[RISCV][InsertVSETVLI] Handle partially transparent instructions in PRE

The PRE implementation was being overly strict when checking to see if a vsetvli was removed in the current block. For instructions which don't use all the fields of VTYPE or VL, we can propagate a changed state past the first instruction with an SEW operand and remove a vsetvli later in the block. We do need to be careful now to ensure that the state convergences before the end of the block or we'd invalidate the cached data flow results.

Taking a step back, we're modeling the effect of the emitVSETVLIs pass which runs just after PRE. This is unfortunate, and makes me think we should probably reevaluate doing the PRE as a post-pass instead of as surgery in the data flow phases. Doing that requires us to get more aggressive about mutating user written vsetvlis which we've tried not to do up to now, but well, maybe it's time? Anyways, that's a thought for the future, not something I'm proposing doing now.

Differential Revision: https://reviews.llvm.org/D142409

Author: preames

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

@@ -1188,27 +1188,40 @@ void RISCVInsertVSETVLI::doPRE(MachineBasicBlock &MBB) {
   if (!hasFixedResult(AvailableInfo, ST))
     return;
 
-  // Does it actually let us remove an implicit transition in MBB?
-  bool Found = false;
-  for (auto &MI : MBB) {
-    if (isVectorConfigInstr(MI))
-      return;
-
-    const uint64_t TSFlags = MI.getDesc().TSFlags;
-    if (RISCVII::hasSEWOp(TSFlags)) {
-      if (AvailableInfo != computeInfoForInstr(MI, TSFlags, MRI))
-        return;
-      Found = true;
+  // Model the effect of changing the input state of the block MBB to
+  // AvailableInfo.  We're looking for two issues here; one legality,
+  // one profitability.
+  // 1) If the block doesn't use some of the fields from VL or VTYPE, we
+  //    may hit the end of the block with a different end state.  We can
+  //    not make this change without reflowing later blocks as well.
+  // 2) If we don't actually remove a transition, inserting a vsetvli
+  //    into the predecessor block would be correct, but unprofitable.
+  VSETVLIInfo OldInfo = BlockInfo[MBB.getNumber()].Pred;
+  VSETVLIInfo CurInfo = AvailableInfo;
+  int TransitionsRemoved = 0;
+  for (const MachineInstr &MI : MBB) {
+    const VSETVLIInfo LastInfo = CurInfo;
+    const VSETVLIInfo LastOldInfo = OldInfo;
+    transferBefore(CurInfo, MI);
+    transferBefore(OldInfo, MI);
+    if (CurInfo == LastInfo)
+      TransitionsRemoved++;
+    if (LastOldInfo == OldInfo)
+      TransitionsRemoved--;
+    transferAfter(CurInfo, MI);
+    transferAfter(OldInfo, MI);
+    if (CurInfo == OldInfo)
+      // Convergence.  All transitions after this must match by construction.
       break;
-    }
   }
-  if (!Found)
+  if (CurInfo != OldInfo || TransitionsRemoved <= 0)
+    // Issues 1 and 2 above
     return;
 
   // Finally, update both data flow state and insert the actual vsetvli.
   // Doing both keeps the code in sync with the dataflow results, which
   // is critical for correctness of phase 3.
-  auto OldInfo = BlockInfo[UnavailablePred->getNumber()].Exit;
+  auto OldExit = BlockInfo[UnavailablePred->getNumber()].Exit;
   LLVM_DEBUG(dbgs() << "PRE VSETVLI from " << MBB.getName() << " to "
                     << UnavailablePred->getName() << " with state "
                     << AvailableInfo << "\n");
@@ -1220,7 +1233,7 @@ void RISCVInsertVSETVLI::doPRE(MachineBasicBlock &MBB) {
   auto InsertPt = UnavailablePred->getFirstInstrTerminator();
   insertVSETVLI(*UnavailablePred, InsertPt,
                 UnavailablePred->findDebugLoc(InsertPt),
-                AvailableInfo, OldInfo);
+                AvailableInfo, OldExit);
 }
 
 static void doUnion(DemandedFields &A, DemandedFields B) {

llvm/test/CodeGen/RISCV/rvv/vsetvli-insert-crossbb.ll

@@ -954,16 +954,15 @@ if.end:
   ret <vscale x 2 x i32> %e
 }
 
-; This case demonstrates a PRE oppurtunity where the first instruction
-; in the block doesn't require a state transition.  Essentially, we need
-; to FRE the transition to the start of the block, and *then* PRE it.
+; This case demonstrates a PRE case where the first instruction in the block
+; doesn't require a state transition.
 define void @pre_over_vle(ptr %A) {
 ; CHECK-LABEL: pre_over_vle:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    li a1, 100
+; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
 ; CHECK-NEXT:  .LBB22_1: # %vector.body
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
 ; CHECK-NEXT:    vle8.v v8, (a0)
 ; CHECK-NEXT:    vsext.vf4 v9, v8
 ; CHECK-NEXT:    vse32.v v9, (a0)