AMDGPU: Don't wait at end of block with a trivial successor

If there is only one successor, and that successor only
has one predecessor the wait can obviously be delayed until
uses or the end of the next block. This avoids code quality
regressions when there are trivial fallthrough blocks inserted
for structurization.

llvm-svn: 297251

1 files changed, 14 insertions, 2 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaits.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaits.cpp
index 3daf13cb4b5..285bc8a8144 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaits.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaits.cpp
@@ -524,6 +524,16 @@ void SIInsertWaits::handleSendMsg(MachineBasicBlock &MBB,
   }
 }
 
+/// Return true if \p MBB has one successor immediately following, and is its
+/// only predecessor
+static bool hasTrivialSuccessor(const MachineBasicBlock &MBB) {
+  if (MBB.succ_size() != 1)
+    return false;
+
+  const MachineBasicBlock *Succ = *MBB.succ_begin();
+  return (Succ->pred_size() == 1) && MBB.isLayoutSuccessor(Succ);
+}
+
 // FIXME: Insert waits listed in Table 4.2 "Required User-Inserted Wait States"
 // around other non-memory instructions.
 bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
@@ -642,8 +652,10 @@ bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
         EndPgmBlocks.push_back(&MBB);
     }
 
-    // Wait for everything at the end of the MBB
-    Changes |= insertWait(MBB, MBB.getFirstTerminator(), LastIssued);
+    // Wait for everything at the end of the MBB. If there is only one
+    // successor, we can defer this until the uses there.
+    if (!hasTrivialSuccessor(MBB))
+      Changes |= insertWait(MBB, MBB.getFirstTerminator(), LastIssued);
   }
 
   if (HaveScalarStores) {