[llvm] [AMDGPU] Insert before and after instructions that always use GDS (PR #131338)

[Jay Foad via llvm-commits]

================
@@ -1982,6 +2003,65 @@ static bool isCacheInvOrWBInst(MachineInstr &Inst) {
          Opc == AMDGPU::GLOBAL_WBINV;
 }
 
+// Return true if the next instruction is S_ENDPGM, following fallthrough
+// blocks if necessary.
+bool SIInsertWaitcnts::isNextENDPGM(MachineBasicBlock::instr_iterator It,
+                                    MachineBasicBlock *Block) const {
+  auto E = Block->instr_end();
+
+  while (true) {
+    if (It == E) {
+      if (auto FallThrough = Block->getFallThrough(false)) {
----------------
jayfoad wrote:

On a related note, `getFallThrough` is far more powerful than we need here. We could just use `std::next`. But maybe that would be less readable.

https://github.com/llvm/llvm-project/pull/131338