[mlir][polynomial] ensure primitive root calculation doesn't overflow

Author: j2kun

mlir/lib/Dialect/Polynomial/IR/PolynomialOps.cpp

@@ -17,6 +17,7 @@
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/Support/LogicalResult.h"
 #include "llvm/ADT/APInt.h"
+#include <iostream>
 
 using namespace mlir;
 using namespace mlir::polynomial;
@@ -107,16 +108,21 @@ LogicalResult MulScalarOp::verify() {
 /// Test if a value is a primitive nth root of unity modulo cmod.
 bool isPrimitiveNthRootOfUnity(const APInt &root, const APInt &n,
                                const APInt &cmod) {
+  // The first or subsequent multiplications, may overflow the input bit width,
+  // so scale them up to ensure they do not overflow.
+  unsigned requiredBitWidth =
+      std::max(root.getActiveBits() * 2, cmod.getActiveBits() * 2);
   // Root bitwidth may be 1 less then cmod.
-  APInt r = APInt(root).zext(cmod.getBitWidth());
-  assert(r.ule(cmod) && "root must be less than cmod");
-  unsigned upperBound = n.getZExtValue();
+  APInt r = APInt(root).zextOrTrunc(requiredBitWidth);
+  APInt cmodExt = APInt(cmod).zextOrTrunc(requiredBitWidth);
+  assert(r.ule(cmodExt) && "root must be less than cmod");
+  uint64_t upperBound = n.getZExtValue();
 
   APInt a = r;
   for (size_t k = 1; k < upperBound; k++) {
     if (a.isOne())
       return false;
-    a = (a * r).urem(cmod);
+    a = (a * r).urem(cmodExt);
   }
   return a.isOne();
 }

mlir/test/Dialect/Polynomial/ops.mlir

@@ -18,6 +18,11 @@
 #ntt_ring = #polynomial.ring<coefficientType=i32, coefficientModulus=256, polynomialModulus=#ntt_poly>
 !ntt_poly_ty = !polynomial.polynomial<ring=#ntt_ring>
 
+#ntt_poly_2 = #polynomial.int_polynomial<1 + x**65536>
+#ntt_ring_2 = #polynomial.ring<coefficientType = i32, coefficientModulus = 786433 : i32, polynomialModulus=#ntt_poly_2>
+#ntt_ring_2_root = #polynomial.primitive_root<value=283965:i32, degree=131072:i32>
+!ntt_poly_ty_2 = !polynomial.polynomial<ring=#ntt_ring_2>
+
 module {
   func.func @test_multiply() -> !polynomial.polynomial<ring=#ring1> {
     %c0 = arith.constant 0 : index
@@ -95,6 +100,11 @@ module {
     return
   }
 
+  func.func @test_ntt_with_overflowing_root(%0 : !ntt_poly_ty_2) {
+    %1 = polynomial.ntt %0 {root=#ntt_ring_2_root} : !ntt_poly_ty_2 -> tensor<65536xi32, #ntt_ring_2>
+    return
+  }
+
   func.func @test_intt(%0 : tensor<8xi32, #ntt_ring>) {
     %1 = polynomial.intt %0 {root=#polynomial.primitive_root<value=31:i32, degree=8:index>} : tensor<8xi32, #ntt_ring> -> !ntt_poly_ty
     return