[llvm] [ValueTracking] ComputeNumSignBitsImpl - add basic handling of BITCAST nodes (PR #127218)

[via llvm-commits]

================
@@ -3922,6 +3922,38 @@ static unsigned ComputeNumSignBitsImpl(const Value *V,
   if (auto *U = dyn_cast<Operator>(V)) {
     switch (Operator::getOpcode(V)) {
     default: break;
+    case Instruction::BitCast: {
+      Value *Src = U->getOperand(0);
+      Type *SrcTy = Src->getType();
+
+      // Skip if the source type is not an integer or integer vector type
+      // This ensures we only process integer-like types
+      if (!SrcTy->isIntOrIntVectorTy())
+        break;
+
+      unsigned SrcBits = SrcTy->getScalarSizeInBits();
+
+      if ((SrcBits % TyBits) != 0)
+        break;
+
+      // Only proceed if the destination type is a fixed-size vector
+      if (isa<FixedVectorType>(Ty)) {
+        // Check if the source is also a fixed-size vector
+        if (isa<FixedVectorType>(SrcTy)) {
+          // Compute the number of sign bits in the source
+          Tmp = ComputeNumSignBits(Src, Depth + 1, Q);
+          if (Tmp == SrcBits)
+            return TyBits;
+        } else {
+          // Handle scalar to vector bitcast case
+          // Compute sign bits for the scalar source
+          Tmp = ComputeNumSignBits(Src, Depth + 1, Q);
+          if (Tmp == SrcBits)
+            return TyBits;
----------------
vortex73 wrote:

Apologies for the oversight.

https://github.com/llvm/llvm-project/pull/127218