Revert "remove hand-coded methods for (U)Int128 on 32 bit systems (#53867)" (#59211)

We need to revert this for now. Resolves
#58911

It seems that the back-end LLVM emits this libcall a lot less often than
it used to, but it still happens sometimes and 32-bit Windows has no
implementation of `__divti3`.

Author: topolarity

base/int.jl

@@ -850,14 +850,166 @@ widemul(x::Bool,y::Number) = x * y
 widemul(x::Number,y::Bool) = x * y
 
 
-# Int128 multiply and divide
-*(x::T, y::T) where {T<:Union{Int128,UInt128}}  = mul_int(x, y)
+## wide multiplication, Int128 multiply and divide ##
+
+if Core.sizeof(Int) == 4
+    function widemul(u::Int64, v::Int64)
+        local u0::UInt64, v0::UInt64, w0::UInt64
+        local u1::Int64, v1::Int64, w1::UInt64, w2::Int64, t::UInt64
+
+        u0 = u & 0xffffffff; u1 = u >> 32
+        v0 = v & 0xffffffff; v1 = v >> 32
+        w0 = u0 * v0
+        t = reinterpret(UInt64, u1) * v0 + (w0 >>> 32)
+        w2 = reinterpret(Int64, t) >> 32
+        w1 = u0 * reinterpret(UInt64, v1) + (t & 0xffffffff)
+        hi = u1 * v1 + w2 + (reinterpret(Int64, w1) >> 32)
+        lo = w0 & 0xffffffff + (w1 << 32)
+        return Int128(hi) << 64 + Int128(lo)
+    end
+
+    function widemul(u::UInt64, v::UInt64)
+        local u0::UInt64, v0::UInt64, w0::UInt64
+        local u1::UInt64, v1::UInt64, w1::UInt64, w2::UInt64, t::UInt64
+
+        u0 = u & 0xffffffff; u1 = u >>> 32
+        v0 = v & 0xffffffff; v1 = v >>> 32
+        w0 = u0 * v0
+        t = u1 * v0 + (w0 >>> 32)
+        w2 = t >>> 32
+        w1 = u0 * v1 + (t & 0xffffffff)
+        hi = u1 * v1 + w2 + (w1 >>> 32)
+        lo = w0 & 0xffffffff + (w1 << 32)
+        return UInt128(hi) << 64 + UInt128(lo)
+    end
+
+    function *(u::Int128, v::Int128)
+        u0 = u % UInt64; u1 = Int64(u >> 64)
+        v0 = v % UInt64; v1 = Int64(v >> 64)
+        lolo = widemul(u0, v0)
+        lohi = widemul(reinterpret(Int64, u0), v1)
+        hilo = widemul(u1, reinterpret(Int64, v0))
+        t = reinterpret(UInt128, hilo) + (lolo >>> 64)
+        w1 = reinterpret(UInt128, lohi) + (t & 0xffffffffffffffff)
+        return Int128(lolo & 0xffffffffffffffff) + reinterpret(Int128, w1) << 64
+    end
+
+    function *(u::UInt128, v::UInt128)
+        u0 = u % UInt64; u1 = UInt64(u>>>64)
+        v0 = v % UInt64; v1 = UInt64(v>>>64)
+        lolo = widemul(u0, v0)
+        lohi = widemul(u0, v1)
+        hilo = widemul(u1, v0)
+        t = hilo + (lolo >>> 64)
+        w1 = lohi + (t & 0xffffffffffffffff)
+        return (lolo & 0xffffffffffffffff) + UInt128(w1) << 64
+    end
+
+    function _setbit(x::UInt128, i)
+        # faster version of `return x | (UInt128(1) << i)`
+        j = i >> 5
+        y = UInt128(one(UInt32) << (i & 0x1f))
+        if j == 0
+            return x | y
+        elseif j == 1
+            return x | (y << 32)
+        elseif j == 2
+            return x | (y << 64)
+        elseif j == 3
+            return x | (y << 96)
+        end
+        return x
+    end
 
-div(x::Int128,  y::Int128)  = checked_sdiv_int(x, y)
-div(x::UInt128, y::UInt128) = checked_udiv_int(x, y)
+    function divrem(x::UInt128, y::UInt128)
+        iszero(y) && throw(DivideError())
+        if (x >> 64) % UInt64 == 0
+            if (y >> 64) % UInt64 == 0
+                # fast path: upper 64 bits are zero, so we can fallback to UInt64 division
+                q64, x64 = divrem(x % UInt64, y % UInt64)
+                return UInt128(q64), UInt128(x64)
+            else
+                # this implies y>x, so
+                return zero(UInt128), x
+            end
+        end
+        n = leading_zeros(y) - leading_zeros(x)
+        q = zero(UInt128)
+        ys = y << n
+        while n >= 0
+            # ys == y * 2^n
+            if ys <= x
+                x -= ys
+                q = _setbit(q, n)
+                if (x >> 64) % UInt64 == 0
+                    # exit early, similar to above fast path
+                    if (y >> 64) % UInt64 == 0
+                        q64, x64 = divrem(x % UInt64, y % UInt64)
+                        q |= q64
+                        x = UInt128(x64)
+                    end
+                    return q, x
+                end
+            end
+            ys >>>= 1
+            n -= 1
+        end
+        return q, x
+    end
 
-rem(x::Int128,  y::Int128)  = checked_srem_int(x, y)
-rem(x::UInt128, y::UInt128) = checked_urem_int(x, y)
+    function div(x::Int128, y::Int128)
+        (x == typemin(Int128)) & (y == -1) && throw(DivideError())
+        return Int128(div(BigInt(x), BigInt(y)))::Int128
+    end
+    div(x::UInt128, y::UInt128) = divrem(x, y)[1]
+
+    function rem(x::Int128, y::Int128)
+        return Int128(rem(BigInt(x), BigInt(y)))::Int128
+    end
+
+    function rem(x::UInt128, y::UInt128)
+        iszero(y) && throw(DivideError())
+        if (x >> 64) % UInt64 == 0
+            if (y >> 64) % UInt64 == 0
+                # fast path: upper 64 bits are zero, so we can fallback to UInt64 division
+                return UInt128(rem(x % UInt64, y % UInt64))
+            else
+                # this implies y>x, so
+                return x
+            end
+        end
+        n = leading_zeros(y) - leading_zeros(x)
+        ys = y << n
+        while n >= 0
+            # ys == y * 2^n
+            if ys <= x
+                x -= ys
+                if (x >> 64) % UInt64 == 0
+                    # exit early, similar to above fast path
+                    if (y >> 64) % UInt64 == 0
+                        x = UInt128(rem(x % UInt64, y % UInt64))
+                    end
+                    return x
+                end
+            end
+            ys >>>= 1
+            n -= 1
+        end
+        return x
+    end
+
+    function mod(x::Int128, y::Int128)
+        return Int128(mod(BigInt(x), BigInt(y)))::Int128
+    end
+else
+    *(x::T, y::T) where {T<:Union{Int128,UInt128}}  = mul_int(x, y)
+
+    div(x::Int128,  y::Int128)  = checked_sdiv_int(x, y)
+    div(x::UInt128, y::UInt128) = checked_udiv_int(x, y)
+
+    rem(x::Int128,  y::Int128)  = checked_srem_int(x, y)
+    rem(x::UInt128, y::UInt128) = checked_urem_int(x, y)
+end
 
 # issue #15489: since integer ops are unchecked, they shouldn't check promotion
 for op in (:+, :-, :*, :&, :|, :xor)