AMDGPU/SI: Remove ending s_endpgm from non-void functions

[• tstellarAMD]

LGTM.

[arsenm]

Shouldn't this just not be emitting in the first place?

[mareko]

In D16035#324101, @arsenm wrote:

Shouldn't this just not be emitting in the first place?

RET_FLAG is translated to S_ENDPGM. The answer to your question is the same as the answer to this: Is it possible to have RET_FLAG that is translated to 0 instructions (some kind of glue), yet those 0 instructions have an arbitrary number of (ret) operands?

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

for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
         Changes |= insertWait(MBB, I, handleOperands(*I));
 
       pushInstruction(MBB, I);
       handleSendMsg(MBB, I);
     }
 
     // Wait for everything at the end of the MBB
     Changes |= insertWait(MBB, MBB.getFirstTerminator(), LastIssued);
+
+    // Functions returning something shouldn't contain S_ENDPGM, because other
+    // bytecode will be appended after it.
+    if (!ReturnsVoid) {
+      MachineBasicBlock::iterator I = MBB.getFirstTerminator();
+      if (I != MBB.end() && I->getOpcode() == AMDGPU::S_ENDPGM)
+        I->eraseFromParent();
+    }
   }
 
   return Changes;
 }