[llvm] [X86][GlobalISel] Support G_FCMP for scalar cases (PR #123598)

[Evgenii Kudriashov via llvm-commits]

================
@@ -1048,17 +1048,29 @@ bool X86InstructionSelector::selectFCmp(MachineInstr &I,
     break;
   }
 
+  assert((LhsReg.isVirtual() && RhsReg.isVirtual()) &&
+         "Both arguments of FCMP need to be virtual!");
+  auto *LhsBank = RBI.getRegBank(LhsReg, MRI, TRI);
+  auto *RhsBank = RBI.getRegBank(RhsReg, MRI, TRI);
+  assert((LhsBank == RhsBank) &&
+         "Both banks assigned to FCMP arguments need to be same!");
+
   // Compute the opcode for the CMP instruction.
   unsigned OpCmp;
   LLT Ty = MRI.getType(LhsReg);
   switch (Ty.getSizeInBits()) {
   default:
     return false;
   case 32:
-    OpCmp = X86::UCOMISSrr;
+    OpCmp = LhsBank->getID() == X86::PSRRegBankID ? X86::UCOM_FpIr32
+                                                  : X86::UCOMISSrr;
     break;
   case 64:
-    OpCmp = X86::UCOMISDrr;
+    OpCmp = LhsBank->getID() == X86::PSRRegBankID ? X86::UCOM_FpIr64
----------------
e-kud wrote:

floats and doubles are supported as well but haven't been covered with tests. Let's take IR from `test/CodeGen/X86/GlobalISel/select-fcmp.mir` and create `llvm/test/CodeGen/X86/isel-fcmp.ll`. We can test it for `x86-64` where `SSE` is present and `i686` where `x87` must be used. They should be supported already.

https://github.com/llvm/llvm-project/pull/123598