[llvm] AMDGPU: Refactor lowering of s_barrier to split barriers (PR #154648)

[Jay Foad via llvm-commits]

================
@@ -2007,18 +2007,6 @@ bool AMDGPUInstructionSelector::selectSBarrier(MachineInstr &MI) const {
     }
   }
 
-  if (STI.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {
-    // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
-    MachineBasicBlock *MBB = MI.getParent();
-    const DebugLoc &DL = MI.getDebugLoc();
-    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_BARRIER_SIGNAL_IMM))
-        .addImm(AMDGPU::Barrier::WORKGROUP);
-    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_BARRIER_WAIT))
-        .addImm(AMDGPU::Barrier::WORKGROUP);
-    MI.eraseFromParent();
----------------
jayfoad wrote:

PreISelIntrinsicLowering doesn't handle target-specific intrinsics. So is it better to teach that pass how to call into target hooks to handle target intrinsics, or to have a target-specific pass like in this patch?

https://github.com/llvm/llvm-project/pull/154648