[PATCH] D31161: [AMDGPU] New Waitcnt Insertion Pass

[Konstantin Zhuravlyov via Phabricator via llvm-commits]

kzhuravl requested changes to this revision.
kzhuravl added a comment.
This revision now requires changes to proceed.

- Your pass does not take into account encoding changes for gfx9.
- New pass name is confusing.
- Make a note somewhere that the new pass is on by default, and you can switch it off by such and such option.



================
Comment at: lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1069-1071
+      if (!OldWaitcnt || ((Imm & 0xF) != (CntVal[VM_CNT] & 0xF)) ||
+          (((Imm >> 4) & 0x7) != (CntVal[EXP_CNT] & 0x7)) ||
+          (((Imm >> 8) & 0xF) != (CntVal[LGKM_CNT] & 0xF))) {
----------------
Use helper functions from AMDGPUBaseInfo.h (they also have logic for gfx9 in it)


================
Comment at: lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1098-1099
+
+      int64_t Val = (CntVal[VM_CNT] & 0xF) | ((CntVal[EXP_CNT] & 0x7) << 4) |
+                    ((CntVal[LGKM_CNT] & 0xF) << 8);
+      const MachineOperand &Op = MachineOperand::CreateImm(Val);
----------------
Use helper functions from AMDGPUBaseInfo.h (they also have logic for gfx9 in it)


https://reviews.llvm.org/D31161