[llvm] f1ee458 - [ValueTracking] improve `isKnownNonZero` precision for `smax`

[Noah Goldstein via llvm-commits]

Author: Noah Goldstein
Date: 2024-04-10T10:40:49-05:00
New Revision: f1ee458ddb45c9887b3df583ce9a4ba12aae8b3b

URL: https://github.com/llvm/llvm-project/commit/f1ee458ddb45c9887b3df583ce9a4ba12aae8b3b
DIFF: https://github.com/llvm/llvm-project/commit/f1ee458ddb45c9887b3df583ce9a4ba12aae8b3b.diff

LOG: [ValueTracking] improve `isKnownNonZero` precision for `smax`

Instead of relying on known-bits for strictly positive, use the
`isKnownPositive` API. This will use `isKnownNonZero` which is more
accurate.

Closes #88170

Added: 
    

Modified: 
    llvm/lib/Analysis/ValueTracking.cpp
    llvm/test/Transforms/InstSimplify/known-non-zero.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index b32dc493ace919..5ef1969893b425 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -2870,23 +2870,44 @@ static bool isKnownNonZeroFromOperator(const Operator *I,
       case Intrinsic::uadd_sat:
         return isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q) ||
                isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q);
-      case Intrinsic::smin:
       case Intrinsic::smax: {
-        auto KnownOpImpliesNonZero = [&](const KnownBits &K) {
-          return II->getIntrinsicID() == Intrinsic::smin
-                     ? K.isNegative()
-                     : K.isStrictlyPositive();
+        // If either arg is strictly positive the result is non-zero. Otherwise
+        // the result is non-zero if both ops are non-zero.
+        auto IsNonZero = [&](Value *Op, std::optional<bool> &OpNonZero,
+                             const KnownBits &OpKnown) {
+          if (!OpNonZero.has_value())
+            OpNonZero = OpKnown.isNonZero() ||
+                        isKnownNonZero(Op, DemandedElts, Depth, Q);
+          return *OpNonZero;
         };
-        KnownBits XKnown =
+        // Avoid re-computing isKnownNonZero.
+        std::optional<bool> Op0NonZero, Op1NonZero;
+        KnownBits Op1Known =
+            computeKnownBits(II->getArgOperand(1), DemandedElts, Depth, Q);
+        if (Op1Known.isNonNegative() &&
+            IsNonZero(II->getArgOperand(1), Op1NonZero, Op1Known))
+          return true;
+        KnownBits Op0Known =
             computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q);
-        if (KnownOpImpliesNonZero(XKnown))
+        if (Op0Known.isNonNegative() &&
+            IsNonZero(II->getArgOperand(0), Op0NonZero, Op0Known))
           return true;
-        KnownBits YKnown =
+        return IsNonZero(II->getArgOperand(1), Op1NonZero, Op1Known) &&
+               IsNonZero(II->getArgOperand(0), Op0NonZero, Op0Known);
+      }
+      case Intrinsic::smin: {
+        // If either arg is negative the result is non-zero. Otherwise
+        // the result is non-zero if both ops are non-zero.
+        KnownBits Op1Known =
             computeKnownBits(II->getArgOperand(1), DemandedElts, Depth, Q);
-        if (KnownOpImpliesNonZero(YKnown))
+        if (Op1Known.isNegative())
+          return true;
+        KnownBits Op0Known =
+            computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q);
+        if (Op0Known.isNegative())
           return true;
 
-        if (XKnown.isNonZero() && YKnown.isNonZero())
+        if (Op1Known.isNonZero() && Op0Known.isNonZero())
           return true;
       }
         [[fallthrough]];

diff  --git a/llvm/test/Transforms/InstSimplify/known-non-zero.ll b/llvm/test/Transforms/InstSimplify/known-non-zero.ll
index 6ebc4e0f31a9cf..51f80f62c2f34c 100644
--- a/llvm/test/Transforms/InstSimplify/known-non-zero.ll
+++ b/llvm/test/Transforms/InstSimplify/known-non-zero.ll
@@ -169,11 +169,7 @@ B:
 
 define i1 @smax_non_zero(i8 %xx, i8 %y) {
 ; CHECK-LABEL: @smax_non_zero(
-; CHECK-NEXT:    [[X0:%.*]] = and i8 [[XX:%.*]], 63
-; CHECK-NEXT:    [[X:%.*]] = add i8 [[X0]], 1
-; CHECK-NEXT:    [[V:%.*]] = call i8 @llvm.smax.i8(i8 [[X]], i8 [[Y:%.*]])
-; CHECK-NEXT:    [[R:%.*]] = icmp eq i8 [[V]], 0
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %x0 = and i8 %xx, 63
   %x = add i8 %x0, 1