[llvm] [AMDGPU][True16][CodeGen] readfirstlane for vgpr16 copy to sgpr32 (PR #118037)

[Matt Arsenault via llvm-commits]

================
@@ -1086,10 +1086,22 @@ void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
         TRI->getRegClassForOperandReg(*MRI, MI->getOperand(1));
     size_t SrcSize = TRI->getRegSizeInBits(*SrcRC);
     if (SrcSize == 16) {
-      // HACK to handle possible 16bit VGPR source
-      auto MIB = BuildMI(*MBB, MI, MI->getDebugLoc(),
-                         TII->get(AMDGPU::V_READFIRSTLANE_B32), DstReg);
-      MIB.addReg(SrcReg, 0, AMDGPU::NoSubRegister);
+      assert(MF.getSubtarget<GCNSubtarget>().useRealTrue16Insts() &&
+             "We do not expect to see 16-bit copies from VGPR to SGPR unless "
+             "we have 16-bit VGPRs");
+      assert(MRI->getRegClass(DstReg) == &AMDGPU::SGPR_LO16RegClass ||
+             MRI->getRegClass(DstReg) == &AMDGPU::SReg_32RegClass ||
+             MRI->getRegClass(DstReg) == &AMDGPU::SReg_32_XM0RegClass);
+      // There is no V_READFIRSTLANE_B16, so legalize the dst/src reg to 32 bits
+      MRI->setRegClass(DstReg, &AMDGPU::SReg_32_XM0RegClass);
+      Register VReg32 = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+      const DebugLoc &DL = MI->getDebugLoc();
+      BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::SUBREG_TO_REG), VReg32)
+          .addImm(0)
+          .addReg(SrcReg, 0)
+          .addImm(AMDGPU::lo16);
----------------
arsenm wrote:

No, as in the operation is fundamentally broken as a representation and should be deleted from the IR. There are 2 issues, first here you are asserting the high bits are 0 when they are not in use here. Second, the problem is it's introducing a hidden dependence on the unnamed super register component which is no good. The coalescer has to insert hacky super register defs to preserve the other subregisters which were unnamed prior to coalescing 

https://github.com/llvm/llvm-project/pull/118037