[InstCombine] foldShiftIntoShiftInAnotherHandOfAndInICmp(): fix miscompile (PR44802)

Much like with reassociateShiftAmtsOfTwoSameDirectionShifts(),
as input, we have the following pattern:
  icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
We want to rewrite that as:
  icmp eq/ne (and (x shift (Q+K)), y), 0  iff (Q+K) u< bitwidth(x)

While we know that originally (Q+K) would not overflow
(because  2 * (N-1) u<= iN -1), we may have looked past extensions of
shift amounts. so it may now overflow in smaller bitwidth.

To ensure that does not happen, we need to ensure that the total maximal
shift amount is still representable in that smaller bitwidth.
If the overflow would happen, (Q+K) u< bitwidth(x) check would be bogus.

https://bugs.llvm.org/show_bug.cgi?id=44802
(cherry picked from commit 2855c8f)

Author: LebedevRI

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

@@ -3494,7 +3494,8 @@ foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
   Instruction *NarrowestShift = XShift;
 
   Type *WidestTy = WidestShift->getType();
-  assert(NarrowestShift->getType() == I.getOperand(0)->getType() &&
+  Type *NarrowestTy = NarrowestShift->getType();
+  assert(NarrowestTy == I.getOperand(0)->getType() &&
          "We did not look past any shifts while matching XShift though.");
   bool HadTrunc = WidestTy != I.getOperand(0)->getType();
 
@@ -3533,6 +3534,23 @@ foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
   if (XShAmt->getType() != YShAmt->getType())
     return nullptr;
 
+  // As input, we have the following pattern:
+  //   icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
+  // We want to rewrite that as:
+  //   icmp eq/ne (and (x shift (Q+K)), y), 0  iff (Q+K) u< bitwidth(x)
+  // While we know that originally (Q+K) would not overflow
+  // (because  2 * (N-1) u<= iN -1), we have looked past extensions of
+  // shift amounts. so it may now overflow in smaller bitwidth.
+  // To ensure that does not happen, we need to ensure that the total maximal
+  // shift amount is still representable in that smaller bit width.
+  unsigned MaximalPossibleTotalShiftAmount =
+      (WidestTy->getScalarSizeInBits() - 1) +
+      (NarrowestTy->getScalarSizeInBits() - 1);
+  APInt MaximalRepresentableShiftAmount =
+      APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
+  if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
+    return nullptr;
+
   // Can we fold (XShAmt+YShAmt) ?
   auto *NewShAmt = dyn_cast_or_null<Constant>(
       SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,

llvm/test/Transforms/InstCombine/shift-amount-reassociation-in-bittest.ll

@@ -688,13 +688,16 @@ entry:
   ret i1 %tobool
 }
 
-; FIXME: this is a miscompile. We should not transform this.
 ; See https://bugs.llvm.org/show_bug.cgi?id=44802
 define i1 @pr44802(i3 %a, i3 %x, i3 %y) {
 ; CHECK-LABEL: @pr44802(
-; CHECK-NEXT:    [[TMP1:%.*]] = and i3 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne i3 [[TMP1]], 0
-; CHECK-NEXT:    ret i1 [[TMP2]]
+; CHECK-NEXT:    [[T0:%.*]] = icmp ne i3 [[A:%.*]], 0
+; CHECK-NEXT:    [[T1:%.*]] = zext i1 [[T0]] to i3
+; CHECK-NEXT:    [[T2:%.*]] = lshr i3 [[X:%.*]], [[T1]]
+; CHECK-NEXT:    [[T3:%.*]] = shl i3 [[Y:%.*]], [[T1]]
+; CHECK-NEXT:    [[T4:%.*]] = and i3 [[T2]], [[T3]]
+; CHECK-NEXT:    [[T5:%.*]] = icmp ne i3 [[T4]], 0
+; CHECK-NEXT:    ret i1 [[T5]]
 ;
   %t0 = icmp ne i3 %a, 0
   %t1 = zext i1 %t0 to i3