[llvm] ac485e4 - [SelectionDAG] Add/Improve cases in `isKnownNeverZero`

[Noah Goldstein via llvm-commits]

Author: Noah Goldstein
Date: 2023-08-16T02:00:15-05:00
New Revision: ac485e40720fcdae3b1f5664143056e3dc89beb1

URL: https://github.com/llvm/llvm-project/commit/ac485e40720fcdae3b1f5664143056e3dc89beb1
DIFF: https://github.com/llvm/llvm-project/commit/ac485e40720fcdae3b1f5664143056e3dc89beb1.diff

LOG: [SelectionDAG] Add/Improve cases in `isKnownNeverZero`

1) Handle casts a bit more cleanly just with a loop rather than with
   recursion.

2) Add additional cases for smin/smax

3 ) For shifts we can also deduce non-zero if the maximum shift amount
    on the known 1s is non-zero.

Differential Revision: https://reviews.llvm.org/D156777

Added: 
    

Modified: 
    llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
    llvm/test/CodeGen/X86/known-never-zero.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 01043920530d9c..0271cad146aeaa 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5088,21 +5088,29 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
     return isKnownNeverZero(Op.getOperand(1), Depth + 1) &&
            isKnownNeverZero(Op.getOperand(2), Depth + 1);
 
-  case ISD::SHL:
+  case ISD::SHL: {
     if (Op->getFlags().hasNoSignedWrap() || Op->getFlags().hasNoUnsignedWrap())
       return isKnownNeverZero(Op.getOperand(0), Depth + 1);
-
-    // 1 << X is never zero. TODO: This can be expanded if we can bound X.
-    // The expression is really !Known.One[BitWidth-MaxLog2(Known):0].isZero()
-    if (computeKnownBits(Op.getOperand(0), Depth + 1).One[0])
+    KnownBits ValKnown = computeKnownBits(Op.getOperand(0), Depth + 1);
+    // 1 << X is never zero.
+    if (ValKnown.One[0])
+      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();
+    if (MaxCnt.ult(ValKnown.getBitWidth()) &&
+        !ValKnown.One.shl(MaxCnt).isZero())
       return true;
     break;
-
+  }
   case ISD::UADDSAT:
   case ISD::UMAX:
     return isKnownNeverZero(Op.getOperand(1), Depth + 1) ||
            isKnownNeverZero(Op.getOperand(0), Depth + 1);
 
+    // TODO for smin/smax: If either operand is known negative/positive
+    // respectively we don't need the other to be known at all.
+  case ISD::SMAX:
+  case ISD::SMIN:
   case ISD::UMIN:
     return isKnownNeverZero(Op.getOperand(1), Depth + 1) &&
            isKnownNeverZero(Op.getOperand(0), Depth + 1);
@@ -5116,16 +5124,19 @@ bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
     return isKnownNeverZero(Op.getOperand(0), Depth + 1);
 
   case ISD::SRA:
-  case ISD::SRL:
+  case ISD::SRL: {
     if (Op->getFlags().hasExact())
       return isKnownNeverZero(Op.getOperand(0), Depth + 1);
-    // Signed >> X is never zero. TODO: This can be expanded if we can bound X.
-    // The expression is really
-    // !Known.One[SignBit:SignBit-(BitWidth-MaxLog2(Known))].isZero()
-    if (computeKnownBits(Op.getOperand(0), Depth + 1).isNegative())
+    KnownBits ValKnown = computeKnownBits(Op.getOperand(0), Depth + 1);
+    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();
+    if (MaxCnt.ult(ValKnown.getBitWidth()) &&
+        !ValKnown.One.lshr(MaxCnt).isZero())
       return true;
     break;
-
+  }
   case ISD::UDIV:
   case ISD::SDIV:
     // div exact can only produce a zero if the dividend is zero.

diff  --git a/llvm/test/CodeGen/X86/known-never-zero.ll b/llvm/test/CodeGen/X86/known-never-zero.ll
index 2c059f72a062c9..cc9862769f2b66 100644
--- a/llvm/test/CodeGen/X86/known-never-zero.ll
+++ b/llvm/test/CodeGen/X86/known-never-zero.ll
@@ -258,9 +258,7 @@ define i32 @smin_known_nonzero(i32 %xx, i32 %yy) {
 ; CHECK-NEXT:    addl $4, %esi
 ; CHECK-NEXT:    cmpl %esi, %eax
 ; CHECK-NEXT:    cmovll %eax, %esi
-; CHECK-NEXT:    bsfl %esi, %ecx
-; CHECK-NEXT:    movl $32, %eax
-; CHECK-NEXT:    cmovnel %ecx, %eax
+; CHECK-NEXT:    rep bsfl %esi, %eax
 ; CHECK-NEXT:    retq
   %x = shl nuw i32 4, %xx
   %y = add nuw nsw i32 %yy, 4
@@ -298,9 +296,7 @@ define i32 @smax_known_nonzero(i32 %xx, i32 %yy) {
 ; CHECK-NEXT:    addl $4, %esi
 ; CHECK-NEXT:    cmpl %esi, %eax
 ; CHECK-NEXT:    cmovgl %eax, %esi
-; CHECK-NEXT:    bsfl %esi, %ecx
-; CHECK-NEXT:    movl $32, %eax
-; CHECK-NEXT:    cmovnel %ecx, %eax
+; CHECK-NEXT:    rep bsfl %esi, %eax
 ; CHECK-NEXT:    retq
   %x = shl nuw i32 4, %xx
   %y = add nuw nsw i32 %yy, 4