AMDGPU/SI: Optimize adjacent s_nop instructions

Use the operand for how long to wait. This is somewhat
distasteful, since it would be better to just emit s_nop
with the right argument in the first place. This would require
changing TII::insertNoop to emit N operands, which would be easy.
Slightly more problematic is the post-RA scheduler and hazard recognizer
represent nops as a single null node, and would require inventing
another way of representing N nops.

llvm-svn: 267456

1 files changed, 27 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
index ad3f63d2cea..346488d38e2 100644
--- a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
+++ b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
@@ -242,6 +242,33 @@ bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
         }
       }
 
+      // Combine adjacent s_nops to use the immediate operand encoding how long
+      // to wait.
+      //
+      // s_nop N
+      // s_nop M
+      //  =>
+      // s_nop (N + M)
+      if (MI.getOpcode() == AMDGPU::S_NOP &&
+          Next != MBB.end() &&
+          (*Next).getOpcode() == AMDGPU::S_NOP) {
+
+        MachineInstr &NextMI = *Next;
+        // The instruction encodes the amount to wait with an offset of 1,
+        // i.e. 0 is wait 1 cycle. Convert both to cycles and then convert back
+        // after adding.
+        uint8_t Nop0 = MI.getOperand(0).getImm() + 1;
+        uint8_t Nop1 = NextMI.getOperand(0).getImm() + 1;
+
+        // Make sure we don't overflow the bounds.
+        if (Nop0 + Nop1 <= 8) {
+          NextMI.getOperand(0).setImm(Nop0 + Nop1 - 1);
+          MI.eraseFromParent();
+        }
+
+        continue;
+      }
+
       // FIXME: We also need to consider movs of constant operands since
       // immediate operands are not folded if they have more than one use, and
       // the operand folding pass is unaware if the immediate will be free since