[AMDGPU] Fix an issue in `op_sel_hi` skipping.

[hliao]

Summary

• Only apply packed literal op_sel_hi skipping on operands requiring packed literals. Even an instruction is packed, it may have operand requiring non-packed literal, such as v_dot2_f32_f16.

[rampitec]

LGTM

llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp

if (MI->getDesc().TSFlags & SIInstrFlags::IsPacked) {
       else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2))
         ModIdx = AMDGPU::OpName::src2_modifiers;
       assert(ModIdx != -1);
       ModIdx = AMDGPU::getNamedOperandIdx(Opcode, ModIdx);
       MachineOperand &Mod = MI->getOperand(ModIdx);
       unsigned Val = Mod.getImm();
       if ((Val & SISrcMods::OP_SEL_0) || !(Val & SISrcMods::OP_SEL_1))
         return false;
+      // Only apply the following transformation if that operand requries
+      // a packed immediate.
+      switch (TII.get(Opcode).OpInfo[OpNo].OperandType) {
+      case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
+      case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
         // If upper part is all zero we do not need op_sel_hi.
         if (!isUInt<16>(Fold.ImmToFold)) {
           if (!(Fold.ImmToFold & 0xffff)) {
             Mod.setImm(Mod.getImm() | SISrcMods::OP_SEL_0);
             Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1);
             Old.ChangeToImmediate((Fold.ImmToFold >> 16) & 0xffff);
             return true;
           }
           Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1);
         }
+        break;
+      default:
+        break;
+      }
     }
 
     if (Fold.needsShrink()) {
       MachineBasicBlock *MBB = MI->getParent();
       auto Liveness = MBB->computeRegisterLiveness(&TRI, AMDGPU::VCC, MI);
       if (Liveness != MachineBasicBlock::LQR_Dead)
         return false;
 

llvm/trunk/test/CodeGen/AMDGPU/llvm.amdgcn.fdot2.ll

 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %c.val = load float, float addrspace(1)* %c
   %r.val = call float @llvm.amdgcn.fdot2(<2 x half> %a.val, <2 x half> %b.val, float %c.val, i1 0)
   store float %r.val, float addrspace(1)* %r
   ret void
 }
+
+; GFX906-LABEL: {{^}}fdot2_inline_literal
+; GFX906: v_dot2_f32_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, 1.0
+define float @fdot2_inline_literal(<2 x half> %a, <2 x half> %b) {
+  %ret = tail call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float 1.0, i1 false)
+  ret float %ret
+}