[llvm] [AMDGPU] Generate more swaps (PR #184164)

[Matt Arsenault via llvm-commits]

================
@@ -724,73 +726,79 @@ MachineInstr *SIShrinkInstructions::matchSwap(MachineInstr &MovT) const {
 
   const unsigned SearchLimit = 16;
   unsigned Count = 0;
-  bool KilledT = false;
+
+  MachineInstr *MovX = nullptr;
+  MachineInstr *InsertionPt = nullptr;
+  MachineInstr *MovY = nullptr;
+
   for (auto Iter = std::next(MovT.getIterator()),
             E = MovT.getParent()->instr_end();
-       Iter != E && Count < SearchLimit && !KilledT; ++Iter) {
-
-    MachineInstr *MovY = &*Iter;
-    KilledT = MovY->killsRegister(T, TRI);
-    if (MovY->isDebugInstr())
+       Iter != E && Count < SearchLimit; ++Iter) {
+    if (Iter->isDebugInstr())
       continue;
     ++Count;
 
-    if ((MovY->getOpcode() != AMDGPU::V_MOV_B32_e32 &&
-         MovY->getOpcode() != AMDGPU::V_MOV_B16_t16_e32 &&
-         MovY->getOpcode() != AMDGPU::COPY) ||
-        !MovY->getOperand(1).isReg() || MovY->getOperand(1).getReg() != T ||
-        MovY->getOperand(1).getSubReg() != Tsub)
-      continue;
-
-    Register Y = MovY->getOperand(0).getReg();
-    unsigned Ysub = MovY->getOperand(0).getSubReg();
-
-    if (!TRI->isVGPR(*MRI, Y))
-      continue;
-
-    MachineInstr *MovX = nullptr;
-    for (auto IY = MovY->getIterator(), I = std::next(MovT.getIterator());
-         I != IY; ++I) {
-      if (I->isDebugInstr())
-        continue;
-      if (instReadsReg(&*I, X, Xsub) || instModifiesReg(&*I, Y, Ysub) ||
-          instModifiesReg(&*I, T, Tsub) ||
-          (MovX && instModifiesReg(&*I, X, Xsub))) {
-        MovX = nullptr;
-        break;
-      }
-      if (!instReadsReg(&*I, Y, Ysub)) {
-        if (!MovX && instModifiesReg(&*I, X, Xsub)) {
-          MovX = nullptr;
-          break;
-        }
-        continue;
+    if (instModifiesReg(&*Iter, T, Tsub))
+      return nullptr;
+
+    if (!MovX) {
+      // Search for mov x, y.
+      if ((Iter->getOpcode() == AMDGPU::V_MOV_B32_e32 ||
+           Iter->getOpcode() == AMDGPU::V_MOV_B16_t16_e32 ||
+           Iter->getOpcode() == AMDGPU::COPY) &&
+          Iter->getOperand(0).getReg() == X &&
+          Iter->getOperand(0).getSubReg() == Xsub &&
+          Iter->getOperand(1).isReg()) {
+        MovX = &*Iter;
+        Y = MovX->getOperand(1).getReg();
+        Ysub = MovX->getOperand(1).getSubReg();
       }
-      if (MovX ||
-          (I->getOpcode() != AMDGPU::V_MOV_B32_e32 &&
-           I->getOpcode() != AMDGPU::V_MOV_B16_t16_e32 &&
-           I->getOpcode() != AMDGPU::COPY) ||
-          I->getOperand(0).getReg() != X ||
-          I->getOperand(0).getSubReg() != Xsub) {
-        MovX = nullptr;
+    } else {
+      // mov x, y has been found.
+      // Search for mov y, t.
+      if ((Iter->getOpcode() == AMDGPU::V_MOV_B32_e32 ||
+           Iter->getOpcode() == AMDGPU::V_MOV_B16_t16_e32 ||
+           Iter->getOpcode() == AMDGPU::COPY) &&
+          Iter->getOperand(0).getReg() == Y &&
+          Iter->getOperand(0).getSubReg() == Ysub &&
+          Iter->getOperand(1).isReg() && Iter->getOperand(1).getReg() == T &&
+          Iter->getOperand(1).getSubReg() == Tsub) {
+        MovY = &*Iter;
         break;
       }
 
-      if (Size > 4 && (I->getNumImplicitOperands() > (I->isCopy() ? 0U : 1U)))
-        continue;
-
-      MovX = &*I;
+      // Effectively, mov x, y must be moved downward
+      // and mov y, t must be moved upward so that they can be fused into a
+      // swap. A write to y creates a barrier that prevents the two moves from
+      // being moved adjacent to each other.
+      if (instModifiesReg(&*Iter, Y, Ysub))
+        return nullptr;
+
+      // Reads or writes to x prevent mov x, y from being moved farther
+      // downward. Select this to be the insertion point.
+      if (!InsertionPt &&
+          (instReadsReg(&*Iter, X, Xsub) || instModifiesReg(&*Iter, X, Xsub))) {
----------------
arsenm wrote:

The MachineInstr reads/writes APIs are a mess. There's a readsWritesVirtualRegister, but why only virtual register? 

https://github.com/llvm/llvm-project/pull/184164