[SelectionDAG]: Add more cases for UDIV, SDIV, SRA, and SRL

Author: AZero13

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -5443,19 +5443,59 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
     if (ValKnown.isNegative())
       return true;
     // If max shift cnt of known ones is non-zero, result is non-zero.
-    APInt MaxCnt = computeKnownBits(Op.getOperand(1), Depth + 1).getMaxValue();
+    const KnownBits Shift = computeKnownBits(Op.getOperand(1), Depth + 1);
+    APInt MaxCnt = Shift.getMaxValue();
     if (MaxCnt.ult(ValKnown.getBitWidth()) &&
         !ValKnown.One.lshr(MaxCnt).isZero())
       return true;
+    // Similar to udiv but we try to see if we can turn it into a division
+    const KnownBits One =
+        KnownBits::makeConstant(APInt(ValKnown.getBitWidth(), 1));
+
+    std::optional<bool> uge =
+        KnownBits::uge(ValKnown, KnownBits::shl(One, Shift));
+    if (uge && *uge)
+      return true;
     break;
   }
-  case ISD::UDIV:
-  case ISD::SDIV:
+  case ISD::UDIV: {
+    if (Op->getFlags().hasExact())
+      return isKnownNeverZero(Op.getOperand(0), Depth + 1);
+    KnownBits Op0 = computeKnownBits(Op.getOperand(0), Depth + 1);
+    KnownBits Op1 = computeKnownBits(Op.getOperand(1), Depth + 1);
+    // True if Op0 u>= Op1
+
+    std::optional<bool> uge = KnownBits::uge(Op0, Op1);
+    if (uge && *uge)
+      return true;
+    break;
+  }
+  case ISD::SDIV: {
     // div exact can only produce a zero if the dividend is zero.
-    // TODO: For udiv this is also true if Op1 u<= Op0
     if (Op->getFlags().hasExact())
       return isKnownNeverZero(Op.getOperand(0), Depth + 1);
+    KnownBits Op0 = computeKnownBits(Op.getOperand(0), Depth + 1);
+    KnownBits Op1 = computeKnownBits(Op.getOperand(1), Depth + 1);
+    if (Op0.isNegative() && Op1.isStrictlyPositive())
+      return true;
+
+    if (Op0.isStrictlyPositive() && Op1.isNegative())
+      return true;
+
+    // For negative numbers, the comparison is reversed. Op0 <= Op1
+    if (Op0.isNegative() && Op1.isNegative()) {
+      std::optional<bool> sle = KnownBits::sle(Op0, Op1);
+      if (sle && *sle)
+        return true;
+    }
+
+    if (Op0.isStrictlyPositive() && Op1.isStrictlyPositive()) {
+      std::optional<bool> uge = KnownBits::uge(Op0, Op1);
+      if (uge && *uge)
+        return true;
+    }
     break;
+  }
 
   case ISD::ADD:
     if (Op->getFlags().hasNoUnsignedWrap())

llvm/lib/Support/KnownBits.cpp

@@ -1009,6 +1009,9 @@ KnownBits KnownBits::udiv(const KnownBits &LHS, const KnownBits &RHS,
 
   Known.Zero.setHighBits(LeadZ);
   Known = divComputeLowBit(Known, LHS, RHS, Exact);
+  std::optional<bool> uge = KnownBits::uge(LHS, RHS);
+  if (uge && *uge)
+    Known.makeGE(APInt(BitWidth, 1));
 
   assert(!Known.hasConflict() && "Bad Output");
   return Known;

llvm/test/CodeGen/X86/known-pow2.ll

@@ -118,12 +118,13 @@ define i1 @pow2_srl_fail0(i32 %x, i32 %y) {
 ; CHECK-LABEL: pow2_srl_fail0:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movl %esi, %ecx
+; CHECK-NEXT:    movl %edi, %eax
 ; CHECK-NEXT:    andb $30, %cl
-; CHECK-NEXT:    notl %edi
 ; CHECK-NEXT:    # kill: def $cl killed $cl killed $ecx
-; CHECK-NEXT:    shll %cl, %edi
-; CHECK-NEXT:    testl $1048576, %edi # imm = 0x100000
-; CHECK-NEXT:    sete %al
+; CHECK-NEXT:    shll %cl, %eax
+; CHECK-NEXT:    shrl $20, %eax
+; CHECK-NEXT:    andl $1, %eax
+; CHECK-NEXT:    # kill: def $al killed $al killed $eax
 ; CHECK-NEXT:    retq
   %yy = and i32 %y, 30
   %d = lshr i32 1048576, %yy
@@ -349,9 +350,8 @@ define i1 @pow2_umax_fail0(i32 %x, i32 %y, i32 %z) {
 ; CHECK-NEXT:    shrl %cl, %esi
 ; CHECK-NEXT:    cmpl %esi, %eax
 ; CHECK-NEXT:    cmoval %eax, %esi
-; CHECK-NEXT:    notl %edi
-; CHECK-NEXT:    testl %edi, %esi
-; CHECK-NEXT:    sete %al
+; CHECK-NEXT:    testl %esi, %edi
+; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %yy = shl i32 1, %y
   %zz = lshr i32 1073741824, %z
@@ -482,9 +482,8 @@ define i1 @pow2_smax_fail0(i32 %x, i32 %y, i32 %z) {
 ; CHECK-NEXT:    shrl %cl, %esi
 ; CHECK-NEXT:    cmpl %esi, %eax
 ; CHECK-NEXT:    cmovgl %eax, %esi
-; CHECK-NEXT:    notl %edi
-; CHECK-NEXT:    testl %edi, %esi
-; CHECK-NEXT:    sete %al
+; CHECK-NEXT:    testl %esi, %edi
+; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %yy = shl i32 1, %y
   %zz = lshr i32 1073741824, %z
@@ -555,9 +554,8 @@ define i1 @pow2_select_fail0(i1 %c, i32 %x, i32 %y, i32 %z) {
 ; CHECK-NEXT:    shrl %cl, %r8d
 ; CHECK-NEXT:    testb $1, %dil
 ; CHECK-NEXT:    cmovnel %edx, %r8d
-; CHECK-NEXT:    notl %esi
-; CHECK-NEXT:    testl %esi, %r8d
-; CHECK-NEXT:    sete %al
+; CHECK-NEXT:    testl %r8d, %esi
+; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %yy = shl i32 1, %y
   %zz = lshr i32 1073741824, %z
@@ -696,10 +694,9 @@ define <4 x i1> @pow2_vselect_fail0_ne(<4 x i1> %c, <4 x i32> %x, <4 x i32> %y,
 ; CHECK-NEXT:    pand %xmm0, %xmm2
 ; CHECK-NEXT:    pandn %xmm7, %xmm0
 ; CHECK-NEXT:    por %xmm2, %xmm0
-; CHECK-NEXT:    pcmpeqd %xmm2, %xmm2
-; CHECK-NEXT:    pand %xmm0, %xmm1
+; CHECK-NEXT:    pand %xmm1, %xmm0
+; CHECK-NEXT:    pxor %xmm1, %xmm1
 ; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
-; CHECK-NEXT:    pxor %xmm2, %xmm0
 ; CHECK-NEXT:    retq
   %yy = shl <4 x i32> <i32 1, i32 1, i32 1, i32 1>, %y
   %zz = lshr <4 x i32> <i32 1073741824, i32 1073741824, i32 1073741824, i32 1073741824>, %z