[InstSimplify] Ignore mask when combinining vp.reverse(vp.reverse). (#171542)

The mask doesn't really affect the reverse. It only poisons the masked
off elements in the results. It should be ok to ignore the mask if we
can eliminate the pair.

I don't have a specific use case for this, but it matches what I had
implemented in our downstream before the current upstream
implementation. Submitting upstream so I can remove the delta
in my downstream.

Author: topperc

llvm/lib/Analysis/InstructionSimplify.cpp

@@ -7302,14 +7302,12 @@ static Value *simplifyIntrinsic(CallBase *Call, Value *Callee,
   }
   case Intrinsic::experimental_vp_reverse: {
     Value *Vec = Call->getArgOperand(0);
-    Value *Mask = Call->getArgOperand(1);
     Value *EVL = Call->getArgOperand(2);
 
     Value *X;
-    // vp.reverse(vp.reverse(X)) == X (with all ones mask and matching EVL)
-    if (match(Mask, m_AllOnes()) &&
-        match(Vec, m_Intrinsic<Intrinsic::experimental_vp_reverse>(
-                       m_Value(X), m_AllOnes(), m_Specific(EVL))))
+    // vp.reverse(vp.reverse(X)) == X (mask doesn't matter)
+    if (match(Vec, m_Intrinsic<Intrinsic::experimental_vp_reverse>(
+                       m_Value(X), m_Value(), m_Specific(EVL))))
       return X;
 
     // vp.reverse(splat(X)) -> splat(X) (regardless of mask and EVL)

llvm/test/Transforms/InstSimplify/vp-reverse.ll

@@ -68,3 +68,12 @@ define <vscale x 4 x i32> @rev_of_splat2(i32 %a, <vscale x 4 x i1> %m, i32 %evl)
   %rev = tail call <vscale x 4 x i32> @llvm.experimental.vp.reverse(<vscale x 4 x i32> %a.vec, <vscale x 4 x i1> %m, i32 %evl)
   ret <vscale x 4 x i32> %rev
 }
+
+define <vscale x 4 x i32> @rev_of_rev_diffmask(<vscale x 4 x i32> %a, <vscale x 4 x i1> %mask1, <vscale x 4 x i1> %mask2, i32 %evl) {
+; CHECK-LABEL: @rev_of_rev_diffmask(
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[RES:%.*]]
+;
+  %a.rev = tail call <vscale x 4 x i32> @llvm.experimental.vp.reverse(<vscale x 4 x i32> %a, <vscale x 4 x i1> %mask1, i32 %evl)
+  %res = tail call <vscale x 4 x i32> @llvm.experimental.vp.reverse(<vscale x 4 x i32> %a.rev, <vscale x 4 x i1> %mask2, i32 %evl)
+  ret <vscale x 4 x i32> %res
+}