[AMDGPU] fix copies between 32 and 16 bit

[rampitec]

Summary

This a hack to fix illegal 32 to 16 bit copies.
The problem is when we make 16 bit subregs legal it creates
a huge amount of failures which can only be resolved at once
without a temporary hack like this.

The next step is to change operands, instruction definitions
and patterns until this hack is not needed.

[arsenm]

This is pretty gross but it will take a while to fix all the tablegen patterns

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

 
 // Must be at least 4 to be able to branch over minimum unconditional branch
 // code. This is only for making it possible to write reasonably small tests for
 // long branches.
 static cl::opt<unsigned>
 BranchOffsetBits("amdgpu-s-branch-bits", cl::ReallyHidden, cl::init(16),
                  cl::desc("Restrict range of branch instructions (DEBUG)"));
 
+static cl::opt<bool> Fix16BitCopies(
+  "amdgpu-fix-16-bit-physreg-copies",
+  cl::desc("Fix copies between 32 and 16 bit registers by extending to 32 bit"),
+  cl::init(true),
+  cl::ReallyHidden);
+
 SIInstrInfo::SIInstrInfo(const GCNSubtarget &ST)
   : AMDGPUGenInstrInfo(AMDGPU::ADJCALLSTACKUP, AMDGPU::ADJCALLSTACKDOWN),
     RI(ST), ST(ST) {
   SchedModel.init(&ST);
 }
 
 //===----------------------------------------------------------------------===//
 // TargetInstrInfo callbacks

 }
 
 void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator MI,
                               const DebugLoc &DL, MCRegister DestReg,
                               MCRegister SrcReg, bool KillSrc) const {
   const TargetRegisterClass *RC = RI.getPhysRegClass(DestReg);
 
+  // FIXME: This is hack to resolve copies between 16 bit and 32 bit
+  // registers until all patterns are fixed.
+  if (Fix16BitCopies &&
+      ((RI.getRegSizeInBits(*RC) == 16) ^
+       (RI.getRegSizeInBits(*RI.getPhysRegClass(SrcReg)) == 16))) {
+    MCRegister &RegToFix = (RI.getRegSizeInBits(*RC) == 16) ? DestReg : SrcReg;
+    MCRegister Super = RI.get32BitRegister(RegToFix);
+    assert(RI.getSubReg(Super, AMDGPU::lo16) == RegToFix);
+    RegToFix = Super;
+
+    if (DestReg == SrcReg) {
+      // Insert empty bundle since ExpandPostRA expects an instruction here.
+      BuildMI(MBB, MI, DL, get(AMDGPU::BUNDLE));
+      return;
+    }
+
+    RC = RI.getPhysRegClass(DestReg);
+  }
+
   if (RC == &AMDGPU::VGPR_32RegClass) {
     assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
            AMDGPU::SReg_32RegClass.contains(SrcReg) ||
            AMDGPU::AGPR_32RegClass.contains(SrcReg));
     unsigned Opc = AMDGPU::AGPR_32RegClass.contains(SrcReg) ?
                      AMDGPU::V_ACCVGPR_READ_B32 : AMDGPU::V_MOV_B32_e32;
     BuildMI(MBB, MI, DL, get(Opc), DestReg)
       .addReg(SrcReg, getKillRegState(KillSrc));

llvm/test/CodeGen/AMDGPU/lo16-32bit-physreg-copy.mir

+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass postrapseudos -amdgpu-fix-16-bit-physreg-copies -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+
+# GCN-LABEL: name: lo16_to_v32
+# GCN: $vgpr1 = V_MOV_B32_e32 $vgpr0, implicit $exec
+name: lo16_to_v32
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1_lo16 = COPY $vgpr0
+    S_ENDPGM 0
+...
+
+# GCN-LABEL: name: v32_to_lo16
+# GCN: $vgpr1 = V_MOV_B32_e32 $vgpr0, implicit $exec
+name: v32_to_lo16
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = COPY $vgpr0_lo16
+    S_ENDPGM 0
+...
+
+# GCN-LABEL: name: samereg
+# GCN:      $vgpr0 = IMPLICIT_DEF
+# GCN-NEXT: BUNDLE
+# GCN-NEXT: S_ENDPGM
+name: samereg
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr0 = COPY $vgpr0_lo16
+    S_ENDPGM 0
+...