[ValueTracking] Add logic for `isKnownNonZero(sadd.sat X, Y)`

[goldstein.w.n]

Summary

The logic here is the same for add so reuse the existing helper
isNonZeroAdd

Alive2 Link:

https://alive2.llvm.org/ce/z/mhKvC5

[StephenFan]

LGTM. But wait for @nikic.

[goldstein.w.n]

use new isNonZeroAdd helper and return unconditionally

[nikic]

llvm/lib/Analysis/ValueTracking.cpp

2948

I think you can set NSW=true here, based on the following reasoning: If the addition were to overflow, then the result would be either SignedMin or SignedMax, neither of which is zero. As such, we can ignore overflow cases.

[goldstein.w.n]

llvm/lib/Analysis/ValueTracking.cpp

2948

I don't think we have NSW I tried:
https://alive2.llvm.org/ce/z/ZBJP5e to see if sadd.sat was a superset and think we might run into some edge cases.
https://alive2.llvm.org/ce/z/9Hz7sn OTOH verifies.

[goldstein.w.n]

Rebase

[nikic]

llvm/lib/Analysis/ValueTracking.cpp

2948

Here's a variant of your proof that works: https://alive2.llvm.org/ce/z/qFwKbT You need to use logical or in this case.

Alternatively, this one also works: https://alive2.llvm.org/ce/z/6px6Ss

We have to exclude undef due to use count increases, but that's just an artifact of the proof.

[goldstein.w.n]

Add NSW flag is isNonZeroAdd

[nikic]

LGTM

llvm/lib/Analysis/ValueTracking.cpp

if (auto *II = dyn_cast<IntrinsicInst>(I)) {
             computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q);
         if (XKnown.isUnknown())
           return false;
         KnownBits YKnown =
             computeKnownBits(II->getArgOperand(1), DemandedElts, Depth, Q);
         std::optional<bool> ugt = KnownBits::ugt(XKnown, YKnown);
         return ugt && *ugt;
       }
+      case Intrinsic::sadd_sat:
+        return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth,
+                            II->getArgOperand(0), II->getArgOperand(1),
+                            /*NSW*/ false);
       case Intrinsic::uadd_sat:
         if (isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q) ||
             isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q))
           return true;
         break;
       case Intrinsic::cttz:
         return computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q)
             .Zero[0];

llvm/test/Analysis/ValueTracking/known-non-zero.ll

 }
 
 define i1 @sadd_sat_nonzero(i8 %xx, i8 %yy, i8 %ind) {
 ; CHECK-LABEL: @sadd_sat_nonzero(
 ; CHECK-NEXT:    [[X_STRICT_POS:%.*]] = icmp sgt i8 [[XX:%.*]], 0
 ; CHECK-NEXT:    [[Y_POS:%.*]] = icmp sge i8 [[YY:%.*]], 0
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[X_STRICT_POS]])
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[Y_POS]])
-; CHECK-NEXT:    [[X:%.*]] = call i8 @llvm.sadd.sat.i8(i8 [[XX]], i8 [[YY]])
-; CHECK-NEXT:    [[Z:%.*]] = or i8 [[X]], [[IND:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = icmp eq i8 [[Z]], 0
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %x_strict_pos = icmp sgt i8 %xx, 0
   %y_pos = icmp sge i8 %yy, 0
   call void @llvm.assume(i1 %x_strict_pos)
   call void @llvm.assume(i1 %y_pos)
   %x = call i8 @llvm.sadd.sat.i8(i8 %xx, i8 %yy)
   %z = or i8 %x, %ind
   %r = icmp eq i8 %z, 0