Reapply commit 143028 with a fix: the problem was casting a ConstantExpr Mul

using BinaryOperator (which only works for instructions) when it should have
been a cast to OverflowingBinaryOperator (which also works for constants).
While there, correct a few other dubious looking uses of BinaryOperator.
Thanks to Chad Rosier for the testcase.  Original commit message:
My super-optimizer noticed that we weren't folding this expression to
true: (x *nsw x) sgt 0, where x = (y | 1).  This occurs in 464.h264ref.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@143125 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 49 insertions, 7 deletions

diff --git a/lib/Analysis/ValueTracking.cpp b/lib/Analysis/ValueTracking.cpp
index 9a234c068b..90757f9798 100644
--- a/lib/Analysis/ValueTracking.cpp
+++ b/lib/Analysis/ValueTracking.cpp
@@ -201,9 +201,36 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
     ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
     ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
                       Depth+1);
-    assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 
-    assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 
-    
+    assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+    assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+    bool isKnownNegative = false;
+    bool isKnownNonNegative = false;
+    // If the multiplication is known not to overflow, compute the sign bit.
+    if (Mask.isNegative() &&
+        cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
+      Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
+      if (Op1 == Op2) {
+        // The product of a number with itself is non-negative.
+        isKnownNonNegative = true;
+      } else {
+        bool isKnownNonNegative1 = KnownZero.isNegative();
+        bool isKnownNonNegative2 = KnownZero2.isNegative();
+        bool isKnownNegative1 = KnownOne.isNegative();
+        bool isKnownNegative2 = KnownOne2.isNegative();
+        // The product of two numbers with the same sign is non-negative.
+        isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
+          (isKnownNonNegative1 && isKnownNonNegative2);
+        // The product of a negative number and a non-negative number is either
+        // negative or zero.
+        if (!isKnownNonNegative)
+          isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
+                             isKnownNonZero(Op2, TD, Depth)) ||
+                            (isKnownNegative2 && isKnownNonNegative1 &&
+                             isKnownNonZero(Op1, TD, Depth));
+      }
+    }
+
     // If low bits are zero in either operand, output low known-0 bits.
     // Also compute a conserative estimate for high known-0 bits.
     // More trickiness is possible, but this is sufficient for the
@@ -220,6 +247,12 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
                 APInt::getHighBitsSet(BitWidth, LeadZ);
     KnownZero &= Mask;
+
+    if (isKnownNonNegative)
+      KnownZero.setBit(BitWidth - 1);
+    else if (isKnownNegative)
+      KnownOne.setBit(BitWidth - 1);
+
     return;
   }
   case Instruction::UDiv: {
@@ -784,7 +817,7 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
   }
 
   // The remaining tests are all recursive, so bail out if we hit the limit.
-  if (Depth++ == MaxDepth)
+  if (Depth++ >= MaxDepth)
     return false;
 
   unsigned BitWidth = getBitWidth(V->getType(), TD);
@@ -802,7 +835,7 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
   // if the lowest bit is shifted off the end.
   if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
     // shl nuw can't remove any non-zero bits.
-    BinaryOperator *BO = cast<BinaryOperator>(V);
+    OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
     if (BO->hasNoUnsignedWrap())
       return isKnownNonZero(X, TD, Depth);
 
@@ -816,7 +849,7 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
   // defined if the sign bit is shifted off the end.
   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
     // shr exact can only shift out zero bits.
-    BinaryOperator *BO = cast<BinaryOperator>(V);
+    PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
     if (BO->isExact())
       return isKnownNonZero(X, TD, Depth);
 
@@ -827,7 +860,7 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
   }
   // div exact can only produce a zero if the dividend is zero.
   else if (match(V, m_IDiv(m_Value(X), m_Value()))) {
-    BinaryOperator *BO = cast<BinaryOperator>(V);
+    PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
     if (BO->isExact())
       return isKnownNonZero(X, TD, Depth);
   }
@@ -868,6 +901,15 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
     if (YKnownNonNegative && isPowerOfTwo(X, TD, /*OrZero*/false, Depth))
       return true;
   }
+  // X * Y.
+  else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
+    OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
+    // If X and Y are non-zero then so is X * Y as long as the multiplication
+    // does not overflow.
+    if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
+        isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
+      return true;
+  }
   // (C ? X : Y) != 0 if X != 0 and Y != 0.
   else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
     if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&