[llvm] [X86][GISel] Fix crash on bitcasting i16 <-> half with gisel enabled. (PR #168456)

[Simon Pilgrim via llvm-commits]

================
@@ -4294,10 +4294,28 @@ static unsigned CopyToFromAsymmetricReg(Register DestReg, Register SrcReg,
 
   if (X86::VR128XRegClass.contains(DestReg) &&
       X86::GR32RegClass.contains(SrcReg))
-    // Copy from a VR128 register to a VR128 register.
+    // Copy from a GR32 register to a VR128 register.
     return HasAVX512 ? X86::VMOVDI2PDIZrr
            : HasAVX  ? X86::VMOVDI2PDIrr
                      : X86::MOVDI2PDIrr;
+
+  // SrcReg(VR128) -> DestReg(GR16)
+  // SrcReg(GR16)  -> DestReg(VR128)
+
+  if (X86::GR16RegClass.contains(DestReg) &&
+      X86::VR128XRegClass.contains(SrcReg))
+    // Copy from a VR128 register to a GR16 register.
+    return HasAVX512 ? X86::VPEXTRWZrri
+           : HasAVX  ? X86::VPEXTRWrri
+                     : X86::PEXTRWrri;
+
+  if (X86::VR128XRegClass.contains(DestReg) &&
+      X86::GR16RegClass.contains(SrcReg))
+    // Copy from a GR16 register to a VR128 register.
+    return HasAVX512 ? X86::VPINSRWZrri
+           : HasAVX  ? X86::VPINSRWrri
+                     : X86::PINSRWrri;
----------------
RKSimon wrote:

It'd be be much better perf if we could use gpr<->xmm moves instead with any gpr32 extension/truncations

https://github.com/llvm/llvm-project/pull/168456