com/llvm/llvm-project/pull/71710