[AMDGPU] Fix wait counts in the presence of 16bit subregisters

[rampitec]

LGTM

[foad]

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

508

I think there's a DivCeil function in MathExtras that you could use here.

[rampitec]

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

508

Please don't. This function does not work. I had to ditch it the other day specifically because of 16 vs 32 bits.

[foad]

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

508

This function does not work.

???

[rampitec]

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

508

Found it: D78772
It was: "divideCeil(getSubRegIdxOffset(SubReg), 32)" changed to :(getSubRegIdxOffset(SubReg) + 31) / 32".
The problem was with offset == 16. If we operate just size this is probably OK.

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

RegInterval WaitcntBrackets::getRegInterval(const MachineInstr *MI,
   }
   // TODO: Handle TTMP
   // else if (TRI->isTTMP(*MRI, Reg.getReg())) ...
   else
     return {-1, -1};
 
   const TargetRegisterClass *RC = TII->getOpRegClass(*MI, OpNo);
   unsigned Size = TRI->getRegSizeInBits(*RC);
-  Result.second = Result.first + (Size / 32);
+  Result.second = Result.first + ((Size + 16) / 32);
 
   return Result;
 }
 
 void WaitcntBrackets::setExpScore(const MachineInstr *MI,
                                   const SIInstrInfo *TII,
                                   const SIRegisterInfo *TRI,
                                   const MachineRegisterInfo *MRI, unsigned OpNo,

llvm/test/CodeGen/AMDGPU/waitcnt.mir

--- |
   define amdgpu_kernel void @exit_bundle() {
     ret void
   }
 
   define amdgpu_kernel void @cross_bundle() {
     ret void
   }
 
+  define amdgpu_kernel void @subregs16bit() {
+    ret void
+  }
 ...
 ---
 
 # CHECK-LABEL: name: flat_zero_waitcnt
 
 # CHECK-LABEL: bb.0:
 # CHECK: FLAT_LOAD_DWORD
 # CHECK: FLAT_LOAD_DWORDX4
bb.0:
     liveins: $vgpr1_vgpr2
     BUNDLE implicit-def $vgpr0, implicit $vgpr1_vgpr2 {
     $vgpr0 = FLAT_LOAD_DWORD $vgpr1_vgpr2, 0, 0, 0, 0, implicit $exec, implicit $flat_scr
     }
     BUNDLE implicit $vgpr0, implicit $vgpr1_vgpr2 {
       FLAT_STORE_DWORD $vgpr1_vgpr2, $vgpr0, 0, 0, 0, 0, implicit $exec, implicit $flat_scr
     }
 ...
+
+---
+# CHECK-LABEL: name: subregs16bit
+# CHECK: S_WAITCNT 112
+# CHECK-NEXT: V_NOP_e32 
+
+name: subregs16bit
+machineFunctionInfo:
+  isEntryFunction: true
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4
+      $vgpr0 = FLAT_LOAD_USHORT killed $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec, implicit $flat_scr
+      $vgpr1 = FLAT_LOAD_USHORT killed $vgpr2_vgpr3, 0, 0, 0, 0, implicit $exec, implicit $flat_scr
+      V_NOP_e32 implicit $exec, implicit $vgpr0_lo16, implicit $vgpr1_lo16
+...