diff --git a/src/fldmod-by-const.jl b/src/fldmod-by-const.jl
index cc15554..36f0885 100644
--- a/src/fldmod-by-const.jl
+++ b/src/fldmod-by-const.jl
@@ -119,7 +119,8 @@ Base.@assume_effects :foldable function calculate_inverse_coeff(::Type{T}, C) wh
     # Now, truncate to only the upper half of invcoeff, after we've shifted. Instead of
     # bitshifting, we round to maintain precision. (This is needed to prevent off-by-ones.)
     # -- This is equivalent to `invcoeff = T(invcoeff >> sizeof(T))`, except rounded. --
-    invcoeff = _round_to_nearest(fldmod(invcoeff, typemax(UT))..., typemax(UT)) % T
+    two_to_N = _widen(typemax(UT)) + UT(1) # Precise value for 2^nbits(T) (doesn't fit in T)
+    invcoeff = _round_to_nearest(fldmod(invcoeff, two_to_N)..., two_to_N ) % T
     return invcoeff, toshift
 end
 
diff --git a/test/fldmod-by-const_tests.jl b/test/fldmod-by-const_tests.jl
index dfcee19..9f1db82 100644
--- a/test/fldmod-by-const_tests.jl
+++ b/test/fldmod-by-const_tests.jl
@@ -57,6 +57,30 @@ end
     end
 end
 
+@testset "fixed decimal multiplication - exhaustive 8-bit" begin
+    @testset for P in (0,1)
+        @testset for T in (Int8, UInt8)
+            FD = FixedDecimal{T,P}
+
+            function test_multiplies_correctly(fd, x)
+                big = FixedDecimal{BigInt, P}(fd)
+                big_mul = big * x
+                # This might overflow: ...
+                mul = fd * x
+                @testset "$fd * $x" begin
+                    # ... so we truncate big to the same size
+                    @test big_mul.i % T == mul.i % T
+                end
+            end
+            @testset for v in typemin(FD) : eps(FD) : typemax(FD)
+                @testset for v2 in typemin(FD) : eps(FD) : typemax(FD)
+                    test_multiplies_correctly(v, v2)
+                end
+            end
+        end
+    end
+end
+
 @testset "fixed decimal multiplication - exhaustive 16-bit" begin
     @testset for P in (0,1,2,3,4)
         @testset for T in (Int16, UInt16)