[AMDGPU] Don't rely on SIAddIMGInit for GlobalISel

[foad]

Summary

Image loads with TFE/LWE need special handling. Do it during instruction
selection rather than fixing it up with another pass. NFCI.

[arsenm]

If the fallback is enabled, this pass will still run. You need to have the pass check if isel failed and skip the function if it suceeded

[foad]

Rebase.

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

bool AMDGPUInstructionSelector::selectImageIntrinsic(
 
   MIB.addImm(TFE); // tfe
   MIB.addImm(LWE); // lwe
   if (!IsGFX10Plus)
     MIB.addImm(DimInfo->DA ? -1 : 0);
   if (BaseOpcode->HasD16)
     MIB.addImm(IsD16 ? -1 : 0);
 
+  if (IsTexFail) {
+    // An image load instruction with TFE/LWE only conditionally writes to its
+    // result registers. Initialize them to zero so that we always get well
+    // defined result values.
+    assert(VDataOut && !VDataIn);
+    Register Tied = MRI->cloneVirtualRegister(VDataOut);
+    Register Zero = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::V_MOV_B32_e32), Zero)
+      .addImm(0);
+    auto Parts = TRI.getRegSplitParts(MRI->getRegClass(Tied), 4);
+    if (STI.usePRTStrictNull()) {
+      // With enable-prt-strict-null enabled, initialize all result registers to
+      // zero.
+      auto RegSeq = BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::REG_SEQUENCE), Tied);
+      for (auto Sub : Parts)
+        RegSeq.addReg(Zero).addImm(Sub);
+    } else {
+      // With enable-prt-strict-null disabled, only initialize the extra TFE/LWE
+      // result register.
+      Register Undef = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+      BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::IMPLICIT_DEF), Undef);
+      auto RegSeq = BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::REG_SEQUENCE), Tied);
+      for (auto Sub : Parts.drop_back(1))
+        RegSeq.addReg(Undef).addImm(Sub);
+      RegSeq.addReg(Zero).addImm(Parts.back());
+    }
+    MIB.addReg(Tied, RegState::Implicit);
+    MIB->tieOperands(0, MIB->getNumOperands() - 1);
+  }
+
   MI.eraseFromParent();
   return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
 }
 
 bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
     MachineInstr &I) const {
   unsigned IntrinsicID = I.getIntrinsicID();
   switch (IntrinsicID) {

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

 
 void GCNPassConfig::addPreGlobalInstructionSelect() {
   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
   addPass(createAMDGPURegBankCombiner(IsOptNone));
 }
 
 bool GCNPassConfig::addGlobalInstructionSelect() {
   addPass(new InstructionSelect(getOptLevel()));
-  // TODO: Fix instruction selection to do the right thing for image
-  // instructions with tfe or lwe in the first place, instead of running a
-  // separate pass to fix them up?
-  addPass(createSIAddIMGInitPass());
   return false;
 }
 
 void GCNPassConfig::addPreRegAlloc() {
   if (LateCFGStructurize) {
     addPass(createAMDGPUMachineCFGStructurizerPass());
   }
 }