[llvm] [AMDGPU] Generate more efficient code to avoid shift64 bug (PR #171871)

[via llvm-commits]

================
@@ -2217,16 +2217,34 @@ bool GCNHazardRecognizer::fixShift64HighRegBug(MachineInstr *MI) {
   if (AmtReg != AMDGPU::VGPR255 && MRI.isPhysRegUsed(AmtReg + 1))
     return false;
 
-  MachineOperand *Src1 = TII.getNamedOperand(*MI, AMDGPU::OpName::src1);
-  bool OverlappedSrc = Src1->isReg() && TRI.regsOverlap(Src1->getReg(), AmtReg);
-  bool OverlappedDst = MI->modifiesRegister(AmtReg, &TRI);
-  bool Overlapped = OverlappedSrc || OverlappedDst;
-
-  assert(!OverlappedDst || !OverlappedSrc ||
-         Src1->getReg() == MI->getOperand(0).getReg());
   assert(ST.needsAlignedVGPRs());
   static_assert(AMDGPU::VGPR0 + 1 == AMDGPU::VGPR1);
 
+  const DebugLoc DL = MI->getDebugLoc();
+  MachineBasicBlock *MBB = MI->getParent();
+  MachineOperand *Src1 = TII.getNamedOperand(*MI, AMDGPU::OpName::src1);
+
+  // In:
+  //
+  //    Dst = shiftrev64 Amt, Src1
+  //
+  // if  Dst!=Src1 then avoid the bug with:
+  //
+  //    Dst.sub0 = Amt
+  //    Dst = shift64 Dst.sub0, Src1
+
+  Register DstReg = MI->getOperand(0).getReg();
+  if (!Src1->isReg() || Src1->getReg() != DstReg) {
+    Register DstLo = TRI.getSubReg(DstReg, AMDGPU::sub0);
+    runOnInstruction(
+        BuildMI(*MBB, MI, DL, TII.get(AMDGPU::V_MOV_B32_e32), DstLo)
+            .addReg(AmtReg, Amt->isKill() ? RegState::Kill : 0));
----------------
LU-JOHN wrote:

Changed.

https://github.com/llvm/llvm-project/pull/171871